Skip to content

GitLab

Unverified Commit 2f234376 authored by kabachuha's avatar kabachuha Committed by GitHub
Browse files

Add (Scheduled) Pseudo-Huber Loss training scripts to research projects (#7527) 

* add scheduled pseudo-huber loss training scripts

See #7488

* add reduction modes to huber loss

* [DB Lora] *2 multiplier to huber loss cause of 1/2 a^2 conv.

pairing of https://github.com/kohya-ss/sd-scripts/pull/1228/commits/c6495def1fbbaf2a0233110d50f976ed61620e83

* [DB Lora] add option for smooth l1 (huber / delta)

Pairing of https://github.com/kohya-ss/sd-scripts/pull/1228/commits/dd22958caa56e4db885324f76188c13bdf504569

* [DB Lora] unify huber scheduling

Pairing of https://github.com/kohya-ss/sd-scripts/pull/1228/commits/19a834c3ab448614e8887b07f2bb4e0aaabf0805

* [DB Lora] add snr huber scheduler

Pairing of https://github.com/kohya-ss/sd-scripts/pull/1228/commits/47fb1a68547e76f33cd54a3da8d2c35b9489c56e



* fixup examples link

* use snr schedule by default in DB

* update all huber scripts with snr

* code quality

* huber: make style && make quality

---------
Co-authored-by: default avatarSayak Paul <spsayakpaul@gmail.com>
parent 2523390c
Hide whitespace changes
Inline Side-by-side
Showing with 10106 additions and 0 deletions
examples/research_projects/scheduled_huber_loss_training/README.md 0 → 100644
View file @ 2f234376
# Scheduled Pseudo-Huber Loss for Diffusers
These are the modifications of to include the possibility of training text2image models with Scheduled Pseudo Huber loss, introduced in https://arxiv.org/abs/2403.16728. (https://github.com/kabachuha/SPHL-for-stable-diffusion)
## Why this might be useful?
- If you suspect that the part of the training dataset might be corrupted, and you don't want these outliers to distort the model's supposed output
- If you want to improve the aesthetic quality of pictures by helping the model disentangle concepts and be less influenced by another sorts of pictures.
See https://github.com/huggingface/diffusers/issues/7488 for the detailed description.
## Instructions
The same usage as in the case of the corresponding vanilla Diffusers scripts https://github.com/huggingface/diffusers/tree/main/examples
examples/research_projects/scheduled_huber_loss_training/dreambooth/train_dreambooth.py 0 → 100644
View file @ 2f234376
#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
import argparse
import copy
import gc
import importlib
import itertools
import logging
import math
import os
import shutil
import warnings
from pathlib import Path
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import ProjectConfiguration, set_seed
from huggingface_hub import create_repo, model_info, upload_folder
from huggingface_hub.utils import insecure_hashlib
from packaging import version
from PIL import Image
from PIL.ImageOps import exif_transpose
from torch.utils.data import Dataset
from torchvision import transforms
from tqdm.auto import tqdm
from transformers import AutoTokenizer, PretrainedConfig
import diffusers
from diffusers import (
AutoencoderKL,
DDPMScheduler,
DiffusionPipeline,
StableDiffusionPipeline,
UNet2DConditionModel,
)
from diffusers.optimization import get_scheduler
from diffusers.training_utils import compute_snr
from diffusers.utils import check_min_version, is_wandb_available
from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.torch_utils import is_compiled_module
if is_wandb_available():
import wandb
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.28.0.dev0")
logger = get_logger(__name__)
def save_model_card(
repo_id: str,
images: list = None,
base_model: str = None,
train_text_encoder=False,
prompt: str = None,
repo_folder: str = None,
pipeline: DiffusionPipeline = None,
):
img_str = ""
if images is not None:
for i, image in enumerate(images):
image.save(os.path.join(repo_folder, f"image_{i}.png"))
img_str += f"![img_{i}](./image_{i}.png)\n"
model_description = f"""
# DreamBooth - {repo_id}
This is a dreambooth model derived from {base_model}. The weights were trained on {prompt} using [DreamBooth](https://dreambooth.github.io/).
You can find some example images in the following. \n
{img_str}
DreamBooth for the text encoder was enabled: {train_text_encoder}.
"""
model_card = load_or_create_model_card(
repo_id_or_path=repo_id,
from_training=True,
license="creativeml-openrail-m",
base_model=base_model,
prompt=prompt,
model_description=model_description,
inference=True,
)
tags = ["text-to-image", "dreambooth", "diffusers-training"]
if isinstance(pipeline, StableDiffusionPipeline):
tags.extend(["stable-diffusion", "stable-diffusion-diffusers"])
else:
tags.extend(["if", "if-diffusers"])
model_card = populate_model_card(model_card, tags=tags)
model_card.save(os.path.join(repo_folder, "README.md"))
def log_validation(
text_encoder,
tokenizer,
unet,
vae,
args,
accelerator,
weight_dtype,
global_step,
prompt_embeds,
negative_prompt_embeds,
):
logger.info(
f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
f" {args.validation_prompt}."
)
pipeline_args = {}
if vae is not None:
pipeline_args["vae"] = vae
# create pipeline (note: unet and vae are loaded again in float32)
pipeline = DiffusionPipeline.from_pretrained(
args.pretrained_model_name_or_path,
tokenizer=tokenizer,
text_encoder=text_encoder,
unet=unet,
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
**pipeline_args,
)
# We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
scheduler_args = {}
if "variance_type" in pipeline.scheduler.config:
variance_type = pipeline.scheduler.config.variance_type
if variance_type in ["learned", "learned_range"]:
variance_type = "fixed_small"
scheduler_args["variance_type"] = variance_type
module = importlib.import_module("diffusers")
scheduler_class = getattr(module, args.validation_scheduler)
pipeline.scheduler = scheduler_class.from_config(pipeline.scheduler.config, **scheduler_args)
pipeline = pipeline.to(accelerator.device)
pipeline.set_progress_bar_config(disable=True)
if args.pre_compute_text_embeddings:
pipeline_args = {
"prompt_embeds": prompt_embeds,
"negative_prompt_embeds": negative_prompt_embeds,
}
else:
pipeline_args = {"prompt": args.validation_prompt}
# run inference
generator = None if args.seed is None else torch.Generator(device=accelerator.device).manual_seed(args.seed)
images = []
if args.validation_images is None:
for _ in range(args.num_validation_images):
with torch.autocast("cuda"):
image = pipeline(**pipeline_args, num_inference_steps=25, generator=generator).images[0]
images.append(image)
else:
for image in args.validation_images:
image = Image.open(image)
image = pipeline(**pipeline_args, image=image, generator=generator).images[0]
images.append(image)
for tracker in accelerator.trackers:
if tracker.name == "tensorboard":
np_images = np.stack([np.asarray(img) for img in images])
tracker.writer.add_images("validation", np_images, global_step, dataformats="NHWC")
if tracker.name == "wandb":
tracker.log(
{
"validation": [
wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images)
]
}
)
del pipeline
torch.cuda.empty_cache()
return images
def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: str, revision: str):
text_encoder_config = PretrainedConfig.from_pretrained(
pretrained_model_name_or_path,
subfolder="text_encoder",
revision=revision,
)
model_class = text_encoder_config.architectures[0]
if model_class == "CLIPTextModel":
from transformers import CLIPTextModel
return CLIPTextModel
elif model_class == "RobertaSeriesModelWithTransformation":
from diffusers.pipelines.alt_diffusion.modeling_roberta_series import RobertaSeriesModelWithTransformation
return RobertaSeriesModelWithTransformation
elif model_class == "T5EncoderModel":
from transformers import T5EncoderModel
return T5EncoderModel
else:
raise ValueError(f"{model_class} is not supported.")
def parse_args(input_args=None):
parser = argparse.ArgumentParser(description="Simple example of a training script.")
parser.add_argument(
"--pretrained_model_name_or_path",
type=str,
default=None,
required=True,
help="Path to pretrained model or model identifier from huggingface.co/models.",
)
parser.add_argument(
"--revision",
type=str,
default=None,
required=False,
help="Revision of pretrained model identifier from huggingface.co/models.",
)
parser.add_argument(
"--variant",
type=str,
default=None,
help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
)
parser.add_argument(
"--tokenizer_name",
type=str,
default=None,
help="Pretrained tokenizer name or path if not the same as model_name",
)
parser.add_argument(
"--instance_data_dir",
type=str,
default=None,
required=True,
help="A folder containing the training data of instance images.",
)
parser.add_argument(
"--class_data_dir",
type=str,
default=None,
required=False,
help="A folder containing the training data of class images.",
)
parser.add_argument(
"--instance_prompt",
type=str,
default=None,
required=True,
help="The prompt with identifier specifying the instance",
)
parser.add_argument(
"--class_prompt",
type=str,
default=None,
help="The prompt to specify images in the same class as provided instance images.",
)
parser.add_argument(
"--with_prior_preservation",
default=False,
action="store_true",
help="Flag to add prior preservation loss.",
)
parser.add_argument("--prior_loss_weight", type=float, default=1.0, help="The weight of prior preservation loss.")
parser.add_argument(
"--num_class_images",
type=int,
default=100,
help=(
"Minimal class images for prior preservation loss. If there are not enough images already present in"
" class_data_dir, additional images will be sampled with class_prompt."
),
)
parser.add_argument(
"--output_dir",
type=str,
default="dreambooth-model",
help="The output directory where the model predictions and checkpoints will be written.",
)
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
parser.add_argument(
"--resolution",
type=int,
default=512,
help=(
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
" resolution"
),
)
parser.add_argument(
"--center_crop",
default=False,
action="store_true",
help=(
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
" cropped. The images will be resized to the resolution first before cropping."
),
)
parser.add_argument(
"--train_text_encoder",
action="store_true",
help="Whether to train the text encoder. If set, the text encoder should be float32 precision.",
)
parser.add_argument(
"--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
)
parser.add_argument(
"--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images."
)
parser.add_argument("--num_train_epochs", type=int, default=1)
parser.add_argument(
"--max_train_steps",
type=int,
default=None,
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
)
parser.add_argument(
"--checkpointing_steps",
type=int,
default=500,
help=(
"Save a checkpoint of the training state every X updates. Checkpoints can be used for resuming training via `--resume_from_checkpoint`. "
"In the case that the checkpoint is better than the final trained model, the checkpoint can also be used for inference."
"Using a checkpoint for inference requires separate loading of the original pipeline and the individual checkpointed model components."
"See https://huggingface.co/docs/diffusers/main/en/training/dreambooth#performing-inference-using-a-saved-checkpoint for step by step"
"instructions."
),
)
parser.add_argument(
"--checkpoints_total_limit",
type=int,
default=None,
help=(
"Max number of checkpoints to store. Passed as `total_limit` to the `Accelerator` `ProjectConfiguration`."
" See Accelerator::save_state https://huggingface.co/docs/accelerate/package_reference/accelerator#accelerate.Accelerator.save_state"
" for more details"
),
)
parser.add_argument(
"--resume_from_checkpoint",
type=str,
default=None,
help=(
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
),
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--gradient_checkpointing",
action="store_true",
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
)
parser.add_argument(
"--learning_rate",
type=float,
default=5e-6,
help="Initial learning rate (after the potential warmup period) to use.",
)
parser.add_argument(
"--scale_lr",
action="store_true",
default=False,
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
)
parser.add_argument(
"--lr_scheduler",
type=str,
default="constant",
help=(
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
' "constant", "constant_with_warmup"]'
),
)
parser.add_argument(
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
)
parser.add_argument(
"--lr_num_cycles",
type=int,
default=1,
help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
)
parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
parser.add_argument(
"--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
)
parser.add_argument(
"--dataloader_num_workers",
type=int,
default=0,
help=(
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
),
)
parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
parser.add_argument(
"--hub_model_id",
type=str,
default=None,
help="The name of the repository to keep in sync with the local `output_dir`.",
)
parser.add_argument(
"--logging_dir",
type=str,
default="logs",
help=(
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
),
)
parser.add_argument(
"--allow_tf32",
action="store_true",
help=(
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
),
)
parser.add_argument(
"--report_to",
type=str,
default="tensorboard",
help=(
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
),
)
parser.add_argument(
"--validation_prompt",
type=str,
default=None,
help="A prompt that is used during validation to verify that the model is learning.",
)
parser.add_argument(
"--num_validation_images",
type=int,
default=4,
help="Number of images that should be generated during validation with `validation_prompt`.",
)
parser.add_argument(
"--validation_steps",
type=int,
default=100,
help=(
"Run validation every X steps. Validation consists of running the prompt"
" `args.validation_prompt` multiple times: `args.num_validation_images`"
" and logging the images."
),
)
parser.add_argument(
"--mixed_precision",
type=str,
default=None,
choices=["no", "fp16", "bf16"],
help=(
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
),
)
parser.add_argument(
"--prior_generation_precision",
type=str,
default=None,
choices=["no", "fp32", "fp16", "bf16"],
help=(
"Choose prior generation precision between fp32, fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
" 1.10.and an Nvidia Ampere GPU. Default to fp16 if a GPU is available else fp32."
),
)
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
parser.add_argument(
"--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
)
parser.add_argument(
"--set_grads_to_none",
action="store_true",
help=(
"Save more memory by using setting grads to None instead of zero. Be aware, that this changes certain"
" behaviors, so disable this argument if it causes any problems. More info:"
" https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html"
),
)
parser.add_argument(
"--offset_noise",
action="store_true",
default=False,
help=(
"Fine-tuning against a modified noise"
" See: https://www.crosslabs.org//blog/diffusion-with-offset-noise for more information."
),
)
parser.add_argument(
"--snr_gamma",
type=float,
default=None,
help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
"More details here: https://arxiv.org/abs/2303.09556.",
)
parser.add_argument(
"--pre_compute_text_embeddings",
action="store_true",
help="Whether or not to pre-compute text embeddings. If text embeddings are pre-computed, the text encoder will not be kept in memory during training and will leave more GPU memory available for training the rest of the model. This is not compatible with `--train_text_encoder`.",
)
parser.add_argument(
"--tokenizer_max_length",
type=int,
default=None,
required=False,
help="The maximum length of the tokenizer. If not set, will default to the tokenizer's max length.",
)
parser.add_argument(
"--text_encoder_use_attention_mask",
action="store_true",
required=False,
help="Whether to use attention mask for the text encoder",
)
parser.add_argument(
"--skip_save_text_encoder", action="store_true", required=False, help="Set to not save text encoder"
)
parser.add_argument(
"--validation_images",
required=False,
default=None,
nargs="+",
help="Optional set of images to use for validation. Used when the target pipeline takes an initial image as input such as when training image variation or superresolution.",
)
parser.add_argument(
"--class_labels_conditioning",
required=False,
default=None,
help="The optional `class_label` conditioning to pass to the unet, available values are `timesteps`.",
)
parser.add_argument(
"--loss_type",
type=str,
default="l2",
choices=["l2", "huber", "smooth_l1"],
help="The type of loss to use and whether it's timestep-scheduled. See Issue #7488 for more info.",
)
parser.add_argument(
"--huber_schedule",
type=str,
default="snr",
choices=["constant", "exponential", "snr"],
help="The schedule to use for the huber losses parameter",
)
parser.add_argument(
"--huber_c",
type=float,
default=0.1,
help="The huber loss parameter. Only used if one of the huber loss modes (huber or smooth l1) is selected with loss_type.",
)
parser.add_argument(
"--validation_scheduler",
type=str,
default="DPMSolverMultistepScheduler",
choices=["DPMSolverMultistepScheduler", "DDPMScheduler"],
help="Select which scheduler to use for validation. DDPMScheduler is recommended for DeepFloyd IF.",
)
if input_args is not None:
args = parser.parse_args(input_args)
else:
args = parser.parse_args()
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
if env_local_rank != -1 and env_local_rank != args.local_rank:
args.local_rank = env_local_rank
if args.with_prior_preservation:
if args.class_data_dir is None:
raise ValueError("You must specify a data directory for class images.")
if args.class_prompt is None:
raise ValueError("You must specify prompt for class images.")
else:
# logger is not available yet
if args.class_data_dir is not None:
warnings.warn("You need not use --class_data_dir without --with_prior_preservation.")
if args.class_prompt is not None:
warnings.warn("You need not use --class_prompt without --with_prior_preservation.")
if args.train_text_encoder and args.pre_compute_text_embeddings:
raise ValueError("`--train_text_encoder` cannot be used with `--pre_compute_text_embeddings`")
return args
class DreamBoothDataset(Dataset):
"""
A dataset to prepare the instance and class images with the prompts for fine-tuning the model.
It pre-processes the images and the tokenizes prompts.
"""
def __init__(
self,
instance_data_root,
instance_prompt,
tokenizer,
class_data_root=None,
class_prompt=None,
class_num=None,
size=512,
center_crop=False,
encoder_hidden_states=None,
class_prompt_encoder_hidden_states=None,
tokenizer_max_length=None,
):
self.size = size
self.center_crop = center_crop
self.tokenizer = tokenizer
self.encoder_hidden_states = encoder_hidden_states
self.class_prompt_encoder_hidden_states = class_prompt_encoder_hidden_states
self.tokenizer_max_length = tokenizer_max_length
self.instance_data_root = Path(instance_data_root)
if not self.instance_data_root.exists():
raise ValueError(f"Instance {self.instance_data_root} images root doesn't exists.")
self.instance_images_path = list(Path(instance_data_root).iterdir())
self.num_instance_images = len(self.instance_images_path)
self.instance_prompt = instance_prompt
self._length = self.num_instance_images
if class_data_root is not None:
self.class_data_root = Path(class_data_root)
self.class_data_root.mkdir(parents=True, exist_ok=True)
self.class_images_path = list(self.class_data_root.iterdir())
if class_num is not None:
self.num_class_images = min(len(self.class_images_path), class_num)
else:
self.num_class_images = len(self.class_images_path)
self._length = max(self.num_class_images, self.num_instance_images)
self.class_prompt = class_prompt
else:
self.class_data_root = None
self.image_transforms = transforms.Compose(
[
transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
]
)
def __len__(self):
return self._length
def __getitem__(self, index):
example = {}
instance_image = Image.open(self.instance_images_path[index % self.num_instance_images])
instance_image = exif_transpose(instance_image)
if not instance_image.mode == "RGB":
instance_image = instance_image.convert("RGB")
example["instance_images"] = self.image_transforms(instance_image)
if self.encoder_hidden_states is not None:
example["instance_prompt_ids"] = self.encoder_hidden_states
else:
text_inputs = tokenize_prompt(
self.tokenizer, self.instance_prompt, tokenizer_max_length=self.tokenizer_max_length
)
example["instance_prompt_ids"] = text_inputs.input_ids
example["instance_attention_mask"] = text_inputs.attention_mask
if self.class_data_root:
class_image = Image.open(self.class_images_path[index % self.num_class_images])
class_image = exif_transpose(class_image)
if not class_image.mode == "RGB":
class_image = class_image.convert("RGB")
example["class_images"] = self.image_transforms(class_image)
if self.class_prompt_encoder_hidden_states is not None:
example["class_prompt_ids"] = self.class_prompt_encoder_hidden_states
else:
class_text_inputs = tokenize_prompt(
self.tokenizer, self.class_prompt, tokenizer_max_length=self.tokenizer_max_length
)
example["class_prompt_ids"] = class_text_inputs.input_ids
example["class_attention_mask"] = class_text_inputs.attention_mask
return example
def collate_fn(examples, with_prior_preservation=False):
has_attention_mask = "instance_attention_mask" in examples[0]
input_ids = [example["instance_prompt_ids"] for example in examples]
pixel_values = [example["instance_images"] for example in examples]
if has_attention_mask:
attention_mask = [example["instance_attention_mask"] for example in examples]
# Concat class and instance examples for prior preservation.
# We do this to avoid doing two forward passes.
if with_prior_preservation:
input_ids += [example["class_prompt_ids"] for example in examples]
pixel_values += [example["class_images"] for example in examples]
if has_attention_mask:
attention_mask += [example["class_attention_mask"] for example in examples]
pixel_values = torch.stack(pixel_values)
pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
input_ids = torch.cat(input_ids, dim=0)
batch = {
"input_ids": input_ids,
"pixel_values": pixel_values,
}
if has_attention_mask:
attention_mask = torch.cat(attention_mask, dim=0)
batch["attention_mask"] = attention_mask
return batch
class PromptDataset(Dataset):
"A simple dataset to prepare the prompts to generate class images on multiple GPUs."
def __init__(self, prompt, num_samples):
self.prompt = prompt
self.num_samples = num_samples
def __len__(self):
return self.num_samples
def __getitem__(self, index):
example = {}
example["prompt"] = self.prompt
example["index"] = index
return example
def model_has_vae(args):
config_file_name = os.path.join("vae", AutoencoderKL.config_name)
if os.path.isdir(args.pretrained_model_name_or_path):
config_file_name = os.path.join(args.pretrained_model_name_or_path, config_file_name)
return os.path.isfile(config_file_name)
else:
files_in_repo = model_info(args.pretrained_model_name_or_path, revision=args.revision).siblings
return any(file.rfilename == config_file_name for file in files_in_repo)
def tokenize_prompt(tokenizer, prompt, tokenizer_max_length=None):
if tokenizer_max_length is not None:
max_length = tokenizer_max_length
else:
max_length = tokenizer.model_max_length
text_inputs = tokenizer(
prompt,
truncation=True,
padding="max_length",
max_length=max_length,
return_tensors="pt",
)
return text_inputs
def encode_prompt(text_encoder, input_ids, attention_mask, text_encoder_use_attention_mask=None):
text_input_ids = input_ids.to(text_encoder.device)
if text_encoder_use_attention_mask:
attention_mask = attention_mask.to(text_encoder.device)
else:
attention_mask = None
prompt_embeds = text_encoder(
text_input_ids,
attention_mask=attention_mask,
return_dict=False,
)
prompt_embeds = prompt_embeds[0]
return prompt_embeds
# NOTE: if you're using the scheduled version, huber_c has to depend on the timesteps already
def conditional_loss(
model_pred: torch.Tensor,
target: torch.Tensor,
reduction: str = "mean",
loss_type: str = "l2",
huber_c: float = 0.1,
):
if loss_type == "l2":
loss = F.mse_loss(model_pred, target, reduction=reduction)
elif loss_type == "huber":
loss = 2 * huber_c * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
elif loss_type == "smooth_l1":
loss = 2 * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
else:
raise NotImplementedError(f"Unsupported Loss Type {loss_type}")
return loss
def main(args):
if args.report_to == "wandb" and args.hub_token is not None:
raise ValueError(
"You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
" Please use `huggingface-cli login` to authenticate with the Hub."
)
logging_dir = Path(args.output_dir, args.logging_dir)
accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
accelerator = Accelerator(
gradient_accumulation_steps=args.gradient_accumulation_steps,
mixed_precision=args.mixed_precision,
log_with=args.report_to,
project_config=accelerator_project_config,
)
if args.report_to == "wandb":
if not is_wandb_available():
raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
# Currently, it's not possible to do gradient accumulation when training two models with accelerate.accumulate
# This will be enabled soon in accelerate. For now, we don't allow gradient accumulation when training two models.
# TODO (patil-suraj): Remove this check when gradient accumulation with two models is enabled in accelerate.
if args.train_text_encoder and args.gradient_accumulation_steps > 1 and accelerator.num_processes > 1:
raise ValueError(
"Gradient accumulation is not supported when training the text encoder in distributed training. "
"Please set gradient_accumulation_steps to 1. This feature will be supported in the future."
)
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state, main_process_only=False)
if accelerator.is_local_main_process:
transformers.utils.logging.set_verbosity_warning()
diffusers.utils.logging.set_verbosity_info()
else:
transformers.utils.logging.set_verbosity_error()
diffusers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Generate class images if prior preservation is enabled.
if args.with_prior_preservation:
class_images_dir = Path(args.class_data_dir)
if not class_images_dir.exists():
class_images_dir.mkdir(parents=True)
cur_class_images = len(list(class_images_dir.iterdir()))
if cur_class_images < args.num_class_images:
torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32
if args.prior_generation_precision == "fp32":
torch_dtype = torch.float32
elif args.prior_generation_precision == "fp16":
torch_dtype = torch.float16
elif args.prior_generation_precision == "bf16":
torch_dtype = torch.bfloat16
pipeline = DiffusionPipeline.from_pretrained(
args.pretrained_model_name_or_path,
torch_dtype=torch_dtype,
safety_checker=None,
revision=args.revision,
variant=args.variant,
)
pipeline.set_progress_bar_config(disable=True)
num_new_images = args.num_class_images - cur_class_images
logger.info(f"Number of class images to sample: {num_new_images}.")
sample_dataset = PromptDataset(args.class_prompt, num_new_images)
sample_dataloader = torch.utils.data.DataLoader(sample_dataset, batch_size=args.sample_batch_size)
sample_dataloader = accelerator.prepare(sample_dataloader)
pipeline.to(accelerator.device)
for example in tqdm(
sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process
):
images = pipeline(example["prompt"]).images
for i, image in enumerate(images):
hash_image = insecure_hashlib.sha1(image.tobytes()).hexdigest()
image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
image.save(image_filename)
del pipeline
if torch.cuda.is_available():
torch.cuda.empty_cache()
# Handle the repository creation
if accelerator.is_main_process:
if args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
if args.push_to_hub:
repo_id = create_repo(
repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
).repo_id
# Load the tokenizer
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=False)
elif args.pretrained_model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
args.pretrained_model_name_or_path,
subfolder="tokenizer",
revision=args.revision,
use_fast=False,
)
# import correct text encoder class
text_encoder_cls = import_model_class_from_model_name_or_path(args.pretrained_model_name_or_path, args.revision)
# Load scheduler and models
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
text_encoder = text_encoder_cls.from_pretrained(
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
)
if model_has_vae(args):
vae = AutoencoderKL.from_pretrained(
args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant
)
else:
vae = None
unet = UNet2DConditionModel.from_pretrained(
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
)
def unwrap_model(model):
model = accelerator.unwrap_model(model)
model = model._orig_mod if is_compiled_module(model) else model
return model
# create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
def save_model_hook(models, weights, output_dir):
if accelerator.is_main_process:
for model in models:
sub_dir = "unet" if isinstance(model, type(unwrap_model(unet))) else "text_encoder"
model.save_pretrained(os.path.join(output_dir, sub_dir))
# make sure to pop weight so that corresponding model is not saved again
weights.pop()
def load_model_hook(models, input_dir):
while len(models) > 0:
# pop models so that they are not loaded again
model = models.pop()
if isinstance(model, type(unwrap_model(text_encoder))):
# load transformers style into model
load_model = text_encoder_cls.from_pretrained(input_dir, subfolder="text_encoder")
model.config = load_model.config
else:
# load diffusers style into model
load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
model.register_to_config(**load_model.config)
model.load_state_dict(load_model.state_dict())
del load_model
accelerator.register_save_state_pre_hook(save_model_hook)
accelerator.register_load_state_pre_hook(load_model_hook)
if vae is not None:
vae.requires_grad_(False)
if not args.train_text_encoder:
text_encoder.requires_grad_(False)
if args.enable_xformers_memory_efficient_attention:
if is_xformers_available():
import xformers
xformers_version = version.parse(xformers.__version__)
if xformers_version == version.parse("0.0.16"):
logger.warning(
"xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
)
unet.enable_xformers_memory_efficient_attention()
else:
raise ValueError("xformers is not available. Make sure it is installed correctly")
if args.gradient_checkpointing:
unet.enable_gradient_checkpointing()
if args.train_text_encoder:
text_encoder.gradient_checkpointing_enable()
# Check that all trainable models are in full precision
low_precision_error_string = (
"Please make sure to always have all model weights in full float32 precision when starting training - even if"
" doing mixed precision training. copy of the weights should still be float32."
)
if unwrap_model(unet).dtype != torch.float32:
raise ValueError(f"Unet loaded as datatype {unwrap_model(unet).dtype}. {low_precision_error_string}")
if args.train_text_encoder and unwrap_model(text_encoder).dtype != torch.float32:
raise ValueError(
f"Text encoder loaded as datatype {unwrap_model(text_encoder).dtype}." f" {low_precision_error_string}"
)
# Enable TF32 for faster training on Ampere GPUs,
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
if args.allow_tf32:
torch.backends.cuda.matmul.allow_tf32 = True
if args.scale_lr:
args.learning_rate = (
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
)
# Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
if args.use_8bit_adam:
try:
import bitsandbytes as bnb
except ImportError:
raise ImportError(
"To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
)
optimizer_class = bnb.optim.AdamW8bit
else:
optimizer_class = torch.optim.AdamW
# Optimizer creation
params_to_optimize = (
itertools.chain(unet.parameters(), text_encoder.parameters()) if args.train_text_encoder else unet.parameters()
)
optimizer = optimizer_class(
params_to_optimize,
lr=args.learning_rate,
betas=(args.adam_beta1, args.adam_beta2),
weight_decay=args.adam_weight_decay,
eps=args.adam_epsilon,
)
if args.pre_compute_text_embeddings:
def compute_text_embeddings(prompt):
with torch.no_grad():
text_inputs = tokenize_prompt(tokenizer, prompt, tokenizer_max_length=args.tokenizer_max_length)
prompt_embeds = encode_prompt(
text_encoder,
text_inputs.input_ids,
text_inputs.attention_mask,
text_encoder_use_attention_mask=args.text_encoder_use_attention_mask,
)
return prompt_embeds
pre_computed_encoder_hidden_states = compute_text_embeddings(args.instance_prompt)
validation_prompt_negative_prompt_embeds = compute_text_embeddings("")
if args.validation_prompt is not None:
validation_prompt_encoder_hidden_states = compute_text_embeddings(args.validation_prompt)
else:
validation_prompt_encoder_hidden_states = None
if args.class_prompt is not None:
pre_computed_class_prompt_encoder_hidden_states = compute_text_embeddings(args.class_prompt)
else:
pre_computed_class_prompt_encoder_hidden_states = None
text_encoder = None
tokenizer = None
gc.collect()
torch.cuda.empty_cache()
else:
pre_computed_encoder_hidden_states = None
validation_prompt_encoder_hidden_states = None
validation_prompt_negative_prompt_embeds = None
pre_computed_class_prompt_encoder_hidden_states = None
# Dataset and DataLoaders creation:
train_dataset = DreamBoothDataset(
instance_data_root=args.instance_data_dir,
instance_prompt=args.instance_prompt,
class_data_root=args.class_data_dir if args.with_prior_preservation else None,
class_prompt=args.class_prompt,
class_num=args.num_class_images,
tokenizer=tokenizer,
size=args.resolution,
center_crop=args.center_crop,
encoder_hidden_states=pre_computed_encoder_hidden_states,
class_prompt_encoder_hidden_states=pre_computed_class_prompt_encoder_hidden_states,
tokenizer_max_length=args.tokenizer_max_length,
)
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.train_batch_size,
shuffle=True,
collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation),
num_workers=args.dataloader_num_workers,
)
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
overrode_max_train_steps = True
lr_scheduler = get_scheduler(
args.lr_scheduler,
optimizer=optimizer,
num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
num_training_steps=args.max_train_steps * accelerator.num_processes,
num_cycles=args.lr_num_cycles,
power=args.lr_power,
)
# Prepare everything with our `accelerator`.
if args.train_text_encoder:
unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
unet, text_encoder, optimizer, train_dataloader, lr_scheduler
)
else:
unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
unet, optimizer, train_dataloader, lr_scheduler
)
# For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) to half-precision
# as these weights are only used for inference, keeping weights in full precision is not required.
weight_dtype = torch.float32
if accelerator.mixed_precision == "fp16":
weight_dtype = torch.float16
elif accelerator.mixed_precision == "bf16":
weight_dtype = torch.bfloat16
# Move vae and text_encoder to device and cast to weight_dtype
if vae is not None:
vae.to(accelerator.device, dtype=weight_dtype)
if not args.train_text_encoder and text_encoder is not None:
text_encoder.to(accelerator.device, dtype=weight_dtype)
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# We need to initialize the trackers we use, and also store our configuration.
