res2net.py

import torch
import torch.nn as nn


__all__ = ['ImageNetRes2Net', 'res2net50', 'res2net101',
           'res2net152', 'res2next50_32x4d', 'se_res2net50',
           'CifarRes2Net', 'res2next29_6cx24wx4scale',
           'res2next29_8cx25wx4scale', 'res2next29_6cx24wx6scale',
           'res2next29_6cx24wx4scale_se', 'res2next29_8cx25wx4scale_se',
           'res2next29_6cx24wx6scale_se']


def conv3x3(in_planes, out_planes, stride=1, groups=1):
    """3x3 convolution with padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=1, groups=groups, 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 SEModule(nn.Module):
    def __init__(self, channels, reduction=16):
        super(SEModule, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.fc1 = nn.Conv2d(channels, channels // reduction, kernel_size=1, padding=0)
        self.relu = nn.ReLU(inplace=True)
        self.fc2 = nn.Conv2d(channels // reduction, channels, kernel_size=1, padding=0)
        self.sigmoid = nn.Sigmoid()

    def forward(self, input):
        x = self.avg_pool(input)
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        x = self.sigmoid(x)
        return input * x


class Res2NetBottleneck(nn.Module):
    expansion = 4

    def __init__(self, inplanes, planes, downsample=None, stride=1, scales=4, groups=1, se=False,  norm_layer=None):
        super(Res2NetBottleneck, self).__init__()
        if planes % scales != 0:
            raise ValueError('Planes must be divisible by scales')
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        bottleneck_planes = groups * planes
        self.conv1 = conv1x1(inplanes, bottleneck_planes, stride)
        self.bn1 = norm_layer(bottleneck_planes)
        self.conv2 = nn.ModuleList([conv3x3(bottleneck_planes // scales, bottleneck_planes // scales, groups=groups) for _ in range(scales-1)])
        self.bn2 = nn.ModuleList([norm_layer(bottleneck_planes // scales) for _ in range(scales-1)])
        self.conv3 = conv1x1(bottleneck_planes, planes * self.expansion)
        self.bn3 = norm_layer(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.se = SEModule(planes * self.expansion) if se else None
        self.downsample = downsample
        self.stride = stride
        self.scales = scales

    def forward(self, x):
        identity = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        xs = torch.chunk(out, self.scales, 1)
        ys = []
        for s in range(self.scales):
            if s == 0:
                ys.append(xs[s])
            elif s == 1:
                ys.append(self.relu(self.bn2[s-1](self.conv2[s-1](xs[s]))))
            else:
                ys.append(self.relu(self.bn2[s-1](self.conv2[s-1](xs[s] + ys[-1]))))
        out = torch.cat(ys, 1)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.se is not None:
            out = self.se(out)

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

        out += identity
        out = self.relu(out)

        return out


class ImageNetRes2Net(nn.Module):
    def __init__(self, layers, num_classes=1000, zero_init_residual=False,
                 groups=1, width=16, scales=4, se=False, norm_layer=None):
        super(ImageNetRes2Net, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        planes = [int(width * scales * 2 ** i) for i in range(4)]
        self.inplanes = planes[0]
        self.conv1 = nn.Conv2d(3, planes[0], kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = norm_layer(planes[0])
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(Res2NetBottleneck, planes[0], layers[0], scales=scales, groups=groups, se=se, norm_layer=norm_layer)
        self.layer2 = self._make_layer(Res2NetBottleneck, planes[1], layers[1], stride=2, scales=scales, groups=groups, se=se, norm_layer=norm_layer)
        self.layer3 = self._make_layer(Res2NetBottleneck, planes[2], layers[2], stride=2, scales=scales, groups=groups, se=se, norm_layer=norm_layer)
        self.layer4 = self._make_layer(Res2NetBottleneck, planes[3], layers[3], stride=2, scales=scales, groups=groups, se=se, norm_layer=norm_layer)
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(planes[3] * Res2NetBottleneck.expansion, num_classes)

        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)

        # 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, Res2NetBottleneck):
                    nn.init.constant_(m.bn3.weight, 0)

    def _make_layer(self, block, planes, blocks, stride=1, scales=4, groups=1, se=False, norm_layer=None):
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                norm_layer(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes, downsample, stride=stride, scales=scales, groups=groups, se=se, norm_layer=norm_layer))
        self.inplanes = planes * block.expansion
        for _ in range(1, blocks):
            layers.append(block(self.inplanes, planes, scales=scales, groups=groups, se=se, norm_layer=norm_layer))

