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

README.md

@@ -6,6 +6,8 @@ Implementation of <a href="https://arxiv.org/abs/2006.11239">Denoising Diffusion
 
 This implementation was transcribed from the official Tensorflow version <a href="https://github.com/hojonathanho/diffusion">here</a>
 
+<a href="https://huggingface.co/blog/annotated-diffusion">Annotated code</a> by Research Scientists / Engineers from <a href="https://huggingface.co/">🤗 Huggingface</a>
+
 <img src="./sample.png" width="500px"><img>
 
 [![PyPI version](https://badge.fury.io/py/denoising-diffusion-pytorch.svg)](https://badge.fury.io/py/denoising-diffusion-pytorch)
@@ -119,3 +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

@@ -59,11 +59,14 @@ class MonotonicLinear(nn.Module):
 
 # log(snr) that approximates the original linear schedule
 
-def beta_linear_log_snr(t):
-    return -torch.log(expm1(1e-4 + 10 * (t ** 2)))
+def log(t, eps = 1e-20):
+    return torch.log(t.clamp(min = eps))
 
-def alpha_cosine_log_snr(t):
-    raise NotImplementedError
+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, 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 """
@@ -73,7 +76,8 @@ class learned_noise_schedule(nn.Module):
         *,
         log_snr_max,
         log_snr_min,
-        hidden_dim = 1024
+        hidden_dim = 1024,
+        frac_gradient = 1.
     ):
         super().__init__()
         self.slope = log_snr_min - log_snr_max
@@ -90,7 +94,10 @@ class learned_noise_schedule(nn.Module):
             Rearrange('... 1 -> ...'),
         )
 
+        self.frac_gradient = frac_gradient
+
     def forward(self, x):
+        frac_gradient = self.frac_gradient
         device = x.device
 
         out_zero = self.net(torch.zeros_like(x))
@@ -98,8 +105,8 @@ class learned_noise_schedule(nn.Module):
 
         x = self.net(x)
 
-        normalized = self.slope * ((x - out_zero) / (out_one - out_zero)) + self.intercept
-        return normalized
+        normed = self.slope * ((x - out_zero) / (out_one - out_zero)) + self.intercept
+        return normed * frac_gradient + normed.detach() * (1 - frac_gradient)
 
 class ContinuousTimeGaussianDiffusion(nn.Module):
     def __init__(
@@ -111,8 +118,11 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         loss_type = 'l1',
         noise_schedule = 'linear',
         num_sample_steps = 500,
-        clip_after_noising_during_sampling = False,
-        learned_schedule_net_hidden_dim = 1024
+        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
+        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
@@ -130,13 +140,16 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
 
         if noise_schedule == 'linear':
             self.log_snr = beta_linear_log_snr
+        elif noise_schedule == 'cosine':
+            self.log_snr = alpha_cosine_log_snr
         elif noise_schedule == 'learned':
             log_snr_max, log_snr_min = [beta_linear_log_snr(torch.tensor([time])).item() for time in (0., 1.)]
 
             self.log_snr = learned_noise_schedule(
                 log_snr_max = log_snr_max,
                 log_snr_min = log_snr_min,
-                hidden_dim = learned_schedule_net_hidden_dim
+                hidden_dim = learned_schedule_net_hidden_dim,
+                frac_gradient = learned_noise_schedule_frac_gradient
             )
         else:
             raise ValueError(f'unknown noise schedule {noise_schedule}')
@@ -144,10 +157,15 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         # sampling
 
         self.num_sample_steps = num_sample_steps
+        self.clip_sample_denoised = clip_sample_denoised
 
-        # clipping related hyperparameters
+        # p2 loss weight
+        # proposed https://arxiv.org/abs/2204.00227
 
-        self.clip_after_noising_during_sampling = clip_after_noising_during_sampling
+        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):
@@ -166,9 +184,6 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         # reviewer found an error in the equation in the paper (missing sigma)
         # following - https://openreview.net/forum?id=2LdBqxc1Yv&noteId=rIQgH0zKsRt
 
-        # todo - derive x_start from the posterior mean and do dynamic thresholding
-        # assumed that is what is going on in Imagen
-
         log_snr = self.log_snr(time)
         log_snr_next = self.log_snr(time_next)
         c = -expm1(log_snr - log_snr_next)
@@ -176,10 +191,21 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         squared_alpha, squared_alpha_next = log_snr.sigmoid(), log_snr_next.sigmoid()
         squared_sigma, squared_sigma_next = (-log_snr).sigmoid(), (-log_snr_next).sigmoid()
 
+        alpha, sigma, alpha_next = map(sqrt, (squared_alpha, squared_sigma, squared_alpha_next))
+
         batch_log_snr = repeat(log_snr, ' -> b', b = x.shape[0])
         pred_noise = self.denoise_fn(x, batch_log_snr)
 
-        model_mean = sqrt(squared_alpha_next / squared_alpha) * (x - c * sqrt(squared_sigma) * pred_noise)
+        if self.clip_sample_denoised:
+            x_start = (x - sigma * pred_noise) / alpha
+
+            # in Imagen, this was changed to dynamic thresholding
+            x_start.clamp_(-1., 1.)
+
+            model_mean = alpha_next * (x * (1 - c) / alpha + c * x_start)
+        else:
+            model_mean = alpha_next / alpha * (x - c * sigma * pred_noise)
+
         posterior_variance = squared_sigma_next * c
 
         return model_mean, posterior_variance
@@ -210,12 +236,9 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
             times_next = steps[i + 1]
             img = self.p_sample(img, times, times_next)
 
-            if self.clip_after_noising_during_sampling:
-                # clip after noise is added. perhaps this is sufficient for Imagen dynamic thresholding?
-                img.clamp_(-1., 1.)
-
+        img.clamp_(-1., 1.)
         img = unnormalize_to_zero_to_one(img)
-        return img.clamp(0., 1.)
+        return img
 
     @torch.no_grad()
     def sample(self, batch_size = 16):
@@ -242,9 +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 = '0.17.2',
+  version = '0.18.0',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',