# The trackers initializes automatically on the main process.
if accelerator.is_main_process:
tracker_config = vars(copy.deepcopy(args))
tracker_config.pop("validation_images")
accelerator.init_trackers("dreambooth", config=tracker_config)
# Train!
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num batches each epoch = {len(train_dataloader)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.train_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
global_step = 0
first_epoch = 0
# Potentially load in the weights and states from a previous save
if args.resume_from_checkpoint:
if args.resume_from_checkpoint != "latest":
path = os.path.basename(args.resume_from_checkpoint)
else:
# Get the most recent checkpoint
dirs = os.listdir(args.output_dir)
dirs = [d for d in dirs if d.startswith("checkpoint")]
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
path = dirs[-1] if len(dirs) > 0 else None
if path is None:
accelerator.print(
f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
)
args.resume_from_checkpoint = None
initial_global_step = 0
else:
accelerator.print(f"Resuming from checkpoint {path}")
accelerator.load_state(os.path.join(args.output_dir, path))
global_step = int(path.split("-")[1])
initial_global_step = global_step
first_epoch = global_step // num_update_steps_per_epoch
else:
initial_global_step = 0
progress_bar = tqdm(
range(0, args.max_train_steps),
initial=initial_global_step,
desc="Steps",
# Only show the progress bar once on each machine.
disable=not accelerator.is_local_main_process,
)
for epoch in range(first_epoch, args.num_train_epochs):
unet.train()
if args.train_text_encoder:
text_encoder.train()
for step, batch in enumerate(train_dataloader):
with accelerator.accumulate(unet):
pixel_values = batch["pixel_values"].to(dtype=weight_dtype)
if vae is not None:
# Convert images to latent space
model_input = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
model_input = model_input * vae.config.scaling_factor
else:
model_input = pixel_values
# Sample noise that we'll add to the model input
if args.offset_noise:
noise = torch.randn_like(model_input) + 0.1 * torch.randn(
model_input.shape[0], model_input.shape[1], 1, 1, device=model_input.device
)
else:
noise = torch.randn_like(model_input)
bsz, channels, height, width = model_input.shape
# Sample a random timestep for each image
if args.loss_type == "huber" or args.loss_type == "smooth_l1":
timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (1,), device="cpu")
timestep = timesteps.item()
if args.huber_schedule == "exponential":
alpha = -math.log(args.huber_c) / noise_scheduler.config.num_train_timesteps
huber_c = math.exp(-alpha * timestep)
elif args.huber_schedule == "snr":
alphas_cumprod = noise_scheduler.alphas_cumprod[timestep]
sigmas = ((1.0 - alphas_cumprod) / alphas_cumprod) ** 0.5
huber_c = (1 - args.huber_c) / (1 + sigmas) ** 2 + args.huber_c
elif args.huber_schedule == "constant":
huber_c = args.huber_c
else:
raise NotImplementedError(f"Unknown Huber loss schedule {args.huber_schedule}!")
timesteps = timesteps.repeat(bsz).to(model_input.device)
elif args.loss_type == "l2":
timesteps = torch.randint(
0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
)
huber_c = 1 # may be anything, as it's not used
else:
raise NotImplementedError(f"Unknown loss type {args.loss_type}")
timesteps = timesteps.long()
# Add noise to the model input according to the noise magnitude at each timestep
# (this is the forward diffusion process)
noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
# Get the text embedding for conditioning
if args.pre_compute_text_embeddings:
encoder_hidden_states = batch["input_ids"]
else:
encoder_hidden_states = encode_prompt(
text_encoder,
batch["input_ids"],
batch["attention_mask"],
text_encoder_use_attention_mask=args.text_encoder_use_attention_mask,
)
if unwrap_model(unet).config.in_channels == channels * 2:
noisy_model_input = torch.cat([noisy_model_input, noisy_model_input], dim=1)
if args.class_labels_conditioning == "timesteps":
class_labels = timesteps
else:
class_labels = None
# Predict the noise residual
model_pred = unet(
noisy_model_input, timesteps, encoder_hidden_states, class_labels=class_labels, return_dict=False
)[0]
if model_pred.shape[1] == 6:
model_pred, _ = torch.chunk(model_pred, 2, dim=1)
# Get the target for loss depending on the prediction type
if noise_scheduler.config.prediction_type == "epsilon":
target = noise
elif noise_scheduler.config.prediction_type == "v_prediction":
target = noise_scheduler.get_velocity(model_input, noise, timesteps)
else:
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
if args.with_prior_preservation:
# Chunk the noise and model_pred into two parts and compute the loss on each part separately.
model_pred, model_pred_prior = torch.chunk(model_pred, 2, dim=0)
target, target_prior = torch.chunk(target, 2, dim=0)
# Compute prior loss
prior_loss = conditional_loss(
model_pred_prior.float(),
target_prior.float(),
reduction="mean",
loss_type=args.loss_type,
huber_c=huber_c,
)
# Compute instance loss
if args.snr_gamma is None:
loss = conditional_loss(
model_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c
)
else:
# Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
# Since we predict the noise instead of x_0, the original formulation is slightly changed.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(noise_scheduler, timesteps)
base_weight = (
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
)
if noise_scheduler.config.prediction_type == "v_prediction":
# Velocity objective needs to be floored to an SNR weight of one.
mse_loss_weights = base_weight + 1
else:
# Epsilon and sample both use the same loss weights.
mse_loss_weights = base_weight
loss = conditional_loss(
model_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
)
loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
loss = loss.mean()
if args.with_prior_preservation:
# Add the prior loss to the instance loss.
loss = loss + args.prior_loss_weight * prior_loss
accelerator.backward(loss)
if accelerator.sync_gradients:
params_to_clip = (
itertools.chain(unet.parameters(), text_encoder.parameters())
if args.train_text_encoder
else unet.parameters()
)
accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad(set_to_none=args.set_grads_to_none)
# Checks if the accelerator has performed an optimization step behind the scenes
if accelerator.sync_gradients:
progress_bar.update(1)
global_step += 1
if accelerator.is_main_process:
if global_step % args.checkpointing_steps == 0:
# _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
if args.checkpoints_total_limit is not None:
checkpoints = os.listdir(args.output_dir)
checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
# before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
if len(checkpoints) >= args.checkpoints_total_limit:
num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
removing_checkpoints = checkpoints[0:num_to_remove]
logger.info(
f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
)
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
for removing_checkpoint in removing_checkpoints:
removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
shutil.rmtree(removing_checkpoint)
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
accelerator.save_state(save_path)
logger.info(f"Saved state to {save_path}")
images = []
if args.validation_prompt is not None and global_step % args.validation_steps == 0:
images = log_validation(
unwrap_model(text_encoder) if text_encoder is not None else text_encoder,
tokenizer,
unwrap_model(unet),
vae,
args,
accelerator,
weight_dtype,
global_step,
validation_prompt_encoder_hidden_states,
validation_prompt_negative_prompt_embeds,
)
logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
progress_bar.set_postfix(**logs)
accelerator.log(logs, step=global_step)
if global_step >= args.max_train_steps:
break
# Create the pipeline using the trained modules and save it.
accelerator.wait_for_everyone()
if accelerator.is_main_process:
pipeline_args = {}
if text_encoder is not None:
pipeline_args["text_encoder"] = unwrap_model(text_encoder)
if args.skip_save_text_encoder:
pipeline_args["text_encoder"] = None
pipeline = DiffusionPipeline.from_pretrained(
args.pretrained_model_name_or_path,
unet=unwrap_model(unet),
revision=args.revision,
variant=args.variant,
**pipeline_args,
)
# We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
scheduler_args = {}
if "variance_type" in pipeline.scheduler.config:
variance_type = pipeline.scheduler.config.variance_type
if variance_type in ["learned", "learned_range"]:
variance_type = "fixed_small"
scheduler_args["variance_type"] = variance_type
pipeline.scheduler = pipeline.scheduler.from_config(pipeline.scheduler.config, **scheduler_args)
pipeline.save_pretrained(args.output_dir)
if args.push_to_hub:
save_model_card(
repo_id,
images=images,
base_model=args.pretrained_model_name_or_path,
train_text_encoder=args.train_text_encoder,
prompt=args.instance_prompt,
repo_folder=args.output_dir,
pipeline=pipeline,
)
upload_folder(
repo_id=repo_id,
folder_path=args.output_dir,
commit_message="End of training",
ignore_patterns=["step_*", "epoch_*"],
)
accelerator.end_training()
if __name__ == "__main__":
args = parse_args()
main(args)
examples/research_projects/scheduled_huber_loss_training/dreambooth/train_dreambooth_lora.py 0 → 100644
View file @ 2f234376
#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
import argparse
import copy
import gc
import logging
import math
import os
import shutil
import warnings
from pathlib import Path
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import ProjectConfiguration, set_seed
from huggingface_hub import create_repo, upload_folder
from huggingface_hub.utils import insecure_hashlib
from packaging import version
from peft import LoraConfig
from peft.utils import get_peft_model_state_dict, set_peft_model_state_dict
from PIL import Image
from PIL.ImageOps import exif_transpose
from torch.utils.data import Dataset
from torchvision import transforms
from tqdm.auto import tqdm
from transformers import AutoTokenizer, PretrainedConfig
import diffusers
from diffusers import (
AutoencoderKL,
DDPMScheduler,
DiffusionPipeline,
DPMSolverMultistepScheduler,
StableDiffusionPipeline,
UNet2DConditionModel,
)
from diffusers.loaders import LoraLoaderMixin
from diffusers.optimization import get_scheduler
from diffusers.training_utils import _set_state_dict_into_text_encoder, cast_training_params
from diffusers.utils import (
check_min_version,
convert_state_dict_to_diffusers,
convert_unet_state_dict_to_peft,
is_wandb_available,
)
from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.torch_utils import is_compiled_module
if is_wandb_available():
import wandb
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.28.0.dev0")
logger = get_logger(__name__)
def save_model_card(
repo_id: str,
images=None,
base_model=str,
train_text_encoder=False,
prompt=str,
repo_folder=None,
pipeline: DiffusionPipeline = None,
):
img_str = ""
for i, image in enumerate(images):
image.save(os.path.join(repo_folder, f"image_{i}.png"))
img_str += f"![img_{i}](./image_{i}.png)\n"
model_description = f"""
# LoRA DreamBooth - {repo_id}
These are LoRA adaption weights for {base_model}. The weights were trained on {prompt} using [DreamBooth](https://dreambooth.github.io/). You can find some example images in the following. \n
{img_str}
LoRA for the text encoder was enabled: {train_text_encoder}.
"""
model_card = load_or_create_model_card(
repo_id_or_path=repo_id,
from_training=True,
license="creativeml-openrail-m",
base_model=base_model,
prompt=prompt,
model_description=model_description,
inference=True,
)
tags = ["text-to-image", "diffusers", "lora", "diffusers-training"]
if isinstance(pipeline, StableDiffusionPipeline):
tags.extend(["stable-diffusion", "stable-diffusion-diffusers"])
else:
tags.extend(["if", "if-diffusers"])
model_card = populate_model_card(model_card, tags=tags)
model_card.save(os.path.join(repo_folder, "README.md"))
def log_validation(
pipeline,
args,
accelerator,
pipeline_args,
epoch,
is_final_validation=False,
):
logger.info(
f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
f" {args.validation_prompt}."
)
# We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
scheduler_args = {}
if "variance_type" in pipeline.scheduler.config:
variance_type = pipeline.scheduler.config.variance_type
if variance_type in ["learned", "learned_range"]:
variance_type = "fixed_small"
scheduler_args["variance_type"] = variance_type
pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, **scheduler_args)
pipeline = pipeline.to(accelerator.device)
pipeline.set_progress_bar_config(disable=True)
# run inference
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
if args.validation_images is None:
images = []
for _ in range(args.num_validation_images):
with torch.cuda.amp.autocast():
image = pipeline(**pipeline_args, generator=generator).images[0]
images.append(image)
else:
images = []
for image in args.validation_images:
image = Image.open(image)
with torch.cuda.amp.autocast():
image = pipeline(**pipeline_args, image=image, generator=generator).images[0]
images.append(image)
for tracker in accelerator.trackers:
phase_name = "test" if is_final_validation else "validation"
if tracker.name == "tensorboard":
np_images = np.stack([np.asarray(img) for img in images])
tracker.writer.add_images(phase_name, np_images, epoch, dataformats="NHWC")
if tracker.name == "wandb":
tracker.log(
{
phase_name: [
wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images)
]
}
)
del pipeline
torch.cuda.empty_cache()
return images
def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: str, revision: str):
text_encoder_config = PretrainedConfig.from_pretrained(
pretrained_model_name_or_path,
subfolder="text_encoder",
revision=revision,
)
model_class = text_encoder_config.architectures[0]
if model_class == "CLIPTextModel":
from transformers import CLIPTextModel
return CLIPTextModel
elif model_class == "RobertaSeriesModelWithTransformation":
from diffusers.pipelines.alt_diffusion.modeling_roberta_series import RobertaSeriesModelWithTransformation
return RobertaSeriesModelWithTransformation
elif model_class == "T5EncoderModel":
from transformers import T5EncoderModel
return T5EncoderModel
else:
raise ValueError(f"{model_class} is not supported.")
def parse_args(input_args=None):
parser = argparse.ArgumentParser(description="Simple example of a training script.")
parser.add_argument(
"--pretrained_model_name_or_path",
type=str,
default=None,
required=True,
help="Path to pretrained model or model identifier from huggingface.co/models.",
)
parser.add_argument(
"--revision",
type=str,
default=None,
required=False,
help="Revision of pretrained model identifier from huggingface.co/models.",
)
parser.add_argument(
"--variant",
type=str,
default=None,
help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
)
parser.add_argument(
"--tokenizer_name",
type=str,
default=None,
help="Pretrained tokenizer name or path if not the same as model_name",
)
parser.add_argument(
"--instance_data_dir",
type=str,
default=None,
required=True,
help="A folder containing the training data of instance images.",
)
parser.add_argument(
"--class_data_dir",
type=str,
default=None,
required=False,
help="A folder containing the training data of class images.",
)
parser.add_argument(
"--instance_prompt",
type=str,
default=None,
required=True,
help="The prompt with identifier specifying the instance",
)
parser.add_argument(
"--class_prompt",
type=str,
default=None,
help="The prompt to specify images in the same class as provided instance images.",
)
parser.add_argument(
"--validation_prompt",
type=str,
default=None,
help="A prompt that is used during validation to verify that the model is learning.",
)
parser.add_argument(
"--num_validation_images",
type=int,
default=4,
help="Number of images that should be generated during validation with `validation_prompt`.",
)
parser.add_argument(
"--validation_epochs",
type=int,
default=50,
help=(
"Run dreambooth validation every X epochs. Dreambooth validation consists of running the prompt"
" `args.validation_prompt` multiple times: `args.num_validation_images`."
),
)
parser.add_argument(
"--with_prior_preservation",
default=False,
action="store_true",
help="Flag to add prior preservation loss.",
)
parser.add_argument("--prior_loss_weight", type=float, default=1.0, help="The weight of prior preservation loss.")
parser.add_argument(
"--num_class_images",
type=int,
default=100,
help=(
"Minimal class images for prior preservation loss. If there are not enough images already present in"
" class_data_dir, additional images will be sampled with class_prompt."
),
)
parser.add_argument(
"--output_dir",
type=str,
default="lora-dreambooth-model",
help="The output directory where the model predictions and checkpoints will be written.",
)
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
parser.add_argument(
"--resolution",
type=int,
default=512,
help=(
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
" resolution"
),
)
parser.add_argument(
"--center_crop",
default=False,
action="store_true",
help=(
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
" cropped. The images will be resized to the resolution first before cropping."
),
)
parser.add_argument(
"--train_text_encoder",
action="store_true",
help="Whether to train the text encoder. If set, the text encoder should be float32 precision.",
)
parser.add_argument(
"--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
)
parser.add_argument(
"--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images."
)
parser.add_argument("--num_train_epochs", type=int, default=1)
parser.add_argument(
"--max_train_steps",
type=int,
default=None,
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
)
parser.add_argument(
"--checkpointing_steps",
type=int,
default=500,
help=(
"Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
" checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
" training using `--resume_from_checkpoint`."
),
)
parser.add_argument(
"--checkpoints_total_limit",
type=int,
default=None,
help=("Max number of checkpoints to store."),
)
parser.add_argument(
"--resume_from_checkpoint",
type=str,
default=None,
help=(
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
),
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--gradient_checkpointing",
action="store_true",
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
)
parser.add_argument(
"--learning_rate",
type=float,
default=5e-4,
help="Initial learning rate (after the potential warmup period) to use.",
)
parser.add_argument(
"--scale_lr",
action="store_true",
default=False,
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
)
parser.add_argument(
"--lr_scheduler",
type=str,
default="constant",
help=(
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
' "constant", "constant_with_warmup"]'
),
)
parser.add_argument(
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
)
parser.add_argument(
"--lr_num_cycles",
type=int,
default=1,
help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
)
parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
parser.add_argument(
"--dataloader_num_workers",
type=int,
default=0,
help=(
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
),
)
parser.add_argument(
"--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
)
parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
parser.add_argument(
"--hub_model_id",
type=str,
default=None,
help="The name of the repository to keep in sync with the local `output_dir`.",
)
parser.add_argument(
"--logging_dir",
type=str,
default="logs",
help=(
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
),
)
parser.add_argument(
"--allow_tf32",
action="store_true",
help=(
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
),
)
parser.add_argument(
"--report_to",
type=str,
default="tensorboard",
help=(
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
),
)
parser.add_argument(
"--mixed_precision",
type=str,
default=None,
choices=["no", "fp16", "bf16"],
help=(
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
),
)
parser.add_argument(
"--prior_generation_precision",
type=str,
default=None,
choices=["no", "fp32", "fp16", "bf16"],
help=(
"Choose prior generation precision between fp32, fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
" 1.10.and an Nvidia Ampere GPU. Default to fp16 if a GPU is available else fp32."
),
)
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
parser.add_argument(
"--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
)
parser.add_argument(
"--pre_compute_text_embeddings",
action="store_true",
help="Whether or not to pre-compute text embeddings. If text embeddings are pre-computed, the text encoder will not be kept in memory during training and will leave more GPU memory available for training the rest of the model. This is not compatible with `--train_text_encoder`.",
)
parser.add_argument(
"--tokenizer_max_length",
type=int,
default=None,
required=False,
help="The maximum length of the tokenizer. If not set, will default to the tokenizer's max length.",
)
parser.add_argument(
"--text_encoder_use_attention_mask",
action="store_true",
required=False,
help="Whether to use attention mask for the text encoder",
)
parser.add_argument(
"--validation_images",
required=False,
default=None,
nargs="+",
help="Optional set of images to use for validation. Used when the target pipeline takes an initial image as input such as when training image variation or superresolution.",
)
parser.add_argument(
"--class_labels_conditioning",
required=False,
default=None,
help="The optional `class_label` conditioning to pass to the unet, available values are `timesteps`.",
)
parser.add_argument(
"--loss_type",
type=str,
default="l2",
choices=["l2", "huber", "smooth_l1"],
help="The type of loss to use and whether it's timestep-scheduled. See Issue #7488 for more info.",
)
parser.add_argument(
"--huber_schedule",
type=str,
default="snr",
choices=["constant", "exponential", "snr"],
help="The schedule to use for the huber losses parameter",
)
parser.add_argument(
"--huber_c",
type=float,
default=0.1,
help="The huber loss parameter. Only used if one of the huber loss modes (huber or smooth l1) is selected with loss_type.",
)
parser.add_argument(
"--rank",
type=int,
default=4,
help=("The dimension of the LoRA update matrices."),
)
if input_args is not None:
args = parser.parse_args(input_args)
else:
args = parser.parse_args()
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
if env_local_rank != -1 and env_local_rank != args.local_rank:
args.local_rank = env_local_rank
if args.with_prior_preservation:
if args.class_data_dir is None:
raise ValueError("You must specify a data directory for class images.")
if args.class_prompt is None:
raise ValueError("You must specify prompt for class images.")
else:
# logger is not available yet
if args.class_data_dir is not None:
warnings.warn("You need not use --class_data_dir without --with_prior_preservation.")
if args.class_prompt is not None:
warnings.warn("You need not use --class_prompt without --with_prior_preservation.")
if args.train_text_encoder and args.pre_compute_text_embeddings:
raise ValueError("`--train_text_encoder` cannot be used with `--pre_compute_text_embeddings`")
return args
class DreamBoothDataset(Dataset):
"""
A dataset to prepare the instance and class images with the prompts for fine-tuning the model.
It pre-processes the images and the tokenizes prompts.
"""
def __init__(
self,
instance_data_root,
instance_prompt,
tokenizer,
class_data_root=None,
class_prompt=None,
class_num=None,
size=512,
center_crop=False,
encoder_hidden_states=None,
class_prompt_encoder_hidden_states=None,
tokenizer_max_length=None,
):
self.size = size
self.center_crop = center_crop
self.tokenizer = tokenizer
self.encoder_hidden_states = encoder_hidden_states
self.class_prompt_encoder_hidden_states = class_prompt_encoder_hidden_states
self.tokenizer_max_length = tokenizer_max_length
self.instance_data_root = Path(instance_data_root)
if not self.instance_data_root.exists():
raise ValueError("Instance images root doesn't exists.")
self.instance_images_path = list(Path(instance_data_root).iterdir())
self.num_instance_images = len(self.instance_images_path)
self.instance_prompt = instance_prompt
self._length = self.num_instance_images
if class_data_root is not None:
self.class_data_root = Path(class_data_root)
self.class_data_root.mkdir(parents=True, exist_ok=True)
self.class_images_path = list(self.class_data_root.iterdir())
if class_num is not None:
self.num_class_images = min(len(self.class_images_path), class_num)
else:
self.num_class_images = len(self.class_images_path)
self._length = max(self.num_class_images, self.num_instance_images)
self.class_prompt = class_prompt
else:
self.class_data_root = None
self.image_transforms = transforms.Compose(
[
transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
]
)
def __len__(self):
return self._length
def __getitem__(self, index):
example = {}
instance_image = Image.open(self.instance_images_path[index % self.num_instance_images])
instance_image = exif_transpose(instance_image)
if not instance_image.mode == "RGB":
instance_image = instance_image.convert("RGB")
example["instance_images"] = self.image_transforms(instance_image)
if self.encoder_hidden_states is not None:
example["instance_prompt_ids"] = self.encoder_hidden_states
else:
text_inputs = tokenize_prompt(
self.tokenizer, self.instance_prompt, tokenizer_max_length=self.tokenizer_max_length
)
example["instance_prompt_ids"] = text_inputs.input_ids
example["instance_attention_mask"] = text_inputs.attention_mask
if self.class_data_root:
class_image = Image.open(self.class_images_path[index % self.num_class_images])
class_image = exif_transpose(class_image)
if not class_image.mode == "RGB":
class_image = class_image.convert("RGB")
example["class_images"] = self.image_transforms(class_image)
if self.class_prompt_encoder_hidden_states is not None:
example["class_prompt_ids"] = self.class_prompt_encoder_hidden_states
else:
class_text_inputs = tokenize_prompt(
self.tokenizer, self.class_prompt, tokenizer_max_length=self.tokenizer_max_length
)
example["class_prompt_ids"] = class_text_inputs.input_ids
example["class_attention_mask"] = class_text_inputs.attention_mask
return example
def collate_fn(examples, with_prior_preservation=False):
has_attention_mask = "instance_attention_mask" in examples[0]
input_ids = [example["instance_prompt_ids"] for example in examples]
pixel_values = [example["instance_images"] for example in examples]
if has_attention_mask:
attention_mask = [example["instance_attention_mask"] for example in examples]
# Concat class and instance examples for prior preservation.
# We do this to avoid doing two forward passes.
if with_prior_preservation:
input_ids += [example["class_prompt_ids"] for example in examples]
pixel_values += [example["class_images"] for example in examples]
if has_attention_mask:
attention_mask += [example["class_attention_mask"] for example in examples]
pixel_values = torch.stack(pixel_values)
pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
input_ids = torch.cat(input_ids, dim=0)
batch = {
"input_ids": input_ids,
"pixel_values": pixel_values,
}
if has_attention_mask:
batch["attention_mask"] = attention_mask
return batch
class PromptDataset(Dataset):
"A simple dataset to prepare the prompts to generate class images on multiple GPUs."
def __init__(self, prompt, num_samples):
self.prompt = prompt
self.num_samples = num_samples
def __len__(self):
return self.num_samples
def __getitem__(self, index):
example = {}
example["prompt"] = self.prompt
example["index"] = index
return example
def tokenize_prompt(tokenizer, prompt, tokenizer_max_length=None):
if tokenizer_max_length is not None:
max_length = tokenizer_max_length
else:
max_length = tokenizer.model_max_length
text_inputs = tokenizer(
prompt,
truncation=True,
padding="max_length",
max_length=max_length,
return_tensors="pt",
)
return text_inputs
def encode_prompt(text_encoder, input_ids, attention_mask, text_encoder_use_attention_mask=None):
text_input_ids = input_ids.to(text_encoder.device)
if text_encoder_use_attention_mask:
attention_mask = attention_mask.to(text_encoder.device)
else:
attention_mask = None
prompt_embeds = text_encoder(
text_input_ids,
attention_mask=attention_mask,
return_dict=False,
)
prompt_embeds = prompt_embeds[0]
return prompt_embeds
# NOTE: if you're using the scheduled version, huber_c has to depend on the timesteps already
def conditional_loss(
model_pred: torch.Tensor,
target: torch.Tensor,
reduction: str = "mean",
loss_type: str = "l2",
huber_c: float = 0.1,
):
if loss_type == "l2":
loss = F.mse_loss(model_pred, target, reduction=reduction)
elif loss_type == "huber":
loss = 2 * huber_c * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
elif loss_type == "smooth_l1":
loss = 2 * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
else:
raise NotImplementedError(f"Unsupported Loss Type {loss_type}")
return loss
def main(args):
if args.report_to == "wandb" and args.hub_token is not None:
raise ValueError(
"You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
" Please use `huggingface-cli login` to authenticate with the Hub."
)
logging_dir = Path(args.output_dir, args.logging_dir)
accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
accelerator = Accelerator(
gradient_accumulation_steps=args.gradient_accumulation_steps,
mixed_precision=args.mixed_precision,
log_with=args.report_to,
project_config=accelerator_project_config,
)
if args.report_to == "wandb":
if not is_wandb_available():
raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
# Currently, it's not possible to do gradient accumulation when training two models with accelerate.accumulate
# This will be enabled soon in accelerate. For now, we don't allow gradient accumulation when training two models.
# TODO (sayakpaul): Remove this check when gradient accumulation with two models is enabled in accelerate.
if args.train_text_encoder and args.gradient_accumulation_steps > 1 and accelerator.num_processes > 1:
raise ValueError(
"Gradient accumulation is not supported when training the text encoder in distributed training. "
"Please set gradient_accumulation_steps to 1. This feature will be supported in the future."
)
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state, main_process_only=False)
if accelerator.is_local_main_process:
transformers.utils.logging.set_verbosity_warning()
diffusers.utils.logging.set_verbosity_info()
else:
transformers.utils.logging.set_verbosity_error()
diffusers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Generate class images if prior preservation is enabled.
if args.with_prior_preservation:
class_images_dir = Path(args.class_data_dir)
if not class_images_dir.exists():
class_images_dir.mkdir(parents=True)
cur_class_images = len(list(class_images_dir.iterdir()))
if cur_class_images < args.num_class_images:
torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32
if args.prior_generation_precision == "fp32":
torch_dtype = torch.float32
elif args.prior_generation_precision == "fp16":
torch_dtype = torch.float16
elif args.prior_generation_precision == "bf16":
torch_dtype = torch.bfloat16
pipeline = DiffusionPipeline.from_pretrained(
args.pretrained_model_name_or_path,
torch_dtype=torch_dtype,
safety_checker=None,
revision=args.revision,
variant=args.variant,
)
pipeline.set_progress_bar_config(disable=True)
num_new_images = args.num_class_images - cur_class_images
logger.info(f"Number of class images to sample: {num_new_images}.")
sample_dataset = PromptDataset(args.class_prompt, num_new_images)
sample_dataloader = torch.utils.data.DataLoader(sample_dataset, batch_size=args.sample_batch_size)
sample_dataloader = accelerator.prepare(sample_dataloader)
pipeline.to(accelerator.device)
for example in tqdm(
sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process
):
images = pipeline(example["prompt"]).images
for i, image in enumerate(images):
hash_image = insecure_hashlib.sha1(image.tobytes()).hexdigest()
image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
image.save(image_filename)
del pipeline
if torch.cuda.is_available():
torch.cuda.empty_cache()
# Handle the repository creation
if accelerator.is_main_process:
if args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
if args.push_to_hub:
repo_id = create_repo(
repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
).repo_id
# Load the tokenizer
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=False)
elif args.pretrained_model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
args.pretrained_model_name_or_path,
subfolder="tokenizer",
revision=args.revision,
use_fast=False,
)
# import correct text encoder class
text_encoder_cls = import_model_class_from_model_name_or_path(args.pretrained_model_name_or_path, args.revision)
# Load scheduler and models
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
text_encoder = text_encoder_cls.from_pretrained(
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
)
try:
vae = AutoencoderKL.from_pretrained(
args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant
)
except OSError:
# IF does not have a VAE so let's just set it to None
# We don't have to error out here
vae = None
unet = UNet2DConditionModel.from_pretrained(
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
)
# We only train the additional adapter LoRA layers
if vae is not None:
vae.requires_grad_(False)
text_encoder.requires_grad_(False)
unet.requires_grad_(False)
# For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) to half-precision
# as these weights are only used for inference, keeping weights in full precision is not required.
weight_dtype = torch.float32
if accelerator.mixed_precision == "fp16":
weight_dtype = torch.float16
elif accelerator.mixed_precision == "bf16":
weight_dtype = torch.bfloat16
# Move unet, vae and text_encoder to device and cast to weight_dtype
unet.to(accelerator.device, dtype=weight_dtype)
if vae is not None:
vae.to(accelerator.device, dtype=weight_dtype)
text_encoder.to(accelerator.device, dtype=weight_dtype)
if args.enable_xformers_memory_efficient_attention:
if is_xformers_available():
import xformers
xformers_version = version.parse(xformers.__version__)
if xformers_version == version.parse("0.0.16"):
logger.warning(
"xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
)
unet.enable_xformers_memory_efficient_attention()
else:
raise ValueError("xformers is not available. Make sure it is installed correctly")
if args.gradient_checkpointing:
unet.enable_gradient_checkpointing()
if args.train_text_encoder:
text_encoder.gradient_checkpointing_enable()
# now we will add new LoRA weights to the attention layers
unet_lora_config = LoraConfig(
r=args.rank,
lora_alpha=args.rank,
init_lora_weights="gaussian",
target_modules=["to_k", "to_q", "to_v", "to_out.0", "add_k_proj", "add_v_proj"],
)
unet.add_adapter(unet_lora_config)
# The text encoder comes from 🤗 transformers, we will also attach adapters to it.
if args.train_text_encoder:
text_lora_config = LoraConfig(
r=args.rank,
lora_alpha=args.rank,
init_lora_weights="gaussian",
target_modules=["q_proj", "k_proj", "v_proj", "out_proj"],
)
text_encoder.add_adapter(text_lora_config)
def unwrap_model(model):
model = accelerator.unwrap_model(model)
model = model._orig_mod if is_compiled_module(model) else model
return model
# create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
def save_model_hook(models, weights, output_dir):
