使用 nn.Transformer 和 TorchText 进行序列到序列建模

原文： https://pytorch.org/tutorials/beginner/transformer_tutorial.html

译者： zanshuxun

本教程展示了如何使用nn.Transformer 模块训练一个seq2seq模型。单击此处下载完整的示例代码

PyTorch 1.2 发布了一个基于论文《Attention is All You Need》的标准transformer模块。transformer模型在很多seq2seq问题上效果更好，且更容易实现并行训练。nn.Transformer模块使用一种注意力机制（最近实现的另一种注意力为 nn.MultiheadAttention）来捕捉输出和输入之间的整体依赖关系。 nn.Transformer做到了高度模块化，其中的单个组件也很容易进行修改和使用(例如本教程中的 nn.TransformerEncoder)。

../_images/transformer_architecture.jpg

定义模型

在本教程中，我们训练了一个nn.TransformerEncoder模型来进行语言建模任务。语言建模任务是指：已有一句话，预测其后续出现某个词或某句话的概率。这句话（一串符号）经过嵌入（embedding）层之后，再使用一个位置编码（positional encoding）层来学习其中的词顺序（详见下一段）。nn.TransformerEncoder由多层 nn.TransformerEncoderLayer 组成。除了输入序列之外，还需要一个正方形的注意力掩码矩阵。因为是用已经出现的词预测后面的词，训练过程中模型不能看到后面已经出现的词，需要用mask矩阵掩盖掉。 为了获得每个单词的预测概率，nn.TransformerEncoder后面会接上一个Linear层和softmax层。

  1. import math
  2. import torch
  3. import torch.nn as nn
  4. import torch.nn.functional as F
  6. class TransformerModel(nn.Module):
  8.     def __init__(self, ntoken, ninp, nhead, nhid, nlayers, dropout=0.5):
  9.         super(TransformerModel, self).__init__()
  10.         from torch.nn import TransformerEncoder, TransformerEncoderLayer
  11.         self.model_type = 'Transformer'
  12.         self.src_mask = None
  13.         self.pos_encoder = PositionalEncoding(ninp, dropout)
  14.         encoder_layers = TransformerEncoderLayer(ninp, nhead, nhid, dropout)
  15.         self.transformer_encoder = TransformerEncoder(encoder_layers, nlayers)
  16.         self.encoder = nn.Embedding(ntoken, ninp)
  17.         self.ninp = ninp
  18.         self.decoder = nn.Linear(ninp, ntoken)
  20.         self.init_weights()
  22.     def _generate_square_subsequent_mask(self, sz):
  23.         mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
  24.         mask = mask.float().masked_fill(mask == 0, float('-inf')).masked_fill(mask == 1, float(0.0))
  25.         return mask
  27.     def init_weights(self):
  28.         initrange = 0.1
  29.         self.encoder.weight.data.uniform_(-initrange, initrange)
  30.         self.decoder.bias.data.zero_()
  31.         self.decoder.weight.data.uniform_(-initrange, initrange)
  33.     def forward(self, src):
  34.         if self.src_mask is None or self.src_mask.size(0) != len(src):
  35.             device = src.device
  36.             mask = self._generate_square_subsequent_mask(len(src)).to(device)
  37.             self.src_mask = mask
  39.         src = self.encoder(src) * math.sqrt(self.ninp)
  40.         src = self.pos_encoder(src)
  41.         output = self.transformer_encoder(src, self.src_mask)
  42.         output = self.decoder(output)
  43.         return output

PositionalEncoding模块能够学到一些序列中符号的相对或绝对位置信息。位置编码层的输出维度与嵌入层相同，两者可以相加。这里我们使用sine和cosine函数来学习单词之间的位置信息。

  1. class PositionalEncoding(nn.Module):
  3.     def __init__(self, d_model, dropout=0.1, max_len=5000):
  4.         super(PositionalEncoding, self).__init__()
  5.         self.dropout = nn.Dropout(p=dropout)
  7.         pe = torch.zeros(max_len, d_model)
  8.         position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
  9.         div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
  10.         pe[:, 0::2] = torch.sin(position * div_term)
  11.         pe[:, 1::2] = torch.cos(position * div_term)
  12.         pe = pe.unsqueeze(0).transpose(0, 1)
  13.         self.register_buffer('pe', pe)
  15.     def forward(self, x):
  16.         x = x + self.pe[:x.size(0), :]
  17.         return self.dropout(x)

