save intermediate

models/vision/ddim/modeling_ddim.py

@@ -55,46 +55,58 @@ class DDIM(DiffusionPipeline):
         for t in tqdm.tqdm(reversed(range(num_inference_steps)), total=num_inference_steps):
             # 1. predict noise residual
             with torch.no_grad():
-                pred_noise_t = self.unet(image, inference_step_times[t])
+                residual = self.unet(image, inference_step_times[t])
 
-            # 2. get actual t and t-1
-            train_step = inference_step_times[t]
-            prev_train_step = inference_step_times[t - 1] if t > 0 else -1
+            # 2. predict previous mean of image x_t-1
+            pred_prev_image = self.noise_scheduler.predict_prev_image_step(residual, image, t, num_inference_steps, eta)
 
-            # 3. compute alphas, betas
-            alpha_prod_t = self.noise_scheduler.get_alpha_prod(train_step)
-            alpha_prod_t_prev = self.noise_scheduler.get_alpha_prod(prev_train_step)
-            beta_prod_t = 1 - alpha_prod_t
-            beta_prod_t_prev = 1 - alpha_prod_t_prev
+            # 3. optionally sample variance
+            variance = 0
+            if eta > 0:
+                noise = self.noise_scheduler.sample_noise(image.shape, device=image.device, generator=generator)
+                variance = self.noise_scheduler.get_variance(t).sqrt() * eta * noise
 
+            # 4. set current image to prev_image: x_t -> x_t-1
+            image = pred_prev_image + variance
+
+            # 2. get actual t and t-1
+#            train_step = inference_step_times[t]
+#            prev_train_step = inference_step_times[t - 1] if t > 0 else -1
+#
+            # 3. compute alphas, betas
+#            alpha_prod_t = self.noise_scheduler.get_alpha_prod(train_step)
+#            alpha_prod_t_prev = self.noise_scheduler.get_alpha_prod(prev_train_step)
+#            beta_prod_t = 1 - alpha_prod_t
+#            beta_prod_t_prev = 1 - alpha_prod_t_prev
+#
             # 4. Compute predicted previous image from predicted noise
             # First: compute predicted original image from predicted noise also called
             # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
-            pred_original_image = (image - beta_prod_t.sqrt() * pred_noise_t) / alpha_prod_t.sqrt()
-
+#            pred_original_image = (image - beta_prod_t.sqrt() * pred_noise_t) / alpha_prod_t.sqrt()
+#
             # Second: Clip "predicted x_0"
-            pred_original_image = torch.clamp(pred_original_image, -1, 1)
-
+#            pred_original_image = torch.clamp(pred_original_image, -1, 1)
+#
             # Third: Compute variance: "sigma_t(η)" -> see formula (16)
             # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
-            std_dev_t = (beta_prod_t_prev / beta_prod_t).sqrt() * (1 - alpha_prod_t / alpha_prod_t_prev).sqrt()
-            std_dev_t = eta * std_dev_t
-
+#            std_dev_t = (beta_prod_t_prev / beta_prod_t).sqrt() * (1 - alpha_prod_t / alpha_prod_t_prev).sqrt()
+#            std_dev_t = eta * std_dev_t
+#
             # Fourth: Compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
-            pred_image_direction = (1 - alpha_prod_t_prev - std_dev_t**2).sqrt() * pred_noise_t
-
+#            pred_image_direction = (1 - alpha_prod_t_prev - std_dev_t**2).sqrt() * pred_noise_t
+#
             # Fifth: Compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
-            pred_prev_image = alpha_prod_t_prev.sqrt() * pred_original_image + pred_image_direction
-
+#            pred_prev_image = alpha_prod_t_prev.sqrt() * pred_original_image + pred_image_direction
+#
             # 5. Sample x_t-1 image optionally if η > 0.0 by adding noise to pred_prev_image
             # Note: eta = 1.0 essentially corresponds to DDPM
-            if eta > 0.0:
-                noise = self.noise_scheduler.sample_noise(image.shape, device=image.device, generator=generator)
-                prev_image = pred_prev_image + std_dev_t * noise
-            else:
-                prev_image = pred_prev_image
-
+#            if eta > 0.0:
+#                noise = self.noise_scheduler.sample_noise(image.shape, device=image.device, generator=generator)
+#                prev_image = pred_prev_image + std_dev_t * noise
+#            else:
+#                prev_image = pred_prev_image
+#
             # 6. Set current image to prev_image: x_t -> x_t-1
-            image = prev_image
+#            image = prev_image
 
         return image

models/vision/ddpm/modeling_ddpm.py

@@ -38,50 +38,23 @@ class DDPM(DiffusionPipeline):
             generator=generator,
         )
 
