# 自然語言期末實驗
## 實驗題目
## 實驗要求
## 實驗代碼解釋
### 將json文件轉換成txt文件
```python3=
import json
def data_synthesis(source, target, new):
src_obj = open(source, 'r')
tar_obj = open(target, 'r')
synthesis_obj = open(new, 'w')
source_lines = []
target_lines = []
for src_line in src_obj.readlines():
source_lines.append(src_line)
for tar_line in tar_obj.readlines():
target_lines.append(tar_line)
for i in range(0, len(source_lines)):
source_data = json.loads(source_lines[i])
target_data = json.loads(target_lines[i])
synthesis_obj.write(source_data['text'] + "\t" + target_data['text'] + '\n')
src_obj.close()
tar_obj.close()
synthesis_obj.close()
data_synthesis("data/train/src.jsonl", "data/train/target.jsonl", "data/train.txt")
data_synthesis("data/test/src.jsonl", "data/test/target.jsonl", "data/test.txt")
data_synthesis("data/validation/src.jsonl", "data/validation/target.jsonl", "data/validation.txt")
```
### 導入庫
```python3=
import os
import sys
import math
from collections import Counter
import numpy as np
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
import nltk
import jieba
# nltk.download('punkt') # 下載 punkt
```
### 1. 数据预处理
#### 1.1 读入中英文数据
* 英文使用nltk的 word_tokenizer 来分词，并且使用小写字母
* 中文直接使用单个汉字作为基本单元
```python3=
def load_data(in_file):
en = []
cn = []
num_examples = 0
with open(in_file, 'r', encoding='utf-8') as lines:
for line in lines:
line = line.strip().split('\t')
en.append(['BOS'] + nltk.word_tokenize(line[0].lower()) + ['EOS'])
cn.append(['BOS'] + jieba.lcut(line[1]) + ['EOS'])
return en, cn
train_file = 'data/train.txt'
dev_file = 'data/test.txt'
train_en, train_cn = load_data(train_file)
dev_en, dev_cn = load_data(dev_file)
```
### 查看返回的数据内容：
```python3=
print(dev_en[:2])
print(dev_cn[:2])
```
### 1.2 构建单词表
```python3=
# 构建词典:利用分词后的结果构建统计词典，可以过滤掉出现频次较低的词语防止词典规模过大
UNK_IDX = 0
PAD_IDX = 1
def build_dict(sentences, max_words=50000):
word_count = Counter()
for sentence in sentences:
for word in sentence:
word_count[word] += 1
ls = word_count.most_common(max_words) # 词频前max_words个单词(降序)
total_words = len(ls) + 2
word_dict = {w[0] : index + 2 for index, w in enumerate(ls)} # {单词:索引}, w[0]:单词, w[1]:词频
word_dict['UNK'] = UNK_IDX
word_dict["PAD"] = PAD_IDX
return word_dict, total_words
en_dict, en_total_words = build_dict(train_en)
cn_dict, cn_total_words = build_dict(train_cn)
inv_en_dict = {v:k for k, v in en_dict.items()}
inv_cn_dict = {v:k for k, v in cn_dict.items()}
```
### 1.3 把单词全部转变成数字
```python3=
def encode(en_sentences, cn_sentences, en_dict, cn_dict, sort_by_len=True):
length = len(en_sentences)
out_en_sentences = [[en_dict.get(w, 0) for w in sent] for sent in en_sentences]
out_cn_sentences = [[cn_dict.get(w, 0) for w in sent] for sent in cn_sentences]
# sort sentences by word
def len_argsort(seq):
return sorted(range(len(seq)), key=lambda x: len(seq[x]))
# 把中文和英文按照同样的顺序排序
if sort_by_len:
sorted_index = len_argsort(out_en_sentences)
out_en_sentences = [out_en_sentences[i] for i in sorted_index]
