FeatureExtractors.py

# Original Author: Jonathan Donnellya ([email protected])
# Modified by Mohammad Reza Taesiri ([email protected])

import os
import torch
import torch.nn as nn
from collections import OrderedDict

model_dir = os.path.dirname(os.path.realpath(__file__))


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)


class BasicBlock(nn.Module):
    # class attribute
    expansion = 1
    num_layers = 2

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(BasicBlock, self).__init__()
        # only conv with possibly not 1 stride
        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)

        # if stride is not 1 then self.downsample cannot be None
        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)

        if self.downsample is not None:
            identity = self.downsample(x)

        # the residual connection
        out += identity
        out = self.relu(out)

        return out

    def block_conv_info(self):
        block_kernel_sizes = [3, 3]
        block_strides = [self.stride, 1]
        block_paddings = [1, 1]

        return block_kernel_sizes, block_strides, block_paddings


class Bottleneck(nn.Module):
    # class attribute
    expansion = 4
    num_layers = 3

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = conv1x1(inplanes, planes)
        self.bn1 = nn.BatchNorm2d(planes)
        # only conv with possibly not 1 stride
        self.conv2 = conv3x3(planes, planes, stride)
        self.bn2 = nn.BatchNorm2d(planes)
        self.conv3 = conv1x1(planes, planes * self.expansion)
        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)

        # if stride is not 1 then self.downsample cannot be None
        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
        out = self.relu(out)

        return out

    def block_conv_info(self):
        block_kernel_sizes = [1, 3, 1]
        block_strides = [1, self.stride, 1]
        block_paddings = [0, 1, 0]

        return block_kernel_sizes, block_strides, block_paddings


class ResNet_features(nn.Module):
    """
    the convolutional layers of ResNet
    the average pooling and final fully convolutional layer is removed
    """

    def __init__(self, block, layers, num_classes=1000, zero_init_residual=False):
        super(ResNet_features, self).__init__()

        self.inplanes = 64

        # the first convolutional layer before the structured sequence of blocks
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)
        self.bn1 = nn.BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        # comes from the first conv and the following max pool
        self.kernel_sizes = [7, 3]
        self.strides = [2, 2]
        self.paddings = [3, 1]

        # the following layers, each layer is a sequence of blocks
        self.block = block
        self.layers = layers
        self.layer1 = self._make_layer(
            block=block, planes=64, num_blocks=self.layers[0]
        )
        self.layer2 = self._make_layer(
            block=block, planes=128, num_blocks=self.layers[1], stride=2
        )
        self.layer3 = self._make_layer(
            block=block, planes=256, num_blocks=self.layers[2], stride=2
        )
        self.layer4 = self._make_layer(
            block=block, planes=512, num_blocks=self.layers[3], stride=2
        )

        # initialize the parameters
        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)

        # Zero-initialize the last BN in each residual branch,
        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, Bottleneck):
                    nn.init.constant_(m.bn3.weight, 0)
                elif isinstance(m, BasicBlock):
                    nn.init.constant_(m.bn2.weight, 0)

    def _make_layer(self, block, planes, num_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 = []
        # only the first block has downsample that is possibly not None
        layers.append(block(self.inplanes, planes, stride, downsample))

        self.inplanes = planes * block.expansion
        for _ in range(1, num_blocks):
            layers.append(block(self.inplanes, planes))

        # keep track of every block's conv size, stride size, and padding size
        for each_block in layers:
            (
                block_kernel_sizes,
                block_strides,
                block_paddings,
            ) = each_block.block_conv_info()
            self.kernel_sizes.extend(block_kernel_sizes)
            self.strides.extend(block_strides)
            self.paddings.extend(block_paddings)