-        for t in tqdm.tqdm(reversed(range(len(self.noise_scheduler))), total=len(self.noise_scheduler)):
+        num_prediction_steps = len(self.noise_scheduler)
+
+        for t in tqdm.tqdm(reversed(range(num_prediction_steps)), total=num_prediction_steps):
             # 1. predict noise residual
             with torch.no_grad():
-                pred_noise_t = self.unet(image, t)
-
-            # 2. compute alphas, betas
-            alpha_prod_t = self.noise_scheduler.get_alpha_prod(t)
-            alpha_prod_t_prev = self.noise_scheduler.get_alpha_prod(t - 1)
-            beta_prod_t = 1 - alpha_prod_t
-            beta_prod_t_prev = 1 - alpha_prod_t_prev
-
-            # 3. compute predicted image from residual
-            # First: compute predicted original image from predicted noise also called
-            # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
-            pred_original_image = (image - beta_prod_t.sqrt() * pred_noise_t) / alpha_prod_t.sqrt()
-
-            # Second: Clip "predicted x_0"
-            pred_original_image = torch.clamp(pred_original_image, -1, 1)
+                residual = self.unet(image, t)
 
-            # Third: Compute coefficients for pred_original_image x_0 and current image x_t
-            # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
-            pred_original_image_coeff = (alpha_prod_t_prev.sqrt() * self.noise_scheduler.get_beta(t)) / beta_prod_t
-            current_image_coeff = self.noise_scheduler.get_alpha(t).sqrt() * beta_prod_t_prev / beta_prod_t
-            # Fourth: Compute predicted previous image µ_t
-            # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
-            pred_prev_image = pred_original_image_coeff * pred_original_image + current_image_coeff * image
+            # 2. predict previous mean of image x_t-1
+            pred_prev_image = self.noise_scheduler.predict_prev_image_step(residual, image, t)
 
-            # 5. For t > 0, compute predicted variance βt (see formala (6) and (7) from https://arxiv.org/pdf/2006.11239.pdf)
-            # and sample from it to get previous image
-            # x_{t-1} ~ N(pred_prev_image, variance) == add variane to pred_image
+            # 3. optionally sample variance
+            variance = 0
             if t > 0:
-                variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * self.noise_scheduler.get_beta(t).sqrt()
-                # TODO(PVP):
-                # This variance seems to be good enough for inference - check if those `fix_small`, `fix_large`
-                # are really only needed for training or also for inference
-                # Also note LDM only uses "fixed_small";
-                # glide seems to use a weird mix of the two: https://github.com/openai/glide-text2im/blob/69b530740eb6cef69442d6180579ef5ba9ef063e/glide_text2im/gaussian_diffusion.py#L246
                 noise = self.noise_scheduler.sample_noise(image.shape, device=image.device, generator=generator)
-                sampled_variance = variance * noise
-                prev_image = pred_prev_image + sampled_variance
-            else:
-                prev_image = pred_prev_image
+                variance = self.noise_scheduler.get_variance(t) * noise
 
-            # 6. Set current image to prev_image: x_t -> x_t-1
-            image = prev_image
+            # 4. set current image to prev_image: x_t -> x_t-1
+            image = pred_prev_image + variance
 
         return image

src/diffusers/__init__.py

@@ -11,4 +11,5 @@ from .models.unet_ldm import UNetLDMModel
 from .pipeline_utils import DiffusionPipeline
 from .schedulers.classifier_free_guidance import ClassifierFreeGuidanceScheduler
 from .schedulers.gaussian_ddpm import GaussianDDPMScheduler
+from .schedulers.ddim import DDIMScheduler
 from .schedulers.glide_ddim import GlideDDIMScheduler