加载数据

训练时使用torchtext中的 Wikitext-2 数据集。下面代码中的 vocab 对象可以将数据集中的符号转为张量，batchify()函数用于生成批次数据，将训练集按照batch_size切分为多个序列，并剔除多余的字符。 例如，当输入序列是字母表时(总长度为 26），设置batch_size为 4，batchify()函数会将输入序列分为 4 个长度为 6 的序列：

使用 nn.Transformer 和 TorchText 进行序列到序列建模 - 图2

不同列对于模型来说是独立的，这意味模型无法学习G和F的依赖性，但可以进行更有效的批次训练。

  1. import torchtext
  2. from torchtext.data.utils import get_tokenizer
  3. TEXT = torchtext.data.Field(tokenize=get_tokenizer("basic_english"),
  4.                             init_token='<sos>',
  5.                             eos_token='<eos>',
  6.                             lower=True)
  7. train_txt, val_txt, test_txt = torchtext.datasets.WikiText2.splits(TEXT)
  8. TEXT.build_vocab(train_txt)
  9. device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
  11. def batchify(data, bsz):
  12.     data = TEXT.numericalize([data.examples[0].text])
  13.     # Divide the dataset into bsz parts.
  14.     nbatch = data.size(0) // bsz
  15.     # Trim off any extra elements that wouldn't cleanly fit (remainders).
  16.     data = data.narrow(0, 0, nbatch * bsz)
  17.     # Evenly divide the data across the bsz batches.
  18.     data = data.view(bsz, -1).t().contiguous()
  19.     return data.to(device)
  21. batch_size = 20
  22. eval_batch_size = 10
  23. train_data = batchify(train_txt, batch_size)
  24. val_data = batchify(val_txt, eval_batch_size)
  25. test_data = batchify(test_txt, eval_batch_size)

输出：

  1. downloading wikitext-2-v1.zip
  2. extracting

生成输入序列和目标序列

get_batch()函数为transformer 模型生成输入和目标序列，将源数据切分为长度为bptt的块。 对于语言建模任务，模型需要后面出现的单词作为Target。 例如，bptt值为 2、i = 0 时，get_batch()函数会得到以下两个变量：

../_images/transformer_input_target.png

注意，每一块数据的第0维与 Transformer 模型中的S维度一致，第1维是批次尺寸N。

  1. bptt = 35
  2. def get_batch(source, i):
  3.     seq_len = min(bptt, len(source) - 1 - i)
  4.     data = source[i:i+seq_len]
  5.     target = source[i+1:i+1+seq_len].view(-1)
  6.     return data, target

初始化模型实例

使用下面的超参数创建模型。 词表大小等于 vocab 对象的长度。

  1. ntokens = len(TEXT.vocab.stoi) # the size of vocabulary
  2. emsize = 200 # embedding dimension
  3. nhid = 200 # the dimension of the feedforward network model in nn.TransformerEncoder
  4. nlayers = 2 # the number of nn.TransformerEncoderLayer in nn.TransformerEncoder
  5. nhead = 2 # the number of heads in the multiheadattention models
  6. dropout = 0.2 # the dropout value
  7. model = TransformerModel(ntokens, emsize, nhead, nhid, nlayers, dropout).to(device)

运行模型

损失函数采用CrossEntropyLoss ， 优化器采用SGD 中实现的随机梯度下降法。初始学习率设置为 5.0。 StepLR 用于在不同的迭代伦次（epochs）中调整学习率。 训练时使用 nn.utils.clipgrad_norm 函数将所有梯度进行缩放，来防止发生梯度爆炸。

