Add models.shallow_cnn

models/shallow_cnn.py

@@ -0,0 +1,52 @@
+import torch.nn as nn
+from .base import PyTorchModel
+
+from torchvision.models.resnet import BasicBlock
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes, out_planes, kernel_size=1,
+                     stride=stride, bias=False)
+
+
+class CNN(nn.Module, PyTorchModel):
+    def __init__(self, hidden_1_size=64, hidden_block_size=64,
+                 input_channels=3, output_size=17, blocks=2):
+        super(CNN, self).__init__()
+
+        self.conv1 = nn.Conv2d(input_channels, hidden_1_size,
+                               kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = nn.BatchNorm2d(hidden_1_size)
+        self.relu = nn.ReLU()
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        blocks_ = []
+        # Fix feature map size
+        self.conv2 = conv1x1(hidden_1_size, hidden_block_size)
+        for block in range(blocks):
+            blocks_.append(BasicBlock(hidden_block_size,
+                                      hidden_block_size, stride=1))
+        self.blocks = nn.Sequential(*blocks_)
+        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc = nn.Linear(hidden_block_size, output_size)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(
+                    m.weight, mode='fan_out', nonlinearity='relu')
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        self.criterion = nn.CrossEntropyLoss()
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+        x = self.conv2(x)
+        x = self.blocks(x)
+        x = self.avgpool(x)
+        output = self.fc(x.view(x.size(0), -1))
+        return output