add weight standardization prior to groupnorm, lessen cosine sim attention scale to 10 for fp16

README.md

@@ -10,8 +10,6 @@ Youtube AI Educators - <a href="https://www.youtube.com/watch?v=W-O7AZNzbzQ">Yan
 
 <a href="https://huggingface.co/blog/annotated-diffusion">Annotated code</a> by Research Scientists / Engineers from <a href="https://huggingface.co/">🤗 Huggingface</a>
 
-Update: Turns out none of the technicalities really matters at all | <a href="https://arxiv.org/abs/2208.09392">"Cold Diffusion" paper</a>
-
 <img src="./images/sample.png" width="500px"><img>
 
 [![PyPI version](https://badge.fury.io/py/denoising-diffusion-pytorch.svg)](https://badge.fury.io/py/denoising-diffusion-pytorch)

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -97,11 +97,16 @@ class WeightStandardizedConv2d(nn.Conv2d):
         eps = 1e-5 if x.dtype == torch.float32 else 1e-3
 
         weight = self.weight
-        mean = reduce(weight, 'o ... -> o 1 1 1', 'mean')
-        var = reduce(weight, 'o ... -> o 1 1 1', partial(torch.var, unbiased = False))
-        normalized_weight = (weight - mean) * (var + eps).rsqrt()
+        flattened_weights = rearrange(weight, 'o ... -> o (...)')
 
-        return F.conv2d(x, normalized_weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
+        mean = reduce(weight, 'o ... -> o 1 1 1', 'mean')
+
+        var = torch.var(flattened_weights, dim = -1, unbiased = False)
+        var = rearrange(var, 'o -> o 1 1 1')
+
+        weight = (weight - mean) * (var + eps).rsqrt()
+
+        return F.conv2d(x, weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
 
 class LayerNorm(nn.Module):
     def __init__(self, dim):
@@ -427,7 +432,6 @@ class GaussianDiffusion(nn.Module):
     ):
         super().__init__()
         assert not (type(self) == GaussianDiffusion and model.channels != model.out_dim)
-        assert not model.learned_sinusoidal_cond
 
         self.model = model
         self.channels = self.model.channels
@@ -570,15 +574,15 @@ class GaussianDiffusion(nn.Module):
 
         times = torch.linspace(0., total_timesteps, steps = sampling_timesteps + 2)[:-1]
         times = list(reversed(times.int().tolist()))
-        time_pairs = list(filter(lambda a: a[0] > a[1], zip(times[:-1], times[1:])))
+        time_pairs = list(zip(times[:-1], times[1:]))
 
         img = torch.randn(shape, device = device)
 
         x_start = None
 
         for time, time_next in tqdm(time_pairs, desc = 'sampling loop time step'):
-            alpha = self.alphas_cumprod[time]
-            alpha_next = self.alphas_cumprod[time_next]
+            alpha = self.alphas_cumprod_prev[time]
+            alpha_next = self.alphas_cumprod_prev[time_next]
 
             time_cond = torch.full((batch,), time, device = device, dtype = torch.long)
 
@@ -846,7 +850,6 @@ class Trainer(object):
 
                 accelerator.wait_for_everyone()
 
-                self.step += 1
                 if accelerator.is_main_process:
                     self.ema.to(device)
                     self.ema.update()
@@ -863,6 +866,7 @@ class Trainer(object):
                         utils.save_image(all_images, str(self.results_folder / f'sample-{milestone}.png'), nrow = int(math.sqrt(self.num_samples)))
                         self.save(milestone)
 
+                self.step += 1
                 pbar.update(1)
 
         accelerator.print('training complete')

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.27.6',
+  version = '0.27.3',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',
@@ -30,4 +30,4 @@ setup(
     'License :: OSI Approved :: MIT License',
     'Programming Language :: Python :: 3.6',
   ],
-)
+)