out_cn_sentences = [out_cn_sentences[i] for i in sorted_index]
return out_en_sentences, out_cn_sentences
train_en, train_cn = encode(train_en, train_cn, en_dict, cn_dict)
dev_en, dev_cn = encode(dev_en, dev_cn, en_dict, cn_dict) # [[2, 168, 201, 4, 3], [], ...., [2, 5, 14, 13, 22, 9, 149, 17, 107, 24, 121, 16, 20, 267, 7, 181, 23, 15, 6, 422, 25, 220, 4, 3]]
```
### 查看返回的数据内容：
```python3=
print(train_cn[2])
print([inv_cn_dict[i] for i in train_cn[2]])
print([inv_en_dict[i] for i in train_en[2]])
```
### 1.4 把全部句子分成batch
```python3=
def get_minibatches(n, minibatch_size, shuffle=True): # n是传进来的句子数
idx_list = np.arange(0, n, minibatch_size) # [0, 1, ..., n-1] 按minibatch_size大小分割
if shuffle:
np.random.shuffle(idx_list)
minibatches = []
for idx in idx_list:
minibatches.append(np.arange(idx, min(idx + minibatch_size, n)))
return minibatches
```
### 查看上面函数的功能：
```python3=
get_minibatches(100, 15)
```
```python3=
def prepare_data(seqs):
lengths = [len(seq) for seq in seqs]
n_samples = len(seqs) # n_samples个句子
max_len = np.max(lengths) # batch_size个句子中最长句子长度
x = np.zeros((n_samples, max_len)).astype('int32')
x_lengths = np.array(lengths).astype('int32') # batch中原始句子长度
for idx, seq in enumerate(seqs):
x[idx, :lengths[idx]] = seq # lengths[idx]: 每个句子的索引, 长度不够补0
return x, x_lengths
def gen_examples(en_sentences, cn_sentences, batch_size):
minibatches = get_minibatches(len(en_sentences), batch_size)
all_ex = []
for minibatch in minibatches:
mb_en_sentences = [en_sentences[t] for t in minibatch] # 一个batch中每个句子的对应编码,[[[2, 982, 8], [14,5,6],...]
mb_cn_sentences = [cn_sentences[t] for t in minibatch]
mb_x, mb_x_len = prepare_data(mb_en_sentences) # 一个batch中每个句子的对应编码，长度不够补0; 一个batch中每个句子长度
mb_y, mb_y_len = prepare_data(mb_cn_sentences)
all_ex.append((mb_x, mb_x_len, mb_y, mb_y_len))
# 返回内容依次是 n / batch_size 个 (batch个句子编码，batch个英文句子长度，batch个中文句子编码，batch个中文句子长度)
return all_ex
batch_size = 64
train_data = gen_examples(train_en, train_cn, batch_size)
dev_data = gen_examples(dev_en, dev_cn, batch_size)
```
## 2. 定义计算损失的函数
```python3=
# masked cross entropy loss
class LanguageModelCriterion(nn.Module):
def __init__(self):
super(LanguageModelCriterion, self).__init__()
def forward(self, input, target, mask):
# input: [64, 12, 3195] target: [64, 12] mask: [64, 12]
# input: (batch_size * seq_len) * vocab_size
input = input.contiguous().view(-1, input.size(2))
# target: batch_size * seq_len
target = target.contiguous().view(-1, 1)
mask = mask.contiguous().view(-1, 1)
output = -input.gather(1, target) * mask # 将input在1维，把target当索引进行取值
#这里算得就是交叉熵损失，前面已经算了F.log_softmax
#output.shape=torch.Size([768, 1])
#因为input.gather时，target为0的地方不是零了，mask作用是把padding为0的地方重置为零，
#因为在volab里0代表的也是一个单词，但是我们这里target尾部的0代表的不是单词
output = torch.sum(output) / torch.sum(mask)
# 均值损失，output前已经加了负号，所以这里还是最小化
return output
```
## 3. 评估模型
```python3=
def evaluate(model, data):
model.eval()
total_num_words = total_loss = 0.
