Add enhanced model

Add ResBlock using ConvBlock;
Use ResBlock in UNet;
Use PReLU instead of LeakyReLU;
Add more channels in lower levels in UNet;
Add more blocks in each level in UNet;

map2map/models/conv.py

@@ -5,7 +5,7 @@ import torch.nn as nn
 class ConvBlock(nn.Module):
     """Convolution blocks of the form specified by `seq`.
     """
-    def __init__(self, in_channels, out_channels, mid_channels=None, seq='CBAC'):
+    def __init__(self, in_channels, out_channels, mid_channels=None, seq='CBA'):
         super().__init__()
 
         self.in_channels = in_channels
@@ -34,7 +34,7 @@ class ConvBlock(nn.Module):
         elif l == 'B':
             return nn.BatchNorm3d(self.bn_channels)
         elif l == 'A':
-            return nn.LeakyReLU(inplace=True)
+            return nn.PReLU()
         else:
             raise NotImplementedError('layer type {} not supported'.format(l))
 
@@ -55,6 +55,26 @@ class ConvBlock(nn.Module):
         return self.convs(x)
 
 
+class ResBlock(ConvBlock):
+    """Residual convolution blocks of the form specified by `seq`. Input is
+    added to the residual followed by an activation.
+    """
+    def __init__(self, channels, seq='CBACB'):
+        super().__init__(in_channels=channels, out_channels=channels, seq=seq)
+
+        self.act = nn.PReLU()
+
+    def forward(self, x):
+        y = x
+
+        x = self.convs(x)
+
+        y = narrow_like(y, x)
+        x += y
+
+        return self.act(x)
+
+
 def narrow_like(a, b):
     """Narrow a to be like b.
 

map2map/models/unet.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 
-from .conv import ConvBlock, narrow_like
+from .conv import ConvBlock, ResBlock, narrow_like
 
 
 class DownBlock(ConvBlock):
@@ -17,17 +17,32 @@ class UNet(nn.Module):
     def __init__(self, in_channels, out_channels):
         super().__init__()
 
-        self.conv_0l = ConvBlock(in_channels, 64, seq='CAC')
+        self.conv_0l = nn.Sequential(
-        self.down_0l = DownBlock(64, 64)
+            ConvBlock(in_channels, 64, seq='CA'),
-        self.conv_1l = ConvBlock(64, 64)
+            ResBlock(64, seq='CBACBACB'),
-        self.down_1l = DownBlock(64, 64)
+        )
+        self.down_0l = ConvBlock(64, 128, seq='DBA')
+        self.conv_1l = nn.Sequential(
+            ResBlock(128, seq='CBACB'),
+            ResBlock(128, seq='CBACB'),
+        )
+        self.down_1l = ConvBlock(128, 256, seq='DBA')
 
-        self.conv_2c = ConvBlock(64, 64)
+        self.conv_2c = nn.Sequential(
+            ResBlock(256, seq='CBACB'),
+            ResBlock(256, seq='CBACB'),
+        )
 
-        self.up_1r = UpBlock(64, 64)
+        self.up_1r = ConvBlock(256, 128, seq='UBA')
-        self.conv_1r = ConvBlock(128, 64)
+        self.conv_1r = nn.Sequential(
-        self.up_0r = UpBlock(64, 64)
+            ResBlock(256, seq='CBACB'),
-        self.conv_0r = ConvBlock(128, out_channels, seq='CAC')
+            ResBlock(256, seq='CBACB'),
+        )
+        self.up_0r = ConvBlock(256, 64, seq='UBA')
+        self.conv_0r = nn.Sequential(
+            ResBlock(128, seq='CBACBAC'),
+            ConvBlock(128, out_channels, seq='C')
+        )
 
     def forward(self, x):
         y0 = self.conv_0l(x)