src/diffusers/schedulers/__init__.py

@@ -18,4 +18,5 @@
 
 from .classifier_free_guidance import ClassifierFreeGuidanceScheduler
 from .gaussian_ddpm import GaussianDDPMScheduler
+from .ddim import DDIMScheduler
 from .glide_ddim import GlideDDIMScheduler

src/diffusers/schedulers/ddim.py

+# Copyright 2022 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.
+import math
+
+import torch
+from torch import nn
+
+from ..configuration_utils import ConfigMixin
+from .schedulers_utils import betas_for_alpha_bar, linear_beta_schedule
+
+
+SAMPLING_CONFIG_NAME = "scheduler_config.json"
+
+
+class DDIMScheduler(nn.Module, ConfigMixin):
+
+    config_name = SAMPLING_CONFIG_NAME
+
+    def __init__(
+        self,
+        timesteps=1000,
+        beta_start=0.0001,
+        beta_end=0.02,
+        beta_schedule="linear",
+        clip_predicted_image=True,
+    ):
+        super().__init__()
+        self.register(
+            timesteps=timesteps,
+            beta_start=beta_start,
+            beta_end=beta_end,
+            beta_schedule=beta_schedule,
+        )
+        self.num_timesteps = int(timesteps)
+        self.clip_image = clip_predicted_image
+
+        if beta_schedule == "linear":
+            betas = linear_beta_schedule(timesteps, beta_start=beta_start, beta_end=beta_end)
+        elif beta_schedule == "squaredcos_cap_v2":
+            # GLIDE cosine schedule
+            betas = betas_for_alpha_bar(
+                timesteps,
+                lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2,
+            )
+        else:
+            raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+        alphas = 1.0 - betas
+        alphas_cumprod = torch.cumprod(alphas, axis=0)
+
+        self.register_buffer("betas", betas.to(torch.float32))
+        self.register_buffer("alphas", alphas.to(torch.float32))
+        self.register_buffer("alphas_cumprod", alphas_cumprod.to(torch.float32))
+
+#        alphas_cumprod_prev = torch.nn.functional.pad(alphas_cumprod[:-1], (1, 0), value=1.0)
+        # TODO(PVP) - check how much of these is actually necessary!
+        # LDM only uses "fixed_small"; glide seems to use a weird mix of the two, ...
+        # https://github.com/openai/glide-text2im/blob/69b530740eb6cef69442d6180579ef5ba9ef063e/glide_text2im/gaussian_diffusion.py#L246
+#        variance = betas * (1.0 - alphas_cumprod_prev) / (1.0 - alphas_cumprod)
+#        if variance_type == "fixed_small":
+#            log_variance = torch.log(variance.clamp(min=1e-20))
+#        elif variance_type == "fixed_large":
+#            log_variance = torch.log(torch.cat([variance[1:2], betas[1:]], dim=0))
+#
+#
+#        self.register_buffer("log_variance", log_variance.to(torch.float32))
+
+    def get_alpha(self, time_step):
+        return self.alphas[time_step]
+
+    def get_beta(self, time_step):
+        return self.betas[time_step]
+
+    def get_alpha_prod(self, time_step):
+        if time_step < 0:
+            return torch.tensor(1.0)
+        return self.alphas_cumprod[time_step]
+
+    def get_variance(self, t, num_inference_steps):
+        orig_t = (self.num_timesteps // num_inference_steps) * t
+        orig_prev_t = (self.num_timesteps // num_inference_steps) * (t - 1) if t > 0 else -1
+
+        alpha_prod_t = self.get_alpha_prod(orig_t)
+        alpha_prod_t_prev = self.get_alpha_prod(orig_prev_t)
+        beta_prod_t = 1 - alpha_prod_t
+        beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+        variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
+
+        return variance
+
+    def predict_prev_image_step(self, residual, image, t, num_inference_steps, eta, output_pred_x_0=False):
+        # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+        # Ideally, read DDIM paper in-detail understanding
+
+        # Notation (<variable name> -> <name in paper>
+        # - pred_noise_t -> e_theta(x_t, t)
+        # - pred_original_image -> f_theta(x_t, t) or x_0
+        # - std_dev_t -> sigma_t
+        # - eta -> η
+        # - pred_image_direction -> "direction pointingc to x_t"
+        # - pred_prev_image -> "x_t-1"
+
+        # 1. get actual t and t-1
+        orig_t = (self.num_timesteps // num_inference_steps) * t
+        orig_prev_t = (self.num_timesteps // num_inference_steps) * (t - 1) if t > 0 else -1
+#        train_step = inference_step_times[t]
+#        prev_train_step = inference_step_times[t - 1] if t > 0 else -1
+
+        # 2. compute alphas, betas
+        alpha_prod_t = self.get_alpha_prod(orig_t)
+        alpha_prod_t_prev = self.get_alpha_prod(orig_prev_t)
+        beta_prod_t = 1 - alpha_prod_t
+
+        # 3. compute predicted original image from predicted noise also called
+        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        pred_original_image = (image - beta_prod_t.sqrt() * residual) / alpha_prod_t.sqrt()
+
+        # 4. Clip "predicted x_0"
+        if self.clip_image:
+            pred_original_image = torch.clamp(pred_original_image, -1, 1)
+
+        # 5. compute variance: "sigma_t(η)" -> see formula (16)
+        # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
+        variance = self.get_variance(t, num_inference_steps)
+        std_dev_t = eta * variance.sqrt()
+
+        # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        pred_image_direction = (1 - alpha_prod_t_prev - std_dev_t**2).sqrt() * residual
+
+        # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        pred_prev_image = alpha_prod_t_prev.sqrt() * pred_original_image + pred_image_direction
+
+        return pred_prev_image
+
+    def sample_noise(self, shape, device, generator=None):
+        # always sample on CPU to be deterministic
+        return torch.randn(shape, generator=generator).to(device)
+
+    def __len__(self):
+        return self.num_timesteps

