paddle实现resnet复现

一、介绍

ResNet（Residual Neural Network）由微软研究院的Kaiming He等四名华人提出，通过使用ResNet Unit成功训练出了152层的神经网络，并在ILSVRC2015比赛中取得冠军，在top5上的错误率为3.57%，同时参数量比VGGNet低，效果非常突出。ResNet的结构可以极快的加速神经网络的训练，模型的准确率也有比较大的提升。

论文地址：“Deep_Residual_Learning_for_Image_Recognition”

二、对总体网络的介绍

1、网络亮点

网络提出了residual结构（残差结构），可有效缓解随网络层数的加深而导致的梯度消失和梯度爆炸现象。从而可以搭建超深的网络结构(突破1000层)。

2、网络构成

网络大体上由以下两种block构成：

Basicblock（用于resnet18，resnet34)

Bottleneckblock（用于resnet50，resnet101，resnet152）

3、网络结构

三、论文复现

3.1导入工具包

import paddle
import paddle.nn as nn
from paddle.nn import Conv2D, MaxPool2D, AdaptiveAvgPool2D, Linear, ReLU, BatchNorm2D
import paddle.nn.functional as F

3.2 建立Basicblock

值得注意的细节

在特征矩阵的大小和通道数发生改变时，应在shortcut中对特征矩阵进行卷积，使shortcut输出的特征矩阵和主分支的大小相同，便于相加。

class Basicblock(paddle.nn.Layer):
    def __init__(self, in_channel, out_channel, stride = 1):
        super(Basicblock, self).__init__()
        self.stride = stride
        self.conv0 = Conv2D(in_channel, out_channel, 3, stride = stride, padding = 1)
        self.conv1 = Conv2D(out_channel, out_channel, 3, stride=1, padding = 1)
        self.conv2 = Conv2D(in_channel, out_channel, 1, stride = stride)
        self.bn0 = BatchNorm2D(out_channel)
        self.bn1 = BatchNorm2D(out_channel)
        self.bn2 = BatchNorm2D(out_channel)

    def forward(self, inputs):
        y = inputs
        x = self.conv0(inputs)
        x = self.bn0(x)
        x = F.relu(x)
        x = self.conv1(x)
        x = self.bn1(x)
        if self.stride == 2:
            y = self.conv2(y)
            y = self.bn2(y)
        z = F.relu(x+y)
        return z

3.3 建立Bottleneckblock

值得注意的细节

基本同上

class Bottleneckblock(paddle.nn.Layer):
    def __init__(self, inplane, in_channel, out_channel, stride = 1, start = False):
        super(Bottleneckblock, self).__init__()
        self.stride = stride
        self.start = start
        self.conv0 = Conv2D(in_channel, inplane, 1, stride = stride)
        self.conv1 = Conv2D(inplane, inplane, 3, stride=1, padding=1)
        self.conv2 = Conv2D(inplane, out_channel, 1, stride=1)
        self.conv3 = Conv2D(in_channel, out_channel, 1, stride = stride)
        self.bn0 = BatchNorm2D(inplane)
        self.bn1 = BatchNorm2D(inplane)
        self.bn2 = BatchNorm2D(out_channel)
        self.bn3 = BatchNorm2D(out_channel)

    def forward(self, inputs):
        y = inputs
        x = self.conv0(inputs)
        x = self.bn0(x)
        x = F.relu(x)
        x = self.conv1(x)
        x = self.bn1(x)
        x = F.relu(x)
        x = self.conv2(x)
        x = self.bn2(x)
        if self.start:
            y = self.conv3(y)
            y = self.bn3(y)
        z = F.relu(x+y)
        return z

3.4 搭建Resnet的主干

值得注意的细节

1.在使用basic_layer的block中，第一个block保留原通道，之后的block因为承接上一个block所以通道数减半。

2.在使用bottleneck_layer的block中，在每一块block内部，输出block时通道深度为输入通道深度乘4，第一个block保留原通道，因为和第一个卷积池化后的通道数和输入通道数一样为64.

class Resnet(paddle.nn.Layer):
    def __init__(self, num, bottlenet):
        super(Resnet, self).__init__()
        self.conv0 = Conv2D(3, 64, 7, stride=2)
        self.bn = BatchNorm2D(64)
        self.pool1 = MaxPool2D(3, stride=2)
        if bottlenet:
            self.layer0 = self.add_bottleneck_layer(num[0], 64, start = True)
            self.layer1 = self.add_bottleneck_layer(num[1], 128)
            self.layer2 = self.add_bottleneck_layer(num[2], 256)
            self.layer3 = self.add_bottleneck_layer(num[3], 512)
        else:
            self.layer0 = self.add_basic_layer(num[0], 64, start = True)
            self.layer1 = self.add_basic_layer(num[1], 128)
            self.layer2 = self.add_basic_layer(num[2], 256)
            self.layer3 = self.add_basic_layer(num[3], 512)
        self.pool2 = AdaptiveAvgPool2D(output_size = (1, 1))

    def add_basic_layer(self, num, inplane, start = False):
        layer = []
        if start:
            layer.append(Basicblock(inplane, inplane))
        else:
            layer.append(Basicblock(inplane//2, inplane, stride = 2))
        for i in range(num-1):
            layer.append(Basicblock(inplane, inplane))
        return nn.Sequential(*layer)

    def add_bottleneck_layer(self, num, inplane, start = False):
        layer = []
        if start:
            layer.append(Bottleneckblock(inplane, inplane, inplane*4, start = True))
        else:
            layer.append(Bottleneckblock(inplane, inplane*2, inplane*4, stride = 2, start = True))
        for i in range(num-1):
            layer.append(Bottleneckblock(inplane, inplane*4, inplane*4))
        return nn.Sequential(*layer)

    def forward(self, inputs):
        x = self.conv0(inputs)
        x = self.bn(x)
        x = self.pool1(x)
        x = self.layer0(x)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)        
        x = self.pool2(x)
        x = paddle.squeeze(x)
        return x

3.5 具象化网络

def resnet18():
    return Resnet([2, 2, 2, 2], bottlenet = False)

def resnet34():
    return Resnet([3, 4, 6, 3], bottlenet = False)

def resnet50():
    return Resnet([3, 4, 6, 3], bottlenet = True)

def resnet101():
    return Resnet([3, 4, 23, 3], bottlenet = True)

def resnet152():
    return Resnet([3, 8, 36, 3], bottlenet = True)

四、查看网络结构

model = resnet18()
paddle.summary(model, (1, 3, 224, 224))

五、测试网络是否可以使用

import paddle
d = paddle.uniform(shape=[1,3,28,28])
print(type(d))
d = model(d)
print(d)