[Scheduler] Move predict epsilon to init (#1155)

* [Scheduler] Move predict epsilon to init

* up

* uP

* uP

* Apply suggestions from code review
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* up
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

examples/unconditional_image_generation/train_unconditional.py

 import argparse
+import inspect
 import math
 import os
 from pathlib import Path
@@ -190,10 +191,10 @@ def parse_args():
     )
 
     parser.add_argument(
-        "--predict_mode",
-        type=str,
-        default="eps",
-        help="What the model should predict. 'eps' to predict error, 'x0' to directly predict reconstruction",
+        "--predict_epsilon",
+        action="store_true",
+        default=True,
+        help="Whether the model should predict the 'epsilon'/noise error or directly the reconstructed image 'x0'.",
     )
 
     parser.add_argument("--ddpm_num_steps", type=int, default=1000)
@@ -252,7 +253,17 @@ def main(args):
             "UpBlock2D",
         ),
     )
+    accepts_predict_epsilon = "predict_epsilon" in set(inspect.signature(DDPMScheduler.__init__).parameters.keys())
+
+    if accepts_predict_epsilon:
+        noise_scheduler = DDPMScheduler(
+            num_train_timesteps=args.ddpm_num_steps,
+            beta_schedule=args.ddpm_beta_schedule,
+            predict_epsilon=args.predict_epsilon,
+        )
+    else:
         noise_scheduler = DDPMScheduler(num_train_timesteps=args.ddpm_num_steps, beta_schedule=args.ddpm_beta_schedule)
+
     optimizer = torch.optim.AdamW(
         model.parameters(),
         lr=args.learning_rate,
@@ -351,9 +362,9 @@ def main(args):
                 # Predict the noise residual
                 model_output = model(noisy_images, timesteps).sample
 
-                if args.predict_mode == "eps":
+                if args.predict_epsilon:
                     loss = F.mse_loss(model_output, noise)  # this could have different weights!
-                elif args.predict_mode == "x0":
+                else:
                     alpha_t = _extract_into_tensor(
                         noise_scheduler.alphas_cumprod, timesteps, (clean_images.shape[0], 1, 1, 1)
                     )
@@ -401,7 +412,6 @@ def main(args):
                     generator=generator,
                     batch_size=args.eval_batch_size,
                     output_type="numpy",
-                    predict_epsilon=args.predict_mode == "eps",
                 ).images
 
                 # denormalize the images and save to tensorboard

src/diffusers/configuration_utils.py

@@ -334,6 +334,11 @@ class ConfigMixin:
         # 3. Create keyword arguments that will be passed to __init__ from expected keyword arguments
         init_dict = {}
         for key in expected_keys:
+            # if config param is passed to kwarg and is present in config dict
+            # it should overwrite existing config dict key
+            if key in kwargs and key in config_dict:
+                config_dict[key] = kwargs.pop(key)
+
             if key in kwargs:
                 # overwrite key
                 init_dict[key] = kwargs.pop(key)

src/diffusers/pipelines/ddpm/pipeline_ddpm.py

@@ -18,7 +18,9 @@ from typing import Optional, Tuple, Union
 
 import torch
 
+from ...configuration_utils import FrozenDict
 from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ...utils import deprecate
 
 
 class DDPMPipeline(DiffusionPipeline):
@@ -45,7 +47,6 @@ class DDPMPipeline(DiffusionPipeline):
         num_inference_steps: int = 1000,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
-        predict_epsilon: bool = True,
         **kwargs,
     ) -> Union[ImagePipelineOutput, Tuple]:
         r"""
@@ -69,6 +70,16 @@ class DDPMPipeline(DiffusionPipeline):
             `return_dict` is True, otherwise a `tuple. When returning a tuple, the first element is a list with the
             generated images.
         """
+        message = (
+            "Please make sure to instantiate your scheduler with `predict_epsilon` instead. E.g. `scheduler ="
+            " DDPMScheduler.from_config(<model_id>, predict_epsilon=True)`."
+        )
+        predict_epsilon = deprecate("predict_epsilon", "0.10.0", message, take_from=kwargs)
+
+        if predict_epsilon is not None:
+            new_config = dict(self.scheduler.config)
+            new_config["predict_epsilon"] = predict_epsilon
+            self.scheduler._internal_dict = FrozenDict(new_config)
 
         # Sample gaussian noise to begin loop
         image = torch.randn(

src/diffusers/schedulers/scheduling_ddpm.py

@@ -21,8 +21,8 @@ from typing import Optional, Tuple, Union
 import numpy as np
 import torch
 
-from ..configuration_utils import ConfigMixin, register_to_config
-from ..utils import BaseOutput
+from ..configuration_utils import ConfigMixin, FrozenDict, register_to_config
+from ..utils import BaseOutput, deprecate
 from .scheduling_utils import SchedulerMixin
 
 
@@ -99,6 +99,8 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
             `fixed_small_log`, `fixed_large`, `fixed_large_log`, `learned` or `learned_range`.
         clip_sample (`bool`, default `True`):
             option to clip predicted sample between -1 and 1 for numerical stability.
+        predict_epsilon (`bool`):
+            optional flag to use when the model predicts the noise (epsilon), or the samples instead of the noise.
 
