参考 http://pytorch123.com/

1. nn模块

import torch
N, D_in, Hidden_size, D_out = 64, 1000, 100, 10

• torch.nn.Sequential 建立模型，跟 keras 很像

x = torch.randn(N, D_in)
y = torch.randn(N, D_out)model = torch.nn.Sequential(torch.nn.Linear(D_in, Hidden_size),torch.nn.ReLU(),torch.nn.Linear(Hidden_size, D_out)
)# 损失函数
loss_fn = torch.nn.MSELoss(reduction='sum')

learning_rate = 1e-4
loss_list = []for t in range(500):y_pred = model(x) # 前向传播loss = loss_fn(y_pred, y) # 损失loss_list.append(loss.item())print(t, loss.item())model.zero_grad() # 清零梯度loss.backward() # 反向传播，计算梯度with torch.no_grad(): # 更新参数，不计入网络图的操作当中for param in model.parameters():param -= learning_rate*param.grad # 更新参数

# 绘制损失
import pandas as pd
loss_curve = pd.DataFrame(loss_list, columns=['loss'])
loss_curve.plot()

2. torch.optim 优化器

• torch.optim.Adam 使用优化器
• optimizer.zero_grad() # 清零梯度
• optimizer.step() # 更新参数

learning_rate = 1e-4optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)loss_list = []
for t in range(500):y_pred = model(x) # 前向传播loss = loss_fn(y_pred, y) # 损失loss_list.append(loss.item())print(t, loss.item())optimizer.zero_grad() # 清零梯度loss.backward() # 反向传播，计算梯度optimizer.step() # 更新参数

3. 自定义nn模块

• 继承 nn.module，并定义 forward 前向传播函数

import torch
class myModel(torch.nn.Module):def __init__(self, D_in, Hidden_size, D_out):super(myModel, self).__init__()self.fc1 = torch.nn.Linear(D_in, Hidden_size)self.fc2 = torch.nn.Linear(Hidden_size, D_out)def forward(self, x):x = self.fc1(x).clamp(min=0) # clamp 修剪数据在 min - max 之间，relu的作用x = self.fc2(x)return x

N, D_in, Hidden_size, D_out = 64, 1000, 100, 10
x = torch.randn(N, D_in)
y = torch.randn(N, D_out)model = myModel(D_in, Hidden_size, D_out) # 自定义模型loss_fn = torch.nn.MSELoss(reduction='sum')
optimizer = torch.optim.SGD(model.parameters(), lr=1e-4)loss_val = []for t in range(500):y_pred = model(x)loss = loss_fn(y_pred, y)loss_val.append(loss.item())optimizer.zero_grad()loss.backward()optimizer.step()import pandas as pd
loss_val = pd.DataFrame(loss_val, columns=['loss'])
loss_val.plot()

4. 权重共享

• 建立一个有3种FC层的玩具模型，中间 shareFC层会被 for 循环重复 0-3 次（随机），这几层（次数随机）的参数是共享的

import random
import torchclass shareParamsModel(torch.nn.Module):def __init__(self, D_in, Hidden_size, D_out):super(shareParamsModel, self).__init__()self.inputFC = torch.nn.Linear(D_in, Hidden_size)self.shareFC = torch.nn.Linear(Hidden_size, Hidden_size)self.outputFC = torch.nn.Linear(Hidden_size, D_out)self.sharelayers = 0 # 记录随机出了多少层def forward(self, x):x = self.inputFC(x).clamp(min=0)self.sharelayers = 0for _ in range(random.randint(0, 3)):x = self.shareFC(x).clamp(min=0)self.sharelayers += 1x = self.outputFC(x)return x

N, D_in, Hidden_size, D_out = 64, 1000, 100, 10
x = torch.randn(N, D_in)
y = torch.randn(N, D_out)model = shareParamsModel(D_in, Hidden_size, D_out)loss_fn = torch.nn.MSELoss(reduction='sum')optimizer = torch.optim.SGD(model.parameters(), lr=1e-4, momentum=0.9)loss_val = []for t in range(500):y_pred = model(x)print('share layers: ', model.sharelayers)loss = loss_fn(y_pred, y)loss_val.append(loss.item())optimizer.zero_grad()loss.backward()optimizer.step()for p in model.parameters():print(p.size())import pandas as pd
loss_val = pd.DataFrame(loss_val, columns=['loss'])
loss_val.plot()

输出：

share layers:  1
share layers:  0
share layers:  2
share layers:  1
share layers:  2
share layers:  1
share layers:  0
share layers:  1
share layers:  0
share layers:  0
share layers:  3
share layers:  3
。。。省略

参数数量，多次运行，均为以下结果

torch.Size([100, 1000])
torch.Size([100])
torch.Size([100, 100])
torch.Size([100])
torch.Size([10, 100])
torch.Size([10])