（刘二大人）PyTorch深度学习实践-卷积网络（基础篇）

目录

1. 简单卷积实现（主要看下shape）

1.1 代码展示

1.2 结果展示

2. 关于padding参数

2.1代码实现

2.2 结果展示

3. 关于stride参数(降低输出的宽和高)

3.1 代码展示（就是修改了padding换成stride）

3.2 结果展示

4. 使用卷积网络实现Minist数据集

4.1 代码展示

4.2 结果展示（5轮，性能提高一个点）

---

1. 简单卷积实现（主要看下shape）

 1.1 代码展示

import torch
input_channels , output_channels = 5,10
batch_size = 1
kernel_size = 3
width , height = 100,100

# BxCxWxH 随机定义一个满足要求的输入
input = torch.randn(batch_size,input_channels,width,height)

#定义卷积层
conv_layer = torch.nn.Conv2d(in_channels=input_channels,out_channels=output_channels,kernel_size=kernel_size)

output = conv_layer(input)

print(input.shape)
print(output.shape)
print(conv_layer.weight.shape)

1.2 结果展示

D:\Anaconda3\envs\pytorch\python.exe E:/learn_pytorch/LE/convolution_base1.py
torch.Size([1, 5, 100, 100])
torch.Size([1, 10, 98, 98])
torch.Size([10, 5, 3, 3])

Process finished with exit code 0
 

2. 关于padding参数

有时为了保证输入和输出size相同，会使用padding参数，若padding=1，则在输入矩阵外层添加一圈数据，一般默认为0（可以扩大输出的宽和高）

 例如：

 2.1代码实现

import torch

input = [
    3,4,6,5,7,
    2,4,6,8,2,
    1,6,7,8,4,
    9,7,4,6,2,
    3,7,5,4,1,
]

#将input转换为1*1*5*5的张量
input  = torch.Tensor(input).view(1,1,5,5)

kernel_size = 3
batch_size = 1
output_channels = 1

#定义卷积层
conv_layer = torch.nn.Conv2d(in_channels=1,out_channels=output_channels,kernel_size=kernel_size,padding=1,bias=False)

#自定义卷积核
kernel = torch.Tensor([1,2,3,4,5,6,7,8,9]).view(1,1,3,3)

#将自定义的卷积核的值赋值给卷积层的权重
conv_layer.weight.data = kernel.data

output = conv_layer(input)

print(input.shape)
print(output.shape)
print(conv_layer.weight.shape)

2.2 结果展示

D:\Anaconda3\envs\pytorch\python.exe E:/learn_pytorch/LE/convolution_padding.py
torch.Size([1, 1, 5, 5])
torch.Size([1, 1, 5, 5])
torch.Size([1, 1, 3, 3])

Process finished with exit code 0

3. 关于stride参数(降低输出的宽和高)

 3.1 代码展示（就是修改了padding换成stride）

import torch

input = [
    3,4,6,5,7,
    2,4,6,8,2,
    1,6,7,8,4,
    9,7,4,6,2,
    3,7,5,4,1,
]

#将input转换为1*1*5*5的张量
input  = torch.Tensor(input).view(1,1,5,5)

kernel_size = 3
batch_size = 1
output_channels = 1

#定义卷积层
conv_layer = torch.nn.Conv2d(in_channels=1,out_channels=output_channels,kernel_size=kernel_size,stride=2,bias=False)

#自定义卷积核
kernel = torch.Tensor([1,2,3,4,5,6,7,8,9]).view(1,1,3,3)

#将自定义的卷积核的值赋值给卷积层的权重
conv_layer.weight.data = kernel.data

output = conv_layer(input)

print(input.shape)
print(output.shape)
print(conv_layer.weight.shape)

3.2 结果展示

D:\Anaconda3\envs\pytorch\python.exe E:/learn_pytorch/LE/convolution_padding.py
torch.Size([1, 1, 5, 5])
torch.Size([1, 1, 2, 2])
torch.Size([1, 1, 3, 3])