with torch.no_grad():
for it, (mb_x, mb_x_len, mb_y, mb_y_len) in enumerate(data):
mb_x = torch.from_numpy(mb_x).to(device).long()
mb_x_len = torch.from_numpy(mb_x_len).to(device).long()
mb_input = torch.from_numpy(mb_y[:, :-1]).to(device).long()
mb_output = torch.from_numpy(mb_y[:, 1:]).to(device).long()
mb_y_len = torch.from_numpy(mb_y_len-1).to(device).long()
mb_y_len[mb_y_len<=0] = 1
mb_pred, attn = model(mb_x, mb_x_len, mb_input, mb_y_len)
mb_out_mask = torch.arange(mb_y_len.max().item(), device=device)[None, :] < mb_y_len[:, None]
mb_out_mask = mb_out_mask.float()
loss = loss_fn(mb_pred, mb_output, mb_out_mask)
num_words = torch.sum(mb_y_len).item()
total_loss += loss.item() * num_words
total_num_words += num_words
print("Evaluation loss", total_loss / total_num_words)
```
## 4. Encoder Decoder模型（含Attention版本）
### 4.1 Encoder
> Encoder模型的任务是把输入文字传入embedding层和GRU层，转换成一些hidden states作为后续的context vectors；
```python3=
class Encoder(nn.Module):
def __init__(self, vocab_size, embed_size, enc_hidden_size, dec_hidden_size, dropout=0.2):
super(Encoder, self).__init__()
self.embed = nn.Embedding(vocab_size, embed_size)
self.rnn = nn.GRU(embed_size, enc_hidden_size, batch_first=True, bidirectional=True)
self.dropout = nn.Dropout(dropout)
self.fc = nn.Linear(enc_hidden_size * 2, dec_hidden_size)
def forward(self, x, lengths):
sorted_len, sorted_idx = lengths.sort(0, descending=True)
x_sorted = x[sorted_idx.long()]
embedded = self.dropout(self.embed(x_sorted))
packed_embedded = nn.utils.rnn.pack_padded_sequence(embedded, sorted_len.long().cpu().data.numpy(), batch_first=True)
packed_out, hid = self.rnn(packed_embedded)
out, _ = nn.utils.rnn.pad_packed_sequence(packed_out, batch_first=True)
_, original_idx = sorted_idx.sort(0, descending=False)
out = out[original_idx.long()].contiguous()
hid = hid[:, original_idx.long()].contiguous()
# hid: [2, batch_size, enc_hidden_size]
hid = torch.cat([hid[-2], hid[-1]], dim=1) # 将最后一层的hid的双向拼接
# hid: [batch_size, 2*enc_hidden_size]
hid = torch.tanh(self.fc(hid)).unsqueeze(0)
# hid: [1, batch_size, dec_hidden_size]
# out: [batch_size, seq_len, 2*enc_hidden_size]
return out, hid
```
### 4.2 Luong Attention
```python3=
class Attention(nn.Module):
def __init__(self, enc_hidden_size, dec_hidden_size):
# enc_hidden_size跟Encoder的一样
super(Attention, self).__init__()
self.enc_hidden_size = enc_hidden_size
self.dec_hidden_size = dec_hidden_size
self.linear_in = nn.Linear(enc_hidden_size*2, dec_hidden_size, bias=False)
self.linear_out = nn.Linear(enc_hidden_size*2 + dec_hidden_size, dec_hidden_size)
def forward(self, output, context, mask):
# mask = batch_size, output_len, context_len # mask在Decoder中创建好了
# output: batch_size, output_len, dec_hidden_size，就是Decoder的output
# context: batch_size, context_len, 2*enc_hidden_size，就是Encoder的output
# 这里Encoder网络是双向的，Decoder是单向的
batch_size = output.size(0)
output_len = output.size(1)
input_len = context.size(1) # input_len = context_len
# 通过decoder的hidden states加上encoder的hidden states来计算一个分数，用于计算权重
# batch_size, context_len, dec_hidden_size
# 第一步，公式里的Wa先与hs做点乘，把Encoder output的enc_hidden_size换成dec_hidden_size。
# Q: W·context
context_in = self.linear_in(context.view(batch_size*input_len, -1)).view(
batch_size, input_len, -1)
# Q·K
# context_in.transpose(1,2): batch_size, dec_hidden_size, context_len
# output: batch_size, output_len, dec_hidden_size
attn = torch.bmm(output, context_in.transpose(1,2))
# batch_size, output_len, context_len
# 第二步，ht与上一步结果点乘，得到score
attn.data.masked_fill(mask.bool(), -1e6)
# .masked_fill作用请看这个链接：https://blog.csdn.net/candy134834/article/details/84594754
# mask的维度必须和attn维度相同，mask为1的位置对应attn的位置的值替换成-1e6，
# mask为1的意义需要看Decoder函数里面的定义
attn = F.softmax(attn, dim=2)
# batch_size, output_len, context_len
# 这个dim=2到底是怎么softmax的看下下面单元格例子
# 第三步，计算每一个encoder的hidden states对应的权重。
# context: batch_size, context_len, 2*enc_hidden_size，
context = torch.bmm(attn, context)
# batch_size, output_len, 2*enc_hidden_size
# 第四步，得出context vector是一个对于encoder输出的hidden states的一个加权平均
# output: batch_size, output_len, dec_hidden_size
output = torch.cat((context, output), dim=2)
