Seq2seq+RNN 的英文翻译

# !/usr/bin/env Python3
# -*- coding: utf-8 -*-
# @version: v1.0
# @Author   : Meng Li
# @contact: [email protected]
# @FILE     : torch_seq2seq.py
# @Time     : 2022/6/8 11:11
# @Software : PyCharm
# @site: 
# @Description :
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchsummary
from torch.utils.data import Dataset, DataLoader
import numpy as np
import os


class Seq2seq(nn.Module):
    def __init__(self, in_features, hidden_size):
        super().__init__()
        self.in_features = in_features
        self.hidden_size = hidden_size
        self.encoder = nn.RNN(input_size=in_features, hidden_size=hidden_size, dropout=0.5)  # encoder
        self.decoder = nn.RNN(input_size=in_features, hidden_size=hidden_size, dropout=0.5)  # 翻译的解码器
        self.crition = nn.CrossEntropyLoss()
        self.fc = nn.Linear(hidden_size, in_features)

    def forward(self, enc_input, dec_input, dec_output):
        # enc_input.size() : [Batch_size,seq_len,embedding_size] -> [seq_len,Batch_size,embedding_size]
        enc_input = enc_input.permute(1, 0, 2)  # [seq_len,Batch_size,embedding_size]
        dec_input = dec_input.permute(1, 0, 2)  # [seq_len,Batch_size,embedding_size]
        # output:[seq_len,Batch_size,hidden_size]
        seq_len, batch_size, embedding_size = enc_input.size()
        h_0 = torch.rand(1, batch_size, self.hidden_size)
        _, ht = self.encoder(enc_input, h_0)  # en_ht:[num_layers * num_directions,Batch_size,hidden_size]
        de_output, _ = self.decoder(dec_input, ht)  # de_output:[seq_len,Batch_size,in_features]
        output = self.fc(de_output)
        output = output.permute(1, 0, 2)
        loss = 0
        for i in range(len(output)):  # 对seq的每一个输出进行二分类损失计算
            loss += self.crition(output[i], dec_output[i])
        return output, loss


class my_dataset(Dataset):
    def __init__(self, enc_input, dec_input, dec_output):
        super().__init__()
        self.enc_input = enc_input
        self.dec_input = dec_input
        self.dec_output = dec_output

    def __getitem__(self, index):
        return self.enc_input[index], self.dec_input[index], self.dec_output[index]

    def __len__(self):
        return self.enc_input.size(0)


def make_data1(seq_data):
    vocab = [i for i in "SE?abcdefghijklmnopqrstuvwxyz"]
    word2idx = {j: i for i, j in enumerate(vocab)}
    V = np.max([len(j) for i in seq_data for j in i])  # 求最长元素的长度
    enc_input = []
    dec_input = []
    dec_output = []
    for seq in seq_data:
        enc_input.append(np.eye(len(word2idx))[[word2idx[i] for i in seq[0] + (V - len(seq[0])) * "?" + 'E']])
        dec_input.append(np.eye(len(word2idx))[[word2idx[i] for i in 'S' + seq[1] + (V - len(seq[1])) * "?"]])
        dec_output.append([word2idx[i] for i in seq[1] + (V - len(seq[1])) * "?" + 'E'])
    return torch.tensor(enc_input).double(), torch.tensor(dec_input).double(), torch.LongTensor(dec_output).double()


def train():
    vocab = [i for i in "SE?abcdefghijklmnopqrstuvwxyz"]
    word2idx = {j: i for i, j in enumerate(vocab)}
    idx2word = {i: j for i, j in enumerate(vocab)}
    seq_data = [['man', 'women'], ['black', 'white'], ['king', 'queen'], ['girl', 'boy'], ['up', 'down'],
                ['high', 'low']]
    enc_input, dec_input, dec_output = make_data(seq_data)
    batch_size = 3
    in_features = 29
    hidden_size = 128

    train_data = my_dataset(enc_input, dec_input, dec_output)
    train_iter = DataLoader(train_data, batch_size, shuffle=True)

    net = Seq2seq(in_features, hidden_size)
    net.train()
    learning_rate = 0.001
    optimizer = optim.Adam(net.parameters(), lr=learning_rate)
    loss = 0

    for i in range(1000):
        for en_input, de_input, de_output in train_iter:
            output, loss = net(en_input, de_input, de_output)
            pre = torch.argmax(output, 2)
            # pre_ques = [[idx2word[j] for j in i] for i in en_input.numpy()]
            pre_ret = [[idx2word[j] for j in i] for i in pre.detach().numpy()]
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
        if i % 100 == 0:
            print("step {0} loss {1}".format(i, loss))
    torch.save(net, "translate.pt")


def make_data(seq_data):
    enc_input_all, dec_input_all, dec_output_all = [], [], []
    vocab = [i for i in "SE?abcdefghijklmnopqrstuvwxyz"]
    word2idx = {j: i for i, j in enumerate(vocab)}
    V = np.max([len(j) for i in seq_data for j in i])  # 求最长元素的长度
    for seq in seq_data:
        for i in range(2):
            seq[i] = seq[i] + '?' * (V - len(seq[i]))  # 'man??', 'women'

        enc_input = [word2idx[n] for n in (seq[0] + 'E')]
        dec_input = [word2idx[i] for i in [i for i in len(enc_input) * '?']]
        dec_output = [word2idx[n] for n in (seq[1] + 'E')]

        enc_input_all.append(np.eye(len(vocab))[enc_input])
        dec_input_all.append(np.eye(len(vocab))[dec_input])
        dec_output_all.append(dec_output)  # not one-hot

    # make tensor
    return torch.Tensor(enc_input_all), torch.Tensor(dec_input_all), torch.LongTensor(dec_output_all)


def translate(word):
    vocab = [i for i in "SE?abcdefghijklmnopqrstuvwxyz"]
    idx2word = {i: j for i, j in enumerate(vocab)}
    V = 5
    x, y, z = make_data([[word, "?" * V]])
    if not os.path.exists("translate.pt"):
        train()
    net = torch.load("translate.pt")
    pre, loss = net(x, y, z)
    pre = torch.argmax(pre, 2)[0]
    pre_word = [idx2word[i] for i in pre.numpy()]
    pre_word = "".join([i.replace("?", "") for i in pre_word])
    print(word, "->  ", pre_word[:pre_word.index('E')])


if __name__ == '__main__':
    translate('man')

 

老规矩，还是直接上代码

本代码主要实现了 采用RNN网络，构建seq2seq模型，对英文进行翻译。

seq2seq模型由编码器和解码器两个部分构成，编码器的输出作为隐含层传递到解码器

解码器的输入为随机初始化的，（这里采用全 ‘？’ 作为初始化内容）

解码器的输出为 [batch_size , seq_len , in_features] , 这里需要轮询seq_len个输出，采用交叉熵计算损失并累加。