2. PyTorch Tutorial 1

Posted 2023-01-06 Shannnon_sun

文章目录

• PyTorch Tutorial 1
• • 1. Dataset&DataLoader
  • 2. Tensors
  • 3. 定义模型：
  • 4. 定义Loss Function
  • 5. 优化
  • 6. Neural Network Training Setup
  • 7. Neural Network Training Loop
  • 8. Neural Network Validation Loop
  • 9. Neural Network Testing Loop
  • 10. Save/Load Trained Models

PyTorch Tutorial 1

An machine learning framework

1. Dataset&DataLoader

from torch.utils.data import Dataset,DataLoader

class MyDataset(Dataset):
  def _init_(self,file):
  def _getitem_(self,index):
  def _len_(self):

2. Tensors

• x.shape()

• 创建Tensor

• 运算：x.transpose(0,1) 转置 x.squeeze(0) torch.cat([x,y,z],dim=0) 在某一维数上连接

• Type：

• Device:

  x=x.to(‘cpu’) x=x.to(‘gpu’)

3. 定义模型：

import torch.nn as nn

class Mymodole(nn.Module):
  def _init_(self):
    super(Mymodel,self)._init_()
    self.net=nn.sequential(
    	nn.Linear(10,32)
    	nn.Sigmoid()
    	nn.Linear(32,1)
    )

4. 定义Loss Function

在torch.nn

nn.MSELoss()
nn.CrossEntropyLoss()

5. 优化

在torch.optim

torch.optim.SGD(model.parameters(),lr,momentum=0)

6. Neural Network Training Setup

dataset = MyDataset(file)
tr_set = DataLoader(dataset,16,shuffle=True) #dataset 放入DataLoader
model = MyModel().to(device)	#模型放到（cpu/cude）
criterion = nn.MSELoss()  #定义损失函数
optimizer = torch.optim.SGD(model.parameters(),0.1)	#优化器

7. Neural Network Training Loop

for epoch in range(n_epoches):
  model.train()			#训练
  for x,y in tr_set:
    optimizer.zero_grad()
    x,y=x.to(device),y.to(device)
    pred=model(x)	#前向传播计算输出
    loss = criterion(pred,y) #计算Loss
    loss.backward() #反向传播计算梯度
    optimizer.step()	#更新参数

8. Neural Network Validation Loop

model.eval()	#模型调到测试模式
total_loss = 0
for x,y in dv_set:
  x,y = x.to(device),y.to(device)
  with torch.no_grad():			#disable gradient calculation
    pred = model(x)
    loss = ceiterion(pred,y)
    total_loss +=loss.cpu().item()*len(x)
    avg_loss = total_loss/len(dv_set.dataset) 

9. Neural Network Testing Loop

为什么把梯度的计算关掉？

• 计算快一点
• 确保不会更新模型

10. Save/Load Trained Models

• Save

  torch.save(model.state_dict(),path)

• Load

  ckpt = torch.load(path)
  model.load_state_dict(ckpt)