python3.6+torch1.2实现Sentiment Analysis（数据集MR）

此文转载自：https://blog.csdn.net/qq_41626059/article/details/111588104#commentBox

总共是下面几个文件：

注意，最后一个是json文件，里面是电影影评数据集MR的划分出来的训练集生成的词典。是个字典文件，也可以自己再弄一个。

在训练集上训练了10个epoch，结果大概是上图这个样子

1、创建model_para.py文件，里面是模型的超参数。

import argparse

class Hpara():
    parser  =  argparse.ArgumentParser()    
    ############# insert paras #############
    parser.add_argument('--batch_size',default = 16, type = int) # batch_size
    parser.add_argument('--epochs',default = 1, type = int) # epochs
    
    parser.add_argument('--lr',default = 3*10**(-5), type = float) # learning rate
    parser.add_argument('--l2_lambda',default = 0.0, type = float) # L2_loss lambda
    parser.add_argument('--seg_net_nums_layer',default = 6, type = int) # segnet Rnn Layers
    parser.add_argument('--seg_net_dropout_rate',default = 0.2, type = float) # segnet dropout rate
    parser.add_argument('--seg_net_hidden_size',default = 64, type = int) # segnet hidden size 
    parser.add_argument('--seg_net_n_best',default = 12, type = int) # sample nums
    parser.add_argument('--seg_net_beam_size',default = 8, type = int) # beam size
    parser.add_argument('--seg_net_rnntype',default = 'GRU', type = str) # segnet Rnn Type
    parser.add_argument('--is_bi_rnn',default = True, type = bool) # segnet bi_Rnn 
    parser.add_argument('--unfreezing_bert_nums',default = 12, type = int)
    
    parser.add_argument('--is_training_bert',default = True, type = bool) 
    parser.add_argument('--is_load_best_model', default = False ,type = bool)
    
    ## MLP model paras
    parser.add_argument('--MLP_hidden_nums_layer',default = 1, type = int)
    parser.add_argument('--Bert_drop_out',default = 0.1, type = float)
    

    ## SA model paras
    parser.add_argument('--SA_hidden_dim',default = 256, type = int)
    parser.add_argument('--SA_is_bi',default = True, type = bool)
    parser.add_argument('--SA_dropout_rate',default = 0.4, type = float)
    parser.add_argument('--SA_word_dim',default = 300, type = int)
    parser.add_argument('--SA_vocab_size',default = 16783, type = int)
    parser.add_argument('--SA_layer_nums',default = 2, type = int)
    parser.add_argument('--SA_Rnn_type',default = 'LSTM', type = str)
    
    
    parser.add_argument('--pretrained_bert_path',default = '/bert-base-uncased', type = str)
    parser.add_argument('--bert_out_size',default = 768, type = int)
    parser.add_argument('--bert_vocab_path',default = /bert-base-uncased/vocab.txt', type = str)
    
    parser.add_argument('--saved_model_path',default = './best_model', type = str)
    parser.add_argument('--final_model_path',default = './final_epoch_saved_model', type = str)
    
    ## SegNet model paras
    parser.add_argument('--pretrained_segnet_model',default = './segnet_model/trained_model.torchsave', type = str)
    parser.add_argument('--seg_net_vocab',default = './segnet_model/all_vocabulary.pickle', type = str)
    parser.add_argument('--seg_net_vocab_size',default = 16703, type = int)
    parser.add_argument('--seg_net_embed_size',default = 300, type = int)
    
    parser.add_argument('--seg_net_use_cuda',default = False, type = bool)
    parser.add_argument('--seg_net_finetuning',default = False, type = bool)
    parser.add_argument('--seg_net_isbanor',default = True, type = bool)
    

    
    ## Other paras
    
    parser.add_argument('--emotion_label_nums',default = 2, type = int)
 
    parser.add_argument('--train_csv',default = r'./data/train_1.csv', type = str)
    parser.add_argument('--test_csv',default = r'./data/test_1.csv', type = str)
    parser.add_argument('--val_csv',default = r'./data/val_1.csv', type = str)

