Adds squeeze and excitation (scSE) modules, resolves mapbox#157

robosat/scse.py

@@ -0,0 +1,63 @@
+"""Squeeze and Excitation blocks - attention for classification and segmentation
+
+See:
+- https://arxiv.org/abs/1709.01507 - Squeeze-and-Excitation Networks
+- https://arxiv.org/abs/1803.02579 - Concurrent Spatial and Channel 'Squeeze & Excitation' in Fully Convolutional Networks
+
+"""
+
+import torch
+import torch.nn as nn
+
+
+class SpatialSqChannelEx(nn.Module):
+    """Spatial Squeeze and Channel Excitation (cSE) block
+       See https://arxiv.org/abs/1803.02579 Figure 1 b
+    """
+
+    def __init__(self, num_in, r):
+        super().__init__()
+        self.fc0 = Conv1x1(num_in, num_in // r)
+        self.fc1 = Conv1x1(num_in // r, num_in)
+
+    def forward(self, x):
+        xx = nn.functional.adaptive_avg_pool2d(x, 1)
+        xx = self.fc0(xx)
+        xx = nn.functional.relu(xx, inplace=True)
+        xx = self.fc1(xx)
+        xx = torch.sigmoid(xx)
+        return x * xx
+
+
+class ChannelSqSpatialEx(nn.Module):
+    """Channel Squeeze and Spatial Excitation (sSE) block
+       See https://arxiv.org/abs/1803.02579 Figure 1 c
+    """
+
+    def __init__(self, num_in):
+        super().__init__()
+        self.conv = Conv1x1(num_in, 1)
+
+    def forward(self, x):
+        xx = self.conv(x)
+        xx = torch.sigmoid(xx)
+        return x * xx
+
+
+class SpatialChannelSqChannelEx(nn.Module):
+    """Concurrent Spatial and Channel Squeeze and Channel Excitation (csSE) block
+       See https://arxiv.org/abs/1803.02579 Figure 1 d
+    """
+
+    def __init__(self, num_in, r=16):
+        super().__init__()
+
+        self.cse = SpatialSqChannelEx(num_in, r)
+        self.sse = ChannelSqSpatialEx(num_in)
+
+    def forward(self, x):
+        return self.cse(x) + self.sse(x)
+
+
+def Conv1x1(num_in, num_out):
+    return nn.Conv2d(num_in, num_out, kernel_size=1, bias=False)

robosat/unet.py

@@ -14,6 +14,8 @@
 
 from torchvision.models import resnet50
 
+from robosat.scse import SpatialChannelSqChannelEx
+
 
 class ConvRelu(nn.Module):
     """3x3 convolution followed by ReLU activation building block.
@@ -91,10 +93,23 @@ def __init__(self, num_classes, num_filters=32, pretrained=True):
 
         # Todo: make input channels configurable, not hard-coded to three channels for RGB
 
-        self.resnet = resnet50(pretrained=pretrained)
-
         # Access resnet directly in forward pass; do not store refs here due to
         # https://github.com/pytorch/pytorch/issues/8392
+        self.resnet = resnet50(pretrained=pretrained)
+
+        # seSE blocks to append to encoder and decoder as recommended by
+        # https://arxiv.org/abs/1803.02579
+        self.scse0 = SpatialChannelSqChannelEx(64)
+        self.scse1 = SpatialChannelSqChannelEx(256)
+        self.scse2 = SpatialChannelSqChannelEx(512)
+        self.scse3 = SpatialChannelSqChannelEx(1024)
+        self.scse4 = SpatialChannelSqChannelEx(2048)
+
+        self.scse5 = SpatialChannelSqChannelEx(num_filters * 8)
+        self.scse6 = SpatialChannelSqChannelEx(num_filters * 8)
+        self.scse7 = SpatialChannelSqChannelEx(num_filters * 2)
+        self.scse8 = SpatialChannelSqChannelEx(num_filters * 2 * 2)
+        self.scse9 = SpatialChannelSqChannelEx(num_filters)
 
         self.center = DecoderBlock(2048, num_filters * 8)
 
@@ -122,20 +137,21 @@ def forward(self, x):
         enc0 = self.resnet.conv1(x)
         enc0 = self.resnet.bn1(enc0)
         enc0 = self.resnet.relu(enc0)
+        enc0 = self.scse0(enc0)
         enc0 = self.resnet.maxpool(enc0)
 
-        enc1 = self.resnet.layer1(enc0)
-        enc2 = self.resnet.layer2(enc1)
-        enc3 = self.resnet.layer3(enc2)
-        enc4 = self.resnet.layer4(enc3)
+        enc1 = self.scse1(self.resnet.layer1(enc0))
+        enc2 = self.scse2(self.resnet.layer2(enc1))
+        enc3 = self.scse3(self.resnet.layer3(enc2))
+        enc4 = self.scse4(self.resnet.layer4(enc3))
 
         center = self.center(nn.functional.max_pool2d(enc4, kernel_size=2, stride=2))
 
-        dec0 = self.dec0(torch.cat([enc4, center], dim=1))
-        dec1 = self.dec1(torch.cat([enc3, dec0], dim=1))
-        dec2 = self.dec2(torch.cat([enc2, dec1], dim=1))
-        dec3 = self.dec3(torch.cat([enc1, dec2], dim=1))
-        dec4 = self.dec4(dec3)
+        dec0 = self.scse5(self.dec0(torch.cat([enc4, center], dim=1)))
+        dec1 = self.scse6(self.dec1(torch.cat([enc3, dec0], dim=1)))
+        dec2 = self.scse7(self.dec2(torch.cat([enc2, dec1], dim=1)))
+        dec3 = self.scse8(self.dec3(torch.cat([enc1, dec2], dim=1)))
+        dec4 = self.scse9(self.dec4(dec3))
         dec5 = self.dec5(dec4)
 
         return self.final(dec5)