bug resnet fpn

simlearner3d/models/modules/resnet_fpn.py

@@ -0,0 +1,325 @@
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution without padding"""
+    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, padding=0, bias=False)
+
+
+def conv3x3(in_planes, out_planes, stride=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False)
+
+
+class BasicBlock(nn.Module):
+    def __init__(self, in_planes, planes, stride=1):
+        super().__init__()
+        self.conv1 = conv3x3(in_planes, planes, stride)
+        self.conv2 = conv3x3(planes, planes)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.relu = nn.ReLU(inplace=True)
+
+        if stride == 1:
+            self.downsample = None
+        else:
+            self.downsample = nn.Sequential(
+                conv1x1(in_planes, planes, stride=stride),
+                nn.BatchNorm2d(planes)
+            )
+
+    def forward(self, x):
+        y = x
+        y = self.relu(self.bn1(self.conv1(y)))
+        y = self.bn2(self.conv2(y))
+
+        if self.downsample is not None:
+            x = self.downsample(x)
+
+        return self.relu(x+y)
+
+
+class ResNetFPN_8_1(nn.Module):
+    """
+    ResNet+FPN, output resolution are 1/8 and 1/2.
+    Each block has 2 layers.
+    """
+
+    def __init__(self, INITIAL_DIM,BLOC_DIMS):
+        super(ResNetFPN_8_1, self).__init__()
+        # Config
+        block = BasicBlock
+        initial_dim = INITIAL_DIM
+        block_dims = BLOC_DIMS
+
+        # Class Variable
+        self.in_planes = initial_dim
+
+        # Networks
+        self.conv1 = nn.Conv2d(1, initial_dim, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = nn.BatchNorm2d(initial_dim)
+        self.relu = nn.ReLU(inplace=True)
+
+        self.layer1 = self._make_layer(block, block_dims[0], stride=1)  # 1/2
+        self.layer2 = self._make_layer(block, block_dims[1], stride=2)  # 1/4
+        self.layer3 = self._make_layer(block, block_dims[2], stride=2)  # 1/8
+
+        # 3. FPN upsample
+        self.layer3_outconv = conv1x1(block_dims[2], block_dims[2])
+        self.layer2_outconv = conv1x1(block_dims[1], block_dims[2])
+        self.layer2_outconv2 = nn.Sequential(
+            conv3x3(block_dims[2], block_dims[2]),
+            nn.BatchNorm2d(block_dims[2]),
+            nn.LeakyReLU(),
+            conv3x3(block_dims[2], block_dims[1]),
+        )
+        self.layer1_outconv = conv1x1(block_dims[0], block_dims[1])
+        self.layer1_outconv2 = nn.Sequential(
+            conv3x3(block_dims[1], block_dims[1]),
+            nn.BatchNorm2d(block_dims[1]),
+            nn.LeakyReLU(),
+            conv3x3(block_dims[1], block_dims[0]),
+        )
+
+        self.layer0_outconv = conv1x1(block_dims[0], initial_dim)
+        self.layer0_outconv1 = nn.Sequential(
+            conv3x3(block_dims[0], block_dims[0]),
+            nn.BatchNorm2d(block_dims[0]),
+            nn.LeakyReLU(),
+            conv3x3(block_dims[0], initial_dim),
+        )
+
+        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.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+    def _make_layer(self, block, dim, stride=1):
+        layer1 = block(self.in_planes, dim, stride=stride)
+        layer2 = block(dim, dim, stride=1)
+        layers = (layer1, layer2)
+
+        self.in_planes = dim
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        # ResNet Backbone
+        x0 = self.relu(self.bn1(self.conv1(x)))
+        x1 = self.layer1(x0)  # 1/2
+        x2 = self.layer2(x1)  # 1/4
+        x3 = self.layer3(x2)  # 1/8
+
+        # FPN
+        x3_out = self.layer3_outconv(x3)
+
+        x3_out_2x = F.interpolate(x3_out, scale_factor=2., mode='bilinear', align_corners=True) # res 4
+        x2_out = self.layer2_outconv(x2)
+        x2_out = self.layer2_outconv2(x2_out+x3_out_2x)
+
+        x2_out_2x = F.interpolate(x2_out, scale_factor=2., mode='bilinear', align_corners=True) # res 2
+        x1_out = self.layer1_outconv(x1)
+        x1_out = self.layer1_outconv2(x1_out+x2_out_2x)
+
+        x1_out_1x = F.interpolate(x1_out, scale_factor=2., mode='bilinear', align_corners=True) # res 1
+
+        x0_out=self.layer0_outconv(x0)
+        x0_out= self.layer0_outconv1(x0_out+x1_out_1x)
+
+        return x0_out
+
+
+
+class ResNetFPN_8_1_Inference(nn.Module):
+    """
+    ResNet+FPN, output resolution are 1/8 and 1/2.
