写点什么

恒源云 _CIFAR-10 数据集实战:构建 ResNet18 神经网络

作者:恒源云
  • 2021 年 12 月 02 日
  • 本文字数:4829 字

    阅读完需:约 16 分钟

恒源云_CIFAR-10数据集实战:构建ResNet18神经网络

文章来源 | 恒源云社区


原文地址 | 数据集实战


原文作者 | Mathor




实不相瞒,小编我对平台社区内的大佬 Mathor 很崇拜!这不,今天又来给大家分享大佬论文笔记了,赶紧看看接下来的内容是否有你们需要的知识点吧!


正文开始:


如果不了解 ResNet 的同学可以先看我的这篇博客ResNet论文阅读



首先实现一个 Residual Block


import torchfrom torch import nnfrom torch.nn import functional as F
class ResBlk(nn.Module): def __init__(self, ch_in, ch_out, stride=1): super(ResBlk, self).__init__() self.conv1 = nn.Conv2d(ch_in, ch_out, kernel_size=3, stride=stride, padding=1) self.bn1 = nn.BatchNorm2d(ch_out) self.conv2 = nn.Conv2d(ch_out, ch_out, kernel_size=3, stride=1, padding=1) self.bn2 = nn.BatchNorm2d(ch_out) if ch_out == ch_in: self.extra = nn.Sequential() else: self.extra = nn.Sequential( # 1×1的卷积作用是修改输入x的channel # [b, ch_in, h, w] => [b, ch_out, h, w] nn.Conv2d(ch_in, ch_out, kernel_size=1, stride=stride), nn.BatchNorm2d(ch_out), ) def forward(self, x): out = F.relu(self.bn1(self.conv1(x))) out = self.bn2(self.conv2(out))
# short cut out = self.extra(x) + out out = F.relu(out) return out
复制代码


Block 中进行了正则化处理,以使 train 过程更快更稳定。同时要考虑,如果两元素的 ch_in 和 ch_out 不匹配,进行加法时会报错,因此需要判断一下,如果不想等,就用 1×1 的卷积调整一下


测试一下


blk = ResBlk(64, 128, stride=2)tmp = torch.randn(2, 64, 32, 32)out = blk(tmp)print(out.shape)
复制代码


输出的 shape 大小是torch.Size([2, 128, 16, 16])


这里解释一下,为什么有的层要专门设置 stride。先不考虑别的层,对于一个 Residual block,channel 从 64 增大到 128,如果所有的 stride 都是 1,padding 也是 1,那么图片的 w 和 h 也不会变,但是 channel 增大了,此时就会导致整个网络的参数增多。而这才仅仅一个 Block,更不用说后面的 FC 以及更多 Block 了,所以 stride 不能全部设置为 1,不要让网络的参数一直增大


然后我们搭建完整的 ResNet-18


class ResNet18(nn.Module):    def __init__(self):        super(ResNet18, self).__init__()                self.conv1 = nn.Sequential(            nn.Conv2d(3, 64, kernel_size=3, stride=3, padding=0),            nn.BatchNorm2d(64),        )        # followed 4 blocks                # [b, 64, h, w] => [b, 128, h, w]        self.blk1 = ResBlk(64, 128, stride=2)        # [b, 128, h, w] => [b, 256, h, w]        self.blk2 = ResBlk(128, 256, stride=2)        # [b, 256, h, w] => [b, 512, h, w]        self.blk3 = ResBlk(256, 512, stride=2)        # [b, 512, h, w] => [b, 512, h, w]        self.blk4 = ResBlk(512, 512, stride=2)                self.outlayer = nn.Linear(512*1*1, 10)        def forward(self, x):        x = F.relu(self.conv1(x))                # 经过四个blk以后 [b, 64, h, w] => [b, 512, h, w]        x = self.blk1(x)        x = self.blk2(x)        x = self.blk3(x)        x = self.blk4(x)                x = self.outlayer(x)                return x
复制代码


测试一下


x = torch.randn(2, 3, 32, 32)model = ResNet18()out = model(x)print("ResNet:", out.shape)
复制代码


结果报错了,错误信息如下


size mismatch, m1: [2048 x 2], m2: [512 x 10] at /pytorch/aten/src/TH/generic/THTensorMath.cpp:961
复制代码


问题在于我们最后定义线性层的输入维度,和上一层 Block 的输出维度不匹配,在 ResNet18 的最后一个 Block 运行结束后打印一下当前 x 的 shape,结果是torch.Size([2, 512, 2, 2])


解决办法有很多,可以修改线性层的输入进行匹配,也可以在最后一层 Block 后面再进行一些操作,使其与 512 匹配


先给出修改后的代码,在做解释


class ResNet18(nn.Module):    def __init__(self):        super(ResNet18, self).__init__()                self.conv1 = nn.Sequential(            nn.Conv2d(3, 64, kernel_size=3, stride=3, padding=0),            nn.BatchNorm2d(64),        )        # followed 4 blocks                # [b, 64, h, w] => [b, 128, h, w]        self.blk1 = ResBlk(64, 128, stride=2)        # [b, 128, h, w] => [b, 256, h, w]        self.blk2 = ResBlk(128, 256, stride=2)        # [b, 256, h, w] => [b, 512, h, w]        self.blk3 = ResBlk(256, 512, stride=2)        # [b, 512, h, w] => [b, 512, h, w]        self.blk4 = ResBlk(512, 512, stride=2)                self.outlayer = nn.Linear(512*1*1, 10)        def forward(self, x):        x = F.relu(self.conv1(x))                # 经过四个blk以后 [b, 64, h, w] => [b, 512, h, w]        x = self.blk1(x)        x = self.blk2(x)        x = self.blk3(x)        x = self.blk4(x)                # print("after conv:", x.shape) # [b, 512, 2, 2]                # [b, 512, h, w] => [b, 512, 1, 1]        x = F.adaptive_avg_pool2d(x, [1, 1])                x = x.view(x.size(0), -1) # [b, 512, 1, 1] => [b, 512*1*1]        x = self.outlayer(x)                return x
复制代码


