转载自：| 03_language_model/02_Transformer语言模型.ipynb | 从头训练Transformer语言模型 |Open In Colab |

Transformer语言模型

本节训练一个 sequence-to-sequence 模型，使用pytorch的
nn.Transformer <https://pytorch.org/docs/master/nn.html?highlight=nn%20transformer#torch.nn.Transformer> module.

PyTorch 1.2 基于论文 Attention is All YouNeed <https://arxiv.org/pdf/1706.03762.pdf> 实现了一个 Transformer 模型， nn.Transformer 模块依赖于 attention 机制实现表达输入和输出文本的关系。

定义模型

基于 nn.TransformerEncoder 模型训练语言模型。

语言模型任务是为句子后跟随单词输出一个似然概率，表征这个单词可能出现的概率。

首先做 embedding，再做 positional encoding, 表征单词位置关系。nn.TransformerEncoder 由多层nn.TransformerEncoderLayer <https://pytorch.org/docs/master/nn.html?highlight=transformerencoderlayer#torch.nn.TransformerEncoderLayer>组成，对于语言模型任务，每个未来可能出现的单词都需要 mask 并预测其概率，为了得到实际的预测单词，nn.TransformerEncoder模型的输出后需要接一个 log-Softmax 函数。

import math
import torch
import torch.nn as nn
import torch.nn.functional as Fclass TransformerModel(nn.Module):def __init__(self, ntoken, ninp, nhead, nhid, nlayers, dropout=0.5):super(TransformerModel, self).__init__()from torch.nn import TransformerEncoder, TransformerEncoderLayerself.model_type = 'Transformer'self.src_mask = Noneself.pos_encoder = PositionalEncoding(ninp, dropout)encoder_layers = TransformerEncoderLayer(ninp, nhead, nhid, dropout)self.transformer_encoder = TransformerEncoder(encoder_layers, nlayers)self.encoder = nn.Embedding(ntoken, ninp)self.ninp = ninpself.decoder = nn.Linear(ninp, ntoken)self.init_weights()def _generate_square_subsequent_mask(self, sz):mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)mask = mask.float().masked_fill(mask == 0, float('-inf')).masked_fill(mask == 1, float(0.0))return maskdef init_weights(self):initrange = 0.1self.encoder.weight.data.uniform_(-initrange, initrange)self.decoder.bias.data.zero_()self.decoder.weight.data.uniform_(-initrange, initrange)def forward(self, src):if self.src_mask is None or self.src_mask.size(0) != len(src):device = src.devicemask = self._generate_square_subsequent_mask(len(src)).to(device)self.src_mask = masksrc = self.encoder(src) * math.sqrt(self.ninp)src = self.pos_encoder(src)output = self.transformer_encoder(src, self.src_mask)output = self.decoder(output)return output

PositionalEncoding 模块包括 relative 和 absolute 位置编码，positional encodings 与 embeddings 的维度是一样的，这样两者可以相加。

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

加载数据

模型训练过程使用来自 torchtext 的Wikitext-2数据集。vocab 基于 train 数据集构建。batchify（）函数将数据集排列成列，在将数据划分为大小为`batch_size``的批次后，删除所有剩余的标记。

例如，将字母表作为序列（总长度为26），批量大小为4，我们将字母表分成4个长度为6的序列：

import os
import torchtext
from torchtext.data.utils import get_tokenizerdevice = torch.device("cuda" if torch.cuda.is_available() else "cpu")TEXT = torchtext.legacy.data.Field(init_token='<sos>',eos_token='<eos>',lower=True)
train_txt, val_txt, test_txt = torchtext.legacy.datasets.language_modeling.WikiText2.splits(TEXT)
TEXT.build_vocab(train_txt)TEXT

len(train_txt.examples[0].text)
# 2088628

def batchify(data, bsz):data = TEXT.numericalize([data.examples[0].text])# Divide the dataset into bsz parts.nbatch = data.size(0) // bsz# Trim off any extra elements that wouldn't cleanly fit (remainders).data = data.narrow(0, 0, nbatch * bsz)# Evenly divide the data across the bsz batches.data = data.view(bsz, -1).t().contiguous()return data.to(device)batch_size = 20
eval_batch_size = 10
train_data = batchify(train_txt, batch_size)
val_data = batchify(val_txt, eval_batch_size)
test_data = batchify(test_txt, eval_batch_size)print(train_data.shape)
print(val_data.shape)
# torch.Size([104431, 20])
# torch.Size([21764, 10])

定义生成target文本

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

试一下模型效果

设置超参：

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

运行模型

import time
criterion = nn.CrossEntropyLoss()
lr = 5.0  # learning rate
optimizer = torch.optim.SGD(model.parameters(), lr=lr)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 1.0, gamma=0.95)def train():model.train()  # Turn on the train modetotal_loss = 0.start_time = time.time()ntokens = len(TEXT.vocab.stoi)for batch, i in enumerate(range(0, train_data.size(0) - 1, bptt)):data, targets = get_batch(train_data, i)optimizer.zero_grad()output = model(data)loss = criterion(output.view(-1, ntokens), targets)loss.backward()torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5)optimizer.step()total_loss += loss.item()log_interval = 200if batch % log_interval == 0 and batch > 0:cur_loss = total_loss / log_intervalelapsed = time.time() - start_timeprint('| epoch {:3d} | {:5d}/{:5d} batches | ''lr {:02.2f} | ms/batch {:5.2f} | ''loss {:5.2f} | ppl {:8.2f}'.format(epoch, batch, len(train_data) // bptt, scheduler.get_lr()[0],elapsed * 1000 / log_interval,cur_loss, math.exp(cur_loss)))total_loss = 0start_time = time.time()def evaluate(eval_model, data_source):eval_model.eval()  # Turn on the evaluation modetotal_loss = 0.ntokens = len(TEXT.vocab.stoi)with torch.no_grad():for i in range(0, data_source.size(0) - 1, bptt):data, targets = get_batch(data_source, i)output = eval_model(data)output_flat = output.view(-1, ntokens)total_loss += len(data) * criterion(output_flat, targets).item()return total_loss / (len(data_source) - 1)

在validation loss最优时保存模型，在每个epoch结束时调整learning rate。

best_val_loss = float("inf")
epochs = 10  # The number of epochs
best_model = None
MODEL_PATH = 'transformer_lm.pth'
for epoch in range(1, epochs + 1):epoch_start_time = time.time()train()val_loss = evaluate(model, val_data)print('-' * 89)print('| end of epoch {:3d} | time: {:5.2f}s | valid loss {:5.2f} | ''valid ppl {:8.2f}'.format(epoch, (time.time() - epoch_start_time),val_loss, math.exp(val_loss)))print('-' * 100)if val_loss < best_val_loss:best_val_loss = val_lossbest_model = modeltorch.save(best_model.state_dict(), MODEL_PATH)scheduler.step()best_model.load_state_dict(torch.load(MODEL_PATH))

Evaluate the model with the test dataset

Apply the best model to check the result with the test dataset.

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

import os
os.remove('transformer_lm.pth')