     """
 
@@ -121,6 +123,7 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
         trained_betas: Optional[np.ndarray] = None,
         variance_type: str = "fixed_small",
         clip_sample: bool = True,
+        predict_epsilon: bool = True,
     ):
         if trained_betas is not None:
             self.betas = torch.from_numpy(trained_betas)
@@ -221,9 +224,9 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
         model_output: torch.FloatTensor,
         timestep: int,
         sample: torch.FloatTensor,
-        predict_epsilon=True,
         generator=None,
         return_dict: bool = True,
+        **kwargs,
     ) -> Union[DDPMSchedulerOutput, Tuple]:
         """
         Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
@@ -234,8 +237,6 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
             timestep (`int`): current discrete timestep in the diffusion chain.
             sample (`torch.FloatTensor`):
                 current instance of sample being created by diffusion process.
-            predict_epsilon (`bool`):
-                optional flag to use when model predicts the samples directly instead of the noise, epsilon.
             generator: random number generator.
             return_dict (`bool`): option for returning tuple rather than DDPMSchedulerOutput class
 
@@ -245,6 +246,16 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
             returning a tuple, the first element is the sample tensor.
 
         """
+        message = (
+            "Please make sure to instantiate your scheduler with `predict_epsilon` instead. E.g. `scheduler ="
+            " DDPMScheduler.from_config(<model_id>, predict_epsilon=True)`."
+        )
+        predict_epsilon = deprecate("predict_epsilon", "0.10.0", message, take_from=kwargs)
+        if predict_epsilon is not None and predict_epsilon != self.config.predict_epsilon:
+            new_config = dict(self.config)
+            new_config["predict_epsilon"] = predict_epsilon
+            self._internal_dict = FrozenDict(new_config)
+
         t = timestep
 
         if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in ["learned", "learned_range"]:
@@ -260,7 +271,7 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
 
         # 2. compute predicted original sample from predicted noise also called
         # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
-        if predict_epsilon:
+        if self.config.predict_epsilon:
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
         else:
             pred_original_sample = model_output

src/diffusers/schedulers/scheduling_ddpm_flax.py

@@ -22,7 +22,8 @@ import flax
 import jax.numpy as jnp
 from jax import random
 
-from ..configuration_utils import ConfigMixin, register_to_config
+from ..configuration_utils import ConfigMixin, FrozenDict, register_to_config
+from ..utils import deprecate
 from .scheduling_utils_flax import FlaxSchedulerMixin, FlaxSchedulerOutput, broadcast_to_shape_from_left
 
 
@@ -97,7 +98,8 @@ class FlaxDDPMScheduler(FlaxSchedulerMixin, ConfigMixin):
             `fixed_small_log`, `fixed_large`, `fixed_large_log`, `learned` or `learned_range`.
         clip_sample (`bool`, default `True`):
             option to clip predicted sample between -1 and 1 for numerical stability.
-        tensor_format (`str`): whether the scheduler expects pytorch or numpy arrays.
+        predict_epsilon (`bool`):
+            optional flag to use when the model predicts the noise (epsilon), or the samples instead of the noise.
 
     """
 
@@ -115,6 +117,7 @@ class FlaxDDPMScheduler(FlaxSchedulerMixin, ConfigMixin):
         trained_betas: Optional[jnp.ndarray] = None,
         variance_type: str = "fixed_small",
         clip_sample: bool = True,
+        predict_epsilon: bool = True,
     ):
         if trained_betas is not None:
             self.betas = jnp.asarray(trained_betas)
@@ -196,6 +199,7 @@ class FlaxDDPMScheduler(FlaxSchedulerMixin, ConfigMixin):
         key: random.KeyArray,
         predict_epsilon: bool = True,
         return_dict: bool = True,
+        **kwargs,
     ) -> Union[FlaxDDPMSchedulerOutput, Tuple]:
         """
         Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
@@ -208,8 +212,6 @@ class FlaxDDPMScheduler(FlaxSchedulerMixin, ConfigMixin):
             sample (`jnp.ndarray`):
                 current instance of sample being created by diffusion process.
             key (`random.KeyArray`): a PRNG key.
-            predict_epsilon (`bool`):
-                optional flag to use when model predicts the samples directly instead of the noise, epsilon.
             return_dict (`bool`): option for returning tuple rather than FlaxDDPMSchedulerOutput class
 
         Returns:
@@ -217,6 +219,16 @@ class FlaxDDPMScheduler(FlaxSchedulerMixin, ConfigMixin):
             `tuple`. When returning a tuple, the first element is the sample tensor.
 
