LSTM模型的讲解与运用

实验目的：

循环神经网络(RNN)是一类以序列数据为输入，在序列的演进方向进行循环且所有循环单元按链式连接的神经网络。目前RNN已经广泛应用于语音识别、文本分类等自然语言处理任务中。本实验通过采用循环神经网络的变体长短期记忆网络(LSTM)，合理设计网络结构和算法来实现自动写诗和藏头诗的功能。

环境配置：

利用pytcharm配置pytroch和opencv；

参数：

epoches = 8

Lr=0.001

Batch=575

实验步骤：（以自动写诗为例）

图 5 加载数据

由图5加载诗的数据，返回标签。

图 6 lstm网络

图 7 训练过程

图 8 自动写诗代码

图 9 根据首字母生成诗句代码

图5到图9分别从加载图片，训练神经网络，到生成诗句的过程。

准确率：

图 10 训练结果

结果分析：

图 11 自动写诗的过程

由该图可以看出，对床前明月光的自动续写结果。

图 12 根据首字生成诗

由test函数可以生成龙碧和弦四个字开头的诗句。效果如图12。

源代码：

import numpy as np      # tang.npz的压缩格式处理
import os       # 打开文件
import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.nn.functional as F
import matplotlib.pyplot as plt
from torchnet import meter
import tqdm

def get_data():
    if os.path.exists(data_path):
        datas = np.load(data_path, allow_pickle=True)      # 加载数据
        data = datas['data']  # numpy.ndarray
        word2ix = datas['word2ix'].item()   # dic
        ix2word = datas['ix2word'].item()  # dic
        return data, word2ix, ix2word


class Net(nn.Module):
    def __init__(self, vocab_size, embedding_dim, hidden_dim):
        super(Net, self).__init__()
        self.hidden_dim = hidden_dim
        self.embeddings = nn.Embedding(vocab_size, embedding_dim)
        self.lstm = nn.LSTM(embedding_dim, self.hidden_dim, num_layers=2, batch_first=False)
        # lstm输入为：seq, batch, input_size
        # lstm输出为：seq * batch * 256; (2 * batch * 256,...)
        self.linear1 = nn.Linear(self.hidden_dim, vocab_size)

    def forward(self, input, hidden=None):
        seq_len, batch_size = input.size()
        if hidden is None:
            h_0 = input.data.new(2, batch_size, self.hidden_dim).fill_(0).float()
            c_0 = input.data.new(2, batch_size, self.hidden_dim).fill_(0).float()
            h_0, c_0 = Variable(h_0), Variable(c_0)
        else:
            h_0, c_0 = hidden
        embeds = self.embeddings(input)     # (seq_len, batch_size, embedding_dim), (124,128,128)
        output, hidden = self.lstm(embeds, (h_0, c_0))      #(seq_len, batch_size, hidden_dim), (124,128,256)
        output = self.linear1(output.view(seq_len*batch_size, -1))      # ((seq_len * batch_size),hidden_dim), (15872,256) → (15872,8293)
        return output, hidden


def train():
    modle = Net(len(word2ix), 128, 256)     # 模型定义：vocab_size, embedding_dim, hidden_dim —— 8293 * 128 * 256
    criterion = nn.CrossEntropyLoss()
    if torch.cuda.is_available() == True:
        print('Cuda is available!')
        modle = modle.cuda()
        optimizer = torch.optim.Adam(modle.parameters(), lr=1e-3)  # 学习率1e-3
        criterion = criterion.cuda()
        loss_meter = meter.AverageValueMeter()

        period = []
        loss2 = []
        for epoch in range(8):     # 最大迭代次数为8
            loss_meter.reset()
            for i, data in tqdm.tqdm(enumerate(dataloader)):    # data: torch.Size([128, 125]), dtype=torch.int32
                data = data.long().transpose(0,1).contiguous()      # long为默认tensor类型，并转置, [125, 128]
                data = data.cuda()
                optimizer.zero_grad()

                input, target = Variable(data[:-1, :]), Variable(data[1:, :])
                output, _ = modle(input)
                loss = criterion(output, target.view(-1))       # torch.Size([15872, 8293]), torch.Size([15872])
                loss.backward()
                optimizer.step()

                loss_meter.add(loss.item())    # loss:tensor(3.3510, device='cuda:0', grad_fn=<NllLossBackward>)loss.data:tensor(3.0183, device='cuda:0')

                period.append(i + epoch * len(dataloader))
                loss2.append(loss_meter.value()[0])

                if (1 + i) % 575 == 0:       # 每575个batch可视化一次
                    print(str(i) +':' + generate(modle,'床前明月光', ix2word, word2ix))

        torch.save(modle.state_dict(), './model_poet.pth')
        plt.plot(period, loss2)
        plt.show()


def generate(model, start_words, ix2word, word2ix):     # 给定几个词，根据这几个词生成一首完整的诗歌
    txt = []
    for word in start_words:
        txt.append(word)
    input = Variable(torch.Tensor([word2ix['<START>']]).view(1,1).long())      # tensor([8291.]) → tensor([[8291.]]) → tensor([[8291]])
    input = input.cuda()
    hidden = None
    num = len(txt)
    for i in range(48):      # 最大生成长度
        output, hidden = model(input, hidden)
        if i < num:
            w = txt[i]
            input = Variable(input.data.new([word2ix[w]])).view(1, 1)
        else:
            top_index = output.data[0].topk(1)[1][0]
            w = ix2word[top_index.item()]
            txt.append(w)
            input = Variable(input.data.new([top_index])).view(1, 1)
        if w == '<EOP>':
            break
    return ''.join(txt)

def gen_acrostic(model, start_words, ix2word, word2ix):
    result = []
    txt = []
    for word in start_words:
        txt.append(word)
    input = Variable(
        torch.Tensor([word2ix['<START>']]).view(1, 1).long())  # tensor([8291.]) → tensor([[8291.]]) → tensor([[8291]])
    input = input.cuda()
    hidden = None

    num = len(txt)
    index = 0
    pre_word = '<START>'
    for i in range(48):
        output, hidden = model(input, hidden)
        top_index = output.data[0].topk(1)[1][0]
        w = ix2word[top_index.item()]

        if (pre_word in {'。', '!', '<START>'}):
            if index == num:
                break
            else:
                w = txt[index]
                index += 1
                input = Variable(input.data.new([word2ix[w]])).view(1,1)
        else:
            input = Variable(input.data.new([word2ix[w]])).view(1,1)
        result.append(w)
        pre_word = w
    return ''.join(result)


def test():
    modle = Net(len(word2ix), 128, 256)     # 模型定义：vocab_size, embedding_dim, hidden_dim —— 8293 * 128 * 256
    if torch.cuda.is_available() == True:
        modle.cuda()
        modle.load_state_dict(torch.load('./model_poet.pth'))
        modle.eval()
        # txt = generate(modle, '床前明月光', ix2word, word2ix)
        # print(txt)
        txt = gen_acrostic(modle, '龙碧和弦', ix2word, word2ix)
        print(txt)

if __name__ == '__main__':
    data_path = './tang.npz'
    data, word2ix, ix2word = get_data()
    data = torch.from_numpy(data)
    dataloader = torch.utils.data.DataLoader(data, batch_size=10, shuffle=True, num_workers=1)  # shuffle=True随机打乱
    #train()
    test()