Merge branch 'main' of https://github.com/huggingface/diffusers into main

examples/README.md

@@ -10,7 +10,7 @@ python -m torch.distributed.launch \
   train_unconditional.py \
   --dataset="huggan/flowers-102-categories" \
   --resolution=64 \
-  --output_path="flowers-ddpm" \
+  --output_dir="flowers-ddpm" \
   --batch_size=16 \
   --num_epochs=100 \
   --gradient_accumulation_steps=1 \
@@ -34,7 +34,7 @@ python -m torch.distributed.launch \
   train_unconditional.py \
   --dataset="huggan/pokemon" \
   --resolution=64 \
-  --output_path="pokemon-ddpm" \
+  --output_dir="pokemon-ddpm" \
   --batch_size=16 \
   --num_epochs=100 \
   --gradient_accumulation_steps=1 \

examples/train_unconditional.py

@@ -39,7 +39,7 @@ def main(args):
         resamp_with_conv=True,
         resolution=args.resolution,
     )
-    noise_scheduler = DDPMScheduler(timesteps=1000)
+    noise_scheduler = DDPMScheduler(timesteps=1000, tensor_format="pt")
     optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
 
     augmentations = Compose(
@@ -93,15 +93,13 @@ def main(args):
             pbar.set_description(f"Epoch {epoch}")
             for step, batch in enumerate(train_dataloader):
                 clean_images = batch["input"]
-                noisy_images = torch.empty_like(clean_images)
+                noise_samples = torch.randn(clean_images.shape).to(clean_images.device)
-                noise_samples = torch.empty_like(clean_images)
                 bsz = clean_images.shape[0]
-
                 timesteps = torch.randint(0, noise_scheduler.timesteps, (bsz,), device=clean_images.device).long()
-                for idx in range(bsz):
+
-                    noise = torch.randn(clean_images.shape[1:]).to(clean_images.device)
+                # add noise onto the clean images according to the noise magnitude at each timestep
-                    noise_samples[idx] = noise
+                # (this is the forward diffusion process)
-                    noisy_images[idx] = noise_scheduler.forward_step(clean_images[idx], noise, timesteps[idx])
+                noisy_images = noise_scheduler.training_step(clean_images, noise_samples, timesteps)
 
                 if step % args.gradient_accumulation_steps != 0:
                     with accelerator.no_sync(model):
@@ -146,7 +144,7 @@ def main(args):
             # save image
             test_dir = os.path.join(args.output_dir, "test_samples")
             os.makedirs(test_dir, exist_ok=True)
-            image_pil.save(f"{test_dir}/{epoch}.png")
+            image_pil.save(f"{test_dir}/{epoch:04d}.png")
 
             # save the model
             if args.push_to_hub:

src/diffusers/schedulers/scheduling_ddpm.py

@@ -17,6 +17,7 @@
 import math
 
 import numpy as np
+import torch
 
 from ..configuration_utils import ConfigMixin
 from .scheduling_utils import SchedulerMixin
@@ -142,11 +143,18 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
 
         return pred_prev_sample
 
-    def forward_step(self, original_sample, noise, t):
+    def training_step(self, original_samples: torch.Tensor, noise: torch.Tensor, timesteps: torch.Tensor):
-        sqrt_alpha_prod = self.alphas_cumprod[t] ** 0.5
+        if timesteps.dim() != 1:
-        sqrt_one_minus_alpha_prod = (1 - self.alphas_cumprod[t]) ** 0.5
+            raise ValueError("`timesteps` must be a 1D tensor")
-        noisy_sample = sqrt_alpha_prod * original_sample + sqrt_one_minus_alpha_prod * noise
+
-        return noisy_sample
+        device = original_samples.device
+        batch_size = original_samples.shape[0]
+        timesteps = timesteps.reshape(batch_size, 1, 1, 1)
+
+        sqrt_alpha_prod = self.alphas_cumprod[timesteps] ** 0.5
+        sqrt_one_minus_alpha_prod = (1 - self.alphas_cumprod[timesteps]) ** 0.5
+        noisy_samples = sqrt_alpha_prod.to(device) * original_samples + sqrt_one_minus_alpha_prod.to(device) * noise
+        return noisy_samples
 
     def __len__(self):
         return self.config.timesteps