resnet系列+mobilenet v2+pytorch代码实现

一.resnet系列backbone

import torch.nn as nn
import math
import torch.utils.model_zoo as model_zoo

BatchNorm2d = nn.BatchNorm2d

__all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'deformable_resnet18', 'deformable_resnet50',
           'resnet152']

model_urls = {
    'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
    'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',
    'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
    'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth',
    'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth',
}


def constant_init(module, constant, bias=0):
    nn.init.constant_(module.weight, constant)
    if hasattr(module, 'bias'):
        nn.init.constant_(module.bias, bias)


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):
    expansion = 1

    def __init__(self, inplanes, planes, stride=1, downsample=None, dcn=None):
        super(BasicBlock, self).__init__()
        self.with_dcn = dcn is not None
        self.conv1 = conv3x3(inplanes, planes, stride)
        self.bn1 = BatchNorm2d(planes)
        self.relu = nn.ReLU(inplace=True)
        self.with_modulated_dcn = False
        if not self.with_dcn:
            self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1, bias=False)
        else:
            from torchvision.ops import DeformConv2d
            deformable_groups = dcn.get('deformable_groups', 1)
            offset_channels = 18
            self.conv2_offset = nn.Conv2d(planes, deformable_groups * offset_channels, kernel_size=3, padding=1)
            self.conv2 = DeformConv2d(planes, planes, kernel_size=3, padding=1, bias=False)
        self.bn2 = BatchNorm2d(planes)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        residual = x

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

        if not self.with_dcn:
            out = self.conv2(out)
        else:
            offset = self.conv2_offset(out)
            out = self.conv2(out, offset)
        out = self.bn2(out)

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

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

        return out


class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, inplanes, planes, stride=1, downsample=None, dcn=None):
        super(Bottleneck, self).__init__()
        self.with_dcn = dcn is not None
        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
        self.bn1 = BatchNorm2d(planes)
        self.with_modulated_dcn = False
        if not self.with_dcn:
            self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
        else:
            deformable_groups = dcn.get('deformable_groups', 1)
            from torchvision.ops import DeformConv2d
            offset_channels = 18
            self.conv2_offset = nn.Conv2d(planes, deformable_groups * offset_channels, stride=stride, kernel_size=3, padding=1)
            self.conv2 = DeformConv2d(planes, planes, kernel_size=3, padding=1, stride=stride, bias=False)
        self.bn2 = BatchNorm2d(planes)
        self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
        self.bn3 = BatchNorm2d(planes * 4)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride
        self.dcn = dcn
        self.with_dcn = dcn is not None

    def forward(self, x):
        residual = x

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

        # out = self.conv2(out)
        if not self.with_dcn:
            out = self.conv2(out)
        else:
            offset = self.conv2_offset(out)
            out = self.conv2(out, offset)
        out = self.bn2(out)
        out = self.relu(out)

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

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

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

        return out


class ResNet(nn.Module):
    def __init__(self, block, layers, in_channels=3, dcn=None):
        self.dcn = dcn
        self.inplanes = 64
        super(ResNet, self).__init__()
        self.out_channels = []
        self.conv1 = nn.Conv2d(in_channels, 64, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, layers[0])
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dcn=dcn)
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, dcn=dcn)
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, dcn=dcn)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
            elif isinstance(m, BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()
        if self.dcn is not None:
            for m in self.modules():
                if isinstance(m, Bottleneck) or isinstance(m, BasicBlock):
                    if hasattr(m, 'conv2_offset'):
                        constant_init(m.conv2_offset, 0)

    def _make_layer(self, block, planes, blocks, stride=1, dcn=None):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion,
                          kernel_size=1, stride=stride, bias=False),
                BatchNorm2d(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample, dcn=dcn))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes, dcn=dcn))
        self.out_channels.append(planes * block.expansion)
        return nn.Sequential(*layers)