if accelerator.is_main_process:
# there are only two options here. Either are just the unet attn processor layers
# or there are the unet and text encoder atten layers
unet_lora_layers_to_save = None
text_encoder_lora_layers_to_save = None
for model in models:
if isinstance(model, type(unwrap_model(unet))):
unet_lora_layers_to_save = convert_state_dict_to_diffusers(get_peft_model_state_dict(model))
elif isinstance(model, type(unwrap_model(text_encoder))):
text_encoder_lora_layers_to_save = convert_state_dict_to_diffusers(
get_peft_model_state_dict(model)
)
else:
raise ValueError(f"unexpected save model: {model.__class__}")
# make sure to pop weight so that corresponding model is not saved again
weights.pop()
LoraLoaderMixin.save_lora_weights(
output_dir,
unet_lora_layers=unet_lora_layers_to_save,
text_encoder_lora_layers=text_encoder_lora_layers_to_save,
)
def load_model_hook(models, input_dir):
unet_ = None
text_encoder_ = None
while len(models) > 0:
model = models.pop()
if isinstance(model, type(unwrap_model(unet))):
unet_ = model
elif isinstance(model, type(unwrap_model(text_encoder))):
text_encoder_ = model
else:
raise ValueError(f"unexpected save model: {model.__class__}")
lora_state_dict, network_alphas = LoraLoaderMixin.lora_state_dict(input_dir)
unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
if incompatible_keys is not None:
# check only for unexpected keys
unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
if unexpected_keys:
logger.warning(
f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
f" {unexpected_keys}. "
)
if args.train_text_encoder:
_set_state_dict_into_text_encoder(lora_state_dict, prefix="text_encoder.", text_encoder=text_encoder_)
# Make sure the trainable params are in float32. This is again needed since the base models
# are in `weight_dtype`. More details:
# https://github.com/huggingface/diffusers/pull/6514#discussion_r1449796804
if args.mixed_precision == "fp16":
models = [unet_]
if args.train_text_encoder:
models.append(text_encoder_)
# only upcast trainable parameters (LoRA) into fp32
cast_training_params(models, dtype=torch.float32)
accelerator.register_save_state_pre_hook(save_model_hook)
accelerator.register_load_state_pre_hook(load_model_hook)
# Enable TF32 for faster training on Ampere GPUs,
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
if args.allow_tf32:
torch.backends.cuda.matmul.allow_tf32 = True
if args.scale_lr:
args.learning_rate = (
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
)
# Make sure the trainable params are in float32.
if args.mixed_precision == "fp16":
models = [unet]
if args.train_text_encoder:
models.append(text_encoder)
# only upcast trainable parameters (LoRA) into fp32
cast_training_params(models, dtype=torch.float32)
# Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
if args.use_8bit_adam:
try:
import bitsandbytes as bnb
except ImportError:
raise ImportError(
"To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
)
optimizer_class = bnb.optim.AdamW8bit
else:
optimizer_class = torch.optim.AdamW
# Optimizer creation
params_to_optimize = list(filter(lambda p: p.requires_grad, unet.parameters()))
if args.train_text_encoder:
params_to_optimize = params_to_optimize + list(filter(lambda p: p.requires_grad, text_encoder.parameters()))
optimizer = optimizer_class(
params_to_optimize,
lr=args.learning_rate,
betas=(args.adam_beta1, args.adam_beta2),
weight_decay=args.adam_weight_decay,
eps=args.adam_epsilon,
)
if args.pre_compute_text_embeddings:
def compute_text_embeddings(prompt):
with torch.no_grad():
text_inputs = tokenize_prompt(tokenizer, prompt, tokenizer_max_length=args.tokenizer_max_length)
prompt_embeds = encode_prompt(
text_encoder,
text_inputs.input_ids,
text_inputs.attention_mask,
text_encoder_use_attention_mask=args.text_encoder_use_attention_mask,
)
return prompt_embeds
pre_computed_encoder_hidden_states = compute_text_embeddings(args.instance_prompt)
validation_prompt_negative_prompt_embeds = compute_text_embeddings("")
if args.validation_prompt is not None:
validation_prompt_encoder_hidden_states = compute_text_embeddings(args.validation_prompt)
else:
validation_prompt_encoder_hidden_states = None
if args.class_prompt is not None:
pre_computed_class_prompt_encoder_hidden_states = compute_text_embeddings(args.class_prompt)
else:
pre_computed_class_prompt_encoder_hidden_states = None
text_encoder = None
tokenizer = None
gc.collect()
torch.cuda.empty_cache()
else:
pre_computed_encoder_hidden_states = None
validation_prompt_encoder_hidden_states = None
validation_prompt_negative_prompt_embeds = None
pre_computed_class_prompt_encoder_hidden_states = None
# Dataset and DataLoaders creation:
train_dataset = DreamBoothDataset(
instance_data_root=args.instance_data_dir,
instance_prompt=args.instance_prompt,
class_data_root=args.class_data_dir if args.with_prior_preservation else None,
class_prompt=args.class_prompt,
class_num=args.num_class_images,
tokenizer=tokenizer,
size=args.resolution,
center_crop=args.center_crop,
encoder_hidden_states=pre_computed_encoder_hidden_states,
class_prompt_encoder_hidden_states=pre_computed_class_prompt_encoder_hidden_states,
tokenizer_max_length=args.tokenizer_max_length,
)
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.train_batch_size,
shuffle=True,
collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation),
num_workers=args.dataloader_num_workers,
)
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
overrode_max_train_steps = True
lr_scheduler = get_scheduler(
args.lr_scheduler,
optimizer=optimizer,
num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
num_training_steps=args.max_train_steps * accelerator.num_processes,
num_cycles=args.lr_num_cycles,
power=args.lr_power,
)
# Prepare everything with our `accelerator`.
if args.train_text_encoder:
unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
unet, text_encoder, optimizer, train_dataloader, lr_scheduler
)
else:
unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
unet, optimizer, train_dataloader, lr_scheduler
)
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# We need to initialize the trackers we use, and also store our configuration.
# The trackers initializes automatically on the main process.
if accelerator.is_main_process:
tracker_config = vars(copy.deepcopy(args))
tracker_config.pop("validation_images")
accelerator.init_trackers("dreambooth-lora", config=tracker_config)
# Train!
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num batches each epoch = {len(train_dataloader)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.train_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
global_step = 0
first_epoch = 0
# Potentially load in the weights and states from a previous save
if args.resume_from_checkpoint:
if args.resume_from_checkpoint != "latest":
path = os.path.basename(args.resume_from_checkpoint)
else:
# Get the mos recent checkpoint
dirs = os.listdir(args.output_dir)
dirs = [d for d in dirs if d.startswith("checkpoint")]
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
path = dirs[-1] if len(dirs) > 0 else None
if path is None:
accelerator.print(
f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
)
args.resume_from_checkpoint = None
initial_global_step = 0
else:
accelerator.print(f"Resuming from checkpoint {path}")
accelerator.load_state(os.path.join(args.output_dir, path))
global_step = int(path.split("-")[1])
initial_global_step = global_step
first_epoch = global_step // num_update_steps_per_epoch
else:
initial_global_step = 0
progress_bar = tqdm(
range(0, args.max_train_steps),
initial=initial_global_step,
desc="Steps",
# Only show the progress bar once on each machine.
disable=not accelerator.is_local_main_process,
)
for epoch in range(first_epoch, args.num_train_epochs):
unet.train()
if args.train_text_encoder:
text_encoder.train()
for step, batch in enumerate(train_dataloader):
with accelerator.accumulate(unet):
pixel_values = batch["pixel_values"].to(dtype=weight_dtype)
if vae is not None:
# Convert images to latent space
model_input = vae.encode(pixel_values).latent_dist.sample()
model_input = model_input * vae.config.scaling_factor
else:
model_input = pixel_values
# Sample noise that we'll add to the latents
noise = torch.randn_like(model_input)
bsz, channels, height, width = model_input.shape
# Sample a random timestep for each image
if args.loss_type == "huber" or args.loss_type == "smooth_l1":
timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (1,), device="cpu")
timestep = timesteps.item()
if args.huber_schedule == "exponential":
alpha = -math.log(args.huber_c) / noise_scheduler.config.num_train_timesteps
huber_c = math.exp(-alpha * timestep)
elif args.huber_schedule == "snr":
alphas_cumprod = noise_scheduler.alphas_cumprod[timestep]
sigmas = ((1.0 - alphas_cumprod) / alphas_cumprod) ** 0.5
huber_c = (1 - args.huber_c) / (1 + sigmas) ** 2 + args.huber_c
elif args.huber_schedule == "constant":
huber_c = args.huber_c
else:
raise NotImplementedError(f"Unknown Huber loss schedule {args.huber_schedule}!")
timesteps = timesteps.repeat(bsz).to(model_input.device)
elif args.loss_type == "l2":
timesteps = torch.randint(
0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
)
huber_c = 1 # may be anything, as it's not used
else:
raise NotImplementedError(f"Unknown loss type {args.loss_type}")
timesteps = timesteps.long()
# Add noise to the model input according to the noise magnitude at each timestep
# (this is the forward diffusion process)
noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
# Get the text embedding for conditioning
if args.pre_compute_text_embeddings:
encoder_hidden_states = batch["input_ids"]
else:
encoder_hidden_states = encode_prompt(
text_encoder,
batch["input_ids"],
batch["attention_mask"],
text_encoder_use_attention_mask=args.text_encoder_use_attention_mask,
)
if unwrap_model(unet).config.in_channels == channels * 2:
noisy_model_input = torch.cat([noisy_model_input, noisy_model_input], dim=1)
if args.class_labels_conditioning == "timesteps":
class_labels = timesteps
else:
class_labels = None
# Predict the noise residual
model_pred = unet(
noisy_model_input,
timesteps,
encoder_hidden_states,
class_labels=class_labels,
return_dict=False,
)[0]
# if model predicts variance, throw away the prediction. we will only train on the
# simplified training objective. This means that all schedulers using the fine tuned
# model must be configured to use one of the fixed variance variance types.
if model_pred.shape[1] == 6:
model_pred, _ = torch.chunk(model_pred, 2, dim=1)
# Get the target for loss depending on the prediction type
if noise_scheduler.config.prediction_type == "epsilon":
target = noise
elif noise_scheduler.config.prediction_type == "v_prediction":
target = noise_scheduler.get_velocity(model_input, noise, timesteps)
else:
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
if args.with_prior_preservation:
# Chunk the noise and model_pred into two parts and compute the loss on each part separately.
model_pred, model_pred_prior = torch.chunk(model_pred, 2, dim=0)
target, target_prior = torch.chunk(target, 2, dim=0)
# Compute instance loss
loss = conditional_loss(
model_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c
)
# Compute prior loss
prior_loss = conditional_loss(
model_pred_prior.float(),
target_prior.float(),
reduction="mean",
loss_type=args.loss_type,
huber_c=huber_c,
)
# Add the prior loss to the instance loss.
loss = loss + args.prior_loss_weight * prior_loss
else:
loss = conditional_loss(
model_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c
)
accelerator.backward(loss)
if accelerator.sync_gradients:
accelerator.clip_grad_norm_(params_to_optimize, args.max_grad_norm)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
# Checks if the accelerator has performed an optimization step behind the scenes
if accelerator.sync_gradients:
progress_bar.update(1)
global_step += 1
if accelerator.is_main_process:
if global_step % args.checkpointing_steps == 0:
# _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
if args.checkpoints_total_limit is not None:
checkpoints = os.listdir(args.output_dir)
checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
# before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
if len(checkpoints) >= args.checkpoints_total_limit:
num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
removing_checkpoints = checkpoints[0:num_to_remove]
logger.info(
f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
)
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
for removing_checkpoint in removing_checkpoints:
removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
shutil.rmtree(removing_checkpoint)
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
accelerator.save_state(save_path)
logger.info(f"Saved state to {save_path}")
logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
progress_bar.set_postfix(**logs)
accelerator.log(logs, step=global_step)
if global_step >= args.max_train_steps:
break
if accelerator.is_main_process:
if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
# create pipeline
pipeline = DiffusionPipeline.from_pretrained(
args.pretrained_model_name_or_path,
unet=unwrap_model(unet),
text_encoder=None if args.pre_compute_text_embeddings else unwrap_model(text_encoder),
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
)
if args.pre_compute_text_embeddings:
pipeline_args = {
"prompt_embeds": validation_prompt_encoder_hidden_states,
"negative_prompt_embeds": validation_prompt_negative_prompt_embeds,
}
else:
pipeline_args = {"prompt": args.validation_prompt}
images = log_validation(
pipeline,
args,
accelerator,
pipeline_args,
epoch,
)
# Save the lora layers
accelerator.wait_for_everyone()
if accelerator.is_main_process:
unet = unwrap_model(unet)
unet = unet.to(torch.float32)
unet_lora_state_dict = convert_state_dict_to_diffusers(get_peft_model_state_dict(unet))
if args.train_text_encoder:
text_encoder = unwrap_model(text_encoder)
text_encoder_state_dict = convert_state_dict_to_diffusers(get_peft_model_state_dict(text_encoder))
else:
text_encoder_state_dict = None
LoraLoaderMixin.save_lora_weights(
save_directory=args.output_dir,
unet_lora_layers=unet_lora_state_dict,
text_encoder_lora_layers=text_encoder_state_dict,
)
# Final inference
# Load previous pipeline
pipeline = DiffusionPipeline.from_pretrained(
args.pretrained_model_name_or_path, revision=args.revision, variant=args.variant, torch_dtype=weight_dtype
)
# load attention processors
pipeline.load_lora_weights(args.output_dir, weight_name="pytorch_lora_weights.safetensors")
# run inference
images = []
if args.validation_prompt and args.num_validation_images > 0:
pipeline_args = {"prompt": args.validation_prompt, "num_inference_steps": 25}
images = log_validation(
pipeline,
args,
accelerator,
pipeline_args,
epoch,
is_final_validation=True,
)
if args.push_to_hub:
save_model_card(
repo_id,
images=images,
base_model=args.pretrained_model_name_or_path,
train_text_encoder=args.train_text_encoder,
prompt=args.instance_prompt,
repo_folder=args.output_dir,
pipeline=pipeline,
)
upload_folder(
repo_id=repo_id,
folder_path=args.output_dir,
commit_message="End of training",
ignore_patterns=["step_*", "epoch_*"],
)
accelerator.end_training()
if __name__ == "__main__":
args = parse_args()
main(args)
examples/research_projects/scheduled_huber_loss_training/dreambooth/train_dreambooth_lora_sdxl.py 0 → 100644
View file @ 2f234376
#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
import argparse
import contextlib
import gc
import itertools
import json
import logging
import math
import os
import random
import shutil
import warnings
from pathlib import Path
from typing import Optional
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration, set_seed
from huggingface_hub import create_repo, hf_hub_download, upload_folder
from huggingface_hub.utils import insecure_hashlib
from packaging import version
from peft import LoraConfig, set_peft_model_state_dict
from peft.utils import get_peft_model_state_dict
from PIL import Image
from PIL.ImageOps import exif_transpose
from safetensors.torch import load_file, save_file
from torch.utils.data import Dataset
from torchvision import transforms
from torchvision.transforms.functional import crop
from tqdm.auto import tqdm
from transformers import AutoTokenizer, PretrainedConfig
import diffusers
from diffusers import (
AutoencoderKL,
DDPMScheduler,
DPMSolverMultistepScheduler,
EDMEulerScheduler,
EulerDiscreteScheduler,
StableDiffusionXLPipeline,
UNet2DConditionModel,
)
from diffusers.loaders import LoraLoaderMixin
from diffusers.optimization import get_scheduler
from diffusers.training_utils import _set_state_dict_into_text_encoder, cast_training_params, compute_snr
from diffusers.utils import (
check_min_version,
convert_all_state_dict_to_peft,
convert_state_dict_to_diffusers,
convert_state_dict_to_kohya,
convert_unet_state_dict_to_peft,
is_wandb_available,
)
from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.torch_utils import is_compiled_module
if is_wandb_available():
import wandb
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.28.0.dev0")
logger = get_logger(__name__)
def determine_scheduler_type(pretrained_model_name_or_path, revision):
model_index_filename = "model_index.json"
if os.path.isdir(pretrained_model_name_or_path):
model_index = os.path.join(pretrained_model_name_or_path, model_index_filename)
else:
model_index = hf_hub_download(
repo_id=pretrained_model_name_or_path, filename=model_index_filename, revision=revision
)
with open(model_index, "r") as f:
scheduler_type = json.load(f)["scheduler"][1]
return scheduler_type
def save_model_card(
repo_id: str,
use_dora: bool,
images=None,
base_model: str = None,
train_text_encoder=False,
instance_prompt=None,
validation_prompt=None,
repo_folder=None,
vae_path=None,
):
widget_dict = []
if images is not None:
for i, image in enumerate(images):
image.save(os.path.join(repo_folder, f"image_{i}.png"))
widget_dict.append(
{"text": validation_prompt if validation_prompt else " ", "output": {"url": f"image_{i}.png"}}
)
model_description = f"""
# {'SDXL' if 'playground' not in base_model else 'Playground'} LoRA DreamBooth - {repo_id}
<Gallery />
## Model description
These are {repo_id} LoRA adaption weights for {base_model}.
The weights were trained using [DreamBooth](https://dreambooth.github.io/).
LoRA for the text encoder was enabled: {train_text_encoder}.
Special VAE used for training: {vae_path}.
## Trigger words
You should use {instance_prompt} to trigger the image generation.
## Download model
Weights for this model are available in Safetensors format.
[Download]({repo_id}/tree/main) them in the Files & versions tab.
"""
if "playground" in base_model:
model_description += """\n
## License
Please adhere to the licensing terms as described [here](https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic/blob/main/LICENSE.md).
"""
model_card = load_or_create_model_card(
repo_id_or_path=repo_id,
from_training=True,
license="openrail++" if "playground" not in base_model else "playground-v2dot5-community",
base_model=base_model,
prompt=instance_prompt,
model_description=model_description,
widget=widget_dict,
)
tags = [
"text-to-image",
"text-to-image",
"diffusers-training",
"diffusers",
"lora" if not use_dora else "dora",
"template:sd-lora",
]
if "playground" in base_model:
tags.extend(["playground", "playground-diffusers"])
else:
tags.extend(["stable-diffusion-xl", "stable-diffusion-xl-diffusers"])
model_card = populate_model_card(model_card, tags=tags)
model_card.save(os.path.join(repo_folder, "README.md"))
def log_validation(
pipeline,
args,
accelerator,
pipeline_args,
epoch,
is_final_validation=False,
):
logger.info(
f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
f" {args.validation_prompt}."
)
# We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
scheduler_args = {}
if not args.do_edm_style_training:
if "variance_type" in pipeline.scheduler.config:
variance_type = pipeline.scheduler.config.variance_type
if variance_type in ["learned", "learned_range"]:
variance_type = "fixed_small"
scheduler_args["variance_type"] = variance_type
pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, **scheduler_args)
pipeline = pipeline.to(accelerator.device)
pipeline.set_progress_bar_config(disable=True)
# run inference
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
# Currently the context determination is a bit hand-wavy. We can improve it in the future if there's a better
# way to condition it. Reference: https://github.com/huggingface/diffusers/pull/7126#issuecomment-1968523051
inference_ctx = (
contextlib.nullcontext() if "playground" in args.pretrained_model_name_or_path else torch.cuda.amp.autocast()
)
with inference_ctx:
images = [pipeline(**pipeline_args, generator=generator).images[0] for _ in range(args.num_validation_images)]
for tracker in accelerator.trackers:
phase_name = "test" if is_final_validation else "validation"
if tracker.name == "tensorboard":
np_images = np.stack([np.asarray(img) for img in images])
tracker.writer.add_images(phase_name, np_images, epoch, dataformats="NHWC")
if tracker.name == "wandb":
tracker.log(
{
phase_name: [
wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images)
]
}
)
del pipeline
torch.cuda.empty_cache()
return images
def import_model_class_from_model_name_or_path(
pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
):
text_encoder_config = PretrainedConfig.from_pretrained(
pretrained_model_name_or_path, subfolder=subfolder, revision=revision
)
model_class = text_encoder_config.architectures[0]
if model_class == "CLIPTextModel":
from transformers import CLIPTextModel
return CLIPTextModel
elif model_class == "CLIPTextModelWithProjection":
from transformers import CLIPTextModelWithProjection
return CLIPTextModelWithProjection
else:
raise ValueError(f"{model_class} is not supported.")
def parse_args(input_args=None):
parser = argparse.ArgumentParser(description="Simple example of a training script.")
parser.add_argument(
"--pretrained_model_name_or_path",
type=str,
default=None,
required=True,
help="Path to pretrained model or model identifier from huggingface.co/models.",
)
parser.add_argument(
"--pretrained_vae_model_name_or_path",
type=str,
default=None,
help="Path to pretrained VAE model with better numerical stability. More details: https://github.com/huggingface/diffusers/pull/4038.",
)
parser.add_argument(
"--revision",
type=str,
default=None,
required=False,
help="Revision of pretrained model identifier from huggingface.co/models.",
)
parser.add_argument(
"--variant",
type=str,
default=None,
help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
)
parser.add_argument(
"--dataset_name",
type=str,
default=None,
help=(
"The name of the Dataset (from the HuggingFace hub) containing the training data of instance images (could be your own, possibly private,"
" dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
" or to a folder containing files that 🤗 Datasets can understand."
),
)
parser.add_argument(
"--dataset_config_name",
type=str,
default=None,
help="The config of the Dataset, leave as None if there's only one config.",
)
parser.add_argument(
"--instance_data_dir",
type=str,
default=None,
help=("A folder containing the training data. "),
)
parser.add_argument(
"--cache_dir",
type=str,
default=None,
help="The directory where the downloaded models and datasets will be stored.",
)
parser.add_argument(
"--image_column",
type=str,
default="image",
help="The column of the dataset containing the target image. By "
"default, the standard Image Dataset maps out 'file_name' "
"to 'image'.",
)
parser.add_argument(
"--caption_column",
type=str,
default=None,
help="The column of the dataset containing the instance prompt for each image",
)
parser.add_argument("--repeats", type=int, default=1, help="How many times to repeat the training data.")
parser.add_argument(
"--class_data_dir",
type=str,
default=None,
required=False,
help="A folder containing the training data of class images.",
)
parser.add_argument(
"--instance_prompt",
type=str,
default=None,
required=True,
help="The prompt with identifier specifying the instance, e.g. 'photo of a TOK dog', 'in the style of TOK'",
)
parser.add_argument(
"--class_prompt",
type=str,
default=None,
help="The prompt to specify images in the same class as provided instance images.",
)
parser.add_argument(
"--validation_prompt",
type=str,
default=None,
help="A prompt that is used during validation to verify that the model is learning.",
)
parser.add_argument(
"--num_validation_images",
type=int,
default=4,
help="Number of images that should be generated during validation with `validation_prompt`.",
)
parser.add_argument(
"--validation_epochs",
type=int,
default=50,
help=(
"Run dreambooth validation every X epochs. Dreambooth validation consists of running the prompt"
" `args.validation_prompt` multiple times: `args.num_validation_images`."
),
)
parser.add_argument(
"--do_edm_style_training",
default=False,
action="store_true",
help="Flag to conduct training using the EDM formulation as introduced in https://arxiv.org/abs/2206.00364.",
)
parser.add_argument(
"--with_prior_preservation",
default=False,
action="store_true",
help="Flag to add prior preservation loss.",
)
parser.add_argument("--prior_loss_weight", type=float, default=1.0, help="The weight of prior preservation loss.")
parser.add_argument(
"--num_class_images",
type=int,
default=100,
help=(
"Minimal class images for prior preservation loss. If there are not enough images already present in"
" class_data_dir, additional images will be sampled with class_prompt."
),
)
parser.add_argument(
"--output_dir",
type=str,
default="lora-dreambooth-model",
help="The output directory where the model predictions and checkpoints will be written.",
)
parser.add_argument(
"--output_kohya_format",
action="store_true",
help="Flag to additionally generate final state dict in the Kohya format so that it becomes compatible with A111, Comfy, Kohya, etc.",
)
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
parser.add_argument(
"--resolution",
type=int,
default=1024,
help=(
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
" resolution"
),
)
parser.add_argument(
"--center_crop",
default=False,
action="store_true",
help=(
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
" cropped. The images will be resized to the resolution first before cropping."
),
)
parser.add_argument(
"--random_flip",
action="store_true",
help="whether to randomly flip images horizontally",
)
parser.add_argument(
"--train_text_encoder",
action="store_true",
help="Whether to train the text encoder. If set, the text encoder should be float32 precision.",
)
parser.add_argument(
"--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
)
parser.add_argument(
"--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images."
)
parser.add_argument("--num_train_epochs", type=int, default=1)
parser.add_argument(
"--max_train_steps",
type=int,
default=None,
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
)
parser.add_argument(
"--checkpointing_steps",
type=int,
default=500,
help=(
"Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
" checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
" training using `--resume_from_checkpoint`."
),
)
parser.add_argument(
"--checkpoints_total_limit",
type=int,
default=None,
help=("Max number of checkpoints to store."),
)
parser.add_argument(
"--resume_from_checkpoint",
type=str,
default=None,
help=(
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
),
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--gradient_checkpointing",
action="store_true",
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
)
parser.add_argument(
"--learning_rate",
type=float,
default=1e-4,
help="Initial learning rate (after the potential warmup period) to use.",
)
parser.add_argument(
"--text_encoder_lr",
type=float,
default=5e-6,
help="Text encoder learning rate to use.",
)
parser.add_argument(
"--scale_lr",
action="store_true",
default=False,
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
)
parser.add_argument(
"--lr_scheduler",
type=str,
default="constant",
help=(
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
' "constant", "constant_with_warmup"]'
),
)
parser.add_argument(
"--snr_gamma",
type=float,
default=None,
help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
"More details here: https://arxiv.org/abs/2303.09556.",
)
parser.add_argument(
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
)
parser.add_argument(
"--lr_num_cycles",
type=int,
default=1,
help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
)
parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
parser.add_argument(
"--dataloader_num_workers",
type=int,
default=0,
help=(
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
),
)
parser.add_argument(
"--optimizer",
type=str,
default="AdamW",
help=('The optimizer type to use. Choose between ["AdamW", "prodigy"]'),
)
parser.add_argument(
"--use_8bit_adam",
action="store_true",
help="Whether or not to use 8-bit Adam from bitsandbytes. Ignored if optimizer is not set to AdamW",
)
parser.add_argument(
"--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam and Prodigy optimizers."
)
parser.add_argument(
"--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam and Prodigy optimizers."
)
parser.add_argument(
"--prodigy_beta3",
type=float,
default=None,
help="coefficients for computing the Prodidy stepsize using running averages. If set to None, "
"uses the value of square root of beta2. Ignored if optimizer is adamW",
)
parser.add_argument("--prodigy_decouple", type=bool, default=True, help="Use AdamW style decoupled weight decay")
parser.add_argument("--adam_weight_decay", type=float, default=1e-04, help="Weight decay to use for unet params")
parser.add_argument(
"--adam_weight_decay_text_encoder", type=float, default=1e-03, help="Weight decay to use for text_encoder"
)
parser.add_argument(
"--adam_epsilon",
type=float,
default=1e-08,
help="Epsilon value for the Adam optimizer and Prodigy optimizers.",
)
parser.add_argument(
"--prodigy_use_bias_correction",
type=bool,
default=True,
help="Turn on Adam's bias correction. True by default. Ignored if optimizer is adamW",
)
parser.add_argument(
"--prodigy_safeguard_warmup",
type=bool,
default=True,
help="Remove lr from the denominator of D estimate to avoid issues during warm-up stage. True by default. "
"Ignored if optimizer is adamW",
)
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
parser.add_argument(
"--hub_model_id",
type=str,
default=None,
help="The name of the repository to keep in sync with the local `output_dir`.",
)
parser.add_argument(
"--logging_dir",
type=str,
default="logs",
help=(
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
),
)
parser.add_argument(
"--allow_tf32",
action="store_true",
help=(
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
),
)
parser.add_argument(
"--report_to",
type=str,
default="tensorboard",
help=(
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
),
)
parser.add_argument(
"--mixed_precision",
type=str,
default=None,
choices=["no", "fp16", "bf16"],
help=(
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
),
)
parser.add_argument(
"--prior_generation_precision",
type=str,
default=None,
choices=["no", "fp32", "fp16", "bf16"],
help=(
"Choose prior generation precision between fp32, fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
" 1.10.and an Nvidia Ampere GPU. Default to fp16 if a GPU is available else fp32."
),
)
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
parser.add_argument(
"--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
)
parser.add_argument(
"--rank",
type=int,
default=4,
help=("The dimension of the LoRA update matrices."),
)
parser.add_argument(
"--use_dora",
action="store_true",
default=False,
help=(
"Wether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
"Note: to use DoRA you need to install peft from main, `pip install git+https://github.com/huggingface/peft.git`"
),
)
parser.add_argument(
"--loss_type",
type=str,
default="l2",
choices=["l2", "huber", "smooth_l1"],
help="The type of loss to use and whether it's timestep-scheduled. See Issue #7488 for more info.",
)
parser.add_argument(
"--huber_schedule",
type=str,
default="snr",
choices=["constant", "exponential", "snr"],
help="The schedule to use for the huber losses parameter",
)
parser.add_argument(
"--huber_c",
type=float,
default=0.1,
help="The huber loss parameter. Only used if one of the huber loss modes (huber or smooth l1) is selected with loss_type.",
)
if input_args is not None:
args = parser.parse_args(input_args)
else:
args = parser.parse_args()
if args.dataset_name is None and args.instance_data_dir is None:
raise ValueError("Specify either `--dataset_name` or `--instance_data_dir`")
if args.dataset_name is not None and args.instance_data_dir is not None:
raise ValueError("Specify only one of `--dataset_name` or `--instance_data_dir`")
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
if env_local_rank != -1 and env_local_rank != args.local_rank:
args.local_rank = env_local_rank
if args.with_prior_preservation:
if args.class_data_dir is None:
raise ValueError("You must specify a data directory for class images.")
if args.class_prompt is None:
raise ValueError("You must specify prompt for class images.")
else:
# logger is not available yet
if args.class_data_dir is not None:
warnings.warn("You need not use --class_data_dir without --with_prior_preservation.")
if args.class_prompt is not None:
warnings.warn("You need not use --class_prompt without --with_prior_preservation.")
return args
class DreamBoothDataset(Dataset):
"""
A dataset to prepare the instance and class images with the prompts for fine-tuning the model.
It pre-processes the images.
"""
def __init__(
self,
instance_data_root,
instance_prompt,
class_prompt,
class_data_root=None,
class_num=None,
size=1024,
repeats=1,
center_crop=False,
):
self.size = size
self.center_crop = center_crop
self.instance_prompt = instance_prompt
self.custom_instance_prompts = None
self.class_prompt = class_prompt
# if --dataset_name is provided or a metadata jsonl file is provided in the local --instance_data directory,
# we load the training data using load_dataset
if args.dataset_name is not None:
try:
from datasets import load_dataset
except ImportError:
raise ImportError(
"You are trying to load your data using the datasets library. If you wish to train using custom "
"captions please install the datasets library: `pip install datasets`. If you wish to load a "
"local folder containing images only, specify --instance_data_dir instead."
)
# Downloading and loading a dataset from the hub.
# See more about loading custom images at
# https://huggingface.co/docs/datasets/v2.0.0/en/dataset_script
dataset = load_dataset(
args.dataset_name,
args.dataset_config_name,
cache_dir=args.cache_dir,
)
# Preprocessing the datasets.
column_names = dataset["train"].column_names
# 6. Get the column names for input/target.
if args.image_column is None:
image_column = column_names[0]
logger.info(f"image column defaulting to {image_column}")
else:
image_column = args.image_column
if image_column not in column_names:
raise ValueError(
f"`--image_column` value '{args.image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
)
instance_images = dataset["train"][image_column]
if args.caption_column is None:
logger.info(
"No caption column provided, defaulting to instance_prompt for all images. If your dataset "
"contains captions/prompts for the images, make sure to specify the "