这个里面的参数是我在别的一个项目里面直接复制过来的，很多参数没有修改，实际使用到的参数很少。可以在代码里面自己看一下

2、创建数据处理文件data_utils.py

该文件主要用来加载数据集，并且分成batch输入到模型里面，还有最重要的对文本按照长度进行排序，这样做的目的是LSTM或者别的RNN在计算的时候不考虑padding。

import numpy as np

import json
import pandas as pd

def del_none(text): # delete ‘’ 删除使用空格切分之后句子里面的长度为0的字符串
    new_text = []
    
    for t in text:
        ws = t.split(' ')
        ws = [w for w in ws if w != '']
        new_text.append(' '.join(ws))
        
    return new_text

def read_csv(para):# read Moive Review DataSet
    
    train_data = pd.read_csv(para.train_csv, delimiter=',')
    test_data = pd.read_csv(para.test_csv, delimiter=',')
    val_data = pd.read_csv(para.val_csv, delimiter=',')
       
    train_text = list(train_data['INPUT'])
    train_label = list(train_data['OUTPUT'])
    
    test_text = list(test_data['INPUT'])
    test_label = list(test_data['OUTPUT'])
    
    val_text = list(val_data['INPUT'])
    val_label = list(val_data['OUTPUT'])
    
    train_text = del_none(train_text)
    test_text = del_none(test_text)
    val_text = del_none(val_text)
    train_label = np.array(train_label, dtype=np.int32)
    test_label = np.array(test_label, dtype=np.int32)
    val_label = np.array(val_label, dtype=np.int32)
    

    
    return train_text, test_text, val_text, train_label, test_label, val_label


def create_dict(train_text): 
    for t in train_text:
        ws = t.split(' ')
        aw.extend(ws)
        
    vocab = list(set(aw))
    
    vd = dict(zip(vocab, range(2,len(vocab)+2)))
    
    vd['UNKNOWN'] = 1
    vd['PADDING'] = 0
    
    with open('vd.json', 'w', encoding='utf-8') as f:
        json.dump(vd, f)
        
 
def get_vd(path):   
    f = open(path,'r', encoding='utf-8')
    vd = json.load(f)
    return vd

def text2id(text, vd):
    bs = len(text)
    lens = [len(t.split(' ')) for t in text]
    ml = max(lens)
    text_id = np.zeros(shape = (bs, ml), dtype=np.int32)
    words = list(vd.keys())
    for i in range(bs):
        ws = text[i].split(' ')
        for j in range(lens[i]):
            if ws[j] in words:
                text_id[i,j] = vd[ws[j]]
            else:
                text_id[i, j] = vd['UNKNOWN']
    return text_id, lens

            
def sort_data_by_length(textid, label, lengths):
   
    lengths = np.array(lengths, dtype=np.int32)
    idx = np.argsort(lengths)
    idx = idx[::-1]
    
    
    lengths = lengths[idx]  
    sc_id = textid[idx]
    label = label[idx]
  
    return sc_id, label, lengths


    
def batch_iter_str(text, label , batch_size = 16):
    
    path = 'vd.json'
    vd = get_vd(path)
    
    data_nums = len(text)
    
    num_batch = (data_nums + batch_size - 1) // batch_size
    indices = np.random.permutation(np.arange(data_nums))

    text = [text[i] for i in indices]
    label = label[indices]
   
    for i in range(num_batch):
        
        start_offset = i * batch_size
        end_offset = min(start_offset + batch_size, data_nums)

        bt_text = text[start_offset:end_offset]
        bt_label = label[start_offset:end_offset]
        
        text_id, lens = text2id(bt_text, vd)

        sc_id, bt_label, lengths = sort_data_by_length(text_id, bt_label, lens)
          
        
        yield i, num_batch, sc_id, bt_label, lengths

3、创建Model.py文件

import torch.nn as nn
import torch


device = torch.device("cuda:4" if torch.cuda.is_available() else "cpu")

class sa_model(nn.Module):
    
    def __init__(self, para):
        
        super(sa_model, self).__init__()
        self.para = para
        
        if para.SA_Rnn_type == 'LSTM': 
            
