[fix] forgot resnets

bioimage_embed/models/bolts/resnets.py

@@ -0,0 +1,341 @@
+# https://github.com/Lightning-Universe/lightning-bolts
+# Credit
+
+import torch
+from torch import nn
+from torch.nn import functional as F  # noqa: N812
+
+
+
+class Interpolate(nn.Module):
+    """nn.Module wrapper for F.interpolate."""
+
+    def __init__(self, size=None, scale_factor=None) -> None:
+        super().__init__()
+        self.size, self.scale_factor = size, scale_factor
+
+    def forward(self, x):
+        return F.interpolate(x, size=self.size, scale_factor=self.scale_factor)
+
+
+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)
+
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution."""
+    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)
+
+
+
+def resize_conv3x3(in_planes, out_planes, scale=1):
+    """Upsample + 3x3 convolution with padding to avoid checkerboard artifact."""
+    if scale == 1:
+        return conv3x3(in_planes, out_planes)
+    return nn.Sequential(Interpolate(scale_factor=scale), conv3x3(in_planes, out_planes))
+
+
+
+def resize_conv1x1(in_planes, out_planes, scale=1):
+    """Upsample + 1x1 convolution with padding to avoid checkerboard artifact."""
+    if scale == 1:
+        return conv1x1(in_planes, out_planes)
+    return nn.Sequential(Interpolate(scale_factor=scale), conv1x1(in_planes, out_planes))
+
+
+
+class EncoderBlock(nn.Module):
+    """ResNet block, copied from https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py#L35."""
+
+    expansion = 1
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None) -> None:
+        super().__init__()
+        self.conv1 = conv3x3(inplanes, planes, stride)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = conv3x3(planes, planes)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.downsample = downsample
+
+    def forward(self, x):
+        identity = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        return self.relu(out)
+
+
+
+class EncoderBottleneck(nn.Module):
+    """ResNet bottleneck, copied from
+    https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py#L75."""
+
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None) -> None:
+        super().__init__()
+        width = planes  # this needs to change if we want wide resnets
+        self.conv1 = conv1x1(inplanes, width)
+        self.bn1 = nn.BatchNorm2d(width)
+        self.conv2 = conv3x3(width, width, stride)
+        self.bn2 = nn.BatchNorm2d(width)
+        self.conv3 = conv1x1(width, planes * self.expansion)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        identity = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        return self.relu(out)
+
+
+
+class DecoderBlock(nn.Module):
+    """ResNet block, but convs replaced with resize convs, and channel increase is in second conv, not first."""
+
+    expansion = 1
+
+    def __init__(self, inplanes, planes, scale=1, upsample=None) -> None:
+        super().__init__()
+        self.conv1 = resize_conv3x3(inplanes, inplanes)
+        self.bn1 = nn.BatchNorm2d(inplanes)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = resize_conv3x3(inplanes, planes, scale)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.upsample = upsample
+
+    def forward(self, x):
+        identity = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+
+        if self.upsample is not None:
+            identity = self.upsample(x)
+
+        out += identity
+        return self.relu(out)
+
+
+
+class DecoderBottleneck(nn.Module):
+    """ResNet bottleneck, but convs replaced with resize convs."""