"column as --caption_column"
)
self.custom_instance_prompts = None
else:
if args.caption_column not in column_names:
raise ValueError(
f"`--caption_column` value '{args.caption_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
)
custom_instance_prompts = dataset["train"][args.caption_column]
# create final list of captions according to --repeats
self.custom_instance_prompts = []
for caption in custom_instance_prompts:
self.custom_instance_prompts.extend(itertools.repeat(caption, repeats))
else:
self.instance_data_root = Path(instance_data_root)
if not self.instance_data_root.exists():
raise ValueError("Instance images root doesn't exists.")
instance_images = [Image.open(path) for path in list(Path(instance_data_root).iterdir())]
self.custom_instance_prompts = None
self.instance_images = []
for img in instance_images:
self.instance_images.extend(itertools.repeat(img, repeats))
# image processing to prepare for using SD-XL micro-conditioning
self.original_sizes = []
self.crop_top_lefts = []
self.pixel_values = []
train_resize = transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR)
train_crop = transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size)
train_flip = transforms.RandomHorizontalFlip(p=1.0)
train_transforms = transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
]
)
for image in self.instance_images:
image = exif_transpose(image)
if not image.mode == "RGB":
image = image.convert("RGB")
self.original_sizes.append((image.height, image.width))
image = train_resize(image)
if args.random_flip and random.random() < 0.5:
# flip
image = train_flip(image)
if args.center_crop:
y1 = max(0, int(round((image.height - args.resolution) / 2.0)))
x1 = max(0, int(round((image.width - args.resolution) / 2.0)))
image = train_crop(image)
else:
y1, x1, h, w = train_crop.get_params(image, (args.resolution, args.resolution))
image = crop(image, y1, x1, h, w)
crop_top_left = (y1, x1)
self.crop_top_lefts.append(crop_top_left)
image = train_transforms(image)
self.pixel_values.append(image)
self.num_instance_images = len(self.instance_images)
self._length = self.num_instance_images
if class_data_root is not None:
self.class_data_root = Path(class_data_root)
self.class_data_root.mkdir(parents=True, exist_ok=True)
self.class_images_path = list(self.class_data_root.iterdir())
if class_num is not None:
self.num_class_images = min(len(self.class_images_path), class_num)
else:
self.num_class_images = len(self.class_images_path)
self._length = max(self.num_class_images, self.num_instance_images)
else:
self.class_data_root = None
self.image_transforms = transforms.Compose(
[
transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
]
)
def __len__(self):
return self._length
def __getitem__(self, index):
example = {}
instance_image = self.pixel_values[index % self.num_instance_images]
original_size = self.original_sizes[index % self.num_instance_images]
crop_top_left = self.crop_top_lefts[index % self.num_instance_images]
example["instance_images"] = instance_image
example["original_size"] = original_size
example["crop_top_left"] = crop_top_left
if self.custom_instance_prompts:
caption = self.custom_instance_prompts[index % self.num_instance_images]
if caption:
example["instance_prompt"] = caption
else:
example["instance_prompt"] = self.instance_prompt
else: # costum prompts were provided, but length does not match size of image dataset
example["instance_prompt"] = self.instance_prompt
if self.class_data_root:
class_image = Image.open(self.class_images_path[index % self.num_class_images])
class_image = exif_transpose(class_image)
if not class_image.mode == "RGB":
class_image = class_image.convert("RGB")
example["class_images"] = self.image_transforms(class_image)
example["class_prompt"] = self.class_prompt
return example
def collate_fn(examples, with_prior_preservation=False):
pixel_values = [example["instance_images"] for example in examples]
prompts = [example["instance_prompt"] for example in examples]
original_sizes = [example["original_size"] for example in examples]
crop_top_lefts = [example["crop_top_left"] for example in examples]
# Concat class and instance examples for prior preservation.
# We do this to avoid doing two forward passes.
if with_prior_preservation:
pixel_values += [example["class_images"] for example in examples]
prompts += [example["class_prompt"] for example in examples]
original_sizes += [example["original_size"] for example in examples]
crop_top_lefts += [example["crop_top_left"] for example in examples]
pixel_values = torch.stack(pixel_values)
pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
batch = {
"pixel_values": pixel_values,
"prompts": prompts,
"original_sizes": original_sizes,
"crop_top_lefts": crop_top_lefts,
}
return batch
class PromptDataset(Dataset):
"A simple dataset to prepare the prompts to generate class images on multiple GPUs."
def __init__(self, prompt, num_samples):
self.prompt = prompt
self.num_samples = num_samples
def __len__(self):
return self.num_samples
def __getitem__(self, index):
example = {}
example["prompt"] = self.prompt
example["index"] = index
return example
def tokenize_prompt(tokenizer, prompt):
text_inputs = tokenizer(
prompt,
padding="max_length",
max_length=tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
return text_input_ids
# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
def encode_prompt(text_encoders, tokenizers, prompt, text_input_ids_list=None):
prompt_embeds_list = []
for i, text_encoder in enumerate(text_encoders):
if tokenizers is not None:
tokenizer = tokenizers[i]
text_input_ids = tokenize_prompt(tokenizer, prompt)
else:
assert text_input_ids_list is not None
text_input_ids = text_input_ids_list[i]
prompt_embeds = text_encoder(
text_input_ids.to(text_encoder.device), output_hidden_states=True, return_dict=False
)
# We are only ALWAYS interested in the pooled output of the final text encoder
pooled_prompt_embeds = prompt_embeds[0]
prompt_embeds = prompt_embeds[-1][-2]
bs_embed, seq_len, _ = prompt_embeds.shape
prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
prompt_embeds_list.append(prompt_embeds)
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
return prompt_embeds, pooled_prompt_embeds
# NOTE: if you're using the scheduled version, huber_c has to depend on the timesteps already
def conditional_loss(
model_pred: torch.Tensor,
target: torch.Tensor,
reduction: str = "mean",
loss_type: str = "l2",
huber_c: float = 0.1,
weighting: Optional[torch.Tensor] = None,
):
if loss_type == "l2":
if weighting is not None:
loss = torch.mean(
(weighting * (model_pred.float() - target.float()) ** 2).reshape(target.shape[0], -1),
1,
)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
else:
loss = F.mse_loss(model_pred.float(), target.float(), reduction=reduction)
elif loss_type == "huber":
if weighting is not None:
loss = torch.mean(
(
2
* huber_c
* (
torch.sqrt(weighting.float() * (model_pred.float() - target.float()) ** 2 + huber_c**2)
- huber_c
)
).reshape(target.shape[0], -1),
1,
)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
else:
loss = 2 * huber_c * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
elif loss_type == "smooth_l1":
if weighting is not None:
loss = torch.mean(
(
2
* (
torch.sqrt(weighting.float() * (model_pred.float() - target.float()) ** 2 + huber_c**2)
- huber_c
)
).reshape(target.shape[0], -1),
1,
)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
else:
loss = 2 * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
else:
raise NotImplementedError(f"Unsupported Loss Type {loss_type}")
return loss
def main(args):
if args.report_to == "wandb" and args.hub_token is not None:
raise ValueError(
"You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
" Please use `huggingface-cli login` to authenticate with the Hub."
)
if args.do_edm_style_training and args.snr_gamma is not None:
raise ValueError("Min-SNR formulation is not supported when conducting EDM-style training.")
logging_dir = Path(args.output_dir, args.logging_dir)
accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
accelerator = Accelerator(
gradient_accumulation_steps=args.gradient_accumulation_steps,
mixed_precision=args.mixed_precision,
log_with=args.report_to,
project_config=accelerator_project_config,
kwargs_handlers=[kwargs],
)
if args.report_to == "wandb":
if not is_wandb_available():
raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state, main_process_only=False)
if accelerator.is_local_main_process:
transformers.utils.logging.set_verbosity_warning()
diffusers.utils.logging.set_verbosity_info()
else:
transformers.utils.logging.set_verbosity_error()
diffusers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Generate class images if prior preservation is enabled.
if args.with_prior_preservation:
class_images_dir = Path(args.class_data_dir)
if not class_images_dir.exists():
class_images_dir.mkdir(parents=True)
cur_class_images = len(list(class_images_dir.iterdir()))
if cur_class_images < args.num_class_images:
torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32
if args.prior_generation_precision == "fp32":
torch_dtype = torch.float32
elif args.prior_generation_precision == "fp16":
torch_dtype = torch.float16
elif args.prior_generation_precision == "bf16":
torch_dtype = torch.bfloat16
pipeline = StableDiffusionXLPipeline.from_pretrained(
args.pretrained_model_name_or_path,
torch_dtype=torch_dtype,
revision=args.revision,
variant=args.variant,
)
pipeline.set_progress_bar_config(disable=True)
num_new_images = args.num_class_images - cur_class_images
logger.info(f"Number of class images to sample: {num_new_images}.")
sample_dataset = PromptDataset(args.class_prompt, num_new_images)
sample_dataloader = torch.utils.data.DataLoader(sample_dataset, batch_size=args.sample_batch_size)
sample_dataloader = accelerator.prepare(sample_dataloader)
pipeline.to(accelerator.device)
for example in tqdm(
sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process
):
images = pipeline(example["prompt"]).images
for i, image in enumerate(images):
hash_image = insecure_hashlib.sha1(image.tobytes()).hexdigest()
image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
image.save(image_filename)
del pipeline
if torch.cuda.is_available():
torch.cuda.empty_cache()
# Handle the repository creation
if accelerator.is_main_process:
if args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
if args.push_to_hub:
repo_id = create_repo(
repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
).repo_id
# Load the tokenizers
tokenizer_one = AutoTokenizer.from_pretrained(
args.pretrained_model_name_or_path,
subfolder="tokenizer",
revision=args.revision,
use_fast=False,
)
tokenizer_two = AutoTokenizer.from_pretrained(
args.pretrained_model_name_or_path,
subfolder="tokenizer_2",
revision=args.revision,
use_fast=False,
)
# import correct text encoder classes
text_encoder_cls_one = import_model_class_from_model_name_or_path(
args.pretrained_model_name_or_path, args.revision
)
text_encoder_cls_two = import_model_class_from_model_name_or_path(
args.pretrained_model_name_or_path, args.revision, subfolder="text_encoder_2"
)
# Load scheduler and models
scheduler_type = determine_scheduler_type(args.pretrained_model_name_or_path, args.revision)
if "EDM" in scheduler_type:
args.do_edm_style_training = True
noise_scheduler = EDMEulerScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
logger.info("Performing EDM-style training!")
elif args.do_edm_style_training:
noise_scheduler = EulerDiscreteScheduler.from_pretrained(
args.pretrained_model_name_or_path, subfolder="scheduler"
)
logger.info("Performing EDM-style training!")
else:
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
text_encoder_one = text_encoder_cls_one.from_pretrained(
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
)
text_encoder_two = text_encoder_cls_two.from_pretrained(
args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
)
vae_path = (
args.pretrained_model_name_or_path
if args.pretrained_vae_model_name_or_path is None
else args.pretrained_vae_model_name_or_path
)
vae = AutoencoderKL.from_pretrained(
vae_path,
subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
revision=args.revision,
variant=args.variant,
)
latents_mean = latents_std = None
if hasattr(vae.config, "latents_mean") and vae.config.latents_mean is not None:
latents_mean = torch.tensor(vae.config.latents_mean).view(1, 4, 1, 1)
if hasattr(vae.config, "latents_std") and vae.config.latents_std is not None:
latents_std = torch.tensor(vae.config.latents_std).view(1, 4, 1, 1)
unet = UNet2DConditionModel.from_pretrained(
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
)
# We only train the additional adapter LoRA layers
vae.requires_grad_(False)
text_encoder_one.requires_grad_(False)
text_encoder_two.requires_grad_(False)
unet.requires_grad_(False)
# For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) to half-precision
# as these weights are only used for inference, keeping weights in full precision is not required.
weight_dtype = torch.float32
if accelerator.mixed_precision == "fp16":
weight_dtype = torch.float16
elif accelerator.mixed_precision == "bf16":
weight_dtype = torch.bfloat16
# Move unet, vae and text_encoder to device and cast to weight_dtype
unet.to(accelerator.device, dtype=weight_dtype)
# The VAE is always in float32 to avoid NaN losses.
vae.to(accelerator.device, dtype=torch.float32)
text_encoder_one.to(accelerator.device, dtype=weight_dtype)
text_encoder_two.to(accelerator.device, dtype=weight_dtype)
if args.enable_xformers_memory_efficient_attention:
if is_xformers_available():
import xformers
xformers_version = version.parse(xformers.__version__)
if xformers_version == version.parse("0.0.16"):
logger.warning(
"xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, "
"please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
)
unet.enable_xformers_memory_efficient_attention()
else:
raise ValueError("xformers is not available. Make sure it is installed correctly")
if args.gradient_checkpointing:
unet.enable_gradient_checkpointing()
if args.train_text_encoder:
text_encoder_one.gradient_checkpointing_enable()
text_encoder_two.gradient_checkpointing_enable()
# now we will add new LoRA weights to the attention layers
unet_lora_config = LoraConfig(
r=args.rank,
use_dora=args.use_dora,
lora_alpha=args.rank,
init_lora_weights="gaussian",
target_modules=["to_k", "to_q", "to_v", "to_out.0"],
)
unet.add_adapter(unet_lora_config)
# The text encoder comes from 🤗 transformers, so we cannot directly modify it.
# So, instead, we monkey-patch the forward calls of its attention-blocks.
if args.train_text_encoder:
text_lora_config = LoraConfig(
r=args.rank,
use_dora=args.use_dora,
lora_alpha=args.rank,
init_lora_weights="gaussian",
target_modules=["q_proj", "k_proj", "v_proj", "out_proj"],
)
text_encoder_one.add_adapter(text_lora_config)
text_encoder_two.add_adapter(text_lora_config)
def unwrap_model(model):
model = accelerator.unwrap_model(model)
model = model._orig_mod if is_compiled_module(model) else model
return model
# create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
def save_model_hook(models, weights, output_dir):
if accelerator.is_main_process:
# there are only two options here. Either are just the unet attn processor layers
# or there are the unet and text encoder atten layers
unet_lora_layers_to_save = None
text_encoder_one_lora_layers_to_save = None
text_encoder_two_lora_layers_to_save = None
for model in models:
if isinstance(model, type(unwrap_model(unet))):
unet_lora_layers_to_save = convert_state_dict_to_diffusers(get_peft_model_state_dict(model))
elif isinstance(model, type(unwrap_model(text_encoder_one))):
text_encoder_one_lora_layers_to_save = convert_state_dict_to_diffusers(
get_peft_model_state_dict(model)
)
elif isinstance(model, type(unwrap_model(text_encoder_two))):
text_encoder_two_lora_layers_to_save = convert_state_dict_to_diffusers(
get_peft_model_state_dict(model)
)
else:
raise ValueError(f"unexpected save model: {model.__class__}")
# make sure to pop weight so that corresponding model is not saved again
weights.pop()
StableDiffusionXLPipeline.save_lora_weights(
output_dir,
unet_lora_layers=unet_lora_layers_to_save,
text_encoder_lora_layers=text_encoder_one_lora_layers_to_save,
text_encoder_2_lora_layers=text_encoder_two_lora_layers_to_save,
)
def load_model_hook(models, input_dir):
unet_ = None
text_encoder_one_ = None
text_encoder_two_ = None
while len(models) > 0:
model = models.pop()
if isinstance(model, type(unwrap_model(unet))):
unet_ = model
elif isinstance(model, type(unwrap_model(text_encoder_one))):
text_encoder_one_ = model
elif isinstance(model, type(unwrap_model(text_encoder_two))):
text_encoder_two_ = model
else:
raise ValueError(f"unexpected save model: {model.__class__}")
lora_state_dict, network_alphas = LoraLoaderMixin.lora_state_dict(input_dir)
unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
if incompatible_keys is not None:
# check only for unexpected keys
unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
if unexpected_keys:
logger.warning(
f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
f" {unexpected_keys}. "
)
if args.train_text_encoder:
# Do we need to call `scale_lora_layers()` here?
_set_state_dict_into_text_encoder(lora_state_dict, prefix="text_encoder.", text_encoder=text_encoder_one_)
_set_state_dict_into_text_encoder(
lora_state_dict, prefix="text_encoder_2.", text_encoder=text_encoder_two_
)
# Make sure the trainable params are in float32. This is again needed since the base models
# are in `weight_dtype`. More details:
# https://github.com/huggingface/diffusers/pull/6514#discussion_r1449796804
if args.mixed_precision == "fp16":
models = [unet_]
if args.train_text_encoder:
models.extend([text_encoder_one_, text_encoder_two_])
# only upcast trainable parameters (LoRA) into fp32
cast_training_params(models)
accelerator.register_save_state_pre_hook(save_model_hook)
accelerator.register_load_state_pre_hook(load_model_hook)
# Enable TF32 for faster training on Ampere GPUs,
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
if args.allow_tf32:
torch.backends.cuda.matmul.allow_tf32 = True
if args.scale_lr:
args.learning_rate = (
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
)
# Make sure the trainable params are in float32.
if args.mixed_precision == "fp16":
models = [unet]
if args.train_text_encoder:
models.extend([text_encoder_one, text_encoder_two])
# only upcast trainable parameters (LoRA) into fp32
cast_training_params(models, dtype=torch.float32)
unet_lora_parameters = list(filter(lambda p: p.requires_grad, unet.parameters()))
if args.train_text_encoder:
text_lora_parameters_one = list(filter(lambda p: p.requires_grad, text_encoder_one.parameters()))
text_lora_parameters_two = list(filter(lambda p: p.requires_grad, text_encoder_two.parameters()))
# Optimization parameters
unet_lora_parameters_with_lr = {"params": unet_lora_parameters, "lr": args.learning_rate}
if args.train_text_encoder:
# different learning rate for text encoder and unet
text_lora_parameters_one_with_lr = {
"params": text_lora_parameters_one,
"weight_decay": args.adam_weight_decay_text_encoder,
"lr": args.text_encoder_lr if args.text_encoder_lr else args.learning_rate,
}
text_lora_parameters_two_with_lr = {
"params": text_lora_parameters_two,
"weight_decay": args.adam_weight_decay_text_encoder,
"lr": args.text_encoder_lr if args.text_encoder_lr else args.learning_rate,
}
params_to_optimize = [
unet_lora_parameters_with_lr,
text_lora_parameters_one_with_lr,
text_lora_parameters_two_with_lr,
]
else:
params_to_optimize = [unet_lora_parameters_with_lr]
# Optimizer creation
if not (args.optimizer.lower() == "prodigy" or args.optimizer.lower() == "adamw"):
logger.warning(
f"Unsupported choice of optimizer: {args.optimizer}.Supported optimizers include [adamW, prodigy]."
"Defaulting to adamW"
)
args.optimizer = "adamw"
if args.use_8bit_adam and not args.optimizer.lower() == "adamw":
logger.warning(
f"use_8bit_adam is ignored when optimizer is not set to 'AdamW'. Optimizer was "
f"set to {args.optimizer.lower()}"
)
if args.optimizer.lower() == "adamw":
if args.use_8bit_adam:
try:
import bitsandbytes as bnb
except ImportError:
raise ImportError(
"To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
)
optimizer_class = bnb.optim.AdamW8bit
else:
optimizer_class = torch.optim.AdamW
optimizer = optimizer_class(
params_to_optimize,
betas=(args.adam_beta1, args.adam_beta2),
weight_decay=args.adam_weight_decay,
eps=args.adam_epsilon,
)
if args.optimizer.lower() == "prodigy":
try:
import prodigyopt
except ImportError:
raise ImportError("To use Prodigy, please install the prodigyopt library: `pip install prodigyopt`")
optimizer_class = prodigyopt.Prodigy
if args.learning_rate <= 0.1:
logger.warning(
"Learning rate is too low. When using prodigy, it's generally better to set learning rate around 1.0"
)
if args.train_text_encoder and args.text_encoder_lr:
logger.warning(
f"Learning rates were provided both for the unet and the text encoder- e.g. text_encoder_lr:"
f" {args.text_encoder_lr} and learning_rate: {args.learning_rate}. "
f"When using prodigy only learning_rate is used as the initial learning rate."
)
# changes the learning rate of text_encoder_parameters_one and text_encoder_parameters_two to be
# --learning_rate
params_to_optimize[1]["lr"] = args.learning_rate
params_to_optimize[2]["lr"] = args.learning_rate
optimizer = optimizer_class(
params_to_optimize,
lr=args.learning_rate,
betas=(args.adam_beta1, args.adam_beta2),
beta3=args.prodigy_beta3,
weight_decay=args.adam_weight_decay,
eps=args.adam_epsilon,
decouple=args.prodigy_decouple,
use_bias_correction=args.prodigy_use_bias_correction,
safeguard_warmup=args.prodigy_safeguard_warmup,
)
# Dataset and DataLoaders creation:
train_dataset = DreamBoothDataset(
instance_data_root=args.instance_data_dir,
instance_prompt=args.instance_prompt,
class_prompt=args.class_prompt,
class_data_root=args.class_data_dir if args.with_prior_preservation else None,
class_num=args.num_class_images,
size=args.resolution,
repeats=args.repeats,
center_crop=args.center_crop,
)
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.train_batch_size,
shuffle=True,
collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation),
num_workers=args.dataloader_num_workers,
)
# Computes additional embeddings/ids required by the SDXL UNet.
# regular text embeddings (when `train_text_encoder` is not True)
# pooled text embeddings
# time ids
def compute_time_ids(original_size, crops_coords_top_left):
# Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
target_size = (args.resolution, args.resolution)
add_time_ids = list(original_size + crops_coords_top_left + target_size)
add_time_ids = torch.tensor([add_time_ids])
add_time_ids = add_time_ids.to(accelerator.device, dtype=weight_dtype)
return add_time_ids
if not args.train_text_encoder:
tokenizers = [tokenizer_one, tokenizer_two]
text_encoders = [text_encoder_one, text_encoder_two]
def compute_text_embeddings(prompt, text_encoders, tokenizers):
with torch.no_grad():
prompt_embeds, pooled_prompt_embeds = encode_prompt(text_encoders, tokenizers, prompt)
prompt_embeds = prompt_embeds.to(accelerator.device)
pooled_prompt_embeds = pooled_prompt_embeds.to(accelerator.device)
return prompt_embeds, pooled_prompt_embeds
# If no type of tuning is done on the text_encoder and custom instance prompts are NOT
# provided (i.e. the --instance_prompt is used for all images), we encode the instance prompt once to avoid
# the redundant encoding.
if not args.train_text_encoder and not train_dataset.custom_instance_prompts:
instance_prompt_hidden_states, instance_pooled_prompt_embeds = compute_text_embeddings(
args.instance_prompt, text_encoders, tokenizers
)
# Handle class prompt for prior-preservation.
if args.with_prior_preservation:
if not args.train_text_encoder:
class_prompt_hidden_states, class_pooled_prompt_embeds = compute_text_embeddings(
args.class_prompt, text_encoders, tokenizers
)
# Clear the memory here
if not args.train_text_encoder and not train_dataset.custom_instance_prompts:
del tokenizers, text_encoders
gc.collect()
torch.cuda.empty_cache()
# If custom instance prompts are NOT provided (i.e. the instance prompt is used for all images),
# pack the statically computed variables appropriately here. This is so that we don't
# have to pass them to the dataloader.
if not train_dataset.custom_instance_prompts:
if not args.train_text_encoder:
prompt_embeds = instance_prompt_hidden_states
unet_add_text_embeds = instance_pooled_prompt_embeds
if args.with_prior_preservation:
prompt_embeds = torch.cat([prompt_embeds, class_prompt_hidden_states], dim=0)
unet_add_text_embeds = torch.cat([unet_add_text_embeds, class_pooled_prompt_embeds], dim=0)
# if we're optmizing the text encoder (both if instance prompt is used for all images or custom prompts) we need to tokenize and encode the
# batch prompts on all training steps
else:
tokens_one = tokenize_prompt(tokenizer_one, args.instance_prompt)
tokens_two = tokenize_prompt(tokenizer_two, args.instance_prompt)
if args.with_prior_preservation:
class_tokens_one = tokenize_prompt(tokenizer_one, args.class_prompt)
class_tokens_two = tokenize_prompt(tokenizer_two, args.class_prompt)
tokens_one = torch.cat([tokens_one, class_tokens_one], dim=0)
tokens_two = torch.cat([tokens_two, class_tokens_two], dim=0)
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
overrode_max_train_steps = True
lr_scheduler = get_scheduler(
args.lr_scheduler,
optimizer=optimizer,
num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
num_training_steps=args.max_train_steps * accelerator.num_processes,
num_cycles=args.lr_num_cycles,
power=args.lr_power,
)
# Prepare everything with our `accelerator`.
if args.train_text_encoder:
unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler
)
else:
unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
unet, optimizer, train_dataloader, lr_scheduler
)
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# We need to initialize the trackers we use, and also store our configuration.
# The trackers initializes automatically on the main process.
if accelerator.is_main_process:
tracker_name = (
"dreambooth-lora-sd-xl"
if "playground" not in args.pretrained_model_name_or_path
else "dreambooth-lora-playground"
)
accelerator.init_trackers(tracker_name, config=vars(args))
# Train!
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num batches each epoch = {len(train_dataloader)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.train_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
global_step = 0
first_epoch = 0
# Potentially load in the weights and states from a previous save
if args.resume_from_checkpoint:
if args.resume_from_checkpoint != "latest":
path = os.path.basename(args.resume_from_checkpoint)
else:
# Get the mos recent checkpoint
dirs = os.listdir(args.output_dir)
dirs = [d for d in dirs if d.startswith("checkpoint")]
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
path = dirs[-1] if len(dirs) > 0 else None
if path is None:
accelerator.print(
f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
)
args.resume_from_checkpoint = None
initial_global_step = 0
else:
accelerator.print(f"Resuming from checkpoint {path}")
accelerator.load_state(os.path.join(args.output_dir, path))
global_step = int(path.split("-")[1])
initial_global_step = global_step
first_epoch = global_step // num_update_steps_per_epoch
else:
initial_global_step = 0
progress_bar = tqdm(
range(0, args.max_train_steps),
initial=initial_global_step,
desc="Steps",
# Only show the progress bar once on each machine.
disable=not accelerator.is_local_main_process,
)
def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
sigmas = noise_scheduler.sigmas.to(device=accelerator.device, dtype=dtype)
schedule_timesteps = noise_scheduler.timesteps.to(accelerator.device)
timesteps = timesteps.to(accelerator.device)
step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
sigma = sigmas[step_indices].flatten()
while len(sigma.shape) < n_dim:
sigma = sigma.unsqueeze(-1)
return sigma
for epoch in range(first_epoch, args.num_train_epochs):
unet.train()
if args.train_text_encoder:
text_encoder_one.train()
text_encoder_two.train()
# set top parameter requires_grad = True for gradient checkpointing works
accelerator.unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
accelerator.unwrap_model(text_encoder_two).text_model.embeddings.requires_grad_(True)
for step, batch in enumerate(train_dataloader):
with accelerator.accumulate(unet):
pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
prompts = batch["prompts"]
# encode batch prompts when custom prompts are provided for each image -
if train_dataset.custom_instance_prompts:
if not args.train_text_encoder:
prompt_embeds, unet_add_text_embeds = compute_text_embeddings(
prompts, text_encoders, tokenizers
)
else:
tokens_one = tokenize_prompt(tokenizer_one, prompts)
tokens_two = tokenize_prompt(tokenizer_two, prompts)
# Convert images to latent space
model_input = vae.encode(pixel_values).latent_dist.sample()
if latents_mean is None and latents_std is None:
model_input = model_input * vae.config.scaling_factor
if args.pretrained_vae_model_name_or_path is None:
model_input = model_input.to(weight_dtype)
else:
latents_mean = latents_mean.to(device=model_input.device, dtype=model_input.dtype)
latents_std = latents_std.to(device=model_input.device, dtype=model_input.dtype)
model_input = (model_input - latents_mean) * vae.config.scaling_factor / latents_std
model_input = model_input.to(dtype=weight_dtype)
# Sample noise that we'll add to the latents
noise = torch.randn_like(model_input)
bsz = model_input.shape[0]
# Sample a random timestep for each image
if not args.do_edm_style_training:
if args.loss_type == "huber" or args.loss_type == "smooth_l1":
timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (1,), device="cpu")
timestep = timesteps.item()
if args.huber_schedule == "exponential":
alpha = -math.log(args.huber_c) / noise_scheduler.config.num_train_timesteps
huber_c = math.exp(-alpha * timestep)
elif args.huber_schedule == "snr":
alphas_cumprod = noise_scheduler.alphas_cumprod[timestep]
sigmas = ((1.0 - alphas_cumprod) / alphas_cumprod) ** 0.5
huber_c = (1 - args.huber_c) / (1 + sigmas) ** 2 + args.huber_c
elif args.huber_schedule == "constant":
huber_c = args.huber_c
else:
raise NotImplementedError(f"Unknown Huber loss schedule {args.huber_schedule}!")
timesteps = timesteps.repeat(bsz).to(model_input.device)
elif args.loss_type == "l2":
timesteps = torch.randint(
0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
)
huber_c = 1 # may be anything, as it's not used
else:
raise NotImplementedError(f"Unknown loss type {args.loss_type}")
timesteps = timesteps.long()
else:
if "huber" in args.loss_type or "l1" in args.loss_type:
raise NotImplementedError("Huber loss is not implemented for EDM training yet!")
# in EDM formulation, the model is conditioned on the pre-conditioned noise levels
# instead of discrete timesteps, so here we sample indices to get the noise levels
# from `scheduler.timesteps`
indices = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,))
timesteps = noise_scheduler.timesteps[indices].to(device=model_input.device)
# Add noise to the model input according to the noise magnitude at each timestep
# (this is the forward diffusion process)
noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
# For EDM-style training, we first obtain the sigmas based on the continuous timesteps.
# We then precondition the final model inputs based on these sigmas instead of the timesteps.
# Follow: Section 5 of https://arxiv.org/abs/2206.00364.
if args.do_edm_style_training:
sigmas = get_sigmas(timesteps, len(noisy_model_input.shape), noisy_model_input.dtype)
if "EDM" in scheduler_type:
inp_noisy_latents = noise_scheduler.precondition_inputs(noisy_model_input, sigmas)
else:
inp_noisy_latents = noisy_model_input / ((sigmas**2 + 1) ** 0.5)
# time ids
add_time_ids = torch.cat(
[
compute_time_ids(original_size=s, crops_coords_top_left=c)
for s, c in zip(batch["original_sizes"], batch["crop_top_lefts"])
]
)
# Calculate the elements to repeat depending on the use of prior-preservation and custom captions.
if not train_dataset.custom_instance_prompts:
elems_to_repeat_text_embeds = bsz // 2 if args.with_prior_preservation else bsz
else:
elems_to_repeat_text_embeds = 1
# Predict the noise residual
if not args.train_text_encoder:
unet_added_conditions = {
"time_ids": add_time_ids,
"text_embeds": unet_add_text_embeds.repeat(elems_to_repeat_text_embeds, 1),
}
prompt_embeds_input = prompt_embeds.repeat(elems_to_repeat_text_embeds, 1, 1)
model_pred = unet(
inp_noisy_latents if args.do_edm_style_training else noisy_model_input,
timesteps,
prompt_embeds_input,
added_cond_kwargs=unet_added_conditions,
return_dict=False,
)[0]
else:
unet_added_conditions = {"time_ids": add_time_ids}
prompt_embeds, pooled_prompt_embeds = encode_prompt(
text_encoders=[text_encoder_one, text_encoder_two],
tokenizers=None,
prompt=None,
text_input_ids_list=[tokens_one, tokens_two],
)
unet_added_conditions.update(
{"text_embeds": pooled_prompt_embeds.repeat(elems_to_repeat_text_embeds, 1)}
)
prompt_embeds_input = prompt_embeds.repeat(elems_to_repeat_text_embeds, 1, 1)
model_pred = unet(
inp_noisy_latents if args.do_edm_style_training else noisy_model_input,
timesteps,
prompt_embeds_input,
added_cond_kwargs=unet_added_conditions,
return_dict=False,
)[0]
weighting = None
if args.do_edm_style_training:
# Similar to the input preconditioning, the model predictions are also preconditioned
# on noised model inputs (before preconditioning) and the sigmas.
# Follow: Section 5 of https://arxiv.org/abs/2206.00364.
if "EDM" in scheduler_type:
model_pred = noise_scheduler.precondition_outputs(noisy_model_input, model_pred, sigmas)
else:
if noise_scheduler.config.prediction_type == "epsilon":
model_pred = model_pred * (-sigmas) + noisy_model_input
elif noise_scheduler.config.prediction_type == "v_prediction":
model_pred = model_pred * (-sigmas / (sigmas**2 + 1) ** 0.5) + (
noisy_model_input / (sigmas**2 + 1)
)
# We are not doing weighting here because it tends result in numerical problems.
# See: https://github.com/huggingface/diffusers/pull/7126#issuecomment-1968523051
# There might be other alternatives for weighting as well:
# https://github.com/huggingface/diffusers/pull/7126#discussion_r1505404686
if "EDM" not in scheduler_type:
weighting = (sigmas**-2.0).float()
# Get the target for loss depending on the prediction type
if noise_scheduler.config.prediction_type == "epsilon":
target = model_input if args.do_edm_style_training else noise
elif noise_scheduler.config.prediction_type == "v_prediction":
target = (
model_input
if args.do_edm_style_training
else noise_scheduler.get_velocity(model_input, noise, timesteps)
)
else:
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
if args.with_prior_preservation:
# Chunk the noise and model_pred into two parts and compute the loss on each part separately.
model_pred, model_pred_prior = torch.chunk(model_pred, 2, dim=0)
target, target_prior = torch.chunk(target, 2, dim=0)
# Compute prior loss
prior_loss = conditional_loss(
model_pred_prior,
target_prior,
reduction="mean",
loss_type=args.loss_type,
huber_c=huber_c,
weighting=weighting,
)
if args.snr_gamma is None:
loss = conditional_loss(
model_pred,
target,
reduction="mean",
loss_type=args.loss_type,
huber_c=huber_c,
weighting=weighting,
)
else:
# Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
