allow for configuring the hidden dimension of the monotonic mlp parameterizing the noise schedule

denoising_diffusion_pytorch/continuous_time_gaussian_diffusion.py

@@ -98,7 +98,7 @@ class learned_noise_schedule(nn.Module):
 
         x = self.net(x)
 
-        normalized = self.slope * ((x - out_one) / (out_zero - out_one)) + self.intercept
+        normalized = self.slope * ((x - out_zero) / (out_one - out_zero)) + self.intercept
         return normalized
 
 class ContinuousTimeGaussianDiffusion(nn.Module):
@@ -110,7 +110,9 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         channels = 3,
         loss_type = 'l1',
         noise_schedule = 'linear',
-        num_sample_steps = 500
+        num_sample_steps = 500,
+        clip_after_noising_during_sampling = False,
+        learned_schedule_net_hidden_dim = 1024
     ):
         super().__init__()
         assert not denoise_fn.sinusoidal_cond_mlp
@@ -133,7 +135,8 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
 
             self.log_snr = learned_noise_schedule(
                 log_snr_max = log_snr_max,
-                log_snr_min = log_snr_min
+                log_snr_min = log_snr_min,
+                hidden_dim = learned_schedule_net_hidden_dim
             )
         else:
             raise ValueError(f'unknown noise schedule {noise_schedule}')
@@ -142,6 +145,10 @@ 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
@@ -203,9 +210,12 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
             times_next = steps[i + 1]
             img = self.p_sample(img, times, times_next)
 
-        img.clamp_(-1., 1.)
+            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 = unnormalize_to_zero_to_one(img)
-        return img
+        return img.clamp(0., 1.)
 
     @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 = '0.17.0',
+  version = '0.17.2',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',