+    Each block has 2 layers.
+    """
+
+    def __init__(self, INITIAL_DIM,BLOC_DIMS):
+        super(ResNetFPN_8_1_Inference, self).__init__()
+        # Config
+        block = BasicBlock
+        initial_dim = INITIAL_DIM
+        block_dims = BLOC_DIMS
+
+        # Class Variable
+        self.in_planes = initial_dim
+
+        # Networks
+        self.conv1 = nn.Conv2d(1, initial_dim, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = nn.BatchNorm2d(initial_dim)
+        self.relu = nn.ReLU(inplace=True)
+
+        self.layer1 = self._make_layer(block, block_dims[0], stride=1)  # 1/2
+        self.layer2 = self._make_layer(block, block_dims[1], stride=2)  # 1/4
+        self.layer3 = self._make_layer(block, block_dims[2], stride=2)  # 1/8
+
+        # 3. FPN upsample
+        self.layer3_outconv = conv1x1(block_dims[2], block_dims[2])
+        self.layer2_outconv = conv1x1(block_dims[1], block_dims[2])
+        self.layer2_outconv2 = nn.Sequential(
+            conv3x3(block_dims[2], block_dims[2]),
+            nn.BatchNorm2d(block_dims[2]),
+            nn.LeakyReLU(),
+            conv3x3(block_dims[2], block_dims[1]),
+        )
+        self.layer1_outconv = conv1x1(block_dims[0], block_dims[1])
+        self.layer1_outconv2 = nn.Sequential(
+            conv3x3(block_dims[1], block_dims[1]),
+            nn.BatchNorm2d(block_dims[1]),
+            nn.LeakyReLU(),
+            conv3x3(block_dims[1], block_dims[0]),
+        )
+
+        self.layer0_outconv = conv1x1(block_dims[0], initial_dim)
+        self.layer0_outconv1 = nn.Sequential(
+            conv3x3(block_dims[0], block_dims[0]),
+            nn.BatchNorm2d(block_dims[0]),
+            nn.LeakyReLU(),
+            conv3x3(block_dims[0], initial_dim),
+        )
+
+        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.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+    def _make_layer(self, block, dim, stride=1):
+        layer1 = block(self.in_planes, dim, stride=stride)
+        layer2 = block(dim, dim, stride=1)
+        layers = (layer1, layer2)
+
+        self.in_planes = dim
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        if x.size()[-2] % 16 != 0:
+            times = x.size()[-2]//16   
+            top_pad = (times+1)*16 - x.size()[-2]
+        else:
+            top_pad = 0
+        if x.size()[-1] % 16 != 0:
+            times = x.size()[-1]//16
+            right_pad = (times+1)*16-x.size()[-1] 
+        else:
+            right_pad = 0    
+
+        x = F.pad(x,(0,right_pad, top_pad,0))
+
+        # ResNet Backbone
+        x0 = self.relu(self.bn1(self.conv1(x)))
+        x1 = self.layer1(x0)  # 1/2
+        x2 = self.layer2(x1)  # 1/4
+        x3 = self.layer3(x2)  # 1/8
+
+        # FPN
+        x3_out = self.layer3_outconv(x3)
+
+        x3_out_2x = F.interpolate(x3_out, scale_factor=2., mode='bilinear', align_corners=True) # res 4
+        x2_out = self.layer2_outconv(x2)
+        x2_out = self.layer2_outconv2(x2_out+x3_out_2x)
+
+        x2_out_2x = F.interpolate(x2_out, scale_factor=2., mode='bilinear', align_corners=True) # res 2
+        x1_out = self.layer1_outconv(x1)
+        x1_out = self.layer1_outconv2(x1_out+x2_out_2x)
+
+        x1_out_1x = F.interpolate(x1_out, scale_factor=2., mode='bilinear', align_corners=True) # res 1
+
+        x0_out=self.layer0_outconv(x0)
+        x0_out= self.layer0_outconv1(x0_out+x1_out_1x)
+
+        if top_pad !=0 and right_pad != 0:
+            out = x0_out[:,:,top_pad:,:-right_pad]
+        elif top_pad ==0 and right_pad != 0:
+            out = x0_out[:,:,:,:-right_pad]
+        elif top_pad !=0 and right_pad == 0:
+            out = x0_out[:,:,top_pad:,:]
+        else:
+            out = x0_out
+        return out
+
+
+class ResNetFPN_16_4(nn.Module):
+    """
+    ResNet+FPN, output resolution are 1/16 and 1/4.