# Since we predict the noise instead of x_0, the original formulation is slightly changed.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(noise_scheduler, timesteps)
base_weight = (
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
)
if noise_scheduler.config.prediction_type == "v_prediction":
# Velocity objective needs to be floored to an SNR weight of one.
mse_loss_weights = base_weight + 1
else:
# Epsilon and sample both use the same loss weights.
mse_loss_weights = base_weight
loss = conditional_loss(
model_pred, target, reduction="none", loss_type=args.loss_type, huber_c=huber_c, weighting=None
)
loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
loss = loss.mean()
if args.with_prior_preservation:
# Add the prior loss to the instance loss.
loss = loss + args.prior_loss_weight * prior_loss
accelerator.backward(loss)
if accelerator.sync_gradients:
params_to_clip = (
itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two)
if args.train_text_encoder
else unet_lora_parameters
)
accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
# Checks if the accelerator has performed an optimization step behind the scenes
if accelerator.sync_gradients:
progress_bar.update(1)
global_step += 1
if accelerator.is_main_process:
if global_step % args.checkpointing_steps == 0:
# _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
if args.checkpoints_total_limit is not None:
checkpoints = os.listdir(args.output_dir)
checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
# before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
if len(checkpoints) >= args.checkpoints_total_limit:
num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
removing_checkpoints = checkpoints[0:num_to_remove]
logger.info(
f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
)
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
for removing_checkpoint in removing_checkpoints:
removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
shutil.rmtree(removing_checkpoint)
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
accelerator.save_state(save_path)
logger.info(f"Saved state to {save_path}")
logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
progress_bar.set_postfix(**logs)
accelerator.log(logs, step=global_step)
if global_step >= args.max_train_steps:
break
if accelerator.is_main_process:
if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
# create pipeline
if not args.train_text_encoder:
text_encoder_one = text_encoder_cls_one.from_pretrained(
args.pretrained_model_name_or_path,
subfolder="text_encoder",
revision=args.revision,
variant=args.variant,
)
text_encoder_two = text_encoder_cls_two.from_pretrained(
args.pretrained_model_name_or_path,
subfolder="text_encoder_2",
revision=args.revision,
variant=args.variant,
)
pipeline = StableDiffusionXLPipeline.from_pretrained(
args.pretrained_model_name_or_path,
vae=vae,
text_encoder=accelerator.unwrap_model(text_encoder_one),
text_encoder_2=accelerator.unwrap_model(text_encoder_two),
unet=accelerator.unwrap_model(unet),
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
)
pipeline_args = {"prompt": args.validation_prompt}
images = log_validation(
pipeline,
args,
accelerator,
pipeline_args,
epoch,
)
# Save the lora layers
accelerator.wait_for_everyone()
if accelerator.is_main_process:
unet = unwrap_model(unet)
unet = unet.to(torch.float32)
unet_lora_layers = convert_state_dict_to_diffusers(get_peft_model_state_dict(unet))
if args.train_text_encoder:
text_encoder_one = unwrap_model(text_encoder_one)
text_encoder_lora_layers = convert_state_dict_to_diffusers(
get_peft_model_state_dict(text_encoder_one.to(torch.float32))
)
text_encoder_two = unwrap_model(text_encoder_two)
text_encoder_2_lora_layers = convert_state_dict_to_diffusers(
get_peft_model_state_dict(text_encoder_two.to(torch.float32))
)
else:
text_encoder_lora_layers = None
text_encoder_2_lora_layers = None
StableDiffusionXLPipeline.save_lora_weights(
save_directory=args.output_dir,
unet_lora_layers=unet_lora_layers,
text_encoder_lora_layers=text_encoder_lora_layers,
text_encoder_2_lora_layers=text_encoder_2_lora_layers,
)
if args.output_kohya_format:
lora_state_dict = load_file(f"{args.output_dir}/pytorch_lora_weights.safetensors")
peft_state_dict = convert_all_state_dict_to_peft(lora_state_dict)
kohya_state_dict = convert_state_dict_to_kohya(peft_state_dict)
save_file(kohya_state_dict, f"{args.output_dir}/pytorch_lora_weights_kohya.safetensors")
# Final inference
# Load previous pipeline
vae = AutoencoderKL.from_pretrained(
vae_path,
subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
)
pipeline = StableDiffusionXLPipeline.from_pretrained(
args.pretrained_model_name_or_path,
vae=vae,
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
)
# load attention processors
pipeline.load_lora_weights(args.output_dir)
# run inference
images = []
if args.validation_prompt and args.num_validation_images > 0:
pipeline_args = {"prompt": args.validation_prompt, "num_inference_steps": 25}
images = log_validation(
pipeline,
args,
accelerator,
pipeline_args,
epoch,
is_final_validation=True,
)
if args.push_to_hub:
save_model_card(
repo_id,
use_dora=args.use_dora,
images=images,
base_model=args.pretrained_model_name_or_path,
train_text_encoder=args.train_text_encoder,
instance_prompt=args.instance_prompt,
validation_prompt=args.validation_prompt,
repo_folder=args.output_dir,
vae_path=args.pretrained_vae_model_name_or_path,
)
upload_folder(
repo_id=repo_id,
folder_path=args.output_dir,
commit_message="End of training",
ignore_patterns=["step_*", "epoch_*"],
)
accelerator.end_training()
if __name__ == "__main__":
args = parse_args()
main(args)
examples/research_projects/scheduled_huber_loss_training/text_to_image/train_text_to_image.py 0 → 100644
View file @ 2f234376
#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import logging
import math
import os
import random
import shutil
from pathlib import Path
import accelerate
import datasets
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.state import AcceleratorState
from accelerate.utils import ProjectConfiguration, set_seed
from datasets import load_dataset
from huggingface_hub import create_repo, upload_folder
from packaging import version
from torchvision import transforms
from tqdm.auto import tqdm
from transformers import CLIPTextModel, CLIPTokenizer
from transformers.utils import ContextManagers
import diffusers
from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionPipeline, UNet2DConditionModel
from diffusers.optimization import get_scheduler
from diffusers.training_utils import EMAModel, compute_snr
from diffusers.utils import check_min_version, deprecate, is_wandb_available, make_image_grid
from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.torch_utils import is_compiled_module
if is_wandb_available():
import wandb
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.28.0.dev0")
logger = get_logger(__name__, log_level="INFO")
DATASET_NAME_MAPPING = {
"lambdalabs/pokemon-blip-captions": ("image", "text"),
}
def save_model_card(
args,
repo_id: str,
images: list = None,
repo_folder: str = None,
):
img_str = ""
if len(images) > 0:
image_grid = make_image_grid(images, 1, len(args.validation_prompts))
image_grid.save(os.path.join(repo_folder, "val_imgs_grid.png"))
img_str += "![val_imgs_grid](./val_imgs_grid.png)\n"
model_description = f"""
# Text-to-image finetuning - {repo_id}
This pipeline was finetuned from **{args.pretrained_model_name_or_path}** on the **{args.dataset_name}** dataset. Below are some example images generated with the finetuned pipeline using the following prompts: {args.validation_prompts}: \n
{img_str}
## Pipeline usage
You can use the pipeline like so:
```python
from diffusers import DiffusionPipeline
import torch
pipeline = DiffusionPipeline.from_pretrained("{repo_id}", torch_dtype=torch.float16)
prompt = "{args.validation_prompts[0]}"
image = pipeline(prompt).images[0]
image.save("my_image.png")
```
## Training info
These are the key hyperparameters used during training:
* Epochs: {args.num_train_epochs}
* Learning rate: {args.learning_rate}
* Batch size: {args.train_batch_size}
* Gradient accumulation steps: {args.gradient_accumulation_steps}
* Image resolution: {args.resolution}
* Mixed-precision: {args.mixed_precision}
"""
wandb_info = ""
if is_wandb_available():
wandb_run_url = None
if wandb.run is not None:
wandb_run_url = wandb.run.url
if wandb_run_url is not None:
wandb_info = f"""
More information on all the CLI arguments and the environment are available on your [`wandb` run page]({wandb_run_url}).
"""
model_description += wandb_info
model_card = load_or_create_model_card(
repo_id_or_path=repo_id,
from_training=True,
license="creativeml-openrail-m",
base_model=args.pretrained_model_name_or_path,
model_description=model_description,
inference=True,
)
tags = ["stable-diffusion", "stable-diffusion-diffusers", "text-to-image", "diffusers", "diffusers-training"]
model_card = populate_model_card(model_card, tags=tags)
model_card.save(os.path.join(repo_folder, "README.md"))
def log_validation(vae, text_encoder, tokenizer, unet, args, accelerator, weight_dtype, epoch):
logger.info("Running validation... ")
pipeline = StableDiffusionPipeline.from_pretrained(
args.pretrained_model_name_or_path,
vae=accelerator.unwrap_model(vae),
text_encoder=accelerator.unwrap_model(text_encoder),
tokenizer=tokenizer,
unet=accelerator.unwrap_model(unet),
safety_checker=None,
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
)
pipeline = pipeline.to(accelerator.device)
pipeline.set_progress_bar_config(disable=True)
if args.enable_xformers_memory_efficient_attention:
pipeline.enable_xformers_memory_efficient_attention()
if args.seed is None:
generator = None
else:
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)
images = []
for i in range(len(args.validation_prompts)):
with torch.autocast("cuda"):
image = pipeline(args.validation_prompts[i], num_inference_steps=20, generator=generator).images[0]
images.append(image)
for tracker in accelerator.trackers:
if tracker.name == "tensorboard":
np_images = np.stack([np.asarray(img) for img in images])
tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
elif tracker.name == "wandb":
tracker.log(
{
"validation": [
wandb.Image(image, caption=f"{i}: {args.validation_prompts[i]}")
for i, image in enumerate(images)
]
}
)
else:
logger.warning(f"image logging not implemented for {tracker.name}")
del pipeline
torch.cuda.empty_cache()
return images
def parse_args():
parser = argparse.ArgumentParser(description="Simple example of a training script.")
parser.add_argument(
"--input_perturbation", type=float, default=0, help="The scale of input perturbation. Recommended 0.1."
)
parser.add_argument(
"--pretrained_model_name_or_path",
type=str,
default=None,
required=True,
help="Path to pretrained model or model identifier from huggingface.co/models.",
)
parser.add_argument(
"--revision",
type=str,
default=None,
required=False,
help="Revision of pretrained model identifier from huggingface.co/models.",
)
parser.add_argument(
"--variant",
type=str,
default=None,
help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
)
parser.add_argument(
"--dataset_name",
type=str,
default=None,
help=(
"The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
" dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
" or to a folder containing files that 🤗 Datasets can understand."
),
)
parser.add_argument(
"--dataset_config_name",
type=str,
default=None,
help="The config of the Dataset, leave as None if there's only one config.",
)
parser.add_argument(
"--train_data_dir",
type=str,
default=None,
help=(
"A folder containing the training data. Folder contents must follow the structure described in"
" https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
" must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
),
)
parser.add_argument(
"--image_column", type=str, default="image", help="The column of the dataset containing an image."
)
parser.add_argument(
"--caption_column",
type=str,
default="text",
help="The column of the dataset containing a caption or a list of captions.",
)
parser.add_argument(
"--max_train_samples",
type=int,
default=None,
help=(
"For debugging purposes or quicker training, truncate the number of training examples to this "
"value if set."
),
)
parser.add_argument(
"--validation_prompts",
type=str,
default=None,
nargs="+",
help=("A set of prompts evaluated every `--validation_epochs` and logged to `--report_to`."),
)
parser.add_argument(
"--output_dir",
type=str,
default="sd-model-finetuned",
help="The output directory where the model predictions and checkpoints will be written.",
)
parser.add_argument(
"--cache_dir",
type=str,
default=None,
help="The directory where the downloaded models and datasets will be stored.",
)
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
parser.add_argument(
"--resolution",
type=int,
default=512,
help=(
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
" resolution"
),
)
parser.add_argument(
"--center_crop",
default=False,
action="store_true",
help=(
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
" cropped. The images will be resized to the resolution first before cropping."
),
)
parser.add_argument(
"--random_flip",
action="store_true",
help="whether to randomly flip images horizontally",
)
parser.add_argument(
"--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
)
parser.add_argument("--num_train_epochs", type=int, default=100)
parser.add_argument(
"--max_train_steps",
type=int,
default=None,
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--gradient_checkpointing",
action="store_true",
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
)
parser.add_argument(
"--learning_rate",
type=float,
default=1e-4,
help="Initial learning rate (after the potential warmup period) to use.",
)
parser.add_argument(
"--scale_lr",
action="store_true",
default=False,
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
)
parser.add_argument(
"--lr_scheduler",
type=str,
default="constant",
help=(
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
' "constant", "constant_with_warmup"]'
),
)
parser.add_argument(
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
)
parser.add_argument(
"--snr_gamma",
type=float,
default=None,
help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
"More details here: https://arxiv.org/abs/2303.09556.",
)
parser.add_argument(
"--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
)
parser.add_argument(
"--allow_tf32",
action="store_true",
help=(
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
),
)
parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.")
parser.add_argument(
"--non_ema_revision",
type=str,
default=None,
required=False,
help=(
"Revision of pretrained non-ema model identifier. Must be a branch, tag or git identifier of the local or"
" remote repository specified with --pretrained_model_name_or_path."
),
)
parser.add_argument(
"--dataloader_num_workers",
type=int,
default=0,
help=(
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
),
)
parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
parser.add_argument(
"--prediction_type",
type=str,
default=None,
help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediction_type` is chosen.",
)
parser.add_argument(
"--hub_model_id",
type=str,
default=None,
help="The name of the repository to keep in sync with the local `output_dir`.",
)
parser.add_argument(
"--logging_dir",
type=str,
default="logs",
help=(
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
),
)
parser.add_argument(
"--mixed_precision",
type=str,
default=None,
choices=["no", "fp16", "bf16"],
help=(
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
),
)
parser.add_argument(
"--report_to",
type=str,
default="tensorboard",
help=(
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
),
)
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
parser.add_argument(
"--checkpointing_steps",
type=int,
default=500,
help=(
"Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming"
" training using `--resume_from_checkpoint`."
),
)
parser.add_argument(
"--checkpoints_total_limit",
type=int,
default=None,
help=("Max number of checkpoints to store."),
)
parser.add_argument(
"--resume_from_checkpoint",
type=str,
default=None,
help=(
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
),
)
parser.add_argument(
"--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
)
parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
parser.add_argument(
"--loss_type",
type=str,
default="l2",
choices=["l2", "huber", "smooth_l1"],
help="The type of loss to use and whether it's timestep-scheduled. See Issue #7488 for more info.",
)
parser.add_argument(
"--huber_schedule",
type=str,
default="snr",
choices=["constant", "exponential", "snr"],
help="The schedule to use for the huber losses parameter",
)
parser.add_argument(
"--huber_c",
type=float,
default=0.1,
help="The huber loss parameter. Only used if one of the huber loss modes (huber or smooth l1) is selected with loss_type.",
)
parser.add_argument(
"--validation_epochs",
type=int,
default=5,
help="Run validation every X epochs.",
)
parser.add_argument(
"--tracker_project_name",
type=str,
default="text2image-fine-tune",
help=(
"The `project_name` argument passed to Accelerator.init_trackers for"
" more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
),
)
args = parser.parse_args()
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
if env_local_rank != -1 and env_local_rank != args.local_rank:
args.local_rank = env_local_rank
# Sanity checks
if args.dataset_name is None and args.train_data_dir is None:
raise ValueError("Need either a dataset name or a training folder.")
# default to using the same revision for the non-ema model if not specified
if args.non_ema_revision is None:
args.non_ema_revision = args.revision
return args
# NOTE: if you're using the scheduled version, huber_c has to depend on the timesteps already
def conditional_loss(
model_pred: torch.Tensor,
target: torch.Tensor,
reduction: str = "mean",
loss_type: str = "l2",
huber_c: float = 0.1,
):
if loss_type == "l2":
loss = F.mse_loss(model_pred, target, reduction=reduction)
elif loss_type == "huber":
loss = 2 * huber_c * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
elif loss_type == "smooth_l1":
loss = 2 * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
else:
raise NotImplementedError(f"Unsupported Loss Type {loss_type}")
return loss
def main():
args = parse_args()
if args.report_to == "wandb" and args.hub_token is not None:
raise ValueError(
"You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
" Please use `huggingface-cli login` to authenticate with the Hub."
)
if args.non_ema_revision is not None:
deprecate(
"non_ema_revision!=None",
"0.15.0",
message=(
"Downloading 'non_ema' weights from revision branches of the Hub is deprecated. Please make sure to"
" use `--variant=non_ema` instead."
),
)
logging_dir = os.path.join(args.output_dir, args.logging_dir)
accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
accelerator = Accelerator(
gradient_accumulation_steps=args.gradient_accumulation_steps,
mixed_precision=args.mixed_precision,
log_with=args.report_to,
project_config=accelerator_project_config,
)
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state, main_process_only=False)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_warning()
diffusers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
diffusers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
if args.push_to_hub:
repo_id = create_repo(
repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
).repo_id
# Load scheduler, tokenizer and models.
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
tokenizer = CLIPTokenizer.from_pretrained(
args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision
)
def deepspeed_zero_init_disabled_context_manager():
"""
returns either a context list that includes one that will disable zero.Init or an empty context list
"""
deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None
if deepspeed_plugin is None:
return []
return [deepspeed_plugin.zero3_init_context_manager(enable=False)]
# Currently Accelerate doesn't know how to handle multiple models under Deepspeed ZeRO stage 3.
# For this to work properly all models must be run through `accelerate.prepare`. But accelerate
# will try to assign the same optimizer with the same weights to all models during
# `deepspeed.initialize`, which of course doesn't work.
#
# For now the following workaround will partially support Deepspeed ZeRO-3, by excluding the 2
# frozen models from being partitioned during `zero.Init` which gets called during
# `from_pretrained` So CLIPTextModel and AutoencoderKL will not enjoy the parameter sharding
# across multiple gpus and only UNet2DConditionModel will get ZeRO sharded.
with ContextManagers(deepspeed_zero_init_disabled_context_manager()):
text_encoder = CLIPTextModel.from_pretrained(
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
)
vae = AutoencoderKL.from_pretrained(
args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant
)
unet = UNet2DConditionModel.from_pretrained(
args.pretrained_model_name_or_path, subfolder="unet", revision=args.non_ema_revision
)
# Freeze vae and text_encoder and set unet to trainable
vae.requires_grad_(False)
text_encoder.requires_grad_(False)
unet.train()
# Create EMA for the unet.
if args.use_ema:
ema_unet = UNet2DConditionModel.from_pretrained(
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
)
ema_unet = EMAModel(ema_unet.parameters(), model_cls=UNet2DConditionModel, model_config=ema_unet.config)
if args.enable_xformers_memory_efficient_attention:
if is_xformers_available():
import xformers
xformers_version = version.parse(xformers.__version__)
if xformers_version == version.parse("0.0.16"):
logger.warning(
"xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
)
unet.enable_xformers_memory_efficient_attention()
else:
raise ValueError("xformers is not available. Make sure it is installed correctly")
# `accelerate` 0.16.0 will have better support for customized saving
if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
# create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
def save_model_hook(models, weights, output_dir):
if accelerator.is_main_process:
if args.use_ema:
ema_unet.save_pretrained(os.path.join(output_dir, "unet_ema"))
for i, model in enumerate(models):
model.save_pretrained(os.path.join(output_dir, "unet"))
# make sure to pop weight so that corresponding model is not saved again
weights.pop()
def load_model_hook(models, input_dir):
if args.use_ema:
load_model = EMAModel.from_pretrained(os.path.join(input_dir, "unet_ema"), UNet2DConditionModel)
ema_unet.load_state_dict(load_model.state_dict())
ema_unet.to(accelerator.device)
del load_model
for _ in range(len(models)):
# pop models so that they are not loaded again
model = models.pop()
# load diffusers style into model
load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
model.register_to_config(**load_model.config)
model.load_state_dict(load_model.state_dict())
del load_model
accelerator.register_save_state_pre_hook(save_model_hook)
accelerator.register_load_state_pre_hook(load_model_hook)
if args.gradient_checkpointing:
unet.enable_gradient_checkpointing()
# Enable TF32 for faster training on Ampere GPUs,
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
if args.allow_tf32:
torch.backends.cuda.matmul.allow_tf32 = True
if args.scale_lr:
args.learning_rate = (
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
)
# Initialize the optimizer
if args.use_8bit_adam:
try:
import bitsandbytes as bnb
except ImportError:
raise ImportError(
"Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
)
optimizer_cls = bnb.optim.AdamW8bit
else:
optimizer_cls = torch.optim.AdamW
optimizer = optimizer_cls(
unet.parameters(),
lr=args.learning_rate,
betas=(args.adam_beta1, args.adam_beta2),
weight_decay=args.adam_weight_decay,
eps=args.adam_epsilon,
)
# Get the datasets: you can either provide your own training and evaluation files (see below)
# or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).
# In distributed training, the load_dataset function guarantees that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
dataset = load_dataset(
args.dataset_name,
args.dataset_config_name,
cache_dir=args.cache_dir,
data_dir=args.train_data_dir,
)
else:
data_files = {}
if args.train_data_dir is not None:
data_files["train"] = os.path.join(args.train_data_dir, "**")
dataset = load_dataset(
"imagefolder",
data_files=data_files,
cache_dir=args.cache_dir,
)
# See more about loading custom images at
# https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
column_names = dataset["train"].column_names
# 6. Get the column names for input/target.
dataset_columns = DATASET_NAME_MAPPING.get(args.dataset_name, None)
if args.image_column is None:
image_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
else:
image_column = args.image_column
if image_column not in column_names:
raise ValueError(
f"--image_column' value '{args.image_column}' needs to be one of: {', '.join(column_names)}"
)
if args.caption_column is None:
caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
else:
caption_column = args.caption_column
if caption_column not in column_names:
raise ValueError(
f"--caption_column' value '{args.caption_column}' needs to be one of: {', '.join(column_names)}"
)
# Preprocessing the datasets.
# We need to tokenize input captions and transform the images.
def tokenize_captions(examples, is_train=True):
captions = []
for caption in examples[caption_column]:
if isinstance(caption, str):
captions.append(caption)
elif isinstance(caption, (list, np.ndarray)):
# take a random caption if there are multiple
captions.append(random.choice(caption) if is_train else caption[0])
else:
raise ValueError(
f"Caption column `{caption_column}` should contain either strings or lists of strings."
)
inputs = tokenizer(
captions, max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
)
return inputs.input_ids
# Preprocessing the datasets.
train_transforms = transforms.Compose(
[
transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
]
)
def preprocess_train(examples):
images = [image.convert("RGB") for image in examples[image_column]]
examples["pixel_values"] = [train_transforms(image) for image in images]
examples["input_ids"] = tokenize_captions(examples)
return examples
with accelerator.main_process_first():
if args.max_train_samples is not None:
dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
# Set the training transforms
train_dataset = dataset["train"].with_transform(preprocess_train)
def collate_fn(examples):
pixel_values = torch.stack([example["pixel_values"] for example in examples])
pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
input_ids = torch.stack([example["input_ids"] for example in examples])
return {"pixel_values": pixel_values, "input_ids": input_ids}
# DataLoaders creation:
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
shuffle=True,
collate_fn=collate_fn,
batch_size=args.train_batch_size,
num_workers=args.dataloader_num_workers,
)
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
overrode_max_train_steps = True
lr_scheduler = get_scheduler(
args.lr_scheduler,
optimizer=optimizer,
num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
num_training_steps=args.max_train_steps * accelerator.num_processes,
)
# Prepare everything with our `accelerator`.
unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
unet, optimizer, train_dataloader, lr_scheduler
)
if args.use_ema:
ema_unet.to(accelerator.device)
# For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) to half-precision
# as these weights are only used for inference, keeping weights in full precision is not required.
weight_dtype = torch.float32
if accelerator.mixed_precision == "fp16":
weight_dtype = torch.float16
args.mixed_precision = accelerator.mixed_precision
elif accelerator.mixed_precision == "bf16":
weight_dtype = torch.bfloat16
args.mixed_precision = accelerator.mixed_precision
# Move text_encode and vae to gpu and cast to weight_dtype
text_encoder.to(accelerator.device, dtype=weight_dtype)
vae.to(accelerator.device, dtype=weight_dtype)
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# We need to initialize the trackers we use, and also store our configuration.
# The trackers initializes automatically on the main process.
if accelerator.is_main_process:
tracker_config = dict(vars(args))
tracker_config.pop("validation_prompts")
accelerator.init_trackers(args.tracker_project_name, tracker_config)
# Function for unwrapping if model was compiled with `torch.compile`.
def unwrap_model(model):
model = accelerator.unwrap_model(model)
model = model._orig_mod if is_compiled_module(model) else model
return model
# Train!
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.train_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
global_step = 0
first_epoch = 0
# Potentially load in the weights and states from a previous save
if args.resume_from_checkpoint:
if args.resume_from_checkpoint != "latest":
path = os.path.basename(args.resume_from_checkpoint)
else:
# Get the most recent checkpoint
dirs = os.listdir(args.output_dir)
dirs = [d for d in dirs if d.startswith("checkpoint")]
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
path = dirs[-1] if len(dirs) > 0 else None
if path is None:
accelerator.print(
f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
)
args.resume_from_checkpoint = None
initial_global_step = 0
else:
accelerator.print(f"Resuming from checkpoint {path}")
accelerator.load_state(os.path.join(args.output_dir, path))
global_step = int(path.split("-")[1])
initial_global_step = global_step
first_epoch = global_step // num_update_steps_per_epoch
else:
initial_global_step = 0
progress_bar = tqdm(
range(0, args.max_train_steps),
initial=initial_global_step,
desc="Steps",
# Only show the progress bar once on each machine.
disable=not accelerator.is_local_main_process,
)
for epoch in range(first_epoch, args.num_train_epochs):
train_loss = 0.0
for step, batch in enumerate(train_dataloader):
with accelerator.accumulate(unet):
# Convert images to latent space
latents = vae.encode(batch["pixel_values"].to(weight_dtype)).latent_dist.sample()
latents = latents * vae.config.scaling_factor
# Sample noise that we'll add to the latents
noise = torch.randn_like(latents)
if args.noise_offset:
# https://www.crosslabs.org//blog/diffusion-with-offset-noise
noise += args.noise_offset * torch.randn(
(latents.shape[0], latents.shape[1], 1, 1), device=latents.device
)
if args.input_perturbation:
new_noise = noise + args.input_perturbation * torch.randn_like(noise)
bsz = latents.shape[0]
# Sample a random timestep for each image
if args.loss_type == "huber" or args.loss_type == "smooth_l1":
timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (1,), device="cpu")
timestep = timesteps.item()
if args.huber_schedule == "exponential":
alpha = -math.log(args.huber_c) / noise_scheduler.config.num_train_timesteps
huber_c = math.exp(-alpha * timestep)
elif args.huber_schedule == "snr":
alphas_cumprod = noise_scheduler.alphas_cumprod[timestep]
sigmas = ((1.0 - alphas_cumprod) / alphas_cumprod) ** 0.5
huber_c = (1 - args.huber_c) / (1 + sigmas) ** 2 + args.huber_c
elif args.huber_schedule == "constant":
huber_c = args.huber_c
else:
raise NotImplementedError(f"Unknown Huber loss schedule {args.huber_schedule}!")
timesteps = timesteps.repeat(bsz).to(latents.device)
elif args.loss_type == "l2":
timesteps = torch.randint(
0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device
)
huber_c = 1 # may be anything, as it's not used
else:
raise NotImplementedError(f"Unknown loss type {args.loss_type}")
timesteps = timesteps.long()
# Add noise to the latents according to the noise magnitude at each timestep
# (this is the forward diffusion process)
if args.input_perturbation:
noisy_latents = noise_scheduler.add_noise(latents, new_noise, timesteps)
else:
noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
# Get the text embedding for conditioning
encoder_hidden_states = text_encoder(batch["input_ids"], return_dict=False)[0]
# Get the target for loss depending on the prediction type
if args.prediction_type is not None:
# set prediction_type of scheduler if defined
noise_scheduler.register_to_config(prediction_type=args.prediction_type)
if noise_scheduler.config.prediction_type == "epsilon":
target = noise
elif noise_scheduler.config.prediction_type == "v_prediction":
target = noise_scheduler.get_velocity(latents, noise, timesteps)
else:
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
# Predict the noise residual and compute loss
model_pred = unet(noisy_latents, timesteps, encoder_hidden_states, return_dict=False)[0]
if args.snr_gamma is None:
loss = conditional_loss(
model_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c
)
else:
# Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
# Since we predict the noise instead of x_0, the original formulation is slightly changed.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(noise_scheduler, timesteps)
mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
dim=1
)[0]
if noise_scheduler.config.prediction_type == "epsilon":
mse_loss_weights = mse_loss_weights / snr
elif noise_scheduler.config.prediction_type == "v_prediction":
mse_loss_weights = mse_loss_weights / (snr + 1)
loss = conditional_loss(
model_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
)
loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
loss = loss.mean()
# Gather the losses across all processes for logging (if we use distributed training).
avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
train_loss += avg_loss.item() / args.gradient_accumulation_steps
# Backpropagate
accelerator.backward(loss)
if accelerator.sync_gradients:
accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
# Checks if the accelerator has performed an optimization step behind the scenes
if accelerator.sync_gradients:
if args.use_ema:
ema_unet.step(unet.parameters())
progress_bar.update(1)
global_step += 1
accelerator.log({"train_loss": train_loss}, step=global_step)
train_loss = 0.0
if global_step % args.checkpointing_steps == 0:
if accelerator.is_main_process:
# _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
if args.checkpoints_total_limit is not None:
checkpoints = os.listdir(args.output_dir)
checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
# before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
if len(checkpoints) >= args.checkpoints_total_limit:
num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
removing_checkpoints = checkpoints[0:num_to_remove]
logger.info(
f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
)
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
for removing_checkpoint in removing_checkpoints:
removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
shutil.rmtree(removing_checkpoint)
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
accelerator.save_state(save_path)
logger.info(f"Saved state to {save_path}")
logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
progress_bar.set_postfix(**logs)
if global_step >= args.max_train_steps:
break
if accelerator.is_main_process:
if args.validation_prompts is not None and epoch % args.validation_epochs == 0:
if args.use_ema:
# Store the UNet parameters temporarily and load the EMA parameters to perform inference.
ema_unet.store(unet.parameters())
ema_unet.copy_to(unet.parameters())
log_validation(
vae,
text_encoder,
tokenizer,
unet,
args,
accelerator,
weight_dtype,
global_step,
)
