resnet.py

import torch
import torch.nn as nn
import torch.nn.functional as F


import torch
import torch.nn as nn
import torch.nn.functional as F


class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, in_planes, planes, norm_type, stride=1):
        super(BasicBlock, self).__init__()
        self.conv1 = nn.Conv2d(
            in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
        
        if norm_type == 'BN':
            self.bn1 = nn.BatchNorm2d(planes)
        else:
            self.bn1 = nn.GroupNorm(1, planes)
            
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3,
                               stride=1, padding=1, bias=False)

        if norm_type == 'BN':
            self.bn2 = nn.BatchNorm2d(planes)
        else:
            self.bn2 = nn.GroupNorm(1, planes)

        self.shortcut = nn.Sequential()
        if stride != 1 or in_planes != self.expansion*planes:
            self.shortcut = nn.Sequential(
                nn.Conv2d(in_planes, self.expansion*planes,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(self.expansion*planes) if norm_type == 'BN' else nn.GroupNorm(1,self.expansion*planes)
            )

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))
        out += self.shortcut(x)
        out = F.relu(out)
        return out

class ResNet(nn.Module):
    def __init__(self, block, num_blocks, norm_type, num_classes=10):
        super(ResNet, self).__init__()
        self.in_planes = 64

        self.conv1 = nn.Conv2d(3, 64, kernel_size=3,
                               stride=1, padding=1, bias=False)
        if norm_type == 'BN':
           self.bn1 = nn.BatchNorm2d(64)
        else:
           self.bn1 = nn.GroupNorm(1,64)
        
        self.layer1 = self._make_layer(block, 64, num_blocks[0], norm_type, stride=1)
        self.layer2 = self._make_layer(block, 128, num_blocks[1], norm_type, stride=1)
        self.layer3 = self._make_layer(block, 256, num_blocks[2], norm_type, stride=2)
        self.layer4 = self._make_layer(block, 512, num_blocks[3], norm_type, stride=2)
        self.linear = nn.Linear(512*block.expansion, num_classes)

    def _make_layer(self, block, planes, num_blocks, norm_type, stride):
        strides = [stride] + [1]*(num_blocks-1)
        layers = []
        for stride in strides:
            layers.append(block(self.in_planes, planes, norm_type, stride))
            self.in_planes = planes * block.expansion
        return nn.Sequential(*layers)

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)
        out = F.avg_pool2d(out, 8)
        out = out.view(out.size(0), -1)
        out = self.linear(out)
        return out


def ResNet18(norm_type='BN'):
    if norm_type == 'BN':
        return ResNet(BasicBlock, [2, 2, 2, 2],"BN")
    else:
        return ResNet(BasicBlock, [2, 2, 2, 2],"LN")

def ResNet34(norm_type='BN'):
    if norm_type == 'BN':
        return ResNet(BasicBlock, [3, 4, 6, 3],"BN")
    else:
        return ResNet(BasicBlock, [3, 4, 6, 3],"LN")