这里我采用的是第二种方法,在最后一个 Block 结束以后,接了一个自适应的 pooling 层,这个 pooling 的作用是将不论输入的宽高是多少,全部输出称宽高都是 1 的 tensor,其他维度保持不变。然后再做一个 reshape 操作,将[batchsize, 512, 1, 1]reshape[batchsize, 512*1*1]大小的 tensor,这样就和接下来的线性层对上了,线性层的输入大小是 512,输出是 10。因此整个网络最终输出的 shape 就是[batchsize, 10]


最后我们把之前训练 LeNet5 的代码拷贝过来,将里面的model=LeNet5()改为model=ResNet18()就行了。完整代码如下


import torchfrom torch import nn, optimimport torch.nn.functional as Ffrom torch.utils.data import DataLoaderfrom torchvision import datasets, transforms

batch_size=32cifar_train = datasets.CIFAR10(root='cifar', train=True, transform=transforms.Compose([ transforms.Resize([32, 32]), transforms.ToTensor(),]), download=True)
cifar_train = DataLoader(cifar_train, batch_size=batch_size, shuffle=True)
cifar_test = datasets.CIFAR10(root='cifar', train=False, transform=transforms.Compose([ transforms.Resize([32, 32]), transforms.ToTensor(),]), download=True) cifar_test = DataLoader(cifar_test, batch_size=batch_size, shuffle=True)
class ResBlk(nn.Module): def __init__(self, ch_in, ch_out, stride=1): super(ResBlk, self).__init__() self.conv1 = nn.Conv2d(ch_in, ch_out, kernel_size=3, stride=stride, padding=1) self.bn1 = nn.BatchNorm2d(ch_out) self.conv2 = nn.Conv2d(ch_out, ch_out, kernel_size=3, stride=1, padding=1) self.bn2 = nn.BatchNorm2d(ch_out) if ch_out == ch_in: self.extra = nn.Sequential() else: self.extra = nn.Sequential( # 1×1的卷积作用是修改输入x的channel # [b, ch_in, h, w] => [b, ch_out, h, w] nn.Conv2d(ch_in, ch_out, kernel_size=1, stride=stride), nn.BatchNorm2d(ch_out), ) def forward(self, x): out = F.relu(self.bn1(self.conv1(x))) out = self.bn2(self.conv2(out))
# short cut out = self.extra(x) + out out = F.relu(out) return out class ResNet18(nn.Module): def __init__(self): super(ResNet18, self).__init__() self.conv1 = nn.Sequential( nn.Conv2d(3, 64, kernel_size=3, stride=3, padding=0), nn.BatchNorm2d(64), ) # followed 4 blocks # [b, 64, h, w] => [b, 128, h, w] self.blk1 = ResBlk(64, 128, stride=2) # [b, 128, h, w] => [b, 256, h, w] self.blk2 = ResBlk(128, 256, stride=2) # [b, 256, h, w] => [b, 512, h, w] self.blk3 = ResBlk(256, 512, stride=2) # [b, 512, h, w] => [b, 512, h, w] self.blk4 = ResBlk(512, 512, stride=2) self.outlayer = nn.Linear(512*1*1, 10) def forward(self, x): x = F.relu(self.conv1(x)) # 经过四个blk以后 [b, 64, h, w] => [b, 512, h, w] x = self.blk1(x) x = self.blk2(x) x = self.blk3(x) x = self.blk4(x) # print("after conv:", x.shape) # [b, 512, 2, 2] # [b, 512, h, w] => [b, 512, 1, 1] x = F.adaptive_avg_pool2d(x, [1, 1]) x = x.view(x.size(0), -1) # [b, 512, 1, 1] => [b, 512*1*1] x = self.outlayer(x) return x
def main():
########## train ########## #device = torch.device('cuda') #model = ResNet18().to(device) criteon = nn.CrossEntropyLoss() model = ResNet18() optimizer = optim.Adam(model.parameters(), 1e-3) for epoch in range(1000): model.train() for batchidx, (x, label) in enumerate(cifar_train): #x, label = x.to(device), label.to(device) logits = model(x) # logits: [b, 10] # label: [b] loss = criteon(logits, label) # backward optimizer.zero_grad() loss.backward() optimizer.step() print('train:', epoch, loss.item()) ########## test ########## model.eval() with torch.no_grad(): total_correct = 0 total_num = 0 for x, label in cifar_test: # x, label = x.to(device), label.to(device)
# [b] logits = model(x) # [b] pred = logits.argmax(dim=1) # [b] vs [b] total_correct += torch.eq(pred, label).float().sum().item() total_num += x.size(0) acc = total_correct / total_num print('test:', epoch, acc)
if __name__ == '__main__': main()
复制代码



ResNet 和 LeNet 相比,准确率提升的很快,但是由于层数增加,不可避免的会导致运行时间增加,如果没有 GPU,运行一个 epoch 大概要 15 分钟。读者同样可以在此基础上修改网络结构,运用一些 tricks,比方说一开始就对图片做一个 Normalize 等

用户头像

恒源云

关注

专注人工智能云GPU服务器训练平台 2020.12.25 加入

还未添加个人简介

评论

发布
暂无评论
恒源云_CIFAR-10数据集实战:构建ResNet18神经网络