if args.use_ema:
# Switch back to the original UNet parameters.
ema_unet.restore(unet.parameters())
# Create the pipeline using the trained modules and save it.
accelerator.wait_for_everyone()
if accelerator.is_main_process:
unet = unwrap_model(unet)
if args.use_ema:
ema_unet.copy_to(unet.parameters())
pipeline = StableDiffusionPipeline.from_pretrained(
args.pretrained_model_name_or_path,
text_encoder=text_encoder,
vae=vae,
unet=unet,
revision=args.revision,
variant=args.variant,
)
pipeline.save_pretrained(args.output_dir)
# Run a final round of inference.
images = []
if args.validation_prompts is not None:
logger.info("Running inference for collecting generated images...")
pipeline = pipeline.to(accelerator.device)
pipeline.torch_dtype = weight_dtype
pipeline.set_progress_bar_config(disable=True)
if args.enable_xformers_memory_efficient_attention:
pipeline.enable_xformers_memory_efficient_attention()
if args.seed is None:
generator = None
else:
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)
for i in range(len(args.validation_prompts)):
with torch.autocast("cuda"):
image = pipeline(args.validation_prompts[i], num_inference_steps=20, generator=generator).images[0]
images.append(image)
if args.push_to_hub:
save_model_card(args, repo_id, images, repo_folder=args.output_dir)
upload_folder(
repo_id=repo_id,
folder_path=args.output_dir,
commit_message="End of training",
ignore_patterns=["step_*", "epoch_*"],
)
accelerator.end_training()
if __name__ == "__main__":
main()
examples/research_projects/scheduled_huber_loss_training/text_to_image/train_text_to_image_lora.py 0 → 100644
View file @ 2f234376
#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Fine-tuning script for Stable Diffusion for text2image with support for LoRA."""
import argparse
import logging
import math
import os
import random
import shutil
from pathlib import Path
import datasets
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import ProjectConfiguration, set_seed
from datasets import load_dataset
from huggingface_hub import create_repo, upload_folder
from packaging import version
from peft import LoraConfig
from peft.utils import get_peft_model_state_dict
from torchvision import transforms
from tqdm.auto import tqdm
from transformers import CLIPTextModel, CLIPTokenizer
import diffusers
from diffusers import AutoencoderKL, DDPMScheduler, DiffusionPipeline, StableDiffusionPipeline, UNet2DConditionModel
from diffusers.optimization import get_scheduler
from diffusers.training_utils import cast_training_params, compute_snr
from diffusers.utils import check_min_version, convert_state_dict_to_diffusers, is_wandb_available
from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.torch_utils import is_compiled_module
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.28.0.dev0")
logger = get_logger(__name__, log_level="INFO")
def save_model_card(
repo_id: str,
images: list = None,
base_model: str = None,
dataset_name: str = None,
repo_folder: str = None,
):
img_str = ""
if images is not None:
for i, image in enumerate(images):
image.save(os.path.join(repo_folder, f"image_{i}.png"))
img_str += f"![img_{i}](./image_{i}.png)\n"
model_description = f"""
# LoRA text2image fine-tuning - {repo_id}
These are LoRA adaption weights for {base_model}. The weights were fine-tuned on the {dataset_name} dataset. You can find some example images in the following. \n
{img_str}
"""
model_card = load_or_create_model_card(
repo_id_or_path=repo_id,
from_training=True,
license="creativeml-openrail-m",
base_model=base_model,
model_description=model_description,
inference=True,
)
tags = [
"stable-diffusion",
"stable-diffusion-diffusers",
"text-to-image",
"diffusers",
"diffusers-training",
"lora",
]
model_card = populate_model_card(model_card, tags=tags)
model_card.save(os.path.join(repo_folder, "README.md"))
def parse_args():
parser = argparse.ArgumentParser(description="Simple example of a training script.")
parser.add_argument(
"--pretrained_model_name_or_path",
type=str,
default=None,
required=True,
help="Path to pretrained model or model identifier from huggingface.co/models.",
)
parser.add_argument(
"--revision",
type=str,
default=None,
required=False,
help="Revision of pretrained model identifier from huggingface.co/models.",
)
parser.add_argument(
"--variant",
type=str,
default=None,
help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
)
parser.add_argument(
"--dataset_name",
type=str,
default=None,
help=(
"The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
" dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
" or to a folder containing files that 🤗 Datasets can understand."
),
)
parser.add_argument(
"--dataset_config_name",
type=str,
default=None,
help="The config of the Dataset, leave as None if there's only one config.",
)
parser.add_argument(
"--train_data_dir",
type=str,
default=None,
help=(
"A folder containing the training data. Folder contents must follow the structure described in"
" https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
" must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
),
)
parser.add_argument(
"--image_column", type=str, default="image", help="The column of the dataset containing an image."
)
parser.add_argument(
"--caption_column",
type=str,
default="text",
help="The column of the dataset containing a caption or a list of captions.",
)
parser.add_argument(
"--validation_prompt", type=str, default=None, help="A prompt that is sampled during training for inference."
)
parser.add_argument(
"--num_validation_images",
type=int,
default=4,
help="Number of images that should be generated during validation with `validation_prompt`.",
)
parser.add_argument(
"--validation_epochs",
type=int,
default=1,
help=(
"Run fine-tuning validation every X epochs. The validation process consists of running the prompt"
" `args.validation_prompt` multiple times: `args.num_validation_images`."
),
)
parser.add_argument(
"--max_train_samples",
type=int,
default=None,
help=(
"For debugging purposes or quicker training, truncate the number of training examples to this "
"value if set."
),
)
parser.add_argument(
"--output_dir",
type=str,
default="sd-model-finetuned-lora",
help="The output directory where the model predictions and checkpoints will be written.",
)
parser.add_argument(
"--cache_dir",
type=str,
default=None,
help="The directory where the downloaded models and datasets will be stored.",
)
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
parser.add_argument(
"--resolution",
type=int,
default=512,
help=(
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
" resolution"
),
)
parser.add_argument(
"--center_crop",
default=False,
action="store_true",
help=(
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
" cropped. The images will be resized to the resolution first before cropping."
),
)
parser.add_argument(
"--random_flip",
action="store_true",
help="whether to randomly flip images horizontally",
)
parser.add_argument(
"--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
)
parser.add_argument("--num_train_epochs", type=int, default=100)
parser.add_argument(
"--max_train_steps",
type=int,
default=None,
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--gradient_checkpointing",
action="store_true",
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
)
parser.add_argument(
"--learning_rate",
type=float,
default=1e-4,
help="Initial learning rate (after the potential warmup period) to use.",
)
parser.add_argument(
"--scale_lr",
action="store_true",
default=False,
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
)
parser.add_argument(
"--lr_scheduler",
type=str,
default="constant",
help=(
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
' "constant", "constant_with_warmup"]'
),
)
parser.add_argument(
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
)
parser.add_argument(
"--snr_gamma",
type=float,
default=None,
help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
"More details here: https://arxiv.org/abs/2303.09556.",
)
parser.add_argument(
"--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
)
parser.add_argument(
"--allow_tf32",
action="store_true",
help=(
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
),
)
parser.add_argument(
"--dataloader_num_workers",
type=int,
default=0,
help=(
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
),
)
parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
parser.add_argument(
"--prediction_type",
type=str,
default=None,
help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediction_type` is chosen.",
)
parser.add_argument(
"--hub_model_id",
type=str,
default=None,
help="The name of the repository to keep in sync with the local `output_dir`.",
)
parser.add_argument(
"--logging_dir",
type=str,
default="logs",
help=(
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
),
)
parser.add_argument(
"--mixed_precision",
type=str,
default=None,
choices=["no", "fp16", "bf16"],
help=(
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
),
)
parser.add_argument(
"--report_to",
type=str,
default="tensorboard",
help=(
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
),
)
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
parser.add_argument(
"--checkpointing_steps",
type=int,
default=500,
help=(
"Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming"
" training using `--resume_from_checkpoint`."
),
)
parser.add_argument(
"--checkpoints_total_limit",
type=int,
default=None,
help=("Max number of checkpoints to store."),
)
parser.add_argument(
"--resume_from_checkpoint",
type=str,
default=None,
help=(
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
),
)
parser.add_argument(
"--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
)
parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
parser.add_argument(
"--loss_type",
type=str,
default="l2",
choices=["l2", "huber", "smooth_l1"],
help="The type of loss to use and whether it's timestep-scheduled. See Issue #7488 for more info.",
)
parser.add_argument(
"--huber_schedule",
type=str,
default="snr",
choices=["constant", "exponential", "snr"],
help="The schedule to use for the huber losses parameter",
)
parser.add_argument(
"--huber_c",
type=float,
default=0.1,
help="The huber loss parameter. Only used if one of the huber loss modes (huber or smooth l1) is selected with loss_type.",
)
parser.add_argument(
"--rank",
type=int,
default=4,
help=("The dimension of the LoRA update matrices."),
)
args = parser.parse_args()
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
if env_local_rank != -1 and env_local_rank != args.local_rank:
args.local_rank = env_local_rank
# Sanity checks
if args.dataset_name is None and args.train_data_dir is None:
raise ValueError("Need either a dataset name or a training folder.")
return args
DATASET_NAME_MAPPING = {
"lambdalabs/pokemon-blip-captions": ("image", "text"),
}
# NOTE: if you're using the scheduled version, huber_c has to depend on the timesteps already
def conditional_loss(
model_pred: torch.Tensor,
target: torch.Tensor,
reduction: str = "mean",
loss_type: str = "l2",
huber_c: float = 0.1,
):
if loss_type == "l2":
loss = F.mse_loss(model_pred, target, reduction=reduction)
elif loss_type == "huber":
loss = 2 * huber_c * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
elif loss_type == "smooth_l1":
loss = 2 * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
else:
raise NotImplementedError(f"Unsupported Loss Type {loss_type}")
return loss
def main():
args = parse_args()
if args.report_to == "wandb" and args.hub_token is not None:
raise ValueError(
"You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
" Please use `huggingface-cli login` to authenticate with the Hub."
)
logging_dir = Path(args.output_dir, args.logging_dir)
accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
accelerator = Accelerator(
gradient_accumulation_steps=args.gradient_accumulation_steps,
mixed_precision=args.mixed_precision,
log_with=args.report_to,
project_config=accelerator_project_config,
)
if args.report_to == "wandb":
if not is_wandb_available():
raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
import wandb
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state, main_process_only=False)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_warning()
diffusers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
diffusers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
if args.push_to_hub:
repo_id = create_repo(
repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
).repo_id
# Load scheduler, tokenizer and models.
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
tokenizer = CLIPTokenizer.from_pretrained(
args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision
)
text_encoder = CLIPTextModel.from_pretrained(
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
)
vae = AutoencoderKL.from_pretrained(
args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant
)
unet = UNet2DConditionModel.from_pretrained(
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
)
# freeze parameters of models to save more memory
unet.requires_grad_(False)
vae.requires_grad_(False)
text_encoder.requires_grad_(False)
# For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) to half-precision
# as these weights are only used for inference, keeping weights in full precision is not required.
weight_dtype = torch.float32
if accelerator.mixed_precision == "fp16":
weight_dtype = torch.float16
elif accelerator.mixed_precision == "bf16":
weight_dtype = torch.bfloat16
# Freeze the unet parameters before adding adapters
for param in unet.parameters():
param.requires_grad_(False)
unet_lora_config = LoraConfig(
r=args.rank,
lora_alpha=args.rank,
init_lora_weights="gaussian",
target_modules=["to_k", "to_q", "to_v", "to_out.0"],
)
# Move unet, vae and text_encoder to device and cast to weight_dtype
unet.to(accelerator.device, dtype=weight_dtype)
vae.to(accelerator.device, dtype=weight_dtype)
text_encoder.to(accelerator.device, dtype=weight_dtype)
# Add adapter and make sure the trainable params are in float32.
unet.add_adapter(unet_lora_config)
if args.mixed_precision == "fp16":
# only upcast trainable parameters (LoRA) into fp32
cast_training_params(unet, dtype=torch.float32)
if args.enable_xformers_memory_efficient_attention:
if is_xformers_available():
import xformers
xformers_version = version.parse(xformers.__version__)
if xformers_version == version.parse("0.0.16"):
logger.warning(
"xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
)
unet.enable_xformers_memory_efficient_attention()
else:
raise ValueError("xformers is not available. Make sure it is installed correctly")
lora_layers = filter(lambda p: p.requires_grad, unet.parameters())
if args.gradient_checkpointing:
unet.enable_gradient_checkpointing()
# Enable TF32 for faster training on Ampere GPUs,
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
if args.allow_tf32:
torch.backends.cuda.matmul.allow_tf32 = True
if args.scale_lr:
args.learning_rate = (
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
)
# Initialize the optimizer
if args.use_8bit_adam:
try:
import bitsandbytes as bnb
except ImportError:
raise ImportError(
"Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
)
optimizer_cls = bnb.optim.AdamW8bit
else:
optimizer_cls = torch.optim.AdamW
optimizer = optimizer_cls(
lora_layers,
lr=args.learning_rate,
betas=(args.adam_beta1, args.adam_beta2),
weight_decay=args.adam_weight_decay,
eps=args.adam_epsilon,
)
# Get the datasets: you can either provide your own training and evaluation files (see below)
# or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).
# In distributed training, the load_dataset function guarantees that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
dataset = load_dataset(
args.dataset_name,
args.dataset_config_name,
cache_dir=args.cache_dir,
data_dir=args.train_data_dir,
)
else:
data_files = {}
if args.train_data_dir is not None:
data_files["train"] = os.path.join(args.train_data_dir, "**")
dataset = load_dataset(
"imagefolder",
data_files=data_files,
cache_dir=args.cache_dir,
)
# See more about loading custom images at
# https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
column_names = dataset["train"].column_names
# 6. Get the column names for input/target.
dataset_columns = DATASET_NAME_MAPPING.get(args.dataset_name, None)
if args.image_column is None:
image_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
else:
image_column = args.image_column
if image_column not in column_names:
raise ValueError(
f"--image_column' value '{args.image_column}' needs to be one of: {', '.join(column_names)}"
)
if args.caption_column is None:
caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
else:
caption_column = args.caption_column
if caption_column not in column_names:
raise ValueError(
f"--caption_column' value '{args.caption_column}' needs to be one of: {', '.join(column_names)}"
)
# Preprocessing the datasets.
# We need to tokenize input captions and transform the images.
def tokenize_captions(examples, is_train=True):
captions = []
for caption in examples[caption_column]:
if isinstance(caption, str):
captions.append(caption)
elif isinstance(caption, (list, np.ndarray)):
# take a random caption if there are multiple
captions.append(random.choice(caption) if is_train else caption[0])
else:
raise ValueError(
f"Caption column `{caption_column}` should contain either strings or lists of strings."
)
inputs = tokenizer(
captions, max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
)
return inputs.input_ids
# Preprocessing the datasets.
train_transforms = transforms.Compose(
[
transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
]
)
def unwrap_model(model):
model = accelerator.unwrap_model(model)
model = model._orig_mod if is_compiled_module(model) else model
return model
def preprocess_train(examples):
images = [image.convert("RGB") for image in examples[image_column]]
examples["pixel_values"] = [train_transforms(image) for image in images]
examples["input_ids"] = tokenize_captions(examples)
return examples
with accelerator.main_process_first():
if args.max_train_samples is not None:
dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
# Set the training transforms
train_dataset = dataset["train"].with_transform(preprocess_train)
def collate_fn(examples):
pixel_values = torch.stack([example["pixel_values"] for example in examples])
pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
input_ids = torch.stack([example["input_ids"] for example in examples])
return {"pixel_values": pixel_values, "input_ids": input_ids}
# DataLoaders creation:
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
shuffle=True,
collate_fn=collate_fn,
batch_size=args.train_batch_size,
num_workers=args.dataloader_num_workers,
)
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
overrode_max_train_steps = True
lr_scheduler = get_scheduler(
args.lr_scheduler,
optimizer=optimizer,
num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
num_training_steps=args.max_train_steps * accelerator.num_processes,
)
# Prepare everything with our `accelerator`.
unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
unet, optimizer, train_dataloader, lr_scheduler
)
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# We need to initialize the trackers we use, and also store our configuration.
# The trackers initializes automatically on the main process.
if accelerator.is_main_process:
accelerator.init_trackers("text2image-fine-tune", config=vars(args))
# Train!
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.train_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
global_step = 0
first_epoch = 0
# Potentially load in the weights and states from a previous save
if args.resume_from_checkpoint:
if args.resume_from_checkpoint != "latest":
path = os.path.basename(args.resume_from_checkpoint)
else:
# Get the most recent checkpoint
dirs = os.listdir(args.output_dir)
dirs = [d for d in dirs if d.startswith("checkpoint")]
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
path = dirs[-1] if len(dirs) > 0 else None
if path is None:
accelerator.print(
f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
)
args.resume_from_checkpoint = None
initial_global_step = 0
else:
accelerator.print(f"Resuming from checkpoint {path}")
accelerator.load_state(os.path.join(args.output_dir, path))
global_step = int(path.split("-")[1])
initial_global_step = global_step
first_epoch = global_step // num_update_steps_per_epoch
else:
initial_global_step = 0
progress_bar = tqdm(
range(0, args.max_train_steps),
initial=initial_global_step,
desc="Steps",
# Only show the progress bar once on each machine.
disable=not accelerator.is_local_main_process,
)
for epoch in range(first_epoch, args.num_train_epochs):
unet.train()
train_loss = 0.0
for step, batch in enumerate(train_dataloader):
with accelerator.accumulate(unet):
# Convert images to latent space
latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
latents = latents * vae.config.scaling_factor
# Sample noise that we'll add to the latents
noise = torch.randn_like(latents)
if args.noise_offset:
# https://www.crosslabs.org//blog/diffusion-with-offset-noise
noise += args.noise_offset * torch.randn(
(latents.shape[0], latents.shape[1], 1, 1), device=latents.device
)
bsz = latents.shape[0]
# Sample a random timestep for each image
if args.loss_type == "huber" or args.loss_type == "smooth_l1":
timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (1,), device="cpu")
timestep = timesteps.item()
if args.huber_schedule == "exponential":
alpha = -math.log(args.huber_c) / noise_scheduler.config.num_train_timesteps
huber_c = math.exp(-alpha * timestep)
elif args.huber_schedule == "snr":
alphas_cumprod = noise_scheduler.alphas_cumprod[timestep]
sigmas = ((1.0 - alphas_cumprod) / alphas_cumprod) ** 0.5
huber_c = (1 - args.huber_c) / (1 + sigmas) ** 2 + args.huber_c
elif args.huber_schedule == "constant":
huber_c = args.huber_c
else:
raise NotImplementedError(f"Unknown Huber loss schedule {args.huber_schedule}!")
timesteps = timesteps.repeat(bsz).to(latents.device)
elif args.loss_type == "l2":
timesteps = torch.randint(
0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device
)
huber_c = 1 # may be anything, as it's not used
else:
raise NotImplementedError(f"Unknown loss type {args.loss_type}")
timesteps = timesteps.long()
# Add noise to the latents according to the noise magnitude at each timestep
# (this is the forward diffusion process)
noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
# Get the text embedding for conditioning
encoder_hidden_states = text_encoder(batch["input_ids"], return_dict=False)[0]
# Get the target for loss depending on the prediction type
if args.prediction_type is not None:
# set prediction_type of scheduler if defined
noise_scheduler.register_to_config(prediction_type=args.prediction_type)
if noise_scheduler.config.prediction_type == "epsilon":
target = noise
elif noise_scheduler.config.prediction_type == "v_prediction":
target = noise_scheduler.get_velocity(latents, noise, timesteps)
else:
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
# Predict the noise residual and compute loss
model_pred = unet(noisy_latents, timesteps, encoder_hidden_states, return_dict=False)[0]
if args.snr_gamma is None:
loss = conditional_loss(
model_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c
)
else:
# Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
# Since we predict the noise instead of x_0, the original formulation is slightly changed.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(noise_scheduler, timesteps)
mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
dim=1
)[0]
if noise_scheduler.config.prediction_type == "epsilon":
mse_loss_weights = mse_loss_weights / snr
elif noise_scheduler.config.prediction_type == "v_prediction":
mse_loss_weights = mse_loss_weights / (snr + 1)
loss = conditional_loss(
model_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
)
loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
loss = loss.mean()
# Gather the losses across all processes for logging (if we use distributed training).
avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
train_loss += avg_loss.item() / args.gradient_accumulation_steps
# Backpropagate
accelerator.backward(loss)
if accelerator.sync_gradients:
params_to_clip = lora_layers
accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
# Checks if the accelerator has performed an optimization step behind the scenes
if accelerator.sync_gradients:
progress_bar.update(1)
global_step += 1
accelerator.log({"train_loss": train_loss}, step=global_step)
train_loss = 0.0
if global_step % args.checkpointing_steps == 0:
if accelerator.is_main_process:
# _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
if args.checkpoints_total_limit is not None:
checkpoints = os.listdir(args.output_dir)
checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
# before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
if len(checkpoints) >= args.checkpoints_total_limit:
num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
removing_checkpoints = checkpoints[0:num_to_remove]
logger.info(
f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
)
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
for removing_checkpoint in removing_checkpoints:
removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
shutil.rmtree(removing_checkpoint)
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
accelerator.save_state(save_path)
unwrapped_unet = unwrap_model(unet)
unet_lora_state_dict = convert_state_dict_to_diffusers(
get_peft_model_state_dict(unwrapped_unet)
)
StableDiffusionPipeline.save_lora_weights(
save_directory=save_path,
unet_lora_layers=unet_lora_state_dict,
safe_serialization=True,
)
logger.info(f"Saved state to {save_path}")
logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
progress_bar.set_postfix(**logs)
if global_step >= args.max_train_steps:
break
if accelerator.is_main_process:
if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
logger.info(
f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
f" {args.validation_prompt}."
)
# create pipeline
pipeline = DiffusionPipeline.from_pretrained(
args.pretrained_model_name_or_path,
unet=unwrap_model(unet),
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
)
pipeline = pipeline.to(accelerator.device)
pipeline.set_progress_bar_config(disable=True)
# run inference
generator = torch.Generator(device=accelerator.device)
if args.seed is not None:
generator = generator.manual_seed(args.seed)
images = []
with torch.cuda.amp.autocast():
for _ in range(args.num_validation_images):
images.append(
pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0]
)
for tracker in accelerator.trackers:
if tracker.name == "tensorboard":
np_images = np.stack([np.asarray(img) for img in images])
tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
if tracker.name == "wandb":
tracker.log(
{
"validation": [
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
for i, image in enumerate(images)
]
}
)
del pipeline
torch.cuda.empty_cache()
# Save the lora layers
accelerator.wait_for_everyone()
if accelerator.is_main_process:
unet = unet.to(torch.float32)
unwrapped_unet = unwrap_model(unet)
unet_lora_state_dict = convert_state_dict_to_diffusers(get_peft_model_state_dict(unwrapped_unet))
StableDiffusionPipeline.save_lora_weights(
save_directory=args.output_dir,
unet_lora_layers=unet_lora_state_dict,
safe_serialization=True,
)
if args.push_to_hub:
save_model_card(
repo_id,
images=images,
base_model=args.pretrained_model_name_or_path,
dataset_name=args.dataset_name,
repo_folder=args.output_dir,
)
upload_folder(
repo_id=repo_id,
folder_path=args.output_dir,
commit_message="End of training",
ignore_patterns=["step_*", "epoch_*"],
)
# Final inference
# Load previous pipeline
if args.validation_prompt is not None:
pipeline = DiffusionPipeline.from_pretrained(
args.pretrained_model_name_or_path,
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
)
pipeline = pipeline.to(accelerator.device)
# load attention processors
pipeline.load_lora_weights(args.output_dir)
# run inference
generator = torch.Generator(device=accelerator.device)
if args.seed is not None:
generator = generator.manual_seed(args.seed)
images = []
with torch.cuda.amp.autocast():
for _ in range(args.num_validation_images):
images.append(
pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0]
)
for tracker in accelerator.trackers:
if len(images) != 0:
if tracker.name == "tensorboard":
np_images = np.stack([np.asarray(img) for img in images])
tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
if tracker.name == "wandb":
tracker.log(
{
"test": [
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
for i, image in enumerate(images)
]
}
)
accelerator.end_training()
if __name__ == "__main__":
main()
examples/research_projects/scheduled_huber_loss_training/text_to_image/train_text_to_image_lora_sdxl.py 0 → 100644
View file @ 2f234376
#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Fine-tuning script for Stable Diffusion XL for text2image with support for LoRA."""
import argparse
import logging
import math
import os
import random
import shutil
from pathlib import Path
import datasets
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration, set_seed
from datasets import load_dataset
from huggingface_hub import create_repo, upload_folder
from packaging import version
from peft import LoraConfig, set_peft_model_state_dict
from peft.utils import get_peft_model_state_dict
from torchvision import transforms
from torchvision.transforms.functional import crop
from tqdm.auto import tqdm
from transformers import AutoTokenizer, PretrainedConfig
import diffusers
from diffusers import (
AutoencoderKL,
DDPMScheduler,
StableDiffusionXLPipeline,
UNet2DConditionModel,
)
from diffusers.loaders import LoraLoaderMixin
from diffusers.optimization import get_scheduler
from diffusers.training_utils import _set_state_dict_into_text_encoder, cast_training_params, compute_snr
from diffusers.utils import (
check_min_version,
convert_state_dict_to_diffusers,
convert_unet_state_dict_to_peft,
is_wandb_available,
)
from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.torch_utils import is_compiled_module
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.28.0.dev0")
logger = get_logger(__name__)
def save_model_card(
repo_id: str,
images: list = None,
base_model: str = None,
dataset_name: str = None,
train_text_encoder: bool = False,
repo_folder: str = None,
vae_path: str = None,
):
img_str = ""
if images is not None:
for i, image in enumerate(images):
image.save(os.path.join(repo_folder, f"image_{i}.png"))
img_str += f"![img_{i}](./image_{i}.png)\n"
model_description = f"""
# LoRA text2image fine-tuning - {repo_id}
These are LoRA adaption weights for {base_model}. The weights were fine-tuned on the {dataset_name} dataset. You can find some example images in the following. \n
{img_str}
LoRA for the text encoder was enabled: {train_text_encoder}.
Special VAE used for training: {vae_path}.
"""
model_card = load_or_create_model_card(
repo_id_or_path=repo_id,
from_training=True,
license="creativeml-openrail-m",
base_model=base_model,
model_description=model_description,
inference=True,
)
tags = [
"stable-diffusion-xl",
"stable-diffusion-xl-diffusers",
"text-to-image",
"diffusers",
"diffusers-training",
"lora",
]
model_card = populate_model_card(model_card, tags=tags)
model_card.save(os.path.join(repo_folder, "README.md"))
def import_model_class_from_model_name_or_path(
pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
):
text_encoder_config = PretrainedConfig.from_pretrained(
pretrained_model_name_or_path, subfolder=subfolder, revision=revision
)
model_class = text_encoder_config.architectures[0]
if model_class == "CLIPTextModel":
from transformers import CLIPTextModel
return CLIPTextModel
elif model_class == "CLIPTextModelWithProjection":
from transformers import CLIPTextModelWithProjection
return CLIPTextModelWithProjection
else:
raise ValueError(f"{model_class} is not supported.")
def parse_args(input_args=None):
parser = argparse.ArgumentParser(description="Simple example of a training script.")
parser.add_argument(
"--pretrained_model_name_or_path",
type=str,
default=None,
required=True,
help="Path to pretrained model or model identifier from huggingface.co/models.",
)
parser.add_argument(
"--pretrained_vae_model_name_or_path",
type=str,
default=None,
help="Path to pretrained VAE model with better numerical stability. More details: https://github.com/huggingface/diffusers/pull/4038.",
)
parser.add_argument(
"--revision",
type=str,
default=None,
required=False,
help="Revision of pretrained model identifier from huggingface.co/models.",
)
parser.add_argument(
"--variant",
type=str,
default=None,
help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
)
parser.add_argument(
"--dataset_name",
type=str,
default=None,
help=(
"The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
" dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
" or to a folder containing files that 🤗 Datasets can understand."
),
)
parser.add_argument(
"--dataset_config_name",
type=str,
default=None,
help="The config of the Dataset, leave as None if there's only one config.",
)
parser.add_argument(
"--train_data_dir",
type=str,
default=None,
help=(
"A folder containing the training data. Folder contents must follow the structure described in"
" https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
" must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
),
)
parser.add_argument(
"--image_column", type=str, default="image", help="The column of the dataset containing an image."
)
parser.add_argument(
"--caption_column",
type=str,
default="text",
help="The column of the dataset containing a caption or a list of captions.",
)
parser.add_argument(
"--validation_prompt",
type=str,
default=None,
help="A prompt that is used during validation to verify that the model is learning.",
)
parser.add_argument(
"--num_validation_images",
type=int,
default=4,
help="Number of images that should be generated during validation with `validation_prompt`.",
)
parser.add_argument(
"--validation_epochs",
type=int,
default=1,
help=(
"Run fine-tuning validation every X epochs. The validation process consists of running the prompt"
" `args.validation_prompt` multiple times: `args.num_validation_images`."
),
)
parser.add_argument(
"--max_train_samples",
type=int,
default=None,
help=(
"For debugging purposes or quicker training, truncate the number of training examples to this "
"value if set."
),
)
parser.add_argument(
"--output_dir",
type=str,
default="sd-model-finetuned-lora",
help="The output directory where the model predictions and checkpoints will be written.",
)
parser.add_argument(
"--cache_dir",
type=str,
default=None,
help="The directory where the downloaded models and datasets will be stored.",
)
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
parser.add_argument(
"--resolution",
type=int,
default=1024,
help=(
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
" resolution"
),
)
parser.add_argument(
"--center_crop",
default=False,
action="store_true",
help=(
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
" cropped. The images will be resized to the resolution first before cropping."
),
)
parser.add_argument(
"--random_flip",
action="store_true",
help="whether to randomly flip images horizontally",
)
parser.add_argument(
"--train_text_encoder",
action="store_true",
help="Whether to train the text encoder. If set, the text encoder should be float32 precision.",
)
parser.add_argument(
