unet needs to be conditioned on log(snr) in p_mean_variance for continuous time gaussian diffusion

denoising_diffusion_pytorch/continuous_time_gaussian_diffusion.py

@@ -59,12 +59,12 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         *,
         image_size,
         channels = 3,
-        cond_scale = 500,
         loss_type = 'l1',
         noise_schedule = 'linear',
         num_sample_steps = 500
     ):
         super().__init__()
+        assert not denoise_fn.sinusoidal_cond_mlp
 
         self.denoise_fn = denoise_fn
 
@@ -75,7 +75,6 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
 
         # continuous noise schedule related stuff
 
-        self.cond_scale = cond_scale # the log(snr) will be scaled by this value
         self.loss_type = loss_type
 
         if noise_schedule == 'linear':
@@ -107,11 +106,6 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
         # todo - derive x_start from the posterior mean and do dynamic thresholding
         # assumed that is what is going on in Imagen
 
-        batch = x.shape[0]
-        batch_time = repeat(time, ' -> b', b = batch)
-
-        pred_noise = self.denoise_fn(x, batch_time * self.cond_scale)
-
         log_snr = self.log_snr(time)
         log_snr_next = self.log_snr(time_next)
         c = -expm1(log_snr - log_snr_next)
@@ -119,6 +113,9 @@ 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()
 
+        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)
         posterior_variance = squared_sigma_next * c
 
@@ -150,6 +147,7 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
             times_next = steps[i + 1]
             img = self.p_sample(img, times, times_next)
 
+        img.clamp_(-1., 1.)
         img = unnormalize_to_zero_to_one(img)
         return img
 
@@ -179,7 +177,7 @@ class ContinuousTimeGaussianDiffusion(nn.Module):
 
         x, log_snr = self.q_sample(x_start = x_start, times = times, noise = noise)
 
-        model_out = self.denoise_fn(x, log_snr * self.cond_scale)
+        model_out = self.denoise_fn(x, log_snr)
         return self.loss_fn(model_out, noise)
 
     def forward(self, img, *args, **kwargs):

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -16,6 +16,7 @@ from PIL import Image
 
 from tqdm import tqdm
 from einops import rearrange
+from einops.layers.torch import Rearrange
 
 # helpers functions
 
@@ -211,6 +212,18 @@ class Attention(nn.Module):
 
 # model
 
+def MLP(dim_in, dim_hidden):
+    return nn.Sequential(
+        Rearrange('... -> ... 1'),
+        nn.Linear(1, dim_hidden),
+        nn.GELU(),
+        nn.LayerNorm(dim_hidden),
+        nn.Linear(dim_hidden, dim_hidden),
+        nn.GELU(),
+        nn.LayerNorm(dim_hidden),
+        nn.Linear(dim_hidden, dim_hidden)
+    )
+
 class Unet(nn.Module):
     def __init__(
         self,
@@ -219,9 +232,9 @@ class Unet(nn.Module):
         out_dim = None,
         dim_mults=(1, 2, 4, 8),
         channels = 3,
-        with_time_emb = True,
         resnet_block_groups = 8,
-        learned_variance = False
+        learned_variance = False,
+        sinusoidal_cond_mlp = True
     ):
         super().__init__()
 
@@ -239,8 +252,11 @@ class Unet(nn.Module):
 
         # time embeddings
 
-        if with_time_emb:
-            time_dim = dim * 4
+        time_dim = dim * 4
+
+        self.sinusoidal_cond_mlp = sinusoidal_cond_mlp
+
+        if sinusoidal_cond_mlp:
             self.time_mlp = nn.Sequential(
                 SinusoidalPosEmb(dim),
                 nn.Linear(dim, time_dim),
@@ -248,8 +264,7 @@ class Unet(nn.Module):
                 nn.Linear(time_dim, time_dim)
             )
         else:
-            time_dim = None
-            self.time_mlp = None
+            self.time_mlp = MLP(1, time_dim)
 
         # layers
 
@@ -292,8 +307,7 @@ class Unet(nn.Module):
 
     def forward(self, x, time):
         x = self.init_conv(x)
-
-        t = self.time_mlp(time) if exists(self.time_mlp) else None
+        t = self.time_mlp(time)
 
         h = []
 

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.16.4',
+  version = '0.16.7',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',