pytorch-01

时间:2024-07-06 07:17:44

加载mnist数据集

one-hot编码实现

import numpy as np
import torch
x_train = np.load("../dataset/mnist/x_train.npy") # 从网站提前下载数据集,并解压缩
y_train_label = np.load("../dataset/mnist/y_train_label.npy")
x = torch.tensor(y_train_label[:5],dtype=torch.int64)  # 获取前5个样本的标签数据
# 定义一个张量输入,因为此时有 5 个数值,且最大值为9,类别数为10
# 所以我们可以得到 y 的输出结果的形状为 shape=(5,10),即5行12列
y = torch.nn.functional.one_hot(x, 10)  # 一个参数张量x,10为类别数
print(y)

对于拥有6000个样本的MNIST数据集来说,标签就是一个6000\times 10大小的矩阵张量。

多层感知机模型

#设定的多层感知机网络模型
class NeuralNetwork(torch.nn.Module):
    def __init__(self):
        super(NeuralNetwork, self).__init__()
        self.flatten = torch.nn.Flatten()  # 拉平图像矩阵
        self.linear_relu_stack = torch.nn.Sequential(
            torch.nn.Linear(28*28,312),   # 输入大小为28*28,输出大小为312维的线性变换层
            torch.nn.ReLU(),   # 激活函数层
            torch.nn.Linear(312, 256),
            torch.nn.ReLU(),
            torch.nn.Linear(256, 10)  # 最终输出大小为10,对应one-hot标签维度
        )
    def forward(self, input):   # 构建网络
        x = self.flatten(input)  #拉平矩阵为1维
        logits = self.linear_relu_stack(x) # 多层感知机

        return logits

损失函数

优化函数

model = NeuralNetwork()
loss_fu = torch.nn.CrossEntropyLoss() # 交叉熵损失函数,内置了softmax函数,
optimizer = torch.optim.Adam(model.parameters(), lr=2e-5)   #设定优化函数

loss = loss_fu(pred,label_batch)  # 计算损失

完整模型

import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0' #指定GPU编
import torch
import numpy as np


batch_size = 320                        #设定每次训练的批次数
epochs = 1024                           #设定训练次数

#device = "cpu"                         #Pytorch的特性,需要指定计算的硬件,如果没有GPU的存在,就使用CPU进行计算
device = "cuda"                         #在这里读者默认使用GPU,如果读者出现运行问题可以将其改成cpu模式


#设定的多层感知机网络模型
class NeuralNetwork(torch.nn.Module):
    def __init__(self):
        super(NeuralNetwork, self).__init__()
        self.flatten = torch.nn.Flatten()
        self.linear_relu_stack = torch.nn.Sequential(
            torch.nn.Linear(28*28,312),
            torch.nn.ReLU(),
            torch.nn.Linear(312, 256),
            torch.nn.ReLU(),
            torch.nn.Linear(256, 10)
        )
    def forward(self, input):
        x = self.flatten(input)
        logits = self.linear_relu_stack(x)

        return logits

model = NeuralNetwork()
model = model.to(device)                #将计算模型传入GPU硬件等待计算
torch.save(model, './model.pth')
#model = torch.compile(model)            #Pytorch2.0的特性,加速计算速度
loss_fu = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=2e-5)   #设定优化函数

#载入数据
x_train = np.load("../../dataset/mnist/x_train.npy")
y_train_label = np.load("../../dataset/mnist/y_train_label.npy")

train_num = len(x_train)//batch_size

#开始计算
for epoch in range(20):
    train_loss = 0
    for i in range(train_num):
        start = i * batch_size
        end = (i + 1) * batch_size

        train_batch = torch.tensor(x_train[start:end]).to(device)
        label_batch = torch.tensor(y_train_label[start:end]).to(device)

        pred = model(train_batch)
        loss = loss_fu(pred,label_batch)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        train_loss += loss.item()  # 记录每个批次的损失值

    # 计算并打印损失值
    train_loss /= train_num
    accuracy = (pred.argmax(1) == label_batch).type(torch.float32).sum().item() / batch_size
    print("epoch:",epoch,"train_loss:", round(train_loss,2),"accuracy:",round(accuracy,2))

可视化模型结构和参数

model = NeuralNetwork()
print(model)

是对模型具体使用的函数及其对应的参数进行打印。

格式化显示:

param = list(model.parameters())
k=0
for i in param:
    l = 1
    print('该层结构:'+str(list(i.size())))
    for j in i.size():
        l*=j
    print('该层参数和:'+str(l))
    k = k+l
print("总参数量:"+str(k))

模型保存

model = NeuralNetwork()
torch.save(model, './model.pth')

netron可视化

安装:pip install netron

运行:命令行输入netron

打开:通过网址http://localhost:8080打开

打开保存的模型文件model.pth:

 

 点击颜色块,可以显示详细信息:

相关文章