add p2 loss weighting to SNR version of denoising diffusion, brought up by @Mut1nyJD, paper is https://arxiv.org/abs/2204.00227

README.md

@@ -121,3 +121,13 @@ Samples and model checkpoints will be logged to `./results` periodically
     url         = {https://openreview.net/forum?id=2LdBqxc1Yv}
 }
 ```
+
+```bibtex
+@article{Choi2022PerceptionPT,
+    title   = {Perception Prioritized Training of Diffusion Models},
+    author  = {Jooyoung Choi and Jungbeom Lee and Chaehun Shin and Sungwon Kim and Hyunwoo J. Kim and Sung-Hoon Yoon},
+    journal = {ArXiv},
+    year    = {2022},
+    volume  = {abs/2204.00227}
+}
+```

denoising_diffusion_pytorch/continuous_time_gaussian_diffusion.py

@@ -66,7 +66,7 @@ def beta_linear_log_snr(t):
     return -log(expm1(1e-4 + 10 * (t ** 2)))
 
 def alpha_cosine_log_snr(t, s = 0.008):
-    return -log((torch.cos((t + s) / (1 + s) * torch.pi * 0.5) ** -2) - 1)
+    return -log((torch.cos((t + s) / (1 + s) * torch.pi * 0.5) ** -2) - 1, eps = 1e-5)
 
 class learned_noise_schedule(nn.Module):
     """ described in section H and then I.2 of the supplementary material for variational ddpm paper """
@@ -120,7 +120,9 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         num_sample_steps = 500,
         clip_sample_denoised = True,
         learned_schedule_net_hidden_dim = 1024,
-        learned_noise_schedule_frac_gradient = 1.   # between 0 and 1, determines what percentage of gradients go back, so one can update the learned noise schedule more slowly
+        learned_noise_schedule_frac_gradient = 1.,   # between 0 and 1, determines what percentage of gradients go back, so one can update the learned noise schedule more slowly
+        p2_loss_weight_gamma = 0.,                   # p2 loss weight, from https://arxiv.org/abs/2204.00227 - 0 is equivalent to weight of 1 across time
+        p2_loss_weight_k = 1
     ):
         super().__init__()
         assert not denoise_fn.sinusoidal_cond_mlp
@@ -157,6 +159,14 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         self.num_sample_steps = num_sample_steps
         self.clip_sample_denoised = clip_sample_denoised
 
+        # p2 loss weight
+        # proposed https://arxiv.org/abs/2204.00227
+
+        assert p2_loss_weight_gamma <= 2, 'in paper, they noticed any gamma greater than 2 is harmful'
+
+        self.p2_loss_weight_gamma = p2_loss_weight_gamma  # recommended to be 0.5 or 1
+        self.p2_loss_weight_k = p2_loss_weight_k
+
     @property
     def device(self):
         return next(self.denoise_fn.parameters()).device
@@ -255,9 +265,17 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
         x, log_snr = self.q_sample(x_start = x_start, times = times, noise = noise)
-
         model_out = self.denoise_fn(x, log_snr)
-        return self.loss_fn(model_out, noise)
+
+        losses = self.loss_fn(model_out, noise, reduction = 'none')
+        losses = losses.mean(dim = tuple(range(1, losses.ndim)))
+
+        if self.p2_loss_weight_gamma >= 0:
+            # following eq 8. in https://arxiv.org/abs/2204.00227
+            loss_weight = (self.p2_loss_weight_k + log_snr.exp()) ** -self.p2_loss_weight_gamma
+            losses = losses * loss_weight
+
+        return losses.mean()
 
     def forward(self, img, *args, **kwargs):
         b, c, h, w, device, img_size, = *img.shape, img.device, self.image_size

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.17.7',
+  version = '0.18.0',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',