pytorch调整优化器学习率的方法

step = [10,20,30,40]
base_lr = 1e-4
sgd_opt = torch.optim.SGD(model.parameters(), lr=base_lr, nesterov=True, momentum=0.9)
def adjust_lr(epoch):
 lr = base_lr * (0.1 ** np.sum(epoch >= np.array(step)))
 for params_group in sgd_opt.param_groups:
  params_group['lr'] = lr
 return lr

for epoch in range(60):
 model.train()
 adjust_lr(epoch)
 for ind, each in enumerate(train_loader):
 mat, label = each
 ... 

import torch
import pandas as pd
import numpy as np
from tensorflow import keras
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.utils.data import TensorDataset
import gensim
from sklearn.model_selection import train_test_split
class word_extract(nn.Module):
 def __init__(self,d_model,embedding_matrix):
  super(word_extract, self).__init__()
  self.d_model=d_model
  self.embedding=nn.Embedding(num_embeddings=len(embedding_matrix),embedding_dim=200)
  self.embedding.weight.data.copy_(embedding_matrix)
  self.embedding.weight.requires_grad=False
  self.lstm1=nn.LSTM(input_size=200,hidden_size=50,bidirectional=True)
  self.lstm2=nn.LSTM(input_size=2*self.lstm1.hidden_size,hidden_size=50,bidirectional=True)
  self.linear=nn.Linear(2*self.lstm2.hidden_size,4)

 def forward(self,x):
  w_x=self.embedding(x)
  first_x,(first_h_x,first_c_x)=self.lstm1(w_x)
  second_x,(second_h_x,second_c_x)=self.lstm2(first_x)
  output_x=self.linear(second_x)
  return output_x

def trans_num(word2idx,text):
 text_list=[]
 for i in text:
  s=i.rstrip().replace('\r','').replace('\n','').split(' ')
  numtext=[word2idx[j] if j in word2idx.keys() else word2idx['_PAD'] for j in s ]
  text_list.append(numtext)
 return text_list

def establish_word2vec_matrix(model): #负责将数值索引转为要输入的数据
 word2idx = {"_PAD": 0} # 初始化 `[word : token]` 字典，后期 tokenize 语料库就是用该词典。
 num2idx = {0: "_PAD"}
 vocab_list = [(k, model.wv[k]) for k, v in model.wv.vocab.items()]

 # 存储所有 word2vec 中所有向量的数组，留意其中多一位，词向量全为 0， 用于 padding
 embeddings_matrix = np.zeros((len(model.wv.vocab.items()) + 1, model.vector_size))
 for i in range(len(vocab_list)):
  word = vocab_list[i][0]
  word2idx[word] = i + 1
  num2idx[i + 1] = word
  embeddings_matrix[i + 1] = vocab_list[i][1]
 embeddings_matrix = torch.Tensor(embeddings_matrix)
 return embeddings_matrix, word2idx, num2idx

def train(model,epoch,learning_rate,batch_size,x, y, val_x, val_y):
 optimizor = optim.Adam(model.parameters(), lr=learning_rate)
 data = TensorDataset(x, y)
 data = DataLoader(data, batch_size=batch_size)
 for i in range(epoch):
  for j, (per_x, per_y) in enumerate(data):
   output_y = model(per_x)
   loss = F.cross_entropy(output_y.view(-1,output_y.size(2)), per_y.view(-1))
   optimizor.zero_grad()
   loss.backward()
   optimizor.step()
   arg_y=output_y.argmax(dim=2)
   fit_correct=(arg_y==per_y).sum()
   fit_acc=fit_correct.item()/(per_y.size(0)*per_y.size(1))
   print('##################################')
   print('第{}次迭代第{}批次的训练误差为{}'.format(i + 1, j + 1, loss), end=' ')
   print('第{}次迭代第{}批次的训练准确度为{}'.format(i + 1, j + 1, fit_acc))
   val_output_y = model(val_x)
   val_loss = F.cross_entropy(val_output_y.view(-1,val_output_y.size(2)), val_y.view(-1))
   arg_val_y=val_output_y.argmax(dim=2)
   val_correct=(arg_val_y==val_y).sum()
   val_acc=val_correct.item()/(val_y.size(0)*val_y.size(1))
   print('第{}次迭代第{}批次的预测误差为{}'.format(i + 1, j + 1, val_loss), end=' ')
   print('第{}次迭代第{}批次的预测准确度为{}'.format(i + 1, j + 1, val_acc))
 torch.save(model,'./extract_model.pkl')#保存模型

if __name__ =='__main__':
 #生成词向量矩阵
 word2vec = gensim.models.Word2Vec.load('./word2vec_model')
 embedding_matrix,word2idx,num2idx=establish_word2vec_matrix(word2vec)#输入的是词向量模型
 #
 train_data=pd.read_csv('./数据.csv')
 x=list(train_data['文本'])
 # 将文本从文字转化为数值，这部分trans_num函数你需要自己改动去适应你自己的数据集
 x=trans_num(word2idx,x)
 #x需要先进行填充，也就是每个句子都是一样长度，不够长度的以0来填充，填充词单独分为一类
 # #也就是说输入的x是固定长度的数值列表,例如[50,123,1850,21,199,0，0，...]
 #输入的y是[2,0,1,0,0,1,3,3,3,3,3,.....]
 #填充代码你自行编写，以下部分是针对我的数据集
 x=keras.preprocessing.sequence.pad_sequences(
   x,maxlen=60,value=0,padding='post',
 )
 y=list(train_data['BIO数值'])
 y_text=[]
 for i in y:
  s=i.rstrip().split(' ')
  numtext=[int(j) for j in s]
  y_text.append(numtext)
 y=y_text
 y=keras.preprocessing.sequence.pad_sequences(
   y,maxlen=60,value=3,padding='post',
  )
 # 将数据进行划分
 fit_x,val_x,fit_y,val_y=train_test_split(x,y,train_size=0.8,test_size=0.2)
 fit_x=torch.LongTensor(fit_x)
 fit_y=torch.LongTensor(fit_y)
 val_x=torch.LongTensor(val_x)
 val_y=torch.LongTensor(val_y)
 #开始应用
 w_extract=word_extract(d_model=200,embedding_matrix=embedding_matrix)
 train(model=w_extract,epoch=5,learning_rate=0.001,batch_size=50,
   x=fit_x,y=fit_y,val_x=val_x,val_y=val_y)#可以自行改动参数,设置学习率，批次，和迭代次数
 w_extract=torch.load('./extract_model.pkl')#加载保存好的模型
 pred_val_y=w_extract(val_x).argmax(dim=2)