            self.RNN = nn.LSTM(input_size = para.SA_word_dim,
                               hidden_size = para.SA_hidden_dim,
                               num_layers = para.SA_layer_nums, 
                               batch_first = True,
                               dropout = para.SA_dropout_rate,
                               bidirectional = para.SA_is_bi)
        elif para.SA_Rnn_type == 'GRU':  
            
           self.RNN = nn.GRU(input_size = para.SA_word_dim, 
                               hidden_size = para.SA_hidden_dim, 
                               num_layers = para.SA_layer_nums, 
                               batch_first = True, 
                               dropout = para.SA_dropout_rate, 
                               bidirectional = para.SA_is_bi)
            
        elif para.SA_Rnn_type == 'RNN':
            
            self.RNN = nn.RNN(input_size = para.SA_word_dim, 
                               hidden_size = para.SA_hidden_dim, 
                               num_layers = para.SA_layer_nums, 
                               batch_first = True, 
                               dropout = para.SA_dropout_rate, 
                               bidirectional = para.SA_is_bi)
            
        self.nnEm = nn.Embedding(para.SA_vocab_size, para.SA_word_dim, padding_idx = 0).to(device)
        
        self.criterion = nn.CrossEntropyLoss()
        
        self.linear = nn.Linear(para.SA_hidden_dim*2, para.emotion_label_nums)

        
    
        
    
    
    def forward(self, x, x_len, label):
        
        x = self.nnEm(x)
        
        packed_embeds = nn.utils.rnn.pack_padded_sequence(x, x_len, batch_first=True)
        
        out, hs = self.RNN(packed_embeds)
        
        out, _ = nn.utils.rnn.pad_packed_sequence(out,
                                                  batch_first=True)
        
        h1 = hs[0][0]
        
        h2 = hs[0][1]
        
        h = torch.cat([h1, h2], dim = 1)
        logits = self.linear(h)
        
        loss = self.criterion(logits, label)
        
        pred = torch.argmax(torch.nn.functional.softmax(logits, dim = 1), dim = 1)
        
        return loss, pred

4、创建train.py文件

from Model import sa_model
from data_utils import batch_iter_str, read_csv
import torch
from model_para import Hpara
import torch.optim as optim
import matplotlib.pyplot as plt
from tqdm import tqdm
import os
from sklearn.metrics import accuracy_score
device = torch.device("cuda:4" if torch.cuda.is_available() else "cpu")
print(device)






def train(text, label, model, epoch):
    
    optimizer = optim.RMSprop(filter(lambda p: p.requires_grad, model.parameters()), lr = para.lr * (0.95)**epoch)

    pred_label = []
    true_label = []
    all_loss = []
    model.train()
    db = batch_iter_str(text, label)
    
    for i, num_batch, sc_id, label, lengths in db:
        sc_id = torch.tensor(sc_id).long().to(device)
        label = torch.tensor(label).long().to(device)
        lens = torch.tensor(lengths).long().to(device)
        
        model.zero_grad()
        loss ,pred = model(sc_id, lens, label)
        
        all_loss.append(loss)
        pred_label.append(pred)
        true_label.append(label)
        
        loss.backward()
        optimizer.step()
        
        if (i+1)%10 == 0:
            print('%d step loss is : %f'%(i+1, loss.item()))
            print('%d acc is : %f'%(i+1, accuracy_score(label.data.cpu().numpy(), pred.data.cpu().numpy())))
     

    mean_loss = sum(all_loss) / num_batch
    pred_label = torch.cat(pred_label, dim = 0)
    true_label = torch.cat(true_label, dim = 0)
    mean_acc = accuracy_score(true_label, pred_label)

    return mean_loss, mean_acc

def test(saved_acc, text, label, model, mode = 'validation'):
     
    pred_label = []
    true_label = []
    all_loss = []
    
    model.eval()
    
    db = batch_iter_str(text, label)
    
    for i, num_batch, sc_id, label, lengths in db:
        
        sc_id = torch.tensor(sc_id).long().to(device)
        label = torch.tensor(label).long().to(device)
        lens = torch.tensor(lengths).long().to(device)
        
