Upload modeling_ddpm.py

modeling_ddpm.py

@@ -14,15 +14,13 @@
 # limitations under the License.
 
 
-from diffusers import DiffusionPipeline
-import tqdm
 import torch
 
+import tqdm
+from diffusers import DiffusionPipeline
 
-class DDPM(DiffusionPipeline):
-
-    modeling_file = "modeling_ddpm.py"
 
+class DDPM(DiffusionPipeline):
     def __init__(self, unet, noise_scheduler):
         super().__init__()
         self.register_modules(unet=unet, noise_scheduler=noise_scheduler)
@@ -32,30 +30,30 @@ class DDPM(DiffusionPipeline):
             torch_device = "cuda" if torch.cuda.is_available() else "cpu"
 
         self.unet.to(torch_device)
-        # 1. Sample gaussian noise
-        image = self.noise_scheduler.sample_noise((batch_size, self.unet.in_channels, self.unet.resolution, self.unet.resolution), device=torch_device, generator=generator)
-        for t in tqdm.tqdm(reversed(range(len(self.noise_scheduler))), total=len(self.noise_scheduler)):
-            # i) define coefficients for time step t
-            clip_image_coeff = 1 / torch.sqrt(self.noise_scheduler.get_alpha_prod(t))
-            clip_noise_coeff = torch.sqrt(1 / self.noise_scheduler.get_alpha_prod(t) - 1)
-            image_coeff = (1 - self.noise_scheduler.get_alpha_prod(t - 1)) * torch.sqrt(self.noise_scheduler.get_alpha(t)) / (1 - self.noise_scheduler.get_alpha_prod(t))
-            clip_coeff = torch.sqrt(self.noise_scheduler.get_alpha_prod(t - 1)) * self.noise_scheduler.get_beta(t) / (1 - self.noise_scheduler.get_alpha_prod(t))
-
-            # ii) predict noise residual
+
+        # Sample gaussian noise to begin loop
+        image = self.noise_scheduler.sample_noise(
+            (batch_size, self.unet.in_channels, self.unet.resolution, self.unet.resolution),
+            device=torch_device,
+            generator=generator,
+        )
+
+        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():
-                noise_residual = self.unet(image, t)
+                residual = self.unet(image, t)
 
-            # iii) compute predicted image from residual
-            # See 2nd formula at https://github.com/hojonathanho/diffusion/issues/5#issue-896554416 for comparison
-            pred_mean = clip_image_coeff * image - clip_noise_coeff * noise_residual
-            pred_mean = torch.clamp(pred_mean, -1, 1)
-            prev_image = clip_coeff * pred_mean + image_coeff * image
+            # 2. predict previous mean of image x_t-1
+            pred_prev_image = self.noise_scheduler.compute_prev_image_step(residual, image, t)
 
-            # iv) sample variance
-            prev_variance = self.noise_scheduler.sample_variance(t, prev_image.shape, device=torch_device, generator=generator)
+            # 3. optionally sample variance
+            variance = 0
+            if t > 0:
+                noise = self.noise_scheduler.sample_noise(image.shape, device=image.device, generator=generator)
+                variance = self.noise_scheduler.get_variance(t).sqrt() * noise
 
-            # v) sample  x_{t-1} ~ N(prev_image, prev_variance)
-            sampled_prev_image = prev_image + prev_variance
-            image = sampled_prev_image
+            # 4. set current image to prev_image: x_t -> x_t-1
+            image = pred_prev_image + variance
 
         return image