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

        x2 = self.layer1(x)
        x3 = self.layer2(x2)
        x4 = self.layer3(x3)
        x5 = self.layer4(x4)

        return x2, x3, x4, x5


def resnet18(pretrained=True, **kwargs):
    """Constructs a ResNet-18 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
    if pretrained:
        assert kwargs['in_channels'] == 3, 'in_channels must be 3 whem pretrained is True'
        print('load from imagenet')
        model.load_state_dict(model_zoo.load_url(model_urls['resnet18']), strict=False)
    return model


def deformable_resnet18(pretrained=True, **kwargs):
    """Constructs a ResNet-18 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(BasicBlock, [2, 2, 2, 2], dcn=dict(deformable_groups=1), **kwargs)
    if pretrained:
        assert kwargs['in_channels'] == 3, 'in_channels must be 3 whem pretrained is True'
        print('load from imagenet')
        model.load_state_dict(model_zoo.load_url(model_urls['resnet18']), strict=False)
    return model


def resnet34(pretrained=True, **kwargs):
    """Constructs a ResNet-34 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs)
    if pretrained:
        assert kwargs['in_channels'] == 3, 'in_channels must be 3 whem pretrained is True'
        model.load_state_dict(model_zoo.load_url(model_urls['resnet34']), strict=False)
    return model


def resnet50(pretrained=True, **kwargs):
    """Constructs a ResNet-50 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)
    if pretrained:
        assert kwargs['in_channels'] == 3, 'in_channels must be 3 whem pretrained is True'
        model.load_state_dict(model_zoo.load_url(model_urls['resnet50']), strict=False)
    return model


def deformable_resnet50(pretrained=True, **kwargs):
    """Constructs a ResNet-50 model with deformable conv.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(Bottleneck, [3, 4, 6, 3], dcn=dict(deformable_groups=1), **kwargs)
    if pretrained:
        assert kwargs['in_channels'] == 3, 'in_channels must be 3 whem pretrained is True'
        model.load_state_dict(model_zoo.load_url(model_urls['resnet50']), strict=False)
    return model


def resnet101(pretrained=True, **kwargs):
    """Constructs a ResNet-101 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(Bottleneck, [3, 4, 23, 3], **kwargs)
    if pretrained:
        assert kwargs['in_channels'] == 3, 'in_channels must be 3 whem pretrained is True'
        model.load_state_dict(model_zoo.load_url(model_urls['resnet101']), strict=False)
    return model


def resnet152(pretrained=True, **kwargs):
    """Constructs a ResNet-152 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs)
    if pretrained:
        assert kwargs['in_channels'] == 3, 'in_channels must be 3 whem pretrained is True'
        model.load_state_dict(model_zoo.load_url(model_urls['resnet152']), strict=False)
    return model


if __name__ == '__main__':
    import torch

    x = torch.zeros(2, 3, 640, 640)
    # net = deformable_resnet50(pretrained=False)
    model_resnet50 = resnet101(pretrained=False)
    y = model_resnet50(x)
    for u in y:
        print(u.shape)

    # print(model_resnet34.out_channels)

二.mobilenet v2

2.1基础知识

１．采用inverted residual，与resnet不一样的是通道１X1卷积先变宽－＞卷积提特征－>１X1卷积变窄,因为经过1x1的卷积扩大通道数以后，可以提升抽取特征的能力，图１所示。

２．最后不采用Relu，而使用Linear代替，因为降维后特征丢失部分，如果采用Relu还会丢失，图２所示.

 

　　　　　　　　　　　　　　　　图１　inverted residual

　　　　　　　　　　　　图２宽窄通道relu丢失信息对比

上图是嵌入高维空间的低维流形的ReLU变换示例。 在这些例子中，使用随机矩阵T和ReLU将第一个图的螺旋嵌入到n维空间中，然后使用T的逆矩阵投影回2D空间。 当n = 2，3导致信息损失很多，恢复后的张量坍缩严重，而对于n = 15到30，不会丢失太多的输入信息。显然当通道数较多时，如果输入流形可嵌入激活空间的显着较低维的子空间，则ReLU变换将保留信息。当通道数较少时，通道的信息很可能被丢弃。

综上所述，在通道数较少的层后，应该用线性激活代替ReLU。MobileNet V2的Linear bottleneck Inverted residual block中，降维后的1X1卷积层后接的是一个线性激活，其他情况用的是ReLU。 

v1使用可分离卷积的计算减少量,如果k=3，也就是３x3卷积核，能后减少９倍左右的计算量．(k^2+dj)/(k^2*dj)

