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):          =  self .features(x)          =  x.view( - 1 16 * 5 * 5 )          =  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。

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

 

    原文作者:pytorch
    原文地址: https://www.cnblogs.com/ranjiewen/p/9240512.html
    本文转自网络文章,转载此文章仅为分享知识,如有侵权,请联系博主进行删除。
点赞