  1. criterion = nn.CrossEntropyLoss()
  2. lr = 5.0 # learning rate
  3. optimizer = torch.optim.SGD(model.parameters(), lr=lr)
  4. scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 1.0, gamma=0.95)
  6. import time
  7. def train():
  8.     model.train() # Turn on the train mode
  9.     total_loss = 0.
  10.     start_time = time.time()
  11.     ntokens = len(TEXT.vocab.stoi)
  12.     for batch, i in enumerate(range(0, train_data.size(0) - 1, bptt)):
  13.         data, targets = get_batch(train_data, i)
  14.         optimizer.zero_grad()
  15.         output = model(data)
  16.         loss = criterion(output.view(-1, ntokens), targets)
  17.         loss.backward()
  18.         torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5)
  19.         optimizer.step()
  21.         total_loss += loss.item()
  22.         log_interval = 200
  23.         if batch % log_interval == 0 and batch > 0:
  24.             cur_loss = total_loss / log_interval
  25.             elapsed = time.time() - start_time
  26.             print('| epoch {:3d} | {:5d}/{:5d} batches | '
  27.                   'lr {:02.2f} | ms/batch {:5.2f} | '
  28.                   'loss {:5.2f} | ppl {:8.2f}'.format(
  29.                     epoch, batch, len(train_data) // bptt, scheduler.get_lr()[0],
  30.                     elapsed * 1000 / log_interval,
  31.                     cur_loss, math.exp(cur_loss)))
  32.             total_loss = 0
  33.             start_time = time.time()
  35. def evaluate(eval_model, data_source):
  36.     eval_model.eval() # Turn on the evaluation mode
  37.     total_loss = 0.
  38.     ntokens = len(TEXT.vocab.stoi)
  39.     with torch.no_grad():
  40.         for i in range(0, data_source.size(0) - 1, bptt):
  41.             data, targets = get_batch(data_source, i)
  42.             output = eval_model(data)
  43.             output_flat = output.view(-1, ntokens)
  44.             total_loss += len(data) * criterion(output_flat, targets).item()
  45.     return total_loss / (len(data_source) - 1)

对训练集最多遍历epochs次。 如果模型在验证集上的损失达到最优，则保存模型。 每一轮训练结束后都会调整学习率。

  1. best_val_loss = float("inf")
  2. epochs = 3 # The number of epochs
  3. best_model = None
  5. for epoch in range(1, epochs + 1):
  6.     epoch_start_time = time.time()
  7.     train()
  8.     val_loss = evaluate(model, val_data)
  9.     print('-' * 89)
  10.     print('| end of epoch {:3d} | time: {:5.2f}s | valid loss {:5.2f} | '
  11.           'valid ppl {:8.2f}'.format(epoch, (time.time() - epoch_start_time),
  12.                                      val_loss, math.exp(val_loss)))
  13.     print('-' * 89)
  15.     if val_loss < best_val_loss:
  16.         best_val_loss = val_loss
  17.         best_model = model
  19.     scheduler.step()

输出:

