06_PyTorch model training [learning rate and optimizer base class]

When the data, model and loss function are determined, the mathematical model of the task has been determined, and then a suitable one must be selected.

The Optimizer optimizes the model.

All optimizers in PyTorch (such as: optim.Adadelta, optim.SGD, optim.RMSprop, etc.) are

A subclass of Optimizer, some commonly used methods are defined in Optimizer, including zero_grad(), step(closure), state_dict(), load_state_dict(state_dict) and add_param_group(param_group)

The optimizer's management of parameters is based on the concept of groups , and specific lr, momentum, weight_decay , etc. can be configured for each group of parameters.

The parameter group is represented as a list (self.param_groups) in the optimizer, where each element is a dict, representing a parameter and its corresponding configuration, including 'params', 'weight_decay', 'lr', 'momentum', etc. in the dict field.

1. Basic concept code

import torch
import torch.optim as optim


w1 = torch.randn(2, 2)
w1.requires_grad = True

w2 = torch.randn(2, 2)
w2.requires_grad = True

w3 = torch.randn(2, 2)
w3.requires_grad = True

print("w1",w1)
print("w2",w2)
print("w3",w3)

# 一个参数组
optimizer_1 = optim.SGD([w1, w3], lr=0.1)
print('len(optimizer.param_groups): ', len(optimizer_1.param_groups))
print(optimizer_1.param_groups, '\n')

# 两个参数组
optimizer_2 = optim.SGD([{'params': w1, 'lr': 0.1},
                         {'params': w2, 'lr': 0.001}])
print('len(optimizer.param_groups): ', len(optimizer_2.param_groups))
print(optimizer_2.param_groups)

2. zero_grad()

Function: clear the gradient to zero. This is done before every update since PyTorch does not automatically zero gradients.

Code and output:

import torch
import torch.optim as optim

# ----------------------------------- zero_grad

w1 = torch.randn(2, 2)
w1.requires_grad = True

w2 = torch.randn(2, 2)
w2.requires_grad = True

optimizer = optim.SGD([w1, w2], lr=0.001, momentum=0.9)

print(optimizer.param_groups)
print("=======================")
print(optimizer.param_groups[0])
print("=======================")
print(optimizer.param_groups[0]['params'])
print("=======================")
print(optimizer.param_groups[0]['params'][0])  #参数w1

optimizer.param_groups[0]['params'][0].grad = torch.randn(2, 2)  

print('参数w1的梯度：')
print(optimizer.param_groups[0]['params'][0].grad, '\n')  # 参数组，第一个参数(w1)的梯度

optimizer.zero_grad()
print('执行zero_grad()之后，参数w1的梯度：')
print(optimizer.param_groups[0]['params'][0].grad)  # 参数组，第一个参数(w1)的梯度

 

3.state_dict()

Function: Get the current parameters of the model and return it in the form of an ordered dictionary.

In this ordered dictionary, the key is the parameter name of each layer, and the value is the parameter.

Code and output:

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


# ----------------------------------- state_dict
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 1, 3)  #输出一个特征图，需要3个 3*3 的矩阵
        self.pool = nn.MaxPool2d(2, 2)
        self.fc1 = nn.Linear(1 * 3 * 3, 2)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = x.view(-1, 1 * 3 * 3)
        x = F.relu(self.fc1(x))
        return x


net = Net()

# 获取网络当前参数
net_state_dict = net.state_dict()

print('net_state_dict类型：', type(net_state_dict))
print('net_state_dict管理的参数: ', net_state_dict.keys())
for key, value in net_state_dict.items():
    print('参数名: ', key, '\t大小: ',  value.shape)

4.add_param_group()  

effect:

Add a set of parameters to the parameter group managed by the optimizer, and you can customize lr, momentum, weight_decay , etc. for this set of parameters , which are commonly used in finetune.

Code and output:

# coding: utf-8

import torch
import torch.optim as optim

# ----------------------------------- add_param_group

w1 = torch.randn(2, 2)
w1.requires_grad = True

w2 = torch.randn(2, 2)
w2.requires_grad = True

w3 = torch.randn(2, 2)
w3.requires_grad = True

# 一个参数组
optimizer_1 = optim.SGD([w1, w2], lr=0.1)
print('当前参数组个数: ', len(optimizer_1.param_groups))
print(optimizer_1.param_groups, '\n')

# 增加一个参数组
print('增加一组参数 w3\n')
optimizer_1.add_param_group({'params': w3, 'lr': 0.001, 'momentum': 0.8})

print('当前参数组个数: ', len(optimizer_1.param_groups))
print(optimizer_1.param_groups, '\n')

print('可以看到，参数组是一个list，一个元素是一个dict，每个dict中都有lr, momentum等参数，这些都是可单独管理，单独设定，十分灵活！')

 5.load_state_dict(state_dict)

effect:

Load the parameters in state_dict to the current network, often used in finetune.

Code and output:

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


# ----------------------------------- load_state_dict

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 1, 3)
        self.pool = nn.MaxPool2d(2, 2)
        self.fc1 = nn.Linear(1 * 3 * 3, 2)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = x.view(-1, 1 * 3 * 3)
        x = F.relu(self.fc1(x))
        return x

    def zero_param(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                torch.nn.init.constant_(m.weight.data, 0)
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, nn.Linear):
                torch.nn.init.constant_(m.weight.data, 0)
                m.bias.data.zero_()
net = Net()

# 保存，并加载模型参数(仅保存模型参数)
torch.save(net.state_dict(), 'net_params.pkl')   # 假设训练好了一个模型net
pretrained_dict = torch.load('net_params.pkl')

# 将net的参数全部置0，方便对比
net.zero_param()
net_state_dict = net.state_dict()
print('conv1层的权值为:\n', net_state_dict['conv1.weight'], '\n')

# 通过load_state_dict 加载参数
net.load_state_dict(pretrained_dict)
print('加载之后，conv1层的权值变为:\n', net_state_dict['conv1.weight'])

 6.step(closure)

Function: Perform a one-step weight update.