2.2 代码实现

１.BottleNeck实现

import torch.nn as nn
import torch
class BottleNeck(nn.Module):
    def __init__(self, inchannles, outchannels, expansion=1, stride=1, downsample=None):
        super(BottleNeck, self).__init__()
        #1*1
        self.conv1 = nn.Conv2d(inchannles, inchannles*expansion, kernel_size=1)
        self.bn1 = nn.BatchNorm2d(inchannles*expansion)
        #3*3 可分离卷积　groups设置
        self.conv2 = nn.Conv2d(inchannles*expansion, inchannles * expansion, kernel_size=3, padding=1, stride=stride,groups=inchannles * expansion)
        self.bn2 = nn.BatchNorm2d(inchannles * expansion)
        #1*1
        self.conv3 = nn.Conv2d(inchannles*expansion, outchannels, kernel_size=1)
        self.bn3 = nn.BatchNorm2d(outchannels)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample

    def forward(self, x):
        residul = 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:
            residul = self.downsample(x)
        out += residul
        out = self.relu(out)
        return out
def BottleNeck_test():
    inchannels = 3
    outchannels = 3
    stride =2
    downsample_ = nn.Sequential(
        nn.Conv2d(inchannels, outchannels, kernel_size=1, stride=stride),
        nn.BatchNorm2d(outchannels)
    )
    bottleneck = BottleNeck(inchannels, outchannels, expansion=2, downsample=downsample_, stride=stride)
    print('bottleneck:', bottleneck)
    x = torch.rand((8, 3, 224, 224))
    out = bottleneck(x)
    print('out.shape', out.shape)



if __name__ == '__main__':
    BottleNeck_test()

２．整体代码，加了权重初始化

import torch.nn as nn
import torch
class BottleNeck(nn.Module):
    def __init__(self, inchannles, outchannels, expansion=1, stride=1, downsample=None):
        super(BottleNeck, self).__init__()
        #1*1
        self.conv1 = nn.Conv2d(inchannles, inchannles*expansion, kernel_size=1)
        self.bn1 = nn.BatchNorm2d(inchannles*expansion)
        #3*3 可分离卷积　groups设置
        self.conv2 = nn.Conv2d(inchannles*expansion, inchannles * expansion, kernel_size=3, padding=1, stride=stride,groups=inchannles * expansion)
        self.bn2 = nn.BatchNorm2d(inchannles * expansion)
        #1*1
        self.conv3 = nn.Conv2d(inchannles*expansion, outchannels, kernel_size=1)
        self.bn3 = nn.BatchNorm2d(outchannels)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample

    def forward(self, x):
        residul = 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:
            residul = self.downsample(x)
        out += residul
        out = self.relu(out)
        return out

class MobileNetV2(nn.Module):
    def __init__(self, n, numclasses=1000):
        super(MobileNetV2, self).__init__()
        self.inchannels = 32
        self.conv1 = nn.Conv2d(3, 32, kernel_size=3, stride=2, padding=1)
        self.bn1 = nn.BatchNorm2d(32)
        self.relu = nn.ReLU(inplace=True)
        self.layer1 = self.make_layer(n[0], outchannels=16, stride=1, expansion=1)
        self.layer2 = self.make_layer(n[1], outchannels=24, stride=2, expansion=6)
        self.layer3 = self.make_layer(n[2], outchannels=32, stride=2, expansion=6)
        self.layer4 = self.make_layer(n[3], outchannels=64, stride=2, expansion=6)
        self.layer5 = self.make_layer(n[4], outchannels=96, stride=1, expansion=6)
        self.layer6 = self.make_layer(n[5], outchannels=160, stride=2, expansion=6)
        self.layer7 = self.make_layer(n[6], outchannels=320, stride=1, expansion=1)
        self.conv8 = nn.Conv2d(320, 1280, kernel_size=1, stride=1)
        self.avegpool = nn.AvgPool2d(7, stride=1)
        self.conv9 = nn.Conv2d(1280, numclasses, kernel_size=1, stride=1)