"--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
)
parser.add_argument("--num_train_epochs", type=int, default=100)
parser.add_argument(
"--max_train_steps",
type=int,
default=None,
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
)
parser.add_argument(
"--checkpointing_steps",
type=int,
default=500,
help=(
"Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
" checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
" training using `--resume_from_checkpoint`."
),
)
parser.add_argument(
"--checkpoints_total_limit",
type=int,
default=None,
help=("Max number of checkpoints to store."),
)
parser.add_argument(
"--resume_from_checkpoint",
type=str,
default=None,
help=(
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
),
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--gradient_checkpointing",
action="store_true",
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
)
parser.add_argument(
"--learning_rate",
type=float,
default=1e-4,
help="Initial learning rate (after the potential warmup period) to use.",
)
parser.add_argument(
"--scale_lr",
action="store_true",
default=False,
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
)
parser.add_argument(
"--lr_scheduler",
type=str,
default="constant",
help=(
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
' "constant", "constant_with_warmup"]'
),
)
parser.add_argument(
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
)
parser.add_argument(
"--snr_gamma",
type=float,
default=None,
help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
"More details here: https://arxiv.org/abs/2303.09556.",
)
parser.add_argument(
"--allow_tf32",
action="store_true",
help=(
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
),
)
parser.add_argument(
"--dataloader_num_workers",
type=int,
default=0,
help=(
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
),
)
parser.add_argument(
"--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
)
parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
parser.add_argument(
"--prediction_type",
type=str,
default=None,
help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediction_type` is chosen.",
)
parser.add_argument(
"--hub_model_id",
type=str,
default=None,
help="The name of the repository to keep in sync with the local `output_dir`.",
)
parser.add_argument(
"--logging_dir",
type=str,
default="logs",
help=(
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
),
)
parser.add_argument(
"--report_to",
type=str,
default="tensorboard",
help=(
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
),
)
parser.add_argument(
"--mixed_precision",
type=str,
default=None,
choices=["no", "fp16", "bf16"],
help=(
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
),
)
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
parser.add_argument(
"--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
)
parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
parser.add_argument(
"--loss_type",
type=str,
default="l2",
choices=["l2", "huber", "smooth_l1"],
help="The type of loss to use and whether it's timestep-scheduled. See Issue #7488 for more info.",
)
parser.add_argument(
"--huber_schedule",
type=str,
default="snr",
choices=["constant", "exponential", "snr"],
help="The schedule to use for the huber losses parameter",
)
parser.add_argument(
"--huber_c",
type=float,
default=0.1,
help="The huber loss parameter. Only used if one of the huber loss modes (huber or smooth l1) is selected with loss_type.",
)
parser.add_argument(
"--rank",
type=int,
default=4,
help=("The dimension of the LoRA update matrices."),
)
parser.add_argument(
"--debug_loss",
action="store_true",
help="debug loss for each image, if filenames are awailable in the dataset",
)
if input_args is not None:
args = parser.parse_args(input_args)
else:
args = parser.parse_args()
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
if env_local_rank != -1 and env_local_rank != args.local_rank:
args.local_rank = env_local_rank
# Sanity checks
if args.dataset_name is None and args.train_data_dir is None:
raise ValueError("Need either a dataset name or a training folder.")
return args
DATASET_NAME_MAPPING = {
"lambdalabs/pokemon-blip-captions": ("image", "text"),
}
def tokenize_prompt(tokenizer, prompt):
text_inputs = tokenizer(
prompt,
padding="max_length",
max_length=tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
return text_input_ids
# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
def encode_prompt(text_encoders, tokenizers, prompt, text_input_ids_list=None):
prompt_embeds_list = []
for i, text_encoder in enumerate(text_encoders):
if tokenizers is not None:
tokenizer = tokenizers[i]
text_input_ids = tokenize_prompt(tokenizer, prompt)
else:
assert text_input_ids_list is not None
text_input_ids = text_input_ids_list[i]
prompt_embeds = text_encoder(
text_input_ids.to(text_encoder.device), output_hidden_states=True, return_dict=False
)
# We are only ALWAYS interested in the pooled output of the final text encoder
pooled_prompt_embeds = prompt_embeds[0]
prompt_embeds = prompt_embeds[-1][-2]
bs_embed, seq_len, _ = prompt_embeds.shape
prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
prompt_embeds_list.append(prompt_embeds)
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
return prompt_embeds, pooled_prompt_embeds
# NOTE: if you're using the scheduled version, huber_c has to depend on the timesteps already
def conditional_loss(
model_pred: torch.Tensor,
target: torch.Tensor,
reduction: str = "mean",
loss_type: str = "l2",
huber_c: float = 0.1,
):
if loss_type == "l2":
loss = F.mse_loss(model_pred, target, reduction=reduction)
elif loss_type == "huber" or loss_type == "huber_scheduled":
loss = huber_c * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
else:
raise NotImplementedError(f"Unsupported Loss Type {loss_type}")
return loss
def main(args):
if args.report_to == "wandb" and args.hub_token is not None:
raise ValueError(
"You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
" Please use `huggingface-cli login` to authenticate with the Hub."
)
logging_dir = Path(args.output_dir, args.logging_dir)
accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
accelerator = Accelerator(
gradient_accumulation_steps=args.gradient_accumulation_steps,
mixed_precision=args.mixed_precision,
log_with=args.report_to,
project_config=accelerator_project_config,
kwargs_handlers=[kwargs],
)
if args.report_to == "wandb":
if not is_wandb_available():
raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
import wandb
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state, main_process_only=False)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_warning()
diffusers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
diffusers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
if args.push_to_hub:
repo_id = create_repo(
repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
).repo_id
# Load the tokenizers
tokenizer_one = AutoTokenizer.from_pretrained(
args.pretrained_model_name_or_path,
subfolder="tokenizer",
revision=args.revision,
use_fast=False,
)
tokenizer_two = AutoTokenizer.from_pretrained(
args.pretrained_model_name_or_path,
subfolder="tokenizer_2",
revision=args.revision,
use_fast=False,
)
# import correct text encoder classes
text_encoder_cls_one = import_model_class_from_model_name_or_path(
args.pretrained_model_name_or_path, args.revision
)
text_encoder_cls_two = import_model_class_from_model_name_or_path(
args.pretrained_model_name_or_path, args.revision, subfolder="text_encoder_2"
)
# Load scheduler and models
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
text_encoder_one = text_encoder_cls_one.from_pretrained(
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
)
text_encoder_two = text_encoder_cls_two.from_pretrained(
args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
)
vae_path = (
args.pretrained_model_name_or_path
if args.pretrained_vae_model_name_or_path is None
else args.pretrained_vae_model_name_or_path
)
vae = AutoencoderKL.from_pretrained(
vae_path,
subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
revision=args.revision,
variant=args.variant,
)
unet = UNet2DConditionModel.from_pretrained(
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
)
# We only train the additional adapter LoRA layers
vae.requires_grad_(False)
text_encoder_one.requires_grad_(False)
text_encoder_two.requires_grad_(False)
unet.requires_grad_(False)
# For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) to half-precision
# as these weights are only used for inference, keeping weights in full precision is not required.
weight_dtype = torch.float32
if accelerator.mixed_precision == "fp16":
weight_dtype = torch.float16
elif accelerator.mixed_precision == "bf16":
weight_dtype = torch.bfloat16
# Move unet, vae and text_encoder to device and cast to weight_dtype
# The VAE is in float32 to avoid NaN losses.
unet.to(accelerator.device, dtype=weight_dtype)
if args.pretrained_vae_model_name_or_path is None:
vae.to(accelerator.device, dtype=torch.float32)
else:
vae.to(accelerator.device, dtype=weight_dtype)
text_encoder_one.to(accelerator.device, dtype=weight_dtype)
text_encoder_two.to(accelerator.device, dtype=weight_dtype)
if args.enable_xformers_memory_efficient_attention:
if is_xformers_available():
import xformers
xformers_version = version.parse(xformers.__version__)
if xformers_version == version.parse("0.0.16"):
logger.warning(
"xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
)
unet.enable_xformers_memory_efficient_attention()
else:
raise ValueError("xformers is not available. Make sure it is installed correctly")
# now we will add new LoRA weights to the attention layers
# Set correct lora layers
unet_lora_config = LoraConfig(
r=args.rank,
lora_alpha=args.rank,
init_lora_weights="gaussian",
target_modules=["to_k", "to_q", "to_v", "to_out.0"],
)
unet.add_adapter(unet_lora_config)
# The text encoder comes from 🤗 transformers, we will also attach adapters to it.
if args.train_text_encoder:
# ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
text_lora_config = LoraConfig(
r=args.rank,
lora_alpha=args.rank,
init_lora_weights="gaussian",
target_modules=["q_proj", "k_proj", "v_proj", "out_proj"],
)
text_encoder_one.add_adapter(text_lora_config)
text_encoder_two.add_adapter(text_lora_config)
def unwrap_model(model):
model = accelerator.unwrap_model(model)
model = model._orig_mod if is_compiled_module(model) else model
return model
# create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
def save_model_hook(models, weights, output_dir):
if accelerator.is_main_process:
# there are only two options here. Either are just the unet attn processor layers
# or there are the unet and text encoder attn layers
unet_lora_layers_to_save = None
text_encoder_one_lora_layers_to_save = None
text_encoder_two_lora_layers_to_save = None
for model in models:
if isinstance(unwrap_model(model), type(unwrap_model(unet))):
unet_lora_layers_to_save = convert_state_dict_to_diffusers(get_peft_model_state_dict(model))
elif isinstance(unwrap_model(model), type(unwrap_model(text_encoder_one))):
text_encoder_one_lora_layers_to_save = convert_state_dict_to_diffusers(
get_peft_model_state_dict(model)
)
elif isinstance(unwrap_model(model), type(unwrap_model(text_encoder_two))):
text_encoder_two_lora_layers_to_save = convert_state_dict_to_diffusers(
get_peft_model_state_dict(model)
)
else:
raise ValueError(f"unexpected save model: {model.__class__}")
# make sure to pop weight so that corresponding model is not saved again
if weights:
weights.pop()
StableDiffusionXLPipeline.save_lora_weights(
output_dir,
unet_lora_layers=unet_lora_layers_to_save,
text_encoder_lora_layers=text_encoder_one_lora_layers_to_save,
text_encoder_2_lora_layers=text_encoder_two_lora_layers_to_save,
)
def load_model_hook(models, input_dir):
unet_ = None
text_encoder_one_ = None
text_encoder_two_ = None
while len(models) > 0:
model = models.pop()
if isinstance(model, type(unwrap_model(unet))):
unet_ = model
elif isinstance(model, type(unwrap_model(text_encoder_one))):
text_encoder_one_ = model
elif isinstance(model, type(unwrap_model(text_encoder_two))):
text_encoder_two_ = model
else:
raise ValueError(f"unexpected save model: {model.__class__}")
lora_state_dict, _ = LoraLoaderMixin.lora_state_dict(input_dir)
unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
if incompatible_keys is not None:
# check only for unexpected keys
unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
if unexpected_keys:
logger.warning(
f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
f" {unexpected_keys}. "
)
if args.train_text_encoder:
_set_state_dict_into_text_encoder(lora_state_dict, prefix="text_encoder.", text_encoder=text_encoder_one_)
_set_state_dict_into_text_encoder(
lora_state_dict, prefix="text_encoder_2.", text_encoder=text_encoder_two_
)
# Make sure the trainable params are in float32. This is again needed since the base models
# are in `weight_dtype`. More details:
# https://github.com/huggingface/diffusers/pull/6514#discussion_r1449796804
if args.mixed_precision == "fp16":
models = [unet_]
if args.train_text_encoder:
models.extend([text_encoder_one_, text_encoder_two_])
cast_training_params(models, dtype=torch.float32)
accelerator.register_save_state_pre_hook(save_model_hook)
accelerator.register_load_state_pre_hook(load_model_hook)
if args.gradient_checkpointing:
unet.enable_gradient_checkpointing()
if args.train_text_encoder:
text_encoder_one.gradient_checkpointing_enable()
text_encoder_two.gradient_checkpointing_enable()
# Enable TF32 for faster training on Ampere GPUs,
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
if args.allow_tf32:
torch.backends.cuda.matmul.allow_tf32 = True
if args.scale_lr:
args.learning_rate = (
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
)
# Make sure the trainable params are in float32.
if args.mixed_precision == "fp16":
models = [unet]
if args.train_text_encoder:
models.extend([text_encoder_one, text_encoder_two])
cast_training_params(models, dtype=torch.float32)
# Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
if args.use_8bit_adam:
try:
import bitsandbytes as bnb
except ImportError:
raise ImportError(
"To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
)
optimizer_class = bnb.optim.AdamW8bit
else:
optimizer_class = torch.optim.AdamW
# Optimizer creation
params_to_optimize = list(filter(lambda p: p.requires_grad, unet.parameters()))
if args.train_text_encoder:
params_to_optimize = (
params_to_optimize
+ list(filter(lambda p: p.requires_grad, text_encoder_one.parameters()))
+ list(filter(lambda p: p.requires_grad, text_encoder_two.parameters()))
)
optimizer = optimizer_class(
params_to_optimize,
lr=args.learning_rate,
betas=(args.adam_beta1, args.adam_beta2),
weight_decay=args.adam_weight_decay,
eps=args.adam_epsilon,
)
# Get the datasets: you can either provide your own training and evaluation files (see below)
# or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).
# In distributed training, the load_dataset function guarantees that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
dataset = load_dataset(
args.dataset_name, args.dataset_config_name, cache_dir=args.cache_dir, data_dir=args.train_data_dir
)
else:
data_files = {}
if args.train_data_dir is not None:
data_files["train"] = os.path.join(args.train_data_dir, "**")
dataset = load_dataset(
"imagefolder",
data_files=data_files,
cache_dir=args.cache_dir,
)
# See more about loading custom images at
# https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
column_names = dataset["train"].column_names
# 6. Get the column names for input/target.
dataset_columns = DATASET_NAME_MAPPING.get(args.dataset_name, None)
if args.image_column is None:
image_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
else:
image_column = args.image_column
if image_column not in column_names:
raise ValueError(
f"--image_column' value '{args.image_column}' needs to be one of: {', '.join(column_names)}"
)
if args.caption_column is None:
caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
else:
caption_column = args.caption_column
if caption_column not in column_names:
raise ValueError(
f"--caption_column' value '{args.caption_column}' needs to be one of: {', '.join(column_names)}"
)
# Preprocessing the datasets.
# We need to tokenize input captions and transform the images.
def tokenize_captions(examples, is_train=True):
captions = []
for caption in examples[caption_column]:
if isinstance(caption, str):
captions.append(caption)
elif isinstance(caption, (list, np.ndarray)):
# take a random caption if there are multiple
captions.append(random.choice(caption) if is_train else caption[0])
else:
raise ValueError(
f"Caption column `{caption_column}` should contain either strings or lists of strings."
)
tokens_one = tokenize_prompt(tokenizer_one, captions)
tokens_two = tokenize_prompt(tokenizer_two, captions)
return tokens_one, tokens_two
# Preprocessing the datasets.
train_resize = transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR)
train_crop = transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution)
train_flip = transforms.RandomHorizontalFlip(p=1.0)
train_transforms = transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
]
)
def preprocess_train(examples):
images = [image.convert("RGB") for image in examples[image_column]]
# image aug
original_sizes = []
all_images = []
crop_top_lefts = []
for image in images:
original_sizes.append((image.height, image.width))
image = train_resize(image)
if args.random_flip and random.random() < 0.5:
# flip
image = train_flip(image)
if args.center_crop:
y1 = max(0, int(round((image.height - args.resolution) / 2.0)))
x1 = max(0, int(round((image.width - args.resolution) / 2.0)))
image = train_crop(image)
else:
y1, x1, h, w = train_crop.get_params(image, (args.resolution, args.resolution))
image = crop(image, y1, x1, h, w)
crop_top_left = (y1, x1)
crop_top_lefts.append(crop_top_left)
image = train_transforms(image)
all_images.append(image)
examples["original_sizes"] = original_sizes
examples["crop_top_lefts"] = crop_top_lefts
examples["pixel_values"] = all_images
tokens_one, tokens_two = tokenize_captions(examples)
examples["input_ids_one"] = tokens_one
examples["input_ids_two"] = tokens_two
if args.debug_loss:
fnames = [os.path.basename(image.filename) for image in examples[image_column] if image.filename]
if fnames:
examples["filenames"] = fnames
return examples
with accelerator.main_process_first():
if args.max_train_samples is not None:
dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
# Set the training transforms
train_dataset = dataset["train"].with_transform(preprocess_train, output_all_columns=True)
def collate_fn(examples):
pixel_values = torch.stack([example["pixel_values"] for example in examples])
pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
original_sizes = [example["original_sizes"] for example in examples]
crop_top_lefts = [example["crop_top_lefts"] for example in examples]
input_ids_one = torch.stack([example["input_ids_one"] for example in examples])
input_ids_two = torch.stack([example["input_ids_two"] for example in examples])
result = {
"pixel_values": pixel_values,
"input_ids_one": input_ids_one,
"input_ids_two": input_ids_two,
"original_sizes": original_sizes,
"crop_top_lefts": crop_top_lefts,
}
filenames = [example["filenames"] for example in examples if "filenames" in example]
if filenames:
result["filenames"] = filenames
return result
# DataLoaders creation:
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
shuffle=True,
collate_fn=collate_fn,
batch_size=args.train_batch_size,
num_workers=args.dataloader_num_workers,
)
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
overrode_max_train_steps = True
lr_scheduler = get_scheduler(
args.lr_scheduler,
optimizer=optimizer,
num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
)
# Prepare everything with our `accelerator`.
if args.train_text_encoder:
unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler
)
else:
unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
unet, optimizer, train_dataloader, lr_scheduler
)
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# We need to initialize the trackers we use, and also store our configuration.
# The trackers initializes automatically on the main process.
if accelerator.is_main_process:
accelerator.init_trackers("text2image-fine-tune", config=vars(args))
# Train!
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.train_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
global_step = 0
first_epoch = 0
# Potentially load in the weights and states from a previous save
if args.resume_from_checkpoint:
if args.resume_from_checkpoint != "latest":
path = os.path.basename(args.resume_from_checkpoint)
else:
# Get the most recent checkpoint
dirs = os.listdir(args.output_dir)
dirs = [d for d in dirs if d.startswith("checkpoint")]
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
path = dirs[-1] if len(dirs) > 0 else None
if path is None:
accelerator.print(
f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
)
args.resume_from_checkpoint = None
initial_global_step = 0
else:
accelerator.print(f"Resuming from checkpoint {path}")
accelerator.load_state(os.path.join(args.output_dir, path))
global_step = int(path.split("-")[1])
initial_global_step = global_step
first_epoch = global_step // num_update_steps_per_epoch
else:
initial_global_step = 0
progress_bar = tqdm(
range(0, args.max_train_steps),
initial=initial_global_step,
desc="Steps",
# Only show the progress bar once on each machine.
disable=not accelerator.is_local_main_process,
)
for epoch in range(first_epoch, args.num_train_epochs):
unet.train()
if args.train_text_encoder:
text_encoder_one.train()
text_encoder_two.train()
train_loss = 0.0
for step, batch in enumerate(train_dataloader):
with accelerator.accumulate(unet):
# Convert images to latent space
if args.pretrained_vae_model_name_or_path is not None:
pixel_values = batch["pixel_values"].to(dtype=weight_dtype)
else:
pixel_values = batch["pixel_values"]
model_input = vae.encode(pixel_values).latent_dist.sample()
model_input = model_input * vae.config.scaling_factor
if args.pretrained_vae_model_name_or_path is None:
model_input = model_input.to(weight_dtype)
# Sample noise that we'll add to the latents
noise = torch.randn_like(model_input)
if args.noise_offset:
# https://www.crosslabs.org//blog/diffusion-with-offset-noise
noise += args.noise_offset * torch.randn(
(model_input.shape[0], model_input.shape[1], 1, 1), device=model_input.device
)
bsz = model_input.shape[0]
# Sample a random timestep for each image
if args.loss_type == "huber" or args.loss_type == "smooth_l1":
timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (1,), device="cpu")
timestep = timesteps.item()
if args.huber_schedule == "exponential":
alpha = -math.log(args.huber_c) / noise_scheduler.config.num_train_timesteps
huber_c = math.exp(-alpha * timestep)
elif args.huber_schedule == "snr":
alphas_cumprod = noise_scheduler.alphas_cumprod[timestep]
sigmas = ((1.0 - alphas_cumprod) / alphas_cumprod) ** 0.5
huber_c = (1 - args.huber_c) / (1 + sigmas) ** 2 + args.huber_c
elif args.huber_schedule == "constant":
huber_c = args.huber_c
else:
raise NotImplementedError(f"Unknown Huber loss schedule {args.huber_schedule}!")
timesteps = timesteps.repeat(bsz).to(model_input.device)
elif args.loss_type == "l2":
timesteps = torch.randint(
0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
)
huber_c = 1 # may be anything, as it's not used
else:
raise NotImplementedError(f"Unknown loss type {args.loss_type}")
timesteps = timesteps.long()
# Add noise to the model input according to the noise magnitude at each timestep
# (this is the forward diffusion process)
noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
# time ids
def compute_time_ids(original_size, crops_coords_top_left):
# Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
target_size = (args.resolution, args.resolution)
add_time_ids = list(original_size + crops_coords_top_left + target_size)
add_time_ids = torch.tensor([add_time_ids])
add_time_ids = add_time_ids.to(accelerator.device, dtype=weight_dtype)
return add_time_ids
add_time_ids = torch.cat(
[compute_time_ids(s, c) for s, c in zip(batch["original_sizes"], batch["crop_top_lefts"])]
)
# Predict the noise residual
unet_added_conditions = {"time_ids": add_time_ids}
prompt_embeds, pooled_prompt_embeds = encode_prompt(
text_encoders=[text_encoder_one, text_encoder_two],
tokenizers=None,
prompt=None,
text_input_ids_list=[batch["input_ids_one"], batch["input_ids_two"]],
)
unet_added_conditions.update({"text_embeds": pooled_prompt_embeds})
model_pred = unet(
noisy_model_input,
timesteps,
prompt_embeds,
added_cond_kwargs=unet_added_conditions,
return_dict=False,
)[0]
# Get the target for loss depending on the prediction type
if args.prediction_type is not None:
# set prediction_type of scheduler if defined
noise_scheduler.register_to_config(prediction_type=args.prediction_type)
if noise_scheduler.config.prediction_type == "epsilon":
target = noise
elif noise_scheduler.config.prediction_type == "v_prediction":
target = noise_scheduler.get_velocity(model_input, noise, timesteps)
else:
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
if args.snr_gamma is None:
loss = conditional_loss(
model_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c
)
else:
# Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
# Since we predict the noise instead of x_0, the original formulation is slightly changed.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(noise_scheduler, timesteps)
mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
dim=1
)[0]
if noise_scheduler.config.prediction_type == "epsilon":
mse_loss_weights = mse_loss_weights / snr
elif noise_scheduler.config.prediction_type == "v_prediction":
mse_loss_weights = mse_loss_weights / (snr + 1)
loss = conditional_loss(
model_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
)
loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
loss = loss.mean()
if args.debug_loss and "filenames" in batch:
for fname in batch["filenames"]:
accelerator.log({"loss_for_" + fname: loss}, step=global_step)
# Gather the losses across all processes for logging (if we use distributed training).
avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
train_loss += avg_loss.item() / args.gradient_accumulation_steps
# Backpropagate
accelerator.backward(loss)
if accelerator.sync_gradients:
accelerator.clip_grad_norm_(params_to_optimize, args.max_grad_norm)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
# Checks if the accelerator has performed an optimization step behind the scenes
if accelerator.sync_gradients:
progress_bar.update(1)
global_step += 1
accelerator.log({"train_loss": train_loss}, step=global_step)
train_loss = 0.0
if accelerator.is_main_process:
if global_step % args.checkpointing_steps == 0:
# _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
if args.checkpoints_total_limit is not None:
checkpoints = os.listdir(args.output_dir)
checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
# before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
if len(checkpoints) >= args.checkpoints_total_limit:
num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
removing_checkpoints = checkpoints[0:num_to_remove]
logger.info(
f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
)
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
for removing_checkpoint in removing_checkpoints:
removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
shutil.rmtree(removing_checkpoint)
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
accelerator.save_state(save_path)
logger.info(f"Saved state to {save_path}")
logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
progress_bar.set_postfix(**logs)
if global_step >= args.max_train_steps:
break
if accelerator.is_main_process:
if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
logger.info(
f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
f" {args.validation_prompt}."
)
# create pipeline
pipeline = StableDiffusionXLPipeline.from_pretrained(
args.pretrained_model_name_or_path,
vae=vae,
text_encoder=unwrap_model(text_encoder_one),
text_encoder_2=unwrap_model(text_encoder_two),
unet=unwrap_model(unet),
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
)
pipeline = pipeline.to(accelerator.device)
pipeline.set_progress_bar_config(disable=True)
# run inference
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
pipeline_args = {"prompt": args.validation_prompt}
with torch.cuda.amp.autocast():
images = [
pipeline(**pipeline_args, generator=generator).images[0]
for _ in range(args.num_validation_images)
]
for tracker in accelerator.trackers:
if tracker.name == "tensorboard":
np_images = np.stack([np.asarray(img) for img in images])
tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
if tracker.name == "wandb":
tracker.log(
{
"validation": [
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
for i, image in enumerate(images)
]
}
)
del pipeline
torch.cuda.empty_cache()
# Save the lora layers
accelerator.wait_for_everyone()
if accelerator.is_main_process:
unet = unwrap_model(unet)
unet_lora_state_dict = convert_state_dict_to_diffusers(get_peft_model_state_dict(unet))
if args.train_text_encoder:
text_encoder_one = unwrap_model(text_encoder_one)
text_encoder_two = unwrap_model(text_encoder_two)
text_encoder_lora_layers = convert_state_dict_to_diffusers(get_peft_model_state_dict(text_encoder_one))
text_encoder_2_lora_layers = convert_state_dict_to_diffusers(get_peft_model_state_dict(text_encoder_two))
else:
text_encoder_lora_layers = None
text_encoder_2_lora_layers = None
StableDiffusionXLPipeline.save_lora_weights(
save_directory=args.output_dir,
unet_lora_layers=unet_lora_state_dict,
text_encoder_lora_layers=text_encoder_lora_layers,
text_encoder_2_lora_layers=text_encoder_2_lora_layers,
)
del unet
del text_encoder_one
del text_encoder_two
del text_encoder_lora_layers
del text_encoder_2_lora_layers
torch.cuda.empty_cache()
# Final inference
# Make sure vae.dtype is consistent with the unet.dtype
if args.mixed_precision == "fp16":
vae.to(weight_dtype)
# Load previous pipeline
pipeline = StableDiffusionXLPipeline.from_pretrained(
args.pretrained_model_name_or_path,
vae=vae,
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
)
pipeline = pipeline.to(accelerator.device)
# load attention processors
pipeline.load_lora_weights(args.output_dir)
# run inference
images = []
if args.validation_prompt and args.num_validation_images > 0:
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
images = [
pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0]
for _ in range(args.num_validation_images)
]
for tracker in accelerator.trackers:
if tracker.name == "tensorboard":
np_images = np.stack([np.asarray(img) for img in images])
tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
if tracker.name == "wandb":
tracker.log(
{
"test": [
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
for i, image in enumerate(images)
]
}
)
if args.push_to_hub:
save_model_card(
repo_id,
images=images,
base_model=args.pretrained_model_name_or_path,
dataset_name=args.dataset_name,
train_text_encoder=args.train_text_encoder,
repo_folder=args.output_dir,
vae_path=args.pretrained_vae_model_name_or_path,
)
upload_folder(
repo_id=repo_id,
folder_path=args.output_dir,
commit_message="End of training",
ignore_patterns=["step_*", "epoch_*"],
)
accelerator.end_training()
if __name__ == "__main__":
args = parse_args()
main(args)
examples/research_projects/scheduled_huber_loss_training/text_to_image/train_text_to_image_sdxl.py 0 → 100644
View file @ 2f234376
#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Fine-tuning script for Stable Diffusion XL for text2image."""
import argparse
import functools
import gc
import logging
import math
import os
import random
import shutil
from pathlib import Path
import accelerate
import datasets
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import ProjectConfiguration, set_seed
from datasets import concatenate_datasets, load_dataset
from huggingface_hub import create_repo, upload_folder
from packaging import version
from torchvision import transforms
from torchvision.transforms.functional import crop
from tqdm.auto import tqdm
from transformers import AutoTokenizer, PretrainedConfig
import diffusers
from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionXLPipeline, UNet2DConditionModel
from diffusers.optimization import get_scheduler
from diffusers.training_utils import EMAModel, compute_snr
from diffusers.utils import check_min_version, is_wandb_available
from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.torch_utils import is_compiled_module
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.28.0.dev0")
logger = get_logger(__name__)
DATASET_NAME_MAPPING = {
"lambdalabs/pokemon-blip-captions": ("image", "text"),
}
def save_model_card(
repo_id: str,
images: list = None,
validation_prompt: str = None,
base_model: str = None,
dataset_name: str = None,
repo_folder: str = None,
vae_path: str = None,
):
img_str = ""
if images is not None:
for i, image in enumerate(images):
image.save(os.path.join(repo_folder, f"image_{i}.png"))
img_str += f"![img_{i}](./image_{i}.png)\n"
model_description = f"""
# Text-to-image finetuning - {repo_id}
This pipeline was finetuned from **{base_model}** on the **{dataset_name}** dataset. Below are some example images generated with the finetuned pipeline using the following prompt: {validation_prompt}: \n
{img_str}
Special VAE used for training: {vae_path}.
"""
model_card = load_or_create_model_card(
repo_id_or_path=repo_id,
from_training=True,
license="creativeml-openrail-m",
base_model=base_model,
model_description=model_description,
inference=True,
)
tags = [
"stable-diffusion-xl",
"stable-diffusion-xl-diffusers",
"text-to-image",
"diffusers-training",
"diffusers",
]
model_card = populate_model_card(model_card, tags=tags)
model_card.save(os.path.join(repo_folder, "README.md"))
def import_model_class_from_model_name_or_path(
pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
):
text_encoder_config = PretrainedConfig.from_pretrained(
pretrained_model_name_or_path, subfolder=subfolder, revision=revision
)
model_class = text_encoder_config.architectures[0]
if model_class == "CLIPTextModel":
from transformers import CLIPTextModel
return CLIPTextModel
elif model_class == "CLIPTextModelWithProjection":
from transformers import CLIPTextModelWithProjection
return CLIPTextModelWithProjection
else:
raise ValueError(f"{model_class} is not supported.")
def parse_args(input_args=None):
parser = argparse.ArgumentParser(description="Simple example of a training script.")
parser.add_argument(
"--pretrained_model_name_or_path",
type=str,
default=None,
required=True,
help="Path to pretrained model or model identifier from huggingface.co/models.",
)
parser.add_argument(
"--pretrained_vae_model_name_or_path",
type=str,
default=None,
help="Path to pretrained VAE model with better numerical stability. More details: https://github.com/huggingface/diffusers/pull/4038.",
)
parser.add_argument(
"--revision",
type=str,
default=None,
required=False,
help="Revision of pretrained model identifier from huggingface.co/models.",
)
parser.add_argument(
"--variant",
type=str,
default=None,
help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
)
parser.add_argument(
"--dataset_name",
type=str,
default=None,
help=(
"The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
" dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
" or to a folder containing files that 🤗 Datasets can understand."
