autoencoder_kl.py

# Copyright 2023 The HuggingFace 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.
from dataclasses import dataclass
from typing import Dict, Optional, Tuple, Union

import torch
import torch.nn as nn

from ..configuration_utils import ConfigMixin, register_to_config
from ..utils import BaseOutput, apply_forward_hook
from .attention_processor import AttentionProcessor, AttnProcessor
from .modeling_utils import ModelMixin
from .vae import Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder


@dataclass
class AutoencoderKLOutput(BaseOutput):
    """
    Output of AutoencoderKL encoding method.

    Args:
        latent_dist (`DiagonalGaussianDistribution`):
            Encoded outputs of `Encoder` represented as the mean and logvar of `DiagonalGaussianDistribution`.
            `DiagonalGaussianDistribution` allows for sampling latents from the distribution.
    """

    latent_dist: "DiagonalGaussianDistribution"


class AutoencoderKL(ModelMixin, ConfigMixin):
    r"""Variational Autoencoder (VAE) model with KL loss from the paper Auto-Encoding Variational Bayes by Diederik P. Kingma
    and Max Welling.

    This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
    implements for all the model (such as downloading or saving, etc.)

    Parameters:
        in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
        out_channels (int,  *optional*, defaults to 3): Number of channels in the output.
        down_block_types (`Tuple[str]`, *optional*, defaults to :
            obj:`("DownEncoderBlock2D",)`): Tuple of downsample block types.
        up_block_types (`Tuple[str]`, *optional*, defaults to :
            obj:`("UpDecoderBlock2D",)`): Tuple of upsample block types.
        block_out_channels (`Tuple[int]`, *optional*, defaults to :
            obj:`(64,)`): Tuple of block output channels.
        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
        latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
        sample_size (`int`, *optional*, defaults to `32`): TODO
        scaling_factor (`float`, *optional*, defaults to 0.18215):
            The component-wise standard deviation of the trained latent space computed using the first batch of the
            training set. This is used to scale the latent space to have unit variance when training the diffusion
            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
    """

    _supports_gradient_checkpointing = True

    @register_to_config
    def __init__(
        self,
        in_channels: int = 3,
        out_channels: int = 3,
        down_block_types: Tuple[str] = ("DownEncoderBlock2D",),
        up_block_types: Tuple[str] = ("UpDecoderBlock2D",),
        block_out_channels: Tuple[int] = (64,),
        layers_per_block: int = 1,
        act_fn: str = "silu",
        latent_channels: int = 4,
        norm_num_groups: int = 32,
        sample_size: int = 32,
        scaling_factor: float = 0.18215,
    ):
        super().__init__()

        # pass init params to Encoder
        self.encoder = Encoder(
            in_channels=in_channels,
            out_channels=latent_channels,
            down_block_types=down_block_types,
            block_out_channels=block_out_channels,
            layers_per_block=layers_per_block,
            act_fn=act_fn,
            norm_num_groups=norm_num_groups,
            double_z=True,
        )

        # pass init params to Decoder
        self.decoder = Decoder(
            in_channels=latent_channels,
            out_channels=out_channels,
            up_block_types=up_block_types,
            block_out_channels=block_out_channels,
            layers_per_block=layers_per_block,
            norm_num_groups=norm_num_groups,
            act_fn=act_fn,
        )

        self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1)
        self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1)

        self.use_slicing = False
        self.use_tiling = False

        # only relevant if vae tiling is enabled
        self.tile_sample_min_size = self.config.sample_size
        sample_size = (
            self.config.sample_size[0]
            if isinstance(self.config.sample_size, (list, tuple))
            else self.config.sample_size
        )
        self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
        self.tile_overlap_factor = 0.25

    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, (Encoder, Decoder)):
            module.gradient_checkpointing = value

    def enable_tiling(self, use_tiling: bool = True):
        r"""
        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
        compute decoding and encoding in several steps. This is useful to save a large amount of memory and to allow
        the processing of larger images.
        """
        self.use_tiling = use_tiling

    def disable_tiling(self):
        r"""
        Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
        computing decoding in one step.
        """
        self.enable_tiling(False)

    def enable_slicing(self):
        r"""
        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
        """
        self.use_slicing = True

    def disable_slicing(self):
        r"""
        Disable sliced VAE decoding. If `enable_slicing` was previously invoked, this method will go back to computing
        decoding in one step.
        """
        self.use_slicing = False

    @property
    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
    def attn_processors(self) -> Dict[str, AttentionProcessor]:
        r"""
        Returns:
            `dict` of attention processors: A dictionary containing all attention processors used in the model with
            indexed by its weight name.
        """
        # set recursively
        processors = {}

        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
            if hasattr(module, "set_processor"):
                processors[f"{name}.processor"] = module.processor

            for sub_name, child in module.named_children():
                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)

            return processors

        for name, module in self.named_children():
            fn_recursive_add_processors(name, module, processors)

        return processors

    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
        r"""
        Parameters:
            `processor (`dict` of `AttentionProcessor` or `AttentionProcessor`):
                The instantiated processor class or a dictionary of processor classes that will be set as the processor
                of **all** `Attention` layers.
            In case `processor` is a dict, the key needs to define the path to the corresponding cross attention processor. This is strongly recommended when setting trainable attention processors.:

        """
        count = len(self.attn_processors.keys())

        if isinstance(processor, dict) and len(processor) != count:
            raise ValueError(
                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
            )

        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
            if hasattr(module, "set_processor"):
                if not isinstance(processor, dict):
                    module.set_processor(processor)
                else:
                    module.set_processor(processor.pop(f"{name}.processor"))

            for sub_name, child in module.named_children():
                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

        for name, module in self.named_children():
            fn_recursive_attn_processor(name, module, processor)