# output：batch_size, output_len, 2*enc_hidden_size+dec_hidden_size
# 第五步，将context vector和 decoder的hidden states 串起来。
output = output.view(batch_size*output_len, -1)
# output.shape = (batch_size*output_len, 2*enc_hidden_size+dec_hidden_size)
output = torch.tanh(self.linear_out(output))
# output.shape=(batch_size*output_len, dec_hidden_size)
output = output.view(batch_size, output_len, -1)
# output.shape=(batch_size, output_len, dec_hidden_size)
# attn.shape = batch_size, output_len, context_len
return output, attn
```
### 4.3 Decoder
> Decoder会根据已经翻译的句子内容和context vectors，来决定下一个输出的单词；
```python3=
class Decoder(nn.Module):
def __init__(self, vocab_size, embed_size, enc_hidden_size, dec_hidden_size, dropout=0.2):
super(Decoder, self).__init__()
self.embed = nn.Embedding(vocab_size, embed_size)
self.attention = Attention(enc_hidden_size, dec_hidden_size)
self.rnn = nn.GRU(embed_size, hidden_size, batch_first=True)
self.out = nn.Linear(dec_hidden_size, vocab_size)
self.dropout = nn.Dropout(dropout)
def create_mask(self, x_len, y_len):
# x_len 是一个batch中文句子的长度列表
# y_len 是一个batch英文句子的长度列表
# a mask of shape x_len * y_len
device = x_len.device
max_x_len = x_len.max()
max_y_len = y_len.max()
x_mask = torch.arange(max_x_len, device=device)[None, :] < x_len[:, None]
# print(x_mask.shape) = (batch_size, output_len) # 中文句子的mask
y_mask = torch.arange(max_y_len, device=device)[None, :] < y_len[:, None]
# print(y_mask.shape) = (batch_size, context_len) # 英文句子的mask
mask = ( ~ x_mask[:, :, None] * y_mask[:, None, :]).byte()
# mask = (1 - x_mask[:, :, None] * y_mask[:, None, :]).byte()
# 1-说明取反
# x_mask[:, :, None] = (batch_size, output_len, 1)
# y_mask[:, None, :] = (batch_size, 1, context_len)
# print(mask.shape) = (batch_size, output_len, context_len)
# 注意这个例子的*相乘不是torch.bmm矩阵点乘，只是用到了广播机制而已。
return mask
def forward(self, encoder_out, x_lengths, y, y_lengths, hid):
sorted_len, sorted_idx = y_lengths.sort(0, descending=True)
y_sorted = y[sorted_idx.long()]
hid = hid[:, sorted_idx.long()]
y_sorted = self.dropout(self.embed(y_sorted)) # batch_size, output_length, embed_size
packed_seq = nn.utils.rnn.pack_padded_sequence(y_sorted, sorted_len.long().cpu().data.numpy(), batch_first=True)
out, hid = self.rnn(packed_seq, hid)
unpacked, _ = nn.utils.rnn.pad_packed_sequence(out, batch_first=True)
_, original_idx = sorted_idx.sort(0, descending=False)
output_seq = unpacked[original_idx.long()].contiguous()
hid = hid[:, original_idx.long()].contiguous()
mask = self.create_mask(y_lengths, x_lengths) # 这里真是坑，第一个参数位置是中文句子的长度列表
output, attn = self.attention(output_seq, encoder_out, mask)
# output.shape=(batch_size, output_len, dec_hidden_size)
# attn.shape = batch_size, output_len, context_len
# self.out = nn.Linear(dec_hidden_size, vocab_size)
output = F.log_softmax(self.out(output), -1) # 计算最后的输出概率
# output =(batch_size, output_len, vocab_size)
# 最后一个vocab_size维度 log_softmax
# hid.shape = (1, batch_size, dec_hidden_size)
return output, hid, attn
```
### 4.4 Seq2Seq
> 最后我们构建Seq2Seq模型把encoder, attention, decoder串到一起
```python3=
class Seq2Seq(nn.Module):
def __init__(self, encoder, decoder):
super(Seq2Seq, self).__init__()
self.encoder = encoder
self.decoder = decoder
def forward(self, x, x_lengths, y, y_lengths):
encoder_out, hid = self.encoder(x, x_lengths)
# print(hid.shape)=torch.Size([1, batch_size, dec_hidden_size])
# print(out.shape)=torch.Size([batch_size, seq_len, 2*enc_hidden_size])
output, hid, attn = self.decoder(encoder_out=encoder_out,
x_lengths=x_lengths,
y=y,
y_lengths=y_lengths,
hid=hid)
# output =(batch_size, output_len, vocab_size)
# hid.shape = (1, batch_size, dec_hidden_size)
# attn.shape = (batch_size, output_len, context_len)
return output, attn
def translate(self, x, x_lengths, y, max_length=100):
encoder_out, hid = self.encoder(x, x_lengths)
preds = []
batch_size = x.shape[0]
attns = []
for i in range(max_length):
output, hid, attn = self.decoder(encoder_out,
x_lengths,
y,
torch.ones(batch_size).long().to(y.device),
hid)
y = output.max(2)[1].view(batch_size, 1)
preds.append(y)
attns.append(attn)
return torch.cat(preds, 1), torch.cat(attns, 1)
```
## 5.1 训练模型
```python3=
def train(model, data, num_epochs=20):
for epoch in range(num_epochs):
model.train() # 训练模式
total_num_words = total_loss = 0.