        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)
        x = x.view(x.size(0), -1)
        x = self.fc(x)

        return x


class CifarRes2Net(nn.Module):
    def __init__(self, layers, num_classes=100, zero_init_residual=False,
                 groups=1, width=64, scales=4, se=False, norm_layer=None):
        super(CifarRes2Net, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d

        planes = [int(width * scales * 2 ** i) for i in range(3)]
        self.inplanes = planes[0]
        self.conv1 = conv3x3(3, planes[0])
        self.bn1 = norm_layer(planes[0])
        self.relu = nn.ReLU(inplace=True)
        self.layer1 = self._make_layer(Res2NetBottleneck, planes[0], layers[0], scales=scales, groups=groups, se=se, norm_layer=norm_layer)
        self.layer2 = self._make_layer(Res2NetBottleneck, planes[1], layers[1], stride=2, scales=scales, groups=groups, se=se, norm_layer=norm_layer)
        self.layer3 = self._make_layer(Res2NetBottleneck, planes[2], layers[2], stride=2, scales=scales, groups=groups, se=se, norm_layer=norm_layer)
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(planes[2] * Res2NetBottleneck.expansion, num_classes)

        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)

        # 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, Res2NetBottleneck):
                    nn.init.constant_(m.bn3.weight, 0)

    def _make_layer(self, block, planes, blocks, stride=1, scales=4, groups=1, se=False, norm_layer=None):
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                norm_layer(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes, downsample, stride=stride, scales=scales, groups=groups, se=se, norm_layer=norm_layer))
        self.inplanes = planes * block.expansion
        for _ in range(1, blocks):
            layers.append(block(self.inplanes, planes, scales=scales, groups=groups, se=se, norm_layer=norm_layer))

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)

        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        x = self.fc(x)

        return x


def res2net50(**kwargs):
    """Constructs a Res2Net-50 model.
    """
    model = ImageNetRes2Net([3, 4, 6, 3], **kwargs)
    return model


def res2net101(**kwargs):
    """Constructs a ResNet-101 model.
    """
    model = ImageNetRes2Net([3, 4, 23, 3], **kwargs)
    return model


def res2net152(**kwargs):
    """Constructs a ResNet-152 model.
    """
    model = ImageNetRes2Net([3, 8, 36, 3], **kwargs)
    return model


def res2next50_32x4d(**kwargs):
    """Constructs a Res2NeXt-50_32x4d model.
    """
    model = ImageNetRes2Net([3, 4, 6, 3], groups=32, width=4, **kwargs)
    return model


def res2next101_32x8d(**kwargs):
    """Constructs a Res2NeXt-101_32x8d model.
    """
    model = ImageNetRes2Net([3, 4, 23, 3], groups=32, width=8, **kwargs)
    return model


def se_res2net50(**kwargs):
    """Constructs a SE-Res2Net-50 model.
    """
    model = ImageNetRes2Net([3, 4, 6, 3], se=True, **kwargs)
    return model


def res2next29_6cx24wx4scale(**kwargs):
    """Constructs a Res2NeXt-29, 6cx24wx4scale model.
    """
    model = CifarRes2Net([3, 3, 3], groups=6, width=24, scales=4, **kwargs)
    return model


def res2next29_8cx25wx4scale(**kwargs):
    """Constructs a Res2NeXt-29, 8cx25wx4scale model.
    """
    model = CifarRes2Net([3, 3, 3], groups=8, width=25, scales=4, **kwargs)
    return model


def res2next29_6cx24wx6scale(**kwargs):
    """Constructs a Res2NeXt-29, 6cx24wx6scale model.
    """
    model = CifarRes2Net([3, 3, 3], groups=6, width=24, scales=6, **kwargs)
    return model

def res2next29_6cx24wx4scale_se(**kwargs):
    """Constructs a Res2NeXt-29, 6cx24wx4scale-SE model.
    """
    model = CifarRes2Net([3, 3, 3], groups=6, width=24, scales=4, se=True, **kwargs)
    return model


def res2next29_8cx25wx4scale_se(**kwargs):
    """Constructs a Res2NeXt-29, 8cx25wx4scale-SE model.
    """
    model = CifarRes2Net([3, 3, 3], groups=8, width=25, scales=4, se=True, **kwargs)
    return model


def res2next29_6cx24wx6scale_se(**kwargs):
    """Constructs a Res2NeXt-29, 6cx24wx6scale-SE model.
    """
    model = CifarRes2Net([3, 3, 3], groups=6, width=24, scales=6, se=True, **kwargs)
    return model