我是一名小白，最近学习pytorch，用pytorch复现一下吴的课后编程作业

一、 导入库

开始之前先导入库

import numpy as np
import h5py
import matplotlib.pyplot as plt
import tf_utils
import time
import torch
from torch import nn
import torch.nn.functional as F

二、 导入数据

X_train_orig , Y_train_orig , X_test_orig , Y_test_orig , classes = tf_utils.load_dataset()
X_train_flatten = X_train_orig.reshape(X_train_orig.shape[0],-1).T #每一列就是一个样本
X_test_flatten = X_test_orig.reshape(X_test_orig.shape[0],-1).T

#归一化数据
X_train = X_train_flatten / 255
X_test = X_test_flatten / 255

#转换成tensor
X_train = torch.tensor(X_train).float()
X_test = torch.tensor(X_test).float()
Y_train_orig = torch.tensor(Y_train_orig)
Y_test_orig = torch.tensor(Y_test_orig)

Pytorch有自带的独热编码函数torch.nn.functional.one_hot()

#转换为独热矩阵
Y_train = F.one_hot(Y_train_orig,6).squeeze()#移除长度为1的维度
Y_test = F.one_hot(Y_test_orig,6).squeeze()

三、模型搭建

X_train=X_train.T
X_test=X_test.T
#网络为：12288-25-12-6
class Net(nn.Module):
    def __init__(self):
        super(Net,self).__init__()
        self.layer=nn.Sequential(
            nn.Linear(12288,25),
            nn.ReLU(),
            nn.Linear(25,12),
            nn.ReLU(),
            nn.Linear(12,6)
        )
    def forward(self,x):
        x=self.layer(x)
        return x

net=Net()
loss=nn.CrossEntropyLoss()
optimizer=torch.optim.SGD(net.parameters(),lr=0.001)

注意

nn.Linear线性变换公式是

nn.CrossEntropyLoss(input,target)损失函数target参数不需要传one_hot向量，否则会报错1D target tensor expected, multi-target not supported

四、模型训练

def train(net,loss,optimizer,X_train,Y_train):
    for epoch in range(num_epochs):
        y=net(X_train)
        l=loss(input = y,target = Y_train_orig.squeeze())#input放的是独热编码，target放的是一维的label值
        optimizer.zero_grad()
        l.backward()
        optimizer.step()
        train_loss = l.item()
        correct=(y.argmax(dim=1)==Y_train_orig).sum().item()

        if (epoch + 1) % 1000 == 0:
            y_test=net(X_test)
            correct_test = (y_test.argmax(dim=1) == Y_test_orig).sum().item()
            print(
                'epoch {:d} loss={:.4f} 训练集正确率：{:.4f}% 测试集正确率：{:.4f}%'.format(epoch+1,train_loss,correct/X_train.shape[0]*100,correct_test/X_test.shape[0]*100)
                )
if __name__ =='__main__':
    num_epochs=10000
    train(net,loss,optimizer,X_train,Y_train)
    torch.save(net, './model.pth') #保存模型

训练到35000次训练集正确率已经达到100%，到时测试集正确率只有88%，说明存在较大过拟合。

训练十万次的结果如下：

五、结果测试

加载训练好的模型用自己的图片进行测试

my_image1 = "1.png"#定义图片名称
fileName1 = "./" + my_image1#图片地址
image1 =cv2.imread(fileName1)#读取图片
# plt.imshow(image1)
# plt.show()#显示图片
my_image1 = image1.reshape(1,64 * 64 * 3)
my_image1 = torch.tensor(my_image1).float()


model = torch.load('./model.pth') #加载模型

out=model(my_image1)
print(out)

运行结果：

测试结果来说不能说是不准确，但基本是毫不相关了......