        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)

        return x

    def conv_info(self):
        return self.kernel_sizes, self.strides, self.paddings

    def num_layers(self):
        """
        the number of conv layers in the network, not counting the number
        of bypass layers
        """

        return (
            self.block.num_layers * self.layers[0]
            + self.block.num_layers * self.layers[1]
            + self.block.num_layers * self.layers[2]
            + self.block.num_layers * self.layers[3]
            + 1
        )

    def __repr__(self):
        template = "resnet{}_features"
        return template.format(self.num_layers() + 1)


def resnet50_features(pretrained=True, inat=True, **kwargs):
    """Constructs a ResNet-50 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet or iNaturalist
        pretrained (bool): If True, returns a model pre-trained on iNaturalst; else, ImageNet
    """
    model = ResNet_features(Bottleneck, [3, 4, 6, 4], **kwargs)
    if pretrained:
        if inat:
            # print('Loading iNat model')
            model_dict = torch.load(
                model_dir
                + "/../../weights/"
                + "BBN.iNaturalist2017.res50.90epoch.best_model.pth.pt"
            )
        else:
            raise

        if inat:
            model_dict.pop("module.classifier.weight")
            model_dict.pop("module.classifier.bias")
            for key in list(model_dict.keys()):
                model_dict[
                    key.replace("module.backbone.", "")
                    .replace("cb_block", "layer4.2")
                    .replace("rb_block", "layer4.3")
                ] = model_dict.pop(key)

        else:
            raise

        model.load_state_dict(model_dict, strict=False)

    return model


class ResNet_classifier(nn.Module):
    """
    A classifier for Deformable ProtoPNet
    """

    def __init__(self, block, layers, num_classes=1000, zero_init_residual=False):
        super(ResNet_classifier, self).__init__()

        self.inplanes = 64

        # the first convolutional layer before the structured sequence of blocks
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)
        self.bn1 = nn.BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        # comes from the first conv and the following max pool
        self.kernel_sizes = [7, 3]
        self.strides = [2, 2]
        self.paddings = [3, 1]

        # the following layers, each layer is a sequence of blocks
        self.block = block
        self.layers = layers
        self.layer1 = self._make_layer(
            block=block, planes=64, num_blocks=self.layers[0]
        )
        self.layer2 = self._make_layer(
            block=block, planes=128, num_blocks=self.layers[1], stride=2
        )
        self.layer3 = self._make_layer(
            block=block, planes=256, num_blocks=self.layers[2], stride=2
        )
        self.layer4 = self._make_layer(
            block=block, planes=512, num_blocks=self.layers[3], stride=2
        )

        self.classifier = nn.Linear(2048 * 7 * 7, 200)

        # initialize the parameters
        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)

        # Zero-initialize the last BN in each residual branch,
        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, Bottleneck):
                    nn.init.constant_(m.bn3.weight, 0)
                elif isinstance(m, BasicBlock):
                    nn.init.constant_(m.bn2.weight, 0)

    def _make_layer(self, block, planes, num_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 = []
        # only the first block has downsample that is possibly not None
        layers.append(block(self.inplanes, planes, stride, downsample))

        self.inplanes = planes * block.expansion
        for _ in range(1, num_blocks):
            layers.append(block(self.inplanes, planes))

        # keep track of every block's conv size, stride size, and padding size
        for each_block in layers:
            (
                block_kernel_sizes,
                block_strides,
                block_paddings,
            ) = each_block.block_conv_info()
            self.kernel_sizes.extend(block_kernel_sizes)
            self.strides.extend(block_strides)
            self.paddings.extend(block_paddings)

        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.classifier(torch.flatten(x, start_dim=1))
        return x

    def conv_info(self):
        return self.kernel_sizes, self.strides, self.paddings

    def num_layers(self):
        """
        the number of conv layers in the network, not counting the number
        of bypass layers
        """

        return (
            self.block.num_layers * self.layers[0]
            + self.block.num_layers * self.layers[1]
            + self.block.num_layers * self.layers[2]
            + self.block.num_layers * self.layers[3]
            + 1
        )

    def __repr__(self):
        template = "resnet{}_features"
        return template.format(self.num_layers() + 1)