PyTorch如何实现ResNet50、ResNet101和ResNet152

小编给大家分享一下PyTorch如何实现ResNet50、ResNet101和ResNet152，相信大部分人都还不怎么了解，因此分享这篇文章给大家参考一下，希望大家阅读完这篇文章后大有收获，下面让我们一起去了解一下吧！

PyTorch: https://github.com/shanglianlm0525/PyTorch-Networks

import torch
import torch.nn as nn
import torchvision
import numpy as np

print("PyTorch Version: ",torch.__version__)
print("Torchvision Version: ",torchvision.__version__)

__all__ = ['ResNet50', 'ResNet101','ResNet152']

def Conv1(in_planes, places, stride=2):
  return nn.Sequential(
    nn.Conv2d(in_channels=in_planes,out_channels=places,kernel_size=7,stride=stride,padding=3, bias=False),
    nn.BatchNorm2d(places),
    nn.ReLU(inplace=True),
    nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
  )

class Bottleneck(nn.Module):
  def __init__(self,in_places,places, stride=1,downsampling=False, expansion = 4):
    super(Bottleneck,self).__init__()
    self.expansion = expansion
    self.downsampling = downsampling

    self.bottleneck = nn.Sequential(
      nn.Conv2d(in_channels=in_places,out_channels=places,kernel_size=1,stride=1, bias=False),
      nn.BatchNorm2d(places),
      nn.ReLU(inplace=True),
      nn.Conv2d(in_channels=places, out_channels=places, kernel_size=3, stride=stride, padding=1, bias=False),
      nn.BatchNorm2d(places),
      nn.ReLU(inplace=True),
      nn.Conv2d(in_channels=places, out_channels=places*self.expansion, kernel_size=1, stride=1, bias=False),
      nn.BatchNorm2d(places*self.expansion),
    )

    if self.downsampling:
      self.downsample = nn.Sequential(
        nn.Conv2d(in_channels=in_places, out_channels=places*self.expansion, kernel_size=1, stride=stride, bias=False),
        nn.BatchNorm2d(places*self.expansion)
      )
    self.relu = nn.ReLU(inplace=True)
  def forward(self, x):
    residual = x
    out = self.bottleneck(x)

    if self.downsampling:
      residual = self.downsample(x)

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

class ResNet(nn.Module):
  def __init__(self,blocks, num_classes=1000, expansion = 4):
    super(ResNet,self).__init__()
    self.expansion = expansion

    self.conv1 = Conv1(in_planes = 3, places= 64)

    self.layer1 = self.make_layer(in_places = 64, places= 64, block=blocks[0], stride=1)
    self.layer2 = self.make_layer(in_places = 256,places=128, block=blocks[1], stride=2)
    self.layer3 = self.make_layer(in_places=512,places=256, block=blocks[2], stride=2)
    self.layer4 = self.make_layer(in_places=1024,places=512, block=blocks[3], stride=2)

    self.avgpool = nn.AvgPool2d(7, stride=1)
    self.fc = nn.Linear(2048,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.init.constant_(m.weight, 1)
        nn.init.constant_(m.bias, 0)

  def make_layer(self, in_places, places, block, stride):
    layers = []
    layers.append(Bottleneck(in_places, places,stride, downsampling =True))
    for i in range(1, block):
      layers.append(Bottleneck(places*self.expansion, places))

    return nn.Sequential(*layers)


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

def ResNet50():
  return ResNet([3, 4, 6, 3])

def ResNet101():
  return ResNet([3, 4, 23, 3])

def ResNet152():
  return ResNet([3, 8, 36, 3])


if __name__=='__main__':
  #model = torchvision.models.resnet50()
  model = ResNet50()
  print(model)

  input = torch.randn(1, 3, 224, 224)
  out = model(input)
  print(out.shape)

pytorch的优点

1.PyTorch是相当简洁且高效快速的框架；2.设计追求最少的封装；3.设计符合人类思维，它让用户尽可能地专注于实现自己的想法；4.与google的Tensorflow类似，FAIR的支持足以确保PyTorch获得持续的开发更新；5.PyTorch作者亲自维护的论坛 供用户交流和求教问题6.入门简单

以上是“PyTorch如何实现ResNet50、ResNet101和ResNet152”这篇文章的所有内容，感谢各位的阅读！相信大家都有了一定的了解，希望分享的内容对大家有所帮助，如果还想学习更多知识，欢迎关注亿速云行业资讯频道！