        model.zero_grad()
        loss ,pred = model(sc_id, lens, label)
        
        all_loss.append(loss)
        pred_label.append(pred)
        true_label.append(label)
        
                  
    mean_loss = sum(all_loss) / num_batch
    pred_label = torch.cat(pred_label, dim = 0)
    true_label = torch.cat(true_label, dim = 0)
    mean_acc = accuracy_score(true_label, pred_label)
    
    if mode == 'validation':# save the best model when the mode is 'validation' 
        if mean_acc >= saved_acc:
            torch.save(model, os.path.join(para.saved_model_path,'My_model_pth'))
            saved_acc = mean_acc
        ## TODO: the next step is to save the model every epoch
        torch.save(model, os.path.join(para.final_model_path,'My_model_pth'))
    
        
    
    return mean_loss, mean_acc, saved_acc
        
def load_model(para, model, best_model = True):
    if best_model:
        model_path = para.saved_model_path
    else:
        model_path = para.final_model_path
        
    if os.path.exists(os.path.join(model_path, 'My_model_pth')):
        print('model path is : ', model_path)
        model = torch.load(os.path.join(model_path, 'My_model_pth'))
        print('load pre-train')
    else:
        model = model

    return model
    

        
        
        
if __name__ == '__main__':
    
    saved_acc = 0.0
    hp=Hpara()
    parser = hp.parser
    para = parser.parse_args()
    
    train_text, test_text, val_text, train_label, test_label, val_label = read_csv(para)
    
    model = sa_model(para)
    model = model.to(device)
    
    if para.is_load_best_model:
        
        print('load the best model...')
        _, _, saved_acc = test(saved_acc, val_text, val_label, model, mode = 'get_saved_acc')
        model = load_model(para, model)
        
    else:
        
        print('load the final epoch model...')
        _, _, saved_acc = test(saved_acc, val_text, val_label, model, mode = 'get_saved_acc') 
        model = load_model(para, model, best_model = False)
    
    print('saved accuracy is :', saved_acc)
    
    train_loss = []
    train_acc = []
    
    val_loss = []
    val_acc = []
    
    test_loss = []
    test_acc = []
    
    for i in range(para.epochs): 
        ########### train ###########
        train_mean_loss_i, train_mean_acc_i = train(train_text, train_label, model, i)
        
        train_loss.append(train_mean_loss_i)
        train_acc.append(train_mean_acc_i)
        
        ########## val #############
        val_loss_i, val_acc_i, saved_acc = test(saved_acc, val_text, val_label, model, )
        print('%d epoch val acc is', val_acc_i)
        val_loss.append(val_loss_i)
        val_acc.append(val_acc_i)
        
        ########## test ###########
        test_loss_i, test_acc_i, saved_acc = test(saved_acc, test_text, test_label, model, mode = 'test')
        
        print('%d epoch test acc is :', test_acc_i)
        
        test_loss.append(test_loss_i)
        test_acc.append(test_acc_i)
        
        
    
    plt.figure(2)
    plt.subplot(231)
    plt.plot(train_loss)
    plt.xlabel('epochs')
    plt.ylabel('emotion_train_loss')
    
    plt.subplot(232)
    plt.plot(test_loss)
    plt.xlabel('epochs')
    plt.ylabel('emotion_test_loss')
    
    plt.subplot(233)
    plt.plot(val_loss)
    plt.xlabel('epochs')
    plt.ylabel('emotion_val_loss')

    plt.subplot(234)
    plt.plot(train_acc)
    plt.xlabel('epochs')
    plt.ylabel('emotion_train_acc')
    
    plt.subplot(235)
    plt.plot(test_acc)
    plt.xlabel('epochs')
    plt.ylabel('emotion_test_acc')

    plt.subplot(236)
    plt.plot(val_acc)
    plt.xlabel('epochs')
    plt.ylabel('emotion_val_acc')        

    plt.show()

5、完整的代码

链接：https://pan.baidu.com/s/1nWMThnz9eK8zVggQJ7Umag
提取码：0rwq
复制这段内容后打开百度网盘手机App，操作更方便哦