src/diffusers/schedulers/gaussian_ddpm.py

@@ -34,6 +34,7 @@ class GaussianDDPMScheduler(nn.Module, ConfigMixin):
         beta_end=0.02,
         beta_schedule="linear",
         variance_type="fixed_small",
+        clip_predicted_image=True,
     ):
         super().__init__()
         self.register(
@@ -42,8 +43,10 @@ class GaussianDDPMScheduler(nn.Module, ConfigMixin):
             beta_end=beta_end,
             beta_schedule=beta_schedule,
             variance_type=variance_type,
+            clip_predicted_image=clip_predicted_image,
         )
         self.num_timesteps = int(timesteps)
+        self.clip_image = clip_predicted_image
 
         if beta_schedule == "linear":
             betas = linear_beta_schedule(timesteps, beta_start=beta_start, beta_end=beta_end)
@@ -58,23 +61,23 @@ class GaussianDDPMScheduler(nn.Module, ConfigMixin):
 
         alphas = 1.0 - betas
         alphas_cumprod = torch.cumprod(alphas, axis=0)
-        alphas_cumprod_prev = torch.nn.functional.pad(alphas_cumprod[:-1], (1, 0), value=1.0)
-
-        variance = betas * (1.0 - alphas_cumprod_prev) / (1.0 - alphas_cumprod)
-
-        # TODO(PVP) - check how much of these is actually necessary!
-        # LDM only uses "fixed_small"; glide seems to use a weird mix of the two, ...
-        # https://github.com/openai/glide-text2im/blob/69b530740eb6cef69442d6180579ef5ba9ef063e/glide_text2im/gaussian_diffusion.py#L246
-        if variance_type == "fixed_small":
-            log_variance = torch.log(variance.clamp(min=1e-20))
-        elif variance_type == "fixed_large":
-            log_variance = torch.log(torch.cat([variance[1:2], betas[1:]], dim=0))
 
         self.register_buffer("betas", betas.to(torch.float32))
         self.register_buffer("alphas", alphas.to(torch.float32))
         self.register_buffer("alphas_cumprod", alphas_cumprod.to(torch.float32))
 