),
)
parser.add_argument(
"--dataset_config_name",
type=str,
default=None,
help="The config of the Dataset, leave as None if there's only one config.",
)
parser.add_argument(
"--train_data_dir",
type=str,
default=None,
help=(
"A folder containing the training data. Folder contents must follow the structure described in"
" https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
" must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
),
)
parser.add_argument(
"--image_column", type=str, default="image", help="The column of the dataset containing an image."
)
parser.add_argument(
"--caption_column",
type=str,
default="text",
help="The column of the dataset containing a caption or a list of captions.",
)
parser.add_argument(
"--validation_prompt",
type=str,
default=None,
help="A prompt that is used during validation to verify that the model is learning.",
)
parser.add_argument(
"--num_validation_images",
type=int,
default=4,
help="Number of images that should be generated during validation with `validation_prompt`.",
)
parser.add_argument(
"--validation_epochs",
type=int,
default=1,
help=(
"Run fine-tuning validation every X epochs. The validation process consists of running the prompt"
" `args.validation_prompt` multiple times: `args.num_validation_images`."
),
)
parser.add_argument(
"--max_train_samples",
type=int,
default=None,
help=(
"For debugging purposes or quicker training, truncate the number of training examples to this "
"value if set."
),
)
parser.add_argument(
"--proportion_empty_prompts",
type=float,
default=0,
help="Proportion of image prompts to be replaced with empty strings. Defaults to 0 (no prompt replacement).",
)
parser.add_argument(
"--output_dir",
type=str,
default="sdxl-model-finetuned",
help="The output directory where the model predictions and checkpoints will be written.",
)
parser.add_argument(
"--cache_dir",
type=str,
default=None,
help="The directory where the downloaded models and datasets will be stored.",
)
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
parser.add_argument(
"--resolution",
type=int,
default=1024,
help=(
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
" resolution"
),
)
parser.add_argument(
"--center_crop",
default=False,
action="store_true",
help=(
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
" cropped. The images will be resized to the resolution first before cropping."
),
)
parser.add_argument(
"--random_flip",
action="store_true",
help="whether to randomly flip images horizontally",
)
parser.add_argument(
"--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
)
parser.add_argument("--num_train_epochs", type=int, default=100)
parser.add_argument(
"--max_train_steps",
type=int,
default=None,
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
)
parser.add_argument(
"--checkpointing_steps",
type=int,
default=500,
help=(
"Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
" checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
" training using `--resume_from_checkpoint`."
),
)
parser.add_argument(
"--checkpoints_total_limit",
type=int,
default=None,
help=("Max number of checkpoints to store."),
)
parser.add_argument(
"--resume_from_checkpoint",
type=str,
default=None,
help=(
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
),
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--gradient_checkpointing",
action="store_true",
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
)
parser.add_argument(
"--learning_rate",
type=float,
default=1e-4,
help="Initial learning rate (after the potential warmup period) to use.",
)
parser.add_argument(
"--scale_lr",
action="store_true",
default=False,
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
)
parser.add_argument(
"--lr_scheduler",
type=str,
default="constant",
help=(
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
' "constant", "constant_with_warmup"]'
),
)
parser.add_argument(
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
)
parser.add_argument(
"--timestep_bias_strategy",
type=str,
default="none",
choices=["earlier", "later", "range", "none"],
help=(
"The timestep bias strategy, which may help direct the model toward learning low or high frequency details."
" Choices: ['earlier', 'later', 'range', 'none']."
" The default is 'none', which means no bias is applied, and training proceeds normally."
" The value of 'later' will increase the frequency of the model's final training timesteps."
),
)
parser.add_argument(
"--timestep_bias_multiplier",
type=float,
default=1.0,
help=(
"The multiplier for the bias. Defaults to 1.0, which means no bias is applied."
" A value of 2.0 will double the weight of the bias, and a value of 0.5 will halve it."
),
)
parser.add_argument(
"--timestep_bias_begin",
type=int,
default=0,
help=(
"When using `--timestep_bias_strategy=range`, the beginning (inclusive) timestep to bias."
" Defaults to zero, which equates to having no specific bias."
),
)
parser.add_argument(
"--timestep_bias_end",
type=int,
default=1000,
help=(
"When using `--timestep_bias_strategy=range`, the final timestep (inclusive) to bias."
" Defaults to 1000, which is the number of timesteps that Stable Diffusion is trained on."
),
)
parser.add_argument(
"--timestep_bias_portion",
type=float,
default=0.25,
help=(
"The portion of timesteps to bias. Defaults to 0.25, which 25% of timesteps will be biased."
" A value of 0.5 will bias one half of the timesteps. The value provided for `--timestep_bias_strategy` determines"
" whether the biased portions are in the earlier or later timesteps."
),
)
parser.add_argument(
"--snr_gamma",
type=float,
default=None,
help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
"More details here: https://arxiv.org/abs/2303.09556.",
)
parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.")
parser.add_argument(
"--allow_tf32",
action="store_true",
help=(
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
),
)
parser.add_argument(
"--dataloader_num_workers",
type=int,
default=0,
help=(
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
),
)
parser.add_argument(
"--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
)
parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
parser.add_argument(
"--prediction_type",
type=str,
default=None,
help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediction_type` is chosen.",
)
parser.add_argument(
"--hub_model_id",
type=str,
default=None,
help="The name of the repository to keep in sync with the local `output_dir`.",
)
parser.add_argument(
"--logging_dir",
type=str,
default="logs",
help=(
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
),
)
parser.add_argument(
"--report_to",
type=str,
default="tensorboard",
help=(
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
),
)
parser.add_argument(
"--mixed_precision",
type=str,
default=None,
choices=["no", "fp16", "bf16"],
help=(
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
),
)
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
parser.add_argument(
"--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
)
parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
parser.add_argument(
"--loss_type",
type=str,
default="l2",
choices=["l2", "huber", "smooth_l1"],
help="The type of loss to use and whether it's timestep-scheduled. See Issue #7488 for more info.",
)
parser.add_argument(
"--huber_schedule",
type=str,
default="snr",
choices=["constant", "exponential", "snr"],
help="The schedule to use for the huber losses parameter",
)
parser.add_argument(
"--huber_c",
type=float,
default=0.1,
help="The huber loss parameter. Only used if one of the huber loss modes (huber or smooth l1) is selected with loss_type.",
)
if input_args is not None:
args = parser.parse_args(input_args)
else:
args = parser.parse_args()
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
if env_local_rank != -1 and env_local_rank != args.local_rank:
args.local_rank = env_local_rank
# Sanity checks
if args.dataset_name is None and args.train_data_dir is None:
raise ValueError("Need either a dataset name or a training folder.")
if args.proportion_empty_prompts < 0 or args.proportion_empty_prompts > 1:
raise ValueError("`--proportion_empty_prompts` must be in the range [0, 1].")
return args
# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
def encode_prompt(batch, text_encoders, tokenizers, proportion_empty_prompts, caption_column, is_train=True):
prompt_embeds_list = []
prompt_batch = batch[caption_column]
captions = []
for caption in prompt_batch:
if random.random() < proportion_empty_prompts:
captions.append("")
elif isinstance(caption, str):
captions.append(caption)
elif isinstance(caption, (list, np.ndarray)):
# take a random caption if there are multiple
captions.append(random.choice(caption) if is_train else caption[0])
with torch.no_grad():
for tokenizer, text_encoder in zip(tokenizers, text_encoders):
text_inputs = tokenizer(
captions,
padding="max_length",
max_length=tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
prompt_embeds = text_encoder(
text_input_ids.to(text_encoder.device),
output_hidden_states=True,
return_dict=False,
)
# We are only ALWAYS interested in the pooled output of the final text encoder
pooled_prompt_embeds = prompt_embeds[0]
prompt_embeds = prompt_embeds[-1][-2]
bs_embed, seq_len, _ = prompt_embeds.shape
prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
prompt_embeds_list.append(prompt_embeds)
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
return {"prompt_embeds": prompt_embeds.cpu(), "pooled_prompt_embeds": pooled_prompt_embeds.cpu()}
def compute_vae_encodings(batch, vae):
images = batch.pop("pixel_values")
pixel_values = torch.stack(list(images))
pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
pixel_values = pixel_values.to(vae.device, dtype=vae.dtype)
with torch.no_grad():
model_input = vae.encode(pixel_values).latent_dist.sample()
model_input = model_input * vae.config.scaling_factor
return {"model_input": model_input.cpu()}
def generate_timestep_weights(args, num_timesteps):
weights = torch.ones(num_timesteps)
# Determine the indices to bias
num_to_bias = int(args.timestep_bias_portion * num_timesteps)
if args.timestep_bias_strategy == "later":
bias_indices = slice(-num_to_bias, None)
elif args.timestep_bias_strategy == "earlier":
bias_indices = slice(0, num_to_bias)
elif args.timestep_bias_strategy == "range":
# Out of the possible 1000 timesteps, we might want to focus on eg. 200-500.
range_begin = args.timestep_bias_begin
range_end = args.timestep_bias_end
if range_begin < 0:
raise ValueError(
"When using the range strategy for timestep bias, you must provide a beginning timestep greater or equal to zero."
)
if range_end > num_timesteps:
raise ValueError(
"When using the range strategy for timestep bias, you must provide an ending timestep smaller than the number of timesteps."
)
bias_indices = slice(range_begin, range_end)
else: # 'none' or any other string
return weights
if args.timestep_bias_multiplier <= 0:
return ValueError(
"The parameter --timestep_bias_multiplier is not intended to be used to disable the training of specific timesteps."
" If it was intended to disable timestep bias, use `--timestep_bias_strategy none` instead."
" A timestep bias multiplier less than or equal to 0 is not allowed."
)
# Apply the bias
weights[bias_indices] *= args.timestep_bias_multiplier
# Normalize
weights /= weights.sum()
return weights
# NOTE: if you're using the scheduled version, huber_c has to depend on the timesteps already
def conditional_loss(
model_pred: torch.Tensor,
target: torch.Tensor,
reduction: str = "mean",
loss_type: str = "l2",
huber_c: float = 0.1,
):
if loss_type == "l2":
loss = F.mse_loss(model_pred, target, reduction=reduction)
elif loss_type == "huber":
loss = 2 * huber_c * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
elif loss_type == "smooth_l1":
loss = 2 * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
if reduction == "mean":
loss = torch.mean(loss)
elif reduction == "sum":
loss = torch.sum(loss)
else:
raise NotImplementedError(f"Unsupported Loss Type {loss_type}")
return loss
def main(args):
if args.report_to == "wandb" and args.hub_token is not None:
raise ValueError(
"You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
" Please use `huggingface-cli login` to authenticate with the Hub."
)
logging_dir = Path(args.output_dir, args.logging_dir)
accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
accelerator = Accelerator(
gradient_accumulation_steps=args.gradient_accumulation_steps,
mixed_precision=args.mixed_precision,
log_with=args.report_to,
project_config=accelerator_project_config,
)
if args.report_to == "wandb":
if not is_wandb_available():
raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
import wandb
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state, main_process_only=False)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_warning()
diffusers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
diffusers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
if args.push_to_hub:
repo_id = create_repo(
repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
).repo_id
# Load the tokenizers
tokenizer_one = AutoTokenizer.from_pretrained(
args.pretrained_model_name_or_path,
subfolder="tokenizer",
revision=args.revision,
use_fast=False,
)
tokenizer_two = AutoTokenizer.from_pretrained(
args.pretrained_model_name_or_path,
subfolder="tokenizer_2",
revision=args.revision,
use_fast=False,
)
# import correct text encoder classes
text_encoder_cls_one = import_model_class_from_model_name_or_path(
args.pretrained_model_name_or_path, args.revision
)
text_encoder_cls_two = import_model_class_from_model_name_or_path(
args.pretrained_model_name_or_path, args.revision, subfolder="text_encoder_2"
)
# Load scheduler and models
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
# Check for terminal SNR in combination with SNR Gamma
text_encoder_one = text_encoder_cls_one.from_pretrained(
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
)
text_encoder_two = text_encoder_cls_two.from_pretrained(
args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
)
vae_path = (
args.pretrained_model_name_or_path
if args.pretrained_vae_model_name_or_path is None
else args.pretrained_vae_model_name_or_path
)
vae = AutoencoderKL.from_pretrained(
vae_path,
subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
revision=args.revision,
variant=args.variant,
)
unet = UNet2DConditionModel.from_pretrained(
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
)
# Freeze vae and text encoders.
vae.requires_grad_(False)
text_encoder_one.requires_grad_(False)
text_encoder_two.requires_grad_(False)
# Set unet as trainable.
unet.train()
# For mixed precision training we cast all non-trainable weights to half-precision
# as these weights are only used for inference, keeping weights in full precision is not required.
weight_dtype = torch.float32
if accelerator.mixed_precision == "fp16":
weight_dtype = torch.float16
elif accelerator.mixed_precision == "bf16":
weight_dtype = torch.bfloat16
# Move unet, vae and text_encoder to device and cast to weight_dtype
# The VAE is in float32 to avoid NaN losses.
vae.to(accelerator.device, dtype=torch.float32)
text_encoder_one.to(accelerator.device, dtype=weight_dtype)
text_encoder_two.to(accelerator.device, dtype=weight_dtype)
# Create EMA for the unet.
if args.use_ema:
ema_unet = UNet2DConditionModel.from_pretrained(
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
)
ema_unet = EMAModel(ema_unet.parameters(), model_cls=UNet2DConditionModel, model_config=ema_unet.config)
if args.enable_xformers_memory_efficient_attention:
if is_xformers_available():
import xformers
xformers_version = version.parse(xformers.__version__)
if xformers_version == version.parse("0.0.16"):
logger.warning(
"xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
)
unet.enable_xformers_memory_efficient_attention()
else:
raise ValueError("xformers is not available. Make sure it is installed correctly")
# `accelerate` 0.16.0 will have better support for customized saving
if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
# create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
def save_model_hook(models, weights, output_dir):
if accelerator.is_main_process:
if args.use_ema:
ema_unet.save_pretrained(os.path.join(output_dir, "unet_ema"))
for i, model in enumerate(models):
model.save_pretrained(os.path.join(output_dir, "unet"))
# make sure to pop weight so that corresponding model is not saved again
weights.pop()
def load_model_hook(models, input_dir):
if args.use_ema:
load_model = EMAModel.from_pretrained(os.path.join(input_dir, "unet_ema"), UNet2DConditionModel)
ema_unet.load_state_dict(load_model.state_dict())
ema_unet.to(accelerator.device)
del load_model
for _ in range(len(models)):
# pop models so that they are not loaded again
model = models.pop()
# load diffusers style into model
load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
model.register_to_config(**load_model.config)
model.load_state_dict(load_model.state_dict())
del load_model
accelerator.register_save_state_pre_hook(save_model_hook)
accelerator.register_load_state_pre_hook(load_model_hook)
if args.gradient_checkpointing:
unet.enable_gradient_checkpointing()
# Enable TF32 for faster training on Ampere GPUs,
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
if args.allow_tf32:
torch.backends.cuda.matmul.allow_tf32 = True
if args.scale_lr:
args.learning_rate = (
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
)
# Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
if args.use_8bit_adam:
try:
import bitsandbytes as bnb
except ImportError:
raise ImportError(
"To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
)
optimizer_class = bnb.optim.AdamW8bit
else:
optimizer_class = torch.optim.AdamW
# Optimizer creation
params_to_optimize = unet.parameters()
optimizer = optimizer_class(
params_to_optimize,
lr=args.learning_rate,
betas=(args.adam_beta1, args.adam_beta2),
weight_decay=args.adam_weight_decay,
eps=args.adam_epsilon,
)
# Get the datasets: you can either provide your own training and evaluation files (see below)
# or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).
# In distributed training, the load_dataset function guarantees that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
dataset = load_dataset(
args.dataset_name,
args.dataset_config_name,
cache_dir=args.cache_dir,
)
else:
data_files = {}
if args.train_data_dir is not None:
data_files["train"] = os.path.join(args.train_data_dir, "**")
dataset = load_dataset(
"imagefolder",
data_files=data_files,
cache_dir=args.cache_dir,
)
# See more about loading custom images at
# https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
column_names = dataset["train"].column_names
# 6. Get the column names for input/target.
dataset_columns = DATASET_NAME_MAPPING.get(args.dataset_name, None)
if args.image_column is None:
image_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
else:
image_column = args.image_column
if image_column not in column_names:
raise ValueError(
f"--image_column' value '{args.image_column}' needs to be one of: {', '.join(column_names)}"
)
if args.caption_column is None:
caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
else:
caption_column = args.caption_column
if caption_column not in column_names:
raise ValueError(
f"--caption_column' value '{args.caption_column}' needs to be one of: {', '.join(column_names)}"
)
# Preprocessing the datasets.
train_resize = transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR)
train_crop = transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution)
train_flip = transforms.RandomHorizontalFlip(p=1.0)
train_transforms = transforms.Compose([transforms.ToTensor(), transforms.Normalize([0.5], [0.5])])
def preprocess_train(examples):
images = [image.convert("RGB") for image in examples[image_column]]
# image aug
original_sizes = []
all_images = []
crop_top_lefts = []
for image in images:
original_sizes.append((image.height, image.width))
image = train_resize(image)
if args.random_flip and random.random() < 0.5:
# flip
image = train_flip(image)
if args.center_crop:
y1 = max(0, int(round((image.height - args.resolution) / 2.0)))
x1 = max(0, int(round((image.width - args.resolution) / 2.0)))
image = train_crop(image)
else:
y1, x1, h, w = train_crop.get_params(image, (args.resolution, args.resolution))
image = crop(image, y1, x1, h, w)
crop_top_left = (y1, x1)
crop_top_lefts.append(crop_top_left)
image = train_transforms(image)
all_images.append(image)
examples["original_sizes"] = original_sizes
examples["crop_top_lefts"] = crop_top_lefts
examples["pixel_values"] = all_images
return examples
with accelerator.main_process_first():
if args.max_train_samples is not None:
dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
# Set the training transforms
train_dataset = dataset["train"].with_transform(preprocess_train)
# Let's first compute all the embeddings so that we can free up the text encoders
# from memory. We will pre-compute the VAE encodings too.
text_encoders = [text_encoder_one, text_encoder_two]
tokenizers = [tokenizer_one, tokenizer_two]
compute_embeddings_fn = functools.partial(
encode_prompt,
text_encoders=text_encoders,
tokenizers=tokenizers,
proportion_empty_prompts=args.proportion_empty_prompts,
caption_column=args.caption_column,
)
compute_vae_encodings_fn = functools.partial(compute_vae_encodings, vae=vae)
with accelerator.main_process_first():
from datasets.fingerprint import Hasher
# fingerprint used by the cache for the other processes to load the result
# details: https://github.com/huggingface/diffusers/pull/4038#discussion_r1266078401
new_fingerprint = Hasher.hash(args)
new_fingerprint_for_vae = Hasher.hash(vae_path)
train_dataset_with_embeddings = train_dataset.map(
compute_embeddings_fn, batched=True, new_fingerprint=new_fingerprint
)
train_dataset_with_vae = train_dataset.map(
compute_vae_encodings_fn,
batched=True,
batch_size=args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps,
new_fingerprint=new_fingerprint_for_vae,
)
precomputed_dataset = concatenate_datasets(
[train_dataset_with_embeddings, train_dataset_with_vae.remove_columns(["image", "text"])], axis=1
)
precomputed_dataset = precomputed_dataset.with_transform(preprocess_train)
del compute_vae_encodings_fn, compute_embeddings_fn, text_encoder_one, text_encoder_two
del text_encoders, tokenizers, vae
gc.collect()
torch.cuda.empty_cache()
def collate_fn(examples):
model_input = torch.stack([torch.tensor(example["model_input"]) for example in examples])
original_sizes = [example["original_sizes"] for example in examples]
crop_top_lefts = [example["crop_top_lefts"] for example in examples]
prompt_embeds = torch.stack([torch.tensor(example["prompt_embeds"]) for example in examples])
pooled_prompt_embeds = torch.stack([torch.tensor(example["pooled_prompt_embeds"]) for example in examples])
return {
"model_input": model_input,
"prompt_embeds": prompt_embeds,
"pooled_prompt_embeds": pooled_prompt_embeds,
"original_sizes": original_sizes,
"crop_top_lefts": crop_top_lefts,
}
# DataLoaders creation:
train_dataloader = torch.utils.data.DataLoader(
precomputed_dataset,
shuffle=True,
collate_fn=collate_fn,
batch_size=args.train_batch_size,
num_workers=args.dataloader_num_workers,
)
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
overrode_max_train_steps = True
lr_scheduler = get_scheduler(
args.lr_scheduler,
optimizer=optimizer,
num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
)
# Prepare everything with our `accelerator`.
unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
unet, optimizer, train_dataloader, lr_scheduler
)
if args.use_ema:
ema_unet.to(accelerator.device)
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# We need to initialize the trackers we use, and also store our configuration.
# The trackers initializes automatically on the main process.
if accelerator.is_main_process:
accelerator.init_trackers("text2image-fine-tune-sdxl", config=vars(args))
# Function for unwrapping if torch.compile() was used in accelerate.
def unwrap_model(model):
model = accelerator.unwrap_model(model)
model = model._orig_mod if is_compiled_module(model) else model
return model
# Train!
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(precomputed_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.train_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
global_step = 0
first_epoch = 0
# Potentially load in the weights and states from a previous save
if args.resume_from_checkpoint:
if args.resume_from_checkpoint != "latest":
path = os.path.basename(args.resume_from_checkpoint)
else:
# Get the most recent checkpoint
dirs = os.listdir(args.output_dir)
dirs = [d for d in dirs if d.startswith("checkpoint")]
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
path = dirs[-1] if len(dirs) > 0 else None
if path is None:
accelerator.print(
f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
)
args.resume_from_checkpoint = None
initial_global_step = 0
else:
accelerator.print(f"Resuming from checkpoint {path}")
accelerator.load_state(os.path.join(args.output_dir, path))
global_step = int(path.split("-")[1])
initial_global_step = global_step
first_epoch = global_step // num_update_steps_per_epoch
else:
initial_global_step = 0
progress_bar = tqdm(
range(0, args.max_train_steps),
initial=initial_global_step,
desc="Steps",
# Only show the progress bar once on each machine.
disable=not accelerator.is_local_main_process,
)
for epoch in range(first_epoch, args.num_train_epochs):
train_loss = 0.0
for step, batch in enumerate(train_dataloader):
with accelerator.accumulate(unet):
# Sample noise that we'll add to the latents
model_input = batch["model_input"].to(accelerator.device)
noise = torch.randn_like(model_input)
if args.noise_offset:
# https://www.crosslabs.org//blog/diffusion-with-offset-noise
noise += args.noise_offset * torch.randn(
(model_input.shape[0], model_input.shape[1], 1, 1), device=model_input.device
)
bsz = model_input.shape[0]
if args.timestep_bias_strategy == "none":
# Sample a random timestep for each image
if args.loss_type == "huber" or args.loss_type == "smooth_l1":
timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (1,), device="cpu")
timestep = timesteps.item()
if args.huber_schedule == "exponential":
alpha = -math.log(args.huber_c) / noise_scheduler.config.num_train_timesteps
huber_c = math.exp(-alpha * timestep)
elif args.huber_schedule == "snr":
alphas_cumprod = noise_scheduler.alphas_cumprod[timestep]
sigmas = ((1.0 - alphas_cumprod) / alphas_cumprod) ** 0.5
huber_c = (1 - args.huber_c) / (1 + sigmas) ** 2 + args.huber_c
elif args.huber_schedule == "constant":
huber_c = args.huber_c
else:
raise NotImplementedError(f"Unknown Huber loss schedule {args.huber_schedule}!")
timesteps = timesteps.repeat(bsz).to(model_input.device)
elif args.loss_type == "l2":
timesteps = torch.randint(
0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
)
huber_c = 1 # may be anything, as it's not used
else:
raise NotImplementedError(f"Unknown loss type {args.loss_type}")
timesteps = timesteps.long()
else:
if "huber_scheduled" in args.loss_type:
raise NotImplementedError(
"Randomly weighted timesteps not implemented yet for scheduled huber loss!"
)
else:
huber_c = args.huber_c
# Sample a random timestep for each image, potentially biased by the timestep weights.
# Biasing the timestep weights allows us to spend less time training irrelevant timesteps.
weights = generate_timestep_weights(args, noise_scheduler.config.num_train_timesteps).to(
model_input.device
)
timesteps = torch.multinomial(weights, bsz, replacement=True).long()
# Add noise to the model input according to the noise magnitude at each timestep
# (this is the forward diffusion process)
noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
# time ids
def compute_time_ids(original_size, crops_coords_top_left):
# Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
target_size = (args.resolution, args.resolution)
add_time_ids = list(original_size + crops_coords_top_left + target_size)
add_time_ids = torch.tensor([add_time_ids])
add_time_ids = add_time_ids.to(accelerator.device, dtype=weight_dtype)
return add_time_ids
add_time_ids = torch.cat(
[compute_time_ids(s, c) for s, c in zip(batch["original_sizes"], batch["crop_top_lefts"])]
)
# Predict the noise residual
unet_added_conditions = {"time_ids": add_time_ids}
prompt_embeds = batch["prompt_embeds"].to(accelerator.device)
pooled_prompt_embeds = batch["pooled_prompt_embeds"].to(accelerator.device)
unet_added_conditions.update({"text_embeds": pooled_prompt_embeds})
model_pred = unet(
noisy_model_input,
timesteps,
prompt_embeds,
added_cond_kwargs=unet_added_conditions,
return_dict=False,
)[0]
# Get the target for loss depending on the prediction type
if args.prediction_type is not None:
# set prediction_type of scheduler if defined
noise_scheduler.register_to_config(prediction_type=args.prediction_type)
if noise_scheduler.config.prediction_type == "epsilon":
target = noise
elif noise_scheduler.config.prediction_type == "v_prediction":
target = noise_scheduler.get_velocity(model_input, noise, timesteps)
elif noise_scheduler.config.prediction_type == "sample":
# We set the target to latents here, but the model_pred will return the noise sample prediction.
target = model_input
# We will have to subtract the noise residual from the prediction to get the target sample.
model_pred = model_pred - noise
else:
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
if args.snr_gamma is None:
loss = conditional_loss(
model_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c
)
else:
# Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
# Since we predict the noise instead of x_0, the original formulation is slightly changed.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(noise_scheduler, timesteps)
mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
dim=1
)[0]
if noise_scheduler.config.prediction_type == "epsilon":
mse_loss_weights = mse_loss_weights / snr
elif noise_scheduler.config.prediction_type == "v_prediction":
mse_loss_weights = mse_loss_weights / (snr + 1)
loss = conditional_loss(
model_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
)
loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
loss = loss.mean()
# Gather the losses across all processes for logging (if we use distributed training).
avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
train_loss += avg_loss.item() / args.gradient_accumulation_steps
# Backpropagate
accelerator.backward(loss)
if accelerator.sync_gradients:
params_to_clip = unet.parameters()
accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
# Checks if the accelerator has performed an optimization step behind the scenes
if accelerator.sync_gradients:
if args.use_ema:
ema_unet.step(unet.parameters())
progress_bar.update(1)
global_step += 1
accelerator.log({"train_loss": train_loss}, step=global_step)
train_loss = 0.0
if accelerator.is_main_process:
if global_step % args.checkpointing_steps == 0:
# _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
if args.checkpoints_total_limit is not None:
checkpoints = os.listdir(args.output_dir)
checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
# before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
if len(checkpoints) >= args.checkpoints_total_limit:
num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
removing_checkpoints = checkpoints[0:num_to_remove]
logger.info(
f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
)
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
for removing_checkpoint in removing_checkpoints:
removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
shutil.rmtree(removing_checkpoint)
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
accelerator.save_state(save_path)
logger.info(f"Saved state to {save_path}")
logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
progress_bar.set_postfix(**logs)
if global_step >= args.max_train_steps:
break
if accelerator.is_main_process:
if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
logger.info(
f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
f" {args.validation_prompt}."
)
if args.use_ema:
# Store the UNet parameters temporarily and load the EMA parameters to perform inference.
ema_unet.store(unet.parameters())
ema_unet.copy_to(unet.parameters())
# create pipeline
vae = AutoencoderKL.from_pretrained(
vae_path,
subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
revision=args.revision,
variant=args.variant,
)
pipeline = StableDiffusionXLPipeline.from_pretrained(
args.pretrained_model_name_or_path,
vae=vae,
unet=accelerator.unwrap_model(unet),
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
)
if args.prediction_type is not None:
scheduler_args = {"prediction_type": args.prediction_type}
pipeline.scheduler = pipeline.scheduler.from_config(pipeline.scheduler.config, **scheduler_args)
pipeline = pipeline.to(accelerator.device)
pipeline.set_progress_bar_config(disable=True)
# run inference
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
pipeline_args = {"prompt": args.validation_prompt}
with torch.cuda.amp.autocast():
images = [
pipeline(**pipeline_args, generator=generator, num_inference_steps=25).images[0]
for _ in range(args.num_validation_images)
]
for tracker in accelerator.trackers:
if tracker.name == "tensorboard":
np_images = np.stack([np.asarray(img) for img in images])
tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
if tracker.name == "wandb":
tracker.log(
{
"validation": [
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
for i, image in enumerate(images)
]
}
)
del pipeline
torch.cuda.empty_cache()
accelerator.wait_for_everyone()
if accelerator.is_main_process:
unet = unwrap_model(unet)
if args.use_ema:
ema_unet.copy_to(unet.parameters())
# Serialize pipeline.
vae = AutoencoderKL.from_pretrained(
vae_path,
subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
)
pipeline = StableDiffusionXLPipeline.from_pretrained(
args.pretrained_model_name_or_path,
unet=unet,
vae=vae,
revision=args.revision,
variant=args.variant,
torch_dtype=weight_dtype,
)
if args.prediction_type is not None:
scheduler_args = {"prediction_type": args.prediction_type}
pipeline.scheduler = pipeline.scheduler.from_config(pipeline.scheduler.config, **scheduler_args)
pipeline.save_pretrained(args.output_dir)
# run inference
images = []
if args.validation_prompt and args.num_validation_images > 0:
pipeline = pipeline.to(accelerator.device)
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
with torch.cuda.amp.autocast():
images = [
pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0]
for _ in range(args.num_validation_images)
]
for tracker in accelerator.trackers:
if tracker.name == "tensorboard":
np_images = np.stack([np.asarray(img) for img in images])
tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
if tracker.name == "wandb":
tracker.log(
{
"test": [
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
for i, image in enumerate(images)
]
}
)
if args.push_to_hub:
save_model_card(
repo_id=repo_id,
images=images,
validation_prompt=args.validation_prompt,
base_model=args.pretrained_model_name_or_path,
dataset_name=args.dataset_name,
repo_folder=args.output_dir,
vae_path=args.pretrained_vae_model_name_or_path,
)
upload_folder(
repo_id=repo_id,
folder_path=args.output_dir,
commit_message="End of training",
ignore_patterns=["step_*", "epoch_*"],
)
accelerator.end_training()
if __name__ == "__main__":
args = parse_args()
main(args)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment