give an initial conv

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -192,6 +192,7 @@ class Unet(nn.Module):
     def __init__(
         self,
         dim,
+        init_dim = None,
         out_dim = None,
         dim_mults=(1, 2, 4, 8),
         channels = 3,
@@ -200,16 +201,19 @@ class Unet(nn.Module):
         super().__init__()
         self.channels = channels
 
-        dims = [channels, *map(lambda m: dim * m, dim_mults)]
+        init_dim = default(init_dim, dim // 3 * 2)
+        self.init_conv = nn.Conv2d(channels, init_dim, 7, padding = 3)
+
+        dims = [init_dim, *map(lambda m: dim * m, dim_mults)]
         in_out = list(zip(dims[:-1], dims[1:]))
 
         if with_time_emb:
-            time_dim = dim
+            time_dim = dim * 4
             self.time_mlp = nn.Sequential(
                 SinusoidalPosEmb(dim),
-                nn.Linear(dim, dim * 4),
+                nn.Linear(dim, time_dim),
                 nn.GELU(),
-                nn.Linear(dim * 4, dim)
+                nn.Linear(time_dim, time_dim)
             )
         else:
             time_dim = None
@@ -251,6 +255,8 @@ 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
 
         h = []

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'denoising-diffusion-pytorch',
   packages = find_packages(),
-  version = '0.9.2',
+  version = '0.10.1',
   license='MIT',
   description = 'Denoising Diffusion Probabilistic Models - Pytorch',
   author = 'Phil Wang',