Process finished with exit code 0

4. 使用卷积网络实现Minist数据集

4.1 代码展示

import torch
from torchvision import datasets
from torchvision import transforms
from torch.utils.data import DataLoader

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

#准备数据集
trans = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.1307,),(0.3801,))])#这里第一个是均值，第二个是标准差
train_datasets = datasets.MNIST(root='E:\learn_pytorch\LE',train=True,transform=trans,download=True)
test_datasets = datasets.MNIST(root='E:\learn_pytorch\LE',train=False,transform=trans,download=True)

#进行数据集的加载
batch_size = 64
train_loader = DataLoader(dataset=train_datasets,batch_size=batch_size,shuffle=True)
test_loader = DataLoader(dataset=test_datasets,batch_size=batch_size,shuffle=False)

class Conv_model(torch.nn.Module):
    def __init__(self):
        super(Conv_model, self).__init__()
        self.conv1 = torch.nn.Conv2d(1,10,(5,5))#输入通道、输出通道、kernel_size
        self.conv2 = torch.nn.Conv2d(10,20,(5,5))
        self.pooling = torch.nn.MaxPool2d(kernel_size=2)#最大池化层
        self.fc = torch.nn.Linear(320,10)#这里就是一个全连接的分类

    def forward(self,x):
        x = F.relu(self.conv1(x))
        x = self.pooling(x)
        x = F.relu(self.conv2(x))
        x = self.pooling(x)
        x = x.view(-1,320) #这里我试了不可以用torch.nn.Flatten，类型不匹配
        x = self.fc(x)
        return x

Hui = Conv_model()

#定义损失函数和优化器
loss_fn = torch.nn.CrossEntropyLoss()
optimizer = optim.SGD(Hui.parameters(),lr=0.01,momentum=0.5)

#我们将一轮epoch单独拿出来作为一个函数
def train(epoch):
    running_loss = 0.0
    for batch_id,data in enumerate(train_loader,0):
        inputs,targets = data
        optimizer.zero_grad()

        # Forward
        outputs = Hui(inputs)
        loss = loss_fn(outputs,targets)
        loss.backward()
        optimizer.step()
        #标签从0开始
        running_loss+=loss.item()
        if batch_id%300 == 299:
            print('[%d,%5d]  loss:%.3f'%(epoch+1,batch_id+1,running_loss/300))
            running_loss = 0.0


#定义测试集
def test():
    correct = 0
    total = 0
    with torch.no_grad():
        for data in test_loader:
            images,labels = data
            outputs = Hui(images)
            #这个torch.max函数可以返回最大值和最大值的下标,那个predict取的是最大值下标
            _,predict = torch.max(outputs.data,dim=1)
            total += labels.size(0)#总共有多少个标签样本
            correct+=(predict==labels).sum().item()#将我们预测的最有可能的下标与真实标签对比，最后将这个标量取出来
        print('Accuracy on test set: %d %%' % (100*correct/total))


#进行训练和测试
if __name__ == '__main__':
    for epoch in range(5):
        train(epoch)
        test()

4.2 结果展示（5轮，性能提高一个点）

D:\Anaconda3\envs\pytorch\python.exe E:/learn_pytorch/LE/conv_minist.py
[1,  300]  loss:0.671
[1,  600]  loss:0.192
[1,  900]  loss:0.145
Accuracy on test set: 96 %
[2,  300]  loss:0.111
[2,  600]  loss:0.106
[2,  900]  loss:0.094
Accuracy on test set: 97 %
[3,  300]  loss:0.086
[3,  600]  loss:0.075
[3,  900]  loss:0.076
Accuracy on test set: 97 %
[4,  300]  loss:0.069
[4,  600]  loss:0.064
[4,  900]  loss:0.065
Accuracy on test set: 98 %
[5,  300]  loss:0.051
[5,  600]  loss:0.064
[5,  900]  loss:0.055
Accuracy on test set: 98 %

Process finished with exit code 0