使用pytorch完成一个CNN入门demo

import

1. torch
   torch的总包
2. torch.nn
   网络层，通常自定义的网络都会继承nn.Module
3. torch.nn.functional
   F里面都包括常用的函数，relu和pooling等
4. torchvision
   常用的数据集，MNIST和FashionMNIST等
5. torchvision.transforms

数据集

使用FashionMNIST

train_set = torchvision.datasets.FashionMNIST(
    root='./data/FashionMNIST',
    train=True,
    download=False,
    transform=transforms.Compose([
        transforms.ToTensor()
    ])
)
train_dataloader = torch.utils.data.DataLoader(
    train_set,
    batch_size=10
)

建立网络

class nopaNet(nn.Module):
    def __init__(self):
        super(nopaNet, self).__init__()
        #输入图像[batch,1,28,28]
        #out_channel 代表有几个filter
        self.conv1 = nn.Conv2d(in_channels=1, out_channels=6, kernel_size=5)
        self.conv2 = nn.Conv2d(in_channels=6, out_channels=12, kernel_size=5)

        self.dense1 = nn.Linear(in_features=12*4*4, out_features=120)
        self.dense2 = nn.Linear(in_features=120, out_features=60)
        self.dense3 = nn.Linear(in_features=60, out_features=10)
    
    def forward(self, data):
        conv1_out = self.conv1(data)
        conv1_out = F.relu(conv1_out)
        conv1_out = F.max_pool2d(conv1_out, kernel_size=2, stride=2)

        conv2_out = self.conv2(conv1_out)
        conv2_out = F.relu(conv2_out)
        conv2_out = F.max_pool2d(conv2_out, kernel_size=2, stride=2)

        # flatten直接变成了1*1920
        # conv2_out = conv2_out.flatten()

        # 当batch是10的时候，实际上应该是10*192，不应该用flatten
        conv2_out = conv2_out.reshape(-1,12*4*4)

        dense1_out = self.dense1(conv2_out)
        dense1_out = F.relu(dense1_out)
        
        dense2_out = self.dense2(dense1_out)
        dense2_out = F.relu(dense2_out)

        final_out = self.dense3(dense2_out)

        return final_out

完整代码

import torch 
import torch.nn as nn
import torch.nn.functional as F
import torchvision
import torchvision.transforms as transforms

train_set = torchvision.datasets.FashionMNIST(
    root='./data/FashionMNIST',
    train=True,
    download=False,
    transform=transforms.Compose([
        transforms.ToTensor()
    ])
)

class nopaNet(nn.Module):
    def __init__(self):
        super(nopaNet, self).__init__()
        #输入图像[batch,1,28,28]
        #out_channel 代表有几个filter
        self.conv1 = nn.Conv2d(in_channels=1, out_channels=6, kernel_size=5)
        self.conv2 = nn.Conv2d(in_channels=6, out_channels=12, kernel_size=5)

        self.dense1 = nn.Linear(in_features=12*4*4, out_features=120)
        self.dense2 = nn.Linear(in_features=120, out_features=60)
        self.dense3 = nn.Linear(in_features=60, out_features=10)
    
    def forward(self, data):
        conv1_out = self.conv1(data)
        conv1_out = F.relu(conv1_out)
        conv1_out = F.max_pool2d(conv1_out, kernel_size=2, stride=2)

        conv2_out = self.conv2(conv1_out)
        conv2_out = F.relu(conv2_out)
        conv2_out = F.max_pool2d(conv2_out, kernel_size=2, stride=2)

        # flatten直接变成了1*1920
        # conv2_out = conv2_out.flatten()

        # 当batch是10的时候，实际上应该是10*192，不应该用flatten
        conv2_out = conv2_out.reshape(-1,12*4*4)

        dense1_out = self.dense1(conv2_out)
        dense1_out = F.relu(dense1_out)
        
        dense2_out = self.dense2(dense1_out)
        dense2_out = F.relu(dense2_out)

        final_out = self.dense3(dense2_out)

        return final_out

def get_num_correct(preds, labels):
    return preds.argmax(dim=1).eq(labels).sum().item()

if __name__ == "__main__":
    nopaNet = nopaNet()

    #batch为1的预测
    # #type(sample) = tuple
    # sample = next(iter(train_set))
    # image, label = sample
    # #shape = [1,28,28]
    # # print(image.shape)
    # #change image to a batch tensor
    # image = image.unsqueeze(0)
    # #shape = [1,1,28,28]
    # # print(image.shape)
    # pre = nopaNet(image)
    # print(pre)
    # #shape:[10]
    # print(pre.shape)
    # # #查看预测结果的第一种方式
    # # print('pre_label:',pre.argmax())
    # # print('true_label:',label)
    # #看预测结果第二种，归一化
    # print(F.softmax(pre))
    # print('pre_label:',pre.argmax())
    # print('true_label:',label)

    #batch为10的预测
    train_dataloader = torch.utils.data.DataLoader(
        train_set,
        batch_size=10
    )
    torch.set_grad_enabled(False)
    batch = next(iter(train_dataloader))
    # #batch是一个list，包括数据和标签
    # print(batch)
    images, labels = batch
    # #[10,1,28,28]
    # print(images.shape)
    # #[10]
    # print(labels.shape)
    pred_labels = nopaNet(images)
    ##[10,10]
    # print(pred_labels.shape)
    print('pred_labels:',pred_labels.argmax(dim=1))
    print('labels:',labels)
    #计算预测准确率
    print(pred_labels.argmax(dim=1).eq(labels))
    #print(pred_labels.argmax(dim=1).eq(labels).sum())
    print(get_num_correct(pred_labels,labels))
    
    


    print('succ')