python pytorch实现RNN,LSTM，GRU，文本情感分类

本文介绍: python pytorch 实现RNN,LSTM，GRU，文本情感分类

python pytorch实现RNN,LSTM，GRU，文本 情感 分类

实现步骤就是：
1.先对句子进行分词并构建词表
2.生成 word2id
3.构建模型
4.训练模型
5.测试模型


import pandas as pd
import torch
import matplotlib.pyplot as plt
import jieba
import numpy as np

"""
作业：
一、完成优化
优化思路

1 jieba
2 取常用的3000字
3 修改model：rnn、lstm、gru

二、完成测试代码
"""

# 了解数据
dd = pd.read_csv(r'E:peixundatatrain.csv')
# print(dd.head())

# print(dd['label'].value_counts())

# 句子长度分析
# 确定输入句子长度为 500
text_len = [len(i) for i in dd['text']]
# plt.hist(text_len)
# plt.show()
# print(max(text_len), min(text_len))

# 基本参数 config
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('my device:', DEVICE)

MAX_LEN = 500
BATCH_SIZE = 16
EPOCH = 1
LR = 3e-4

# 构建词表 word2id
vocab = []
for i in dd['text']:
    vocab.extend(jieba.lcut(i, cut_all=True))  # 使用 jieba 分词
    # vocab.extend(list(i))

vocab_se = pd.Series(vocab)
print(vocab_se.head())
print(vocab_se.value_counts().head())

vocab = vocab_se.value_counts().index.tolist()[:3000]  # 取频率最高的 3000 token
# print(vocab[:10])
# exit()

WORD_PAD = "<PAD&gt;"
WORD_UNK = "<UNK&gt;"
WORD_PAD_ID = 0
WORD_UNK_ID = 1

vocab = [WORD_PAD, WORD_UNK] + list(set(vocab))

print(vocab[:10])
print(len(vocab))

vocab_dict = {k: v for v, k in enumerate(vocab)}

# 词表大小，vocab_dict: word2id; vocab: id2word
print(len(vocab_dict))

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils import data
import pandas as pd


# 定义数据集 Dataset
class Dataset(data.Dataset):
    def __init__(self, split='train'):
        # ChnSentiCorp 情感分类数据集
        path =  r'E:/peixun/data/' + str(split) + '.csv'
        self.data = pd.read_csv(path)

    def __len__(self):
        return len(self.data)

    def __getitem__(self, i):
        text = self.data.loc[i, 'text']
        label = self.data.loc[i, 'label']

        return text, label


# 实例化 Dataset
dataset = Dataset('train')

# 样本数量
print(len(dataset))
print(dataset[0])


# 句子批处理函数
def collate_fn(batch):
    # [(text1, label1), (text2, label2), (3, 3)...]
    sents = [i[0][:MAX_LEN] for i in batch]
    labels = [i[1] for i in batch]

    inputs = []
    # masks = []

    for sent in sents:
        sent = [vocab_dict.get(i, WORD_UNK_ID) for i in list(sent)]
        pad_len = MAX_LEN - len(sent)

        # mask = len(sent) * [1] + pad_len * [0]
        # masks.append(mask)

        sent += pad_len * [WORD_PAD_ID]

        inputs.append(sent)

    # 只使用 lstm 不需要用 masks
    # masks = torch.tensor(masks)
   # print(inputs)
    inputs = torch.tensor(inputs)
    labels = torch.LongTensor(labels)

    return inputs.to(DEVICE), labels.to(DEVICE)


# 测试 loader
loader = data.DataLoader(dataset,
                         batch_size=BATCH_SIZE,
                         collate_fn=collate_fn,
                         shuffle=True,
                         drop_last=False)

inputs, labels = iter(loader).__next__()
print(inputs.shape, labels)


# 定义模型
class Model(nn.Module):
    def __init__(self, vocab_size=5000):
        super().__init__()
        self.embed = nn.Embedding(vocab_size, 100, padding_idx=WORD_PAD_ID)

        # 多种 rnn
        self.rnn = nn.RNN(100, 100, 1, batch_first=True, bidirectional=True)
        self.gru = nn.GRU(100, 100, 1, batch_first=True, bidirectional=True)
        self.lstm = nn.LSTM(100, 100, 1, batch_first=True, bidirectional=True)

        self.l1 = nn.Linear(500 * 100 * 2, 100)
        self.l2 = nn.Linear(100, 2)

    def forward(self, inputs):
        out = self.embed(inputs)
        out, _ = self.lstm(out)
        out = out.reshape(BATCH_SIZE, -1)  # 16 * 100000
        out = F.relu(self.l1(out))  # 16 * 100
        out = F.softmax(self.l2(out))  # 16 * 2

        return out


# 测试 Model
model = Model()
print(model)

# 模型训练
dataset = Dataset()
loader = data.DataLoader(dataset,
                         batch_size=BATCH_SIZE,
                         collate_fn=collate_fn,
                         shuffle=True)

model = Model().to(DEVICE)

# 交叉熵损失
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.AdamW(model.parameters(), lr=LR)

model.train()
for e in range(EPOCH):
    for idx, (inputs, labels) in enumerate(loader):
        # 前向传播，计算预测值
        out = model(inputs)
        # 计算损失
        loss = loss_fn(out, labels)
        # 反向传播，计算梯度
        loss.backward()
        # 参数更新
        optimizer.step()
        # 梯度清零
        optimizer.zero_grad()

        if idx % 10 == 0:
            out = out.argmax(dim=-1)
            acc = (out == labels).sum().item() / len(labels)

            print('>>epoch:', e,
                  'tbatch:', idx,
                  'tloss:', loss.item(),
                  'tacc:', acc)

# 模型测试
test_dataset = Dataset('test')
test_loader = data.DataLoader(test_dataset,
                              batch_size=BATCH_SIZE,
                              collate_fn=collate_fn,
                              shuffle=False)

loss_fn = nn.CrossEntropyLoss()

out_total = []
labels_total = []

model.eval()
for idx, (inputs, labels) in enumerate(test_loader):
    out = model(inputs)
    loss = loss_fn(out, labels)

    out_total.append(out)
    labels_total.append(labels)

    if idx % 50 == 0:
        print('>>batch:', idx, 'tloss:', loss.item())
        
correct=0
sumz=0
for i in range(len(out_total)):
   out = out_total[i].argmax(dim=-1)
   correct = (out == labels_total[i]).sum().item() +correct
   sumz=sumz+len(labels_total[i])
    #acc = (out_total == labels_total).sum().item() / len(labels_total)

print('>>acc:', correct/sumz)