  1. | epoch   1 |   200/ 2981 batches | lr 5.00 | ms/batch 29.47 | loss  8.04 | ppl  3112.50
  2. | epoch   1 |   400/ 2981 batches | lr 5.00 | ms/batch 28.38 | loss  6.78 | ppl   882.16
  3. | epoch   1 |   600/ 2981 batches | lr 5.00 | ms/batch 28.38 | loss  6.38 | ppl   589.27
  4. | epoch   1 |   800/ 2981 batches | lr 5.00 | ms/batch 28.40 | loss  6.23 | ppl   508.15
  5. | epoch   1 |  1000/ 2981 batches | lr 5.00 | ms/batch 28.41 | loss  6.12 | ppl   454.63
  6. | epoch   1 |  1200/ 2981 batches | lr 5.00 | ms/batch 28.40 | loss  6.09 | ppl   441.65
  7. | epoch   1 |  1400/ 2981 batches | lr 5.00 | ms/batch 28.42 | loss  6.04 | ppl   418.77
  8. | epoch   1 |  1600/ 2981 batches | lr 5.00 | ms/batch 28.41 | loss  6.04 | ppl   421.53
  9. | epoch   1 |  1800/ 2981 batches | lr 5.00 | ms/batch 28.40 | loss  5.96 | ppl   387.98
  10. | epoch   1 |  2000/ 2981 batches | lr 5.00 | ms/batch 28.41 | loss  5.96 | ppl   386.42
  11. | epoch   1 |  2200/ 2981 batches | lr 5.00 | ms/batch 28.42 | loss  5.85 | ppl   346.77
  12. | epoch   1 |  2400/ 2981 batches | lr 5.00 | ms/batch 28.42 | loss  5.89 | ppl   362.54
  13. | epoch   1 |  2600/ 2981 batches | lr 5.00 | ms/batch 28.42 | loss  5.90 | ppl   364.01
  14. | epoch   1 |  2800/ 2981 batches | lr 5.00 | ms/batch 28.43 | loss  5.80 | ppl   329.20
  15. -----------------------------------------------------------------------------------------
  16. | end of epoch   1 | time: 88.26s | valid loss  5.73 | valid ppl   307.01
  17. -----------------------------------------------------------------------------------------
  18. | epoch   2 |   200/ 2981 batches | lr 4.51 | ms/batch 28.58 | loss  5.79 | ppl   328.13
  19. | epoch   2 |   400/ 2981 batches | lr 4.51 | ms/batch 28.42 | loss  5.77 | ppl   319.25
  20. | epoch   2 |   600/ 2981 batches | lr 4.51 | ms/batch 28.42 | loss  5.60 | ppl   270.79
  21. | epoch   2 |   800/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.63 | ppl   279.91
  22. | epoch   2 |  1000/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.58 | ppl   265.99
  23. | epoch   2 |  1200/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.61 | ppl   273.55
  24. | epoch   2 |  1400/ 2981 batches | lr 4.51 | ms/batch 28.42 | loss  5.63 | ppl   277.59
  25. | epoch   2 |  1600/ 2981 batches | lr 4.51 | ms/batch 28.45 | loss  5.66 | ppl   287.09
  26. | epoch   2 |  1800/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.58 | ppl   266.00
  27. | epoch   2 |  2000/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.61 | ppl   272.58
  28. | epoch   2 |  2200/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.50 | ppl   244.59
  29. | epoch   2 |  2400/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.57 | ppl   262.87
  30. | epoch   2 |  2600/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.58 | ppl   265.65
  31. | epoch   2 |  2800/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.51 | ppl   246.48
  32. -----------------------------------------------------------------------------------------
  33. | end of epoch   2 | time: 88.16s | valid loss  5.53 | valid ppl   253.40
  34. -----------------------------------------------------------------------------------------
  35. | epoch   3 |   200/ 2981 batches | lr 4.29 | ms/batch 28.58 | loss  5.55 | ppl   256.02
  36. | epoch   3 |   400/ 2981 batches | lr 4.29 | ms/batch 28.43 | loss  5.55 | ppl   256.76
  37. | epoch   3 |   600/ 2981 batches | lr 4.29 | ms/batch 28.46 | loss  5.36 | ppl   212.31
  38. | epoch   3 |   800/ 2981 batches | lr 4.29 | ms/batch 28.44 | loss  5.42 | ppl   225.88
  39. | epoch   3 |  1000/ 2981 batches | lr 4.29 | ms/batch 28.46 | loss  5.38 | ppl   217.24
  40. | epoch   3 |  1200/ 2981 batches | lr 4.29 | ms/batch 28.45 | loss  5.41 | ppl   223.82
  41. | epoch   3 |  1400/ 2981 batches | lr 4.29 | ms/batch 28.43 | loss  5.42 | ppl   226.87
  42. | epoch   3 |  1600/ 2981 batches | lr 4.29 | ms/batch 28.44 | loss  5.47 | ppl   238.34
  43. | epoch   3 |  1800/ 2981 batches | lr 4.29 | ms/batch 28.45 | loss  5.41 | ppl   223.13
  44. | epoch   3 |  2000/ 2981 batches | lr 4.29 | ms/batch 28.45 | loss  5.44 | ppl   230.23
  45. | epoch   3 |  2200/ 2981 batches | lr 4.29 | ms/batch 28.44 | loss  5.32 | ppl   205.28
  46. | epoch   3 |  2400/ 2981 batches | lr 4.29 | ms/batch 28.44 | loss  5.40 | ppl   221.60
  47. | epoch   3 |  2600/ 2981 batches | lr 4.29 | ms/batch 28.45 | loss  5.42 | ppl   224.76
  48. | epoch   3 |  2800/ 2981 batches | lr 4.29 | ms/batch 28.44 | loss  5.34 | ppl   209.38
  49. -----------------------------------------------------------------------------------------
  50. | end of epoch   3 | time: 88.18s | valid loss  5.48 | valid ppl   240.75
  51. -----------------------------------------------------------------------------------------

在测试集上评估模型

在测试集上使用保存的最优模型来评估效果。

  1. test_loss = evaluate(best_model, test_data)
  2. print('=' * 89)
  3. print('| End of training | test loss {:5.2f} | test ppl {:8.2f}'.format(
  4.     test_loss, math.exp(test_loss)))
  5. print('=' * 89)

输出:

  1. =========================================================================================
  2. | End of training | test loss  5.39 | test ppl   219.13
  3. =========================================================================================

脚本的总运行时间：(4 分钟 42.167 秒）