-        self.register_buffer("log_variance", log_variance.to(torch.float32))
+#        alphas_cumprod_prev = torch.nn.functional.pad(alphas_cumprod[:-1], (1, 0), value=1.0)
+        # TODO(PVP) - check how much of these is actually necessary!
+        # LDM only uses "fixed_small"; glide seems to use a weird mix of the two, ...
+        # https://github.com/openai/glide-text2im/blob/69b530740eb6cef69442d6180579ef5ba9ef063e/glide_text2im/gaussian_diffusion.py#L246
+#        variance = betas * (1.0 - alphas_cumprod_prev) / (1.0 - alphas_cumprod)
+#        if variance_type == "fixed_small":
+#            log_variance = torch.log(variance.clamp(min=1e-20))
+#        elif variance_type == "fixed_large":
+#            log_variance = torch.log(torch.cat([variance[1:2], betas[1:]], dim=0))
+#
+#
+#        self.register_buffer("log_variance", log_variance.to(torch.float32))
 
     def get_alpha(self, time_step):
         return self.alphas[time_step]
@@ -87,6 +90,43 @@ class GaussianDDPMScheduler(nn.Module, ConfigMixin):
             return torch.tensor(1.0)
         return self.alphas_cumprod[time_step]
 
+    def get_variance(self, t):
+        alpha_prod_t = self.get_alpha_prod(t)
+        alpha_prod_t_prev = self.get_alpha_prod(t - 1)
+
+        # For t > 0, compute predicted variance βt (see formala (6) and (7) from https://arxiv.org/pdf/2006.11239.pdf)
+        # and sample from it to get previous image
+        # x_{t-1} ~ N(pred_prev_image, variance) == add variane to pred_image
+        variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * self.get_beta(t).sqrt()
+
+        return variance
+
+    def predict_prev_image_step(self, residual, image, t, output_pred_x_0=False):
+        # 1. compute alphas, betas
+        alpha_prod_t = self.get_alpha_prod(t)
+        alpha_prod_t_prev = self.get_alpha_prod(t - 1)
+        beta_prod_t = 1 - alpha_prod_t
+        beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+        # 2. compute predicted original image from predicted noise also called
+        # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
+        pred_original_image = (image - beta_prod_t.sqrt() * residual) / alpha_prod_t.sqrt()
+
+        # 3. Clip "predicted x_0"
+        if self.clip_predicted_image:
+            pred_original_image = torch.clamp(pred_original_image, -1, 1)
+
+        # 4. Compute coefficients for pred_original_image x_0 and current image x_t
+        # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+        pred_original_image_coeff = (alpha_prod_t_prev.sqrt() * self.get_beta(t)) / beta_prod_t
+        current_image_coeff = self.get_alpha(t).sqrt() * beta_prod_t_prev / beta_prod_t
+
+        # 5. Compute predicted previous image µ_t
+        # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+        pred_prev_image = pred_original_image_coeff * pred_original_image + current_image_coeff * image
+
+        return pred_prev_image
+
     def sample_noise(self, shape, device, generator=None):
         # always sample on CPU to be deterministic
         return torch.randn(shape, generator=generator).to(device)

tests/test_modeling_utils.py

@@ -22,7 +22,7 @@ from distutils.util import strtobool
 
 import torch
 
-from diffusers import GaussianDDPMScheduler, UNetModel
+from diffusers import GaussianDDPMScheduler, UNetModel, DDIMScheduler
 from diffusers.configuration_utils import ConfigMixin
 from diffusers.pipeline_utils import DiffusionPipeline
 from models.vision.ddim.modeling_ddim import DDIM
@@ -304,7 +304,10 @@ class PipelineTesterMixin(unittest.TestCase):
         generator = torch.manual_seed(0)
         model_id = "fusing/ddpm-cifar10"
 
-        ddpm = DDPM.from_pretrained(model_id)
+        unet = UNetModel.from_pretrained(model_id)
+        noise_scheduler = GaussianDDPMScheduler.from_config(model_id)
+
+        ddpm = DDPM(unet=unet, noise_scheduler=noise_scheduler)
         image = ddpm(generator=generator)
 
         image_slice = image[0, -1, -3:, -3:].cpu()
@@ -318,7 +321,10 @@ class PipelineTesterMixin(unittest.TestCase):
         generator = torch.manual_seed(0)
         model_id = "fusing/ddpm-cifar10"
 
-        ddim = DDIM.from_pretrained(model_id)
+        unet = UNetModel.from_pretrained(model_id)
+        noise_scheduler = DDIMScheduler()
+
+        ddim = DDIM(unet=unet, noise_scheduler=noise_scheduler)
         image = ddim(generator=generator, eta=0.0)
 
         image_slice = image[0, -1, -3:, -3:].cpu()