clipping for continuous time diffusion not working

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)

denoising_diffusion_pytorch/continuous_time_gaussian_diffusion.py

@@ -73,7 +73,8 @@ 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 +91,10 @@ 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 +102,8 @@ 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 +115,8 @@ 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
+        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
     ):
         super().__init__()
         assert not denoise_fn.sinusoidal_cond_mlp
@@ -136,7 +140,8 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
             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}')
@@ -145,10 +150,6 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
 
         self.num_sample_steps = num_sample_steps
 
-        # clipping related hyperparameters
-
-        self.clip_after_noising_during_sampling = clip_after_noising_during_sampling
-
     @property
     def device(self):
         return next(self.denoise_fn.parameters()).device
@@ -210,12 +211,9 @@ 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):

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 = '0.17.4',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',