         """
+        message = (
+            "Please make sure to instantiate your scheduler with `predict_epsilon` instead. E.g. `scheduler ="
+            " DDPMScheduler.from_config(<model_id>, predict_epsilon=True)`."
+        )
+        predict_epsilon = deprecate("predict_epsilon", "0.10.0", message, take_from=kwargs)
+        if predict_epsilon is not None and predict_epsilon != self.config.predict_epsilon:
+            new_config = dict(self.config)
+            new_config["predict_epsilon"] = predict_epsilon
+            self._internal_dict = FrozenDict(new_config)
+
         t = timestep
 
         if model_output.shape[1] == sample.shape[1] * 2 and self.config.variance_type in ["learned", "learned_range"]:
@@ -232,7 +244,7 @@ class FlaxDDPMScheduler(FlaxSchedulerMixin, ConfigMixin):
 
         # 2. compute predicted original sample from predicted noise also called
         # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
-        if predict_epsilon:
+        if self.config.predict_epsilon:
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
         else:
             pred_original_sample = model_output

tests/fixtures/custom_pipeline/pipeline.py

@@ -42,7 +42,6 @@ class CustomLocalPipeline(DiffusionPipeline):
         self,
         batch_size: int = 1,
         generator: Optional[torch.Generator] = None,
-        eta: float = 0.0,
         num_inference_steps: int = 50,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
@@ -89,7 +88,7 @@ class CustomLocalPipeline(DiffusionPipeline):
             # 2. predict previous mean of image x_t-1 and add variance depending on eta
             # eta corresponds to η in paper and should be between [0, 1]
             # do x_t -> x_t-1
-            image = self.scheduler.step(model_output, t, image, eta).prev_sample
+            image = self.scheduler.step(model_output, t, image).prev_sample
 
         image = (image / 2 + 0.5).clamp(0, 1)
         image = image.cpu().permute(0, 2, 3, 1).numpy()

tests/pipelines/ddpm/test_ddpm.py

@@ -19,6 +19,7 @@ import numpy as np
 import torch
 
 from diffusers import DDPMPipeline, DDPMScheduler, UNet2DModel
+from diffusers.utils import deprecate
 from diffusers.utils.testing_utils import require_torch, slow, torch_device
 
 from ...test_pipelines_common import PipelineTesterMixin
@@ -28,8 +29,74 @@ torch.backends.cuda.matmul.allow_tf32 = False
 
 
 class DDPMPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
-    # FIXME: add fast tests
-    pass
+    @property
+    def dummy_uncond_unet(self):
+        torch.manual_seed(0)
+        model = UNet2DModel(
+            block_out_channels=(32, 64),
+            layers_per_block=2,
+            sample_size=32,
+            in_channels=3,
+            out_channels=3,
+            down_block_types=("DownBlock2D", "AttnDownBlock2D"),
+            up_block_types=("AttnUpBlock2D", "UpBlock2D"),
+        )
+        return model
+
+    def test_inference(self):
+        unet = self.dummy_uncond_unet
+        scheduler = DDPMScheduler()
+
+        ddpm = DDPMPipeline(unet=unet, scheduler=scheduler)
+        ddpm.to(torch_device)
+        ddpm.set_progress_bar_config(disable=None)
+
+        # Warmup pass when using mps (see #372)
+        if torch_device == "mps":
+            _ = ddpm(num_inference_steps=1)
+
+        generator = torch.manual_seed(0)
+        image = ddpm(generator=generator, num_inference_steps=2, output_type="numpy").images
+
+        generator = torch.manual_seed(0)
+        image_from_tuple = ddpm(generator=generator, num_inference_steps=2, output_type="numpy", return_dict=False)[0]
+
+        image_slice = image[0, -3:, -3:, -1]
+        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]
+
+        assert image.shape == (1, 32, 32, 3)
+        expected_slice = np.array(
+            [5.589e-01, 7.089e-01, 2.632e-01, 6.841e-01, 1.000e-04, 9.999e-01, 1.973e-01, 1.000e-04, 8.010e-02]
+        )
+        tolerance = 1e-2 if torch_device != "mps" else 3e-2
+        assert np.abs(image_slice.flatten() - expected_slice).max() < tolerance
+        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < tolerance
+
+    def test_inference_predict_epsilon(self):
+        deprecate("remove this test", "0.10.0", "remove")
+        unet = self.dummy_uncond_unet
+        scheduler = DDPMScheduler(predict_epsilon=False)
+
+        ddpm = DDPMPipeline(unet=unet, scheduler=scheduler)
+        ddpm.to(torch_device)
+        ddpm.set_progress_bar_config(disable=None)
+
+        # Warmup pass when using mps (see #372)
+        if torch_device == "mps":
+            _ = ddpm(num_inference_steps=1)
+
+        generator = torch.manual_seed(0)
+        image = ddpm(generator=generator, num_inference_steps=2, output_type="numpy").images
+
+        generator = torch.manual_seed(0)
+        image_eps = ddpm(generator=generator, num_inference_steps=2, output_type="numpy", predict_epsilon=False)[0]
+
+        image_slice = image[0, -3:, -3:, -1]
+        image_eps_slice = image_eps[0, -3:, -3:, -1]
+
+        assert image.shape == (1, 32, 32, 3)
+        tolerance = 1e-2 if torch_device != "mps" else 3e-2
+        assert np.abs(image_slice.flatten() - image_eps_slice.flatten()).max() < tolerance
 