+
+    expansion = 4
+
+    def __init__(self, inplanes, planes, scale=1, upsample=None) -> None:
+        super().__init__()
+        width = planes  # this needs to change if we want wide resnets
+        self.conv1 = resize_conv1x1(inplanes, width)
+        self.bn1 = nn.BatchNorm2d(width)
+        self.conv2 = resize_conv3x3(width, width, scale)
+        self.bn2 = nn.BatchNorm2d(width)
+        self.conv3 = conv1x1(width, planes * self.expansion)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+        self.relu = nn.ReLU(inplace=True)
+        self.upsample = upsample
+        self.scale = scale
+
+    def forward(self, x):
+        identity = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.upsample is not None:
+            identity = self.upsample(x)
+
+        out += identity
+        return self.relu(out)
+
+
+
+class ResNetEncoder(nn.Module):
+    def __init__(self, block, layers, first_conv=False, maxpool1=False) -> None:
+        super().__init__()
+
+        self.inplanes = 64
+        self.first_conv = first_conv
+        self.maxpool1 = maxpool1
+
+        if self.first_conv:
+            self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False)
+        else:
+            self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False)
+
+        self.bn1 = nn.BatchNorm2d(self.inplanes)
+        self.relu = nn.ReLU(inplace=True)
+
+        if self.maxpool1:
+            self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        else:
+            self.maxpool = nn.MaxPool2d(kernel_size=1, stride=1)
+
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
+        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+
+    def _make_layer(self, block, planes, blocks, stride=1):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                nn.BatchNorm2d(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(block(self.inplanes, planes, stride, downsample))
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(block(self.inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.avgpool(x)
+        return torch.flatten(x, 1)
+
+
+
+class ResNetDecoder(nn.Module):
+    """Resnet in reverse order."""
+
+    def __init__(self, block, layers, latent_dim, input_height, first_conv=False, maxpool1=False) -> None:
+        super().__init__()
+
+        self.expansion = block.expansion
+        self.inplanes = 512 * block.expansion
+        self.first_conv = first_conv
+        self.maxpool1 = maxpool1
+        self.input_height = input_height
+
+        self.upscale_factor = 8
+
+        self.linear = nn.Linear(latent_dim, self.inplanes * 4 * 4)
+
+        self.layer1 = self._make_layer(block, 256, layers[0], scale=2)
+        self.layer2 = self._make_layer(block, 128, layers[1], scale=2)
+        self.layer3 = self._make_layer(block, 64, layers[2], scale=2)
+
+        if self.maxpool1:
+            self.layer4 = self._make_layer(block, 64, layers[3], scale=2)
+            self.upscale_factor *= 2
+        else:
+            self.layer4 = self._make_layer(block, 64, layers[3])
+
+        if self.first_conv:
+            self.upscale = Interpolate(scale_factor=2)
+            self.upscale_factor *= 2
+        else:
+            self.upscale = Interpolate(scale_factor=1)
+
+        # interpolate after linear layer using scale factor
+        self.upscale1 = Interpolate(size=input_height // self.upscale_factor)
+
+        self.conv1 = nn.Conv2d(64 * block.expansion, 3, kernel_size=3, stride=1, padding=1, bias=False)
+
+    def _make_layer(self, block, planes, blocks, scale=1):
+        upsample = None
+        if scale != 1 or self.inplanes != planes * block.expansion:
+            upsample = nn.Sequential(
+                resize_conv1x1(self.inplanes, planes * block.expansion, scale),
+                nn.BatchNorm2d(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(block(self.inplanes, planes, scale, upsample))
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(block(self.inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.linear(x)
+
+        # NOTE: replaced this by Linear(in_channels, 514 * 4 * 4)
+        # x = F.interpolate(x, scale_factor=4)
+
+        x = x.view(x.size(0), 512 * self.expansion, 4, 4)
+        x = self.upscale1(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.upscale(x)
+
+        return self.conv1(x)
+
+
+
+def resnet18_encoder(first_conv, maxpool1):
+    return ResNetEncoder(EncoderBlock, [2, 2, 2, 2], first_conv, maxpool1)
+
+
+
+def resnet18_decoder(latent_dim, input_height, first_conv, maxpool1):
+    return ResNetDecoder(DecoderBlock, [2, 2, 2, 2], latent_dim, input_height, first_conv, maxpool1)
+
+
+
+def resnet50_encoder(first_conv, maxpool1):
+    return ResNetEncoder(EncoderBottleneck, [3, 4, 6, 3], first_conv, maxpool1)
+
+
+
+def resnet50_decoder(latent_dim, input_height, first_conv, maxpool1):
+    return ResNetDecoder(DecoderBottleneck, [3, 4, 6, 3], latent_dim, input_height, first_conv, maxpool1)