    def make_layer(self, blocks_num, outchannels, stride, expansion):
        downsample_ = nn.Sequential(
            nn.Conv2d(self.inchannels, outchannels, kernel_size=1, stride=stride),
            nn.BatchNorm2d(outchannels)
        )
        layers = []
        #下采样的shortcut有downsample
        temp = BottleNeck(self.inchannels, outchannels, expansion=expansion, stride=stride, downsample=downsample_)
        layers.append(temp)
        #剩下的shortcut干净
        self.inchannels = outchannels
        for i in range(1, blocks_num):
            layers.append(BottleNeck(self.inchannels, outchannels, expansion=expansion, stride=1))
        return nn.Sequential(*layers)#取出每一层
    def forward(self, x):
        x = self.conv1(x)
        print('conv1.shape:', x.shape)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.layer1(x)
        print('layer1.shape:', x.shape)
        x = self.layer2(x)
        print('layer2.shape:', x.shape)
        x = self.layer3(x)
        print('layer3.shape:', x.shape)
        x = self.layer4(x)
        print('layer4.shape:', x.shape)
        x = self.layer5(x)
        print('layer5.shape:', x.shape)
        x = self.layer6(x)
        print('layer6.shape:', x.shape)
        x = self.layer7(x)
        print('layer7.shape:', x.shape)
        x = self.conv8(x)
        print('conv8.shape:', x.shape)
        x = self.avegpool(x)
        print('avegpool:', x.shape)
        x = self.conv9(x)
        print('conv9.shape:', x.shape)
        x = x.view(x.size(0), -1)
        return x

def BottleNeck_test():
    inchannels = 3
    outchannels = 3
    stride =2
    downsample_ = nn.Sequential(
        nn.Conv2d(inchannels, outchannels, kernel_size=1, stride=stride),
        nn.BatchNorm2d(outchannels)
    )
    bottleneck = BottleNeck(inchannels, outchannels, expansion=2, downsample=downsample_, stride=stride)
    print('bottleneck:', bottleneck)
    x = torch.rand((8, 3, 224, 224))
    out = bottleneck(x)
    print('out.shape', out.shape)

def weigth_init(m):
    if isinstance(m, nn.Conv2d):
        nn.init.xavier_uniform_(m.weight.data)
        nn.init.constant_(m.bias.data, 0.1)
    elif isinstance(m, nn.BatchNorm2d):
        m.weight.data.fill_(1)
        m.bias.data.zero_()
    elif isinstance(m, nn.Linear):
        m.weight.data.normal_(0,0.01)
        m.bias.data.zero_()
    # print('weigth_init!')
def MobileNetV2_test():
    model = MobileNetV2(n=[1, 2, 3, 4, 3, 3, 1], numclasses=10)
    model.apply(weigth_init)
    # print('model:', model)
    x = torch.rand((8, 3, 224, 224))
    out = model(x)
    print('out.shape', out.shape)
if __name__ == '__main__':
    # BottleNeck_test()
    MobileNetV2_test()

　　

1.0
==img.shape: torch.Size([1, 3, 256, 192])
==x.shape: torch.Size([1, 32, 128, 96])
==i, x.shape==: 0 torch.Size([1, 16, 128, 96])
==i, x.shape==: 1 torch.Size([1, 24, 64, 48])
==i, x.shape==: 2 torch.Size([1, 32, 32, 24])
==i, x.shape==: 3 torch.Size([1, 64, 16, 12])
==i, x.shape==: 4 torch.Size([1, 96, 16, 12])
==i, x.shape==: 5 torch.Size([1, 160, 8, 6])
==i, x.shape==: 6 torch.Size([1, 320, 8, 6])
==i, x.shape==: 7 torch.Size([1, 1280, 8, 6])

0.5

==img.shape: torch.Size([1, 3, 256, 192])
==x.shape: torch.Size([1, 16, 128, 96])
==i, x.shape==: 0 torch.Size([1, 8, 128, 96])
==i, x.shape==: 1 torch.Size([1, 16, 64, 48])
==i, x.shape==: 2 torch.Size([1, 16, 32, 24])
==i, x.shape==: 3 torch.Size([1, 32, 16, 12])
==i, x.shape==: 4 torch.Size([1, 48, 16, 12])
==i, x.shape==: 5 torch.Size([1, 80, 8, 6])
==i, x.shape==: 6 torch.Size([1, 160, 8, 6])
==i, x.shape==: 7 torch.Size([1, 640, 8, 6])

参考：

从MobileNet V1到MobileNet V2 - 知乎