 
 @slow

tests/test_config.py

@@ -21,6 +21,7 @@ import unittest
 import diffusers
 from diffusers import (
     DDIMScheduler,
+    DDPMScheduler,
     DPMSolverMultistepScheduler,
     EulerAncestralDiscreteScheduler,
     EulerDiscreteScheduler,
@@ -291,6 +292,29 @@ class ConfigTester(unittest.TestCase):
         # no warning should be thrown
         assert cap_logger.out == ""
 
+    def test_overwrite_config_on_load(self):
+        logger = logging.get_logger("diffusers.configuration_utils")
+
+        with CaptureLogger(logger) as cap_logger:
+            ddpm = DDPMScheduler.from_config(
+                "hf-internal-testing/tiny-stable-diffusion-torch",
+                subfolder="scheduler",
+                predict_epsilon=False,
+                beta_end=8,
+            )
+
+        with CaptureLogger(logger) as cap_logger_2:
+            ddpm_2 = DDPMScheduler.from_config("google/ddpm-celebahq-256", beta_start=88)
+
+        assert ddpm.__class__ == DDPMScheduler
+        assert ddpm.config.predict_epsilon is False
+        assert ddpm.config.beta_end == 8
+        assert ddpm_2.config.beta_start == 88
+
+        # no warning should be thrown
+        assert cap_logger.out == ""
+        assert cap_logger_2.out == ""
+
     def test_load_dpmsolver(self):
         logger = logging.get_logger("diffusers.configuration_utils")
 

tests/test_pipelines.py

@@ -107,6 +107,7 @@ class CustomPipelineTests(unittest.TestCase):
         images, output_str = pipeline(num_inference_steps=2, output_type="np")
 
         assert images[0].shape == (1, 32, 32, 3)
+
         # compare output to https://huggingface.co/hf-internal-testing/diffusers-dummy-pipeline/blob/main/pipeline.py#L102
         assert output_str == "This is a test"
 

tests/test_scheduler.py

@@ -33,7 +33,7 @@ from diffusers import (
     ScoreSdeVeScheduler,
     VQDiffusionScheduler,
 )
-from diffusers.utils import torch_device
+from diffusers.utils import deprecate, torch_device
 
 
 torch.backends.cuda.matmul.allow_tf32 = False
@@ -393,6 +393,34 @@ class DDPMSchedulerTest(SchedulerCommonTest):
         for clip_sample in [True, False]:
             self.check_over_configs(clip_sample=clip_sample)
 
+    def test_predict_epsilon(self):
+        for predict_epsilon in [True, False]:
+            self.check_over_configs(predict_epsilon=predict_epsilon)
+
+    def test_deprecated_epsilon(self):
+        deprecate("remove this test", "0.10.0", "remove")
+        scheduler_class = self.scheduler_classes[0]
+        scheduler_config = self.get_scheduler_config()
+
+        sample = self.dummy_sample_deter
+        residual = 0.1 * self.dummy_sample_deter
+        time_step = 4
+
+        scheduler = scheduler_class(**scheduler_config)
+        scheduler_eps = scheduler_class(predict_epsilon=False, **scheduler_config)
+
+        kwargs = {}
+        if "generator" in set(inspect.signature(scheduler.step).parameters.keys()):
+            kwargs["generator"] = torch.Generator().manual_seed(0)
+        output = scheduler.step(residual, time_step, sample, predict_epsilon=False, **kwargs).prev_sample
+
+        kwargs = {}
+        if "generator" in set(inspect.signature(scheduler.step).parameters.keys()):
+            kwargs["generator"] = torch.Generator().manual_seed(0)
+        output_eps = scheduler_eps.step(residual, time_step, sample, predict_epsilon=False, **kwargs).prev_sample
+
+        assert (output - output_eps).abs().sum() < 1e-5
+
     def test_time_indices(self):
         for t in [0, 500, 999]:
             self.check_over_forward(time_step=t)