for it, (mb_x, mb_x_len, mb_y, mb_y_len) in enumerate(data):
mb_x = torch.from_numpy(mb_x).to(device).long()
mb_x_len = torch.from_numpy(mb_x_len).to(device).long()
mb_input = torch.from_numpy(mb_y[:, :-1]).to(device).long() # EOS之前
mb_output = torch.from_numpy(mb_y[:, 1:]).to(device).long() # BOS之后
mb_y_len = torch.from_numpy(mb_y_len - 1).to(device).long()
mb_y_len[mb_y_len <= 0] = 1
mb_pred, attn = model(mb_x, mb_x_len, mb_input, mb_y_len)
# [mb_y_len.max()]->[1, mb_y_len.max()]
mb_out_mask = torch.arange(mb_y_len.max().item(), device=device)[None, :] < mb_y_len[:, None]
mb_out_mask = mb_out_mask.float()
# (pre, target, mask)
# mb_output是句子单词的索引
loss = loss_fn(mb_pred, mb_output, mb_out_mask)
num_words = torch.sum(mb_y_len).item()
total_loss += loss.item() * num_words
total_num_words += num_words
# 更新模型
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 5.)
optimizer.step()
if it % 100 == 0:
print("Epoch: ", epoch, 'iteration', it, 'loss:', loss.item())
print("Epoch", epoch, "Training loss", total_loss / total_num_words)
if epoch % 5 == 0:
evaluate(model, dev_data)
torch.save(model.state_dict(), 'translate_model2.pt')
```
### 5.2 训练函数并调用上面的train函数
```python3=
device = 'cpu'
dropout = 0.2
embed_size = hidden_size = 100
encoder = Encoder(vocab_size=en_total_words,
embed_size=embed_size,
enc_hidden_size=hidden_size,
dec_hidden_size=hidden_size,
dropout=dropout)
decoder = Decoder(vocab_size=cn_total_words,
embed_size=embed_size,
enc_hidden_size=hidden_size,
dec_hidden_size=hidden_size,
dropout=dropout)
model = Seq2Seq(encoder, decoder)
model = model.to(device)
loss_fn = LanguageModelCriterion().to(device)
optimizer = torch.optim.Adam(model.parameters())
train(model, train_data, num_epochs=100)
```
## 6. 翻譯
```python3=
def translate_dev(i):
en_sent = " ".join([inv_en_dict[w] for w in dev_en[i]]) #原来的英文
print(en_sent)
cn_sent = "".join([inv_cn_dict[w] for w in dev_cn[i]]) #原来的中文
print("".join(cn_sent))
# 一条句子
mb_x = torch.from_numpy(np.array(dev_en[i]).reshape(1, -1)).long().to(device)
mb_x_len = torch.from_numpy(np.array([len(dev_en[i])])).long().to(device)
bos = torch.Tensor([[cn_dict["BOS"]]]).long().to(device) # shape:[1,1], [[2]]
# y_lengths: [[2]], 一个句子
translation, attn = model.translate(mb_x, mb_x_len, bos) # [1, 10]
# 映射成中文
translation = [inv_cn_dict[i] for i in translation.data.cpu().numpy().reshape(-1)]
trans = []
for word in translation:
if word != "EOS":
trans.append(word)
else:
break
result = "".join(trans)
print(result) #翻译后的中文
return cn_sent, result
```
```python3=
from nltk.translate.bleu_score import sentence_bleu
def BLEU_evaluation(cn_sent, result):
reference = cn_sent
candidate = result
score = sentence_bleu(reference, candidate, weights=(0, 0, 0, 1))
print('Individual 4-gram: %f' % score)
```
```python3=
model.load_state_dict(torch.load('translate_model2.pt', map_location=device))
for i in range(100, 120):
(cn_sent, result) = translate_dev(i)
score = BLEU_evaluation(cn_sent, result)
```
### 6.2 結果輸出
Sign in with Wallet
Connect another wallet