自学教程：Pytorch torch.optim优化器个性化使用

一、简化前馈网络LeNet

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

import torch as t     class LeNet(t.nn.Module):     def __init__(self):         super(LeNet, self).__init__()         self.features = t.nn.Sequential(             t.nn.Conv2d(3, 6, 5),             t.nn.ReLU(),             t.nn.MaxPool2d(2, 2),             t.nn.Conv2d(6, 16, 5),             t.nn.ReLU(),             t.nn.MaxPool2d(2, 2)         )         # 由于调整shape并不是一个class层，         # 所以在涉及这种操作（非nn.Module操作）需要拆分为多个模型         self.classifiter = t.nn.Sequential(             t.nn.Linear(16*5*5, 120),             t.nn.ReLU(),             t.nn.Linear(120, 84),             t.nn.ReLU(),             t.nn.Linear(84, 10)         )       def forward(self, x):         x = self.features(x)         x = x.view(-1, 16*5*5)         x = self.classifiter(x)         return x   net = LeNet()

二、优化器基本使用方法

1. 建立优化器实例
2. 循环：
   1. 清空梯度
   2. 向前传播
   3. 计算Loss
   4. 反向传播
   5. 更新参数

1 2 3 4 5 6 7 8 9 10 11

from torch import optim   # 通常的step优化过程 optimizer = optim.SGD(params=net.parameters(), lr=1) optimizer.zero_grad()  # net.zero_grad()   input_ = t.autograd.Variable(t.randn(1, 3, 32, 32)) output = net(input_) output.backward(output)   optimizer.step()

三、网络模块参数定制

为不同的子网络参数不同的学习率，finetune常用，使分类器学习率参数更高，学习速度更快（理论上）。

1.经由构建网络时划分好的模组进行学习率设定，

1 2 3

# # 直接对不同的网络模块制定不同学习率 optimizer = optim.SGD([{'params': net.features.parameters()}, # 默认lr是1e-5                        {'params': net.classifiter.parameters(), 'lr': 1e-2}], lr=1e-5)

 2.以网络层对象为单位进行分组，并设定学习率

1 2 3 4 5 6 7 8 9 10

# # 以层为单位，为不同层指定不同的学习率 # ## 提取指定层对象 special_layers = t.nn.ModuleList([net.classifiter[0], net.classifiter[3]]) # ## 获取指定层参数id special_layers_params = list(map(id, special_layers.parameters())) print(special_layers_params) # ## 获取非指定层的参数id base_params = filter(lambda p: id(p) not in special_layers_params, net.parameters()) optimizer = t.optim.SGD([{'params': base_params},                          {'params': special_layers.parameters(), 'lr': 0.01}], lr=0.001)

四、在训练中动态的调整学习率

1 2 3 4 5 6 7 8 9

'''调整学习率''' # 新建optimizer或者修改optimizer.params_groups对应的学习率 # # 新建optimizer更简单也更推荐，optimizer十分轻量级，所以开销很小 # # 但是新的优化器会初始化动量等状态信息，这对于使用动量的优化器（momentum参数的sgd）可能会造成收敛中的震荡 # ## optimizer.param_groups:长度2的list，optimizer.param_groups[0]：长度6的字典 print(optimizer.param_groups[0]['lr']) old_lr = 0.1 optimizer = optim.SGD([{'params': net.features.parameters()},                        {'params': net.classifiter.parameters(), 'lr': old_lr*0.1}], lr=1e-5)

 可以看到optimizer.param_groups结构，[{'params','lr', 'momentum', 'dampening', 'weight_decay', 'nesterov'},{……}]，集合了优化器的各项参数。

• torch.optim的灵活使用

• 重写sgd优化器

import torch
from torch.optim.optimizer import Optimizer, required

class SGD(Optimizer):
    def __init__(self, params, lr=required, momentum=0, dampening=0, weight_decay1=0, weight_decay2=0, nesterov=False):
        defaults = dict(lr=lr, momentum=momentum, dampening=dampening,
                        weight_decay1=weight_decay1, weight_decay2=weight_decay2, nesterov=nesterov)
        if nesterov and (momentum <= 0 or dampening != 0):
            raise ValueError("Nesterov momentum requires a momentum and zero dampening")
        super(SGD, self).__init__(params, defaults)

    def __setstate__(self, state):
        super(SGD, self).__setstate__(state)
        for group in self.param_groups:
            group.setdefault('nesterov', False)

    def step(self, closure=None):
        """Performs a single optimization step. Arguments: closure (callable, optional): A closure that reevaluates the model and returns the loss. """
        loss = None
        if closure is not None:
            loss = closure()

        for group in self.param_groups:
            weight_decay1 = group['weight_decay1']
            weight_decay2 = group['weight_decay2']
            momentum = group['momentum']
            dampening = group['dampening']
            nesterov = group['nesterov']

            for p in group['params']:
                if p.grad is None:
                    continue
                d_p = p.grad.data
                if weight_decay1 != 0:
                    d_p.add_(weight_decay1, torch.sign(p.data))
                if weight_decay2 != 0:
                    d_p.add_(weight_decay2, p.data)
                if momentum != 0:
                    param_state = self.state[p]
                    if 'momentum_buffer' not in param_state:
                        buf = param_state['momentum_buffer'] = torch.zeros_like(p.data)
                        buf.mul_(momentum).add_(d_p)
                    else:
                        buf = param_state['momentum_buffer']
                        buf.mul_(momentum).add_(1 - dampening, d_p)
                    if nesterov:
                        d_p = d_p.add(momentum, buf)
                    else:
                        d_p = buf

                p.data.add_(-group['lr'], d_p)

        return loss