    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
    def set_default_attn_processor(self):
        """
        Disables custom attention processors and sets the default attention implementation.
        """
        self.set_attn_processor(AttnProcessor())

    @apply_forward_hook
    def encode(self, x: torch.FloatTensor, return_dict: bool = True) -> AutoencoderKLOutput:
        if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size):
            return self.tiled_encode(x, return_dict=return_dict)

        h = self.encoder(x)
        moments = self.quant_conv(h)
        posterior = DiagonalGaussianDistribution(moments)

        if not return_dict:
            return (posterior,)

        return AutoencoderKLOutput(latent_dist=posterior)

    def _decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
        if self.use_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size):
            return self.tiled_decode(z, return_dict=return_dict)

        z = self.post_quant_conv(z)
        dec = self.decoder(z)

        if not return_dict:
            return (dec,)

        return DecoderOutput(sample=dec)

    @apply_forward_hook
    def decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
        if self.use_slicing and z.shape[0] > 1:
            decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)]
            decoded = torch.cat(decoded_slices)
        else:
            decoded = self._decode(z).sample

        if not return_dict:
            return (decoded,)

        return DecoderOutput(sample=decoded)

    def blend_v(self, a, b, blend_extent):
        blend_extent = min(a.shape[2], b.shape[2], blend_extent)
        for y in range(blend_extent):
            b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent)
        return b

    def blend_h(self, a, b, blend_extent):
        blend_extent = min(a.shape[3], b.shape[3], blend_extent)
        for x in range(blend_extent):
            b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent)
        return b

    def tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True) -> AutoencoderKLOutput:
        r"""Encode a batch of images using a tiled encoder.

        Args:
        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
        steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is:
        different from non-tiled encoding due to each tile using a different encoder. To avoid tiling artifacts, the
        tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
        look of the output, but they should be much less noticeable.
            x (`torch.FloatTensor`): Input batch of images. return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`AutoencoderKLOutput`] instead of a plain tuple.
        """
        overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
        blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
        row_limit = self.tile_latent_min_size - blend_extent

        # Split the image into 512x512 tiles and encode them separately.
        rows = []
        for i in range(0, x.shape[2], overlap_size):
            row = []
            for j in range(0, x.shape[3], overlap_size):
                tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
                tile = self.encoder(tile)
                tile = self.quant_conv(tile)
                row.append(tile)
            rows.append(row)
        result_rows = []
        for i, row in enumerate(rows):
            result_row = []
            for j, tile in enumerate(row):
                # blend the above tile and the left tile
                # to the current tile and add the current tile to the result row
                if i > 0:
                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
                if j > 0:
                    tile = self.blend_h(row[j - 1], tile, blend_extent)
                result_row.append(tile[:, :, :row_limit, :row_limit])
            result_rows.append(torch.cat(result_row, dim=3))

        moments = torch.cat(result_rows, dim=2)
        posterior = DiagonalGaussianDistribution(moments)

        if not return_dict:
            return (posterior,)

        return AutoencoderKLOutput(latent_dist=posterior)

    def tiled_decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
        r"""Decode a batch of images using a tiled decoder.

        Args:
        When this option is enabled, the VAE will split the input tensor into tiles to compute decoding in several
        steps. This is useful to keep memory use constant regardless of image size. The end result of tiled decoding is:
        different from non-tiled decoding due to each tile using a different decoder. To avoid tiling artifacts, the
        tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
        look of the output, but they should be much less noticeable.
            z (`torch.FloatTensor`): Input batch of latent vectors. return_dict (`bool`, *optional*, defaults to
            `True`):
                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
        """
        overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor))
        blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor)
        row_limit = self.tile_sample_min_size - blend_extent

        # Split z into overlapping 64x64 tiles and decode them separately.
        # The tiles have an overlap to avoid seams between tiles.
        rows = []
        for i in range(0, z.shape[2], overlap_size):
            row = []
            for j in range(0, z.shape[3], overlap_size):
                tile = z[:, :, i : i + self.tile_latent_min_size, j : j + self.tile_latent_min_size]
                tile = self.post_quant_conv(tile)
                decoded = self.decoder(tile)
                row.append(decoded)
            rows.append(row)
        result_rows = []
        for i, row in enumerate(rows):
            result_row = []
            for j, tile in enumerate(row):
                # blend the above tile and the left tile
                # to the current tile and add the current tile to the result row
                if i > 0:
                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
                if j > 0:
                    tile = self.blend_h(row[j - 1], tile, blend_extent)
                result_row.append(tile[:, :, :row_limit, :row_limit])
            result_rows.append(torch.cat(result_row, dim=3))

        dec = torch.cat(result_rows, dim=2)
        if not return_dict:
            return (dec,)

        return DecoderOutput(sample=dec)

    def forward(
        self,
        sample: torch.FloatTensor,
        sample_posterior: bool = False,
        return_dict: bool = True,
        generator: Optional[torch.Generator] = None,
    ) -> Union[DecoderOutput, torch.FloatTensor]:
        r"""
        Args:
            sample (`torch.FloatTensor`): Input sample.
            sample_posterior (`bool`, *optional*, defaults to `False`):
                Whether to sample from the posterior.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
        """
        x = sample
        posterior = self.encode(x).latent_dist
        if sample_posterior:
            z = posterior.sample(generator=generator)
        else:
            z = posterior.mode()
        dec = self.decode(z).sample

        if not return_dict:
            return (dec,)

        return DecoderOutput(sample=dec)