+    Each block has 2 layers.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        # Config
+        block = BasicBlock
+        initial_dim = config['initial_dim']
+        block_dims = config['block_dims']
+
+        # Class Variable
+        self.in_planes = initial_dim
+
+        # Networks
+        self.conv1 = nn.Conv2d(1, initial_dim, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = nn.BatchNorm2d(initial_dim)
+        self.relu = nn.ReLU(inplace=True)
+
+        self.layer1 = self._make_layer(block, block_dims[0], stride=1)  # 1/2
+        self.layer2 = self._make_layer(block, block_dims[1], stride=2)  # 1/4
+        self.layer3 = self._make_layer(block, block_dims[2], stride=2)  # 1/8
+        self.layer4 = self._make_layer(block, block_dims[3], stride=2)  # 1/16
+
+        # 3. FPN upsample
+        self.layer4_outconv = conv1x1(block_dims[3], block_dims[3])
+        self.layer3_outconv = conv1x1(block_dims[2], block_dims[3])
+        self.layer3_outconv2 = nn.Sequential(
+            conv3x3(block_dims[3], block_dims[3]),
+            nn.BatchNorm2d(block_dims[3]),
+            nn.LeakyReLU(),
+            conv3x3(block_dims[3], block_dims[2]),
+        )
+
+        self.layer2_outconv = conv1x1(block_dims[1], block_dims[2])
+        self.layer2_outconv2 = nn.Sequential(
+            conv3x3(block_dims[2], block_dims[2]),
+            nn.BatchNorm2d(block_dims[2]),
+            nn.LeakyReLU(),
+            conv3x3(block_dims[2], block_dims[1]),
+        )
+
+        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.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+    def _make_layer(self, block, dim, stride=1):
+        layer1 = block(self.in_planes, dim, stride=stride)
+        layer2 = block(dim, dim, stride=1)
+        layers = (layer1, layer2)
+
+        self.in_planes = dim
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        # ResNet Backbone
+        x0 = self.relu(self.bn1(self.conv1(x)))
+        x1 = self.layer1(x0)  # 1/2
+        x2 = self.layer2(x1)  # 1/4
+        x3 = self.layer3(x2)  # 1/8
+        x4 = self.layer4(x3)  # 1/16
+
+        # FPN
+        x4_out = self.layer4_outconv(x4)
+
+        x4_out_2x = F.interpolate(x4_out, scale_factor=2., mode='bilinear', align_corners=True)
+        x3_out = self.layer3_outconv(x3)
+        x3_out = self.layer3_outconv2(x3_out+x4_out_2x)
+
+        x3_out_2x = F.interpolate(x3_out, scale_factor=2., mode='bilinear', align_corners=True)
+        x2_out = self.layer2_outconv(x2)
+        x2_out = self.layer2_outconv2(x2_out+x3_out_2x)
+
+        return [x4_out, x2_out]