move accelerator backward outside of autocast context, also calculate total loss correctly across gradient accumulated steps

README.md

@@ -69,7 +69,7 @@ trainer = Trainer(
     diffusion,
     'path/to/your/images',
     train_batch_size = 32,
-    train_lr = 1e-4,
+    train_lr = 8e-5,
     train_num_steps = 700000,         # total training steps
     gradient_accumulate_every = 2,    # gradient accumulation steps
     ema_decay = 0.995,                # exponential moving average decay

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -537,7 +537,6 @@ class GaussianDiffusion(nn.Module):
         batch, device, total_timesteps, sampling_timesteps, eta, objective = shape[0], self.betas.device, self.num_timesteps, self.sampling_timesteps, self.ddim_sampling_eta, self.objective
 
         times = torch.linspace(0., total_timesteps, steps = sampling_timesteps + 2)[:-1]
-
         times = list(reversed(times.int().tolist()))
         time_pairs = list(zip(times[:-1], times[1:]))
 
@@ -554,12 +553,14 @@ class GaussianDiffusion(nn.Module):
             if clip_denoised:
                 x_start.clamp_(-1., 1.)
 
-            c1 = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
-            c2 = ((1 - alpha_next) - torch.square(c1)).sqrt()
+            sigma = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
+            c = ((1 - alpha_next) - sigma ** 2).sqrt()
+
+            noise = torch.randn_like(img) if time_next > 0 else 0.
 
             img = x_start * alpha_next.sqrt() + \
-                  c1 * torch.randn_like(img) + \
-                  c2 * pred_noise
+                  c * pred_noise + \
+                  sigma * noise
 
         img = unnormalize_to_zero_to_one(img)
         return img
@@ -680,6 +681,7 @@ class Trainer(object):
         train_num_steps = 100000,
         ema_update_every = 10,
         ema_decay = 0.995,
+        adam_betas = (0.9, 0.99),
         save_and_sample_every = 1000,
         num_samples = 25,
         results_folder = './results',
@@ -718,7 +720,7 @@ class Trainer(object):
 
         # optimizer
 
-        self.opt = Adam(diffusion_model.parameters(), lr = train_lr)
+        self.opt = Adam(diffusion_model.parameters(), lr = train_lr, betas = adam_betas)
 
         # for logging results in a folder periodically
 
@@ -771,14 +773,19 @@ class Trainer(object):
 
             while self.step < self.train_num_steps:
 
+                total_loss = 0.
+
                 for _ in range(self.gradient_accumulate_every):
                     data = next(self.dl).to(device)
 
                     with self.accelerator.autocast():
                         loss = self.model(data)
-                        self.accelerator.backward(loss / self.gradient_accumulate_every)
+                        loss = loss / self.gradient_accumulate_every
+                        total_loss += loss.item()
 
-                pbar.set_description(f'loss: {loss.item():.4f}')
+                    self.accelerator.backward(loss)
+
+                pbar.set_description(f'loss: {total_loss:.4f}')
 
                 accelerator.wait_for_everyone()
 

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.25.1',
+  version = '0.25.3',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',