官网链接

Language Modeling with nn.Transformer and torchtext — PyTorch Tutorials 2.0.1+cu117 documentation

使用 NN.TRANSFORMER 和 TORCHTEXT进行语言建模

这是一个关于训练模型使用nn.Transformer来预测序列中的下一个单词的教程。

PyTorch 1.2版本包含了一个基于论文Attention is All You Need的标准transformer模块。与循环神经网络(RNNs)相比，transformer模型已被证明在许多序列对序列任务中具有更高的质量，同时具有更高的并发性。

nn.Transformer 模块完全依赖于注意力机制(作为nn.MultiheadAttention实现)来绘制输入和输出之间的全局依赖关系。nn.Transformer 模块是高度模块化的，这样一个单一的组件(例如，nn.TransformerEncoder)可以很容易地使用/组合。

定义模型

在本教程中，我们在语言建模任务上训练一个nn.TransformerEncoder模型。请注意，本教程不包括nn.TransformerDecoder的训练，如上图右半部分所示。语言建模任务是为给定单词(或单词序列)跟随单词序列的可能性分配一个概率。首先将一系列标记传递给源文本嵌入层，然后是一个位置编码器来解释单词的顺序(请参阅下一段了解更多细节)。nn.TransformerEncoder 由nn.TransformerEncoderLayer的多个层组成。除了输入序列外，还需要一个方形注意掩码，因为在nn.TransformerDecoder中的自注意层只允许参与序列中较早的位置。对于语言建模任务，应掩盖未来位置上的任何标记。为了生成输出词的概率分布，nn.TransformerEncoder 模型的输出通过一个线性层来输出非规范化的对数。这里没有应用log-softmax函数，因为稍后会使用CrossEntropyLoss，它要求输入是非标准化的对数。

import math
import os
from tempfile import TemporaryDirectory
from typing import Tupleimport torch
from torch import nn, Tensor
from torch.nn import TransformerEncoder, TransformerEncoderLayer
from torch.utils.data import datasetclass TransformerModel(nn.Module):def __init__(self, ntoken: int, d_model: int, nhead: int, d_hid: int,nlayers: int, dropout: float = 0.5):super().__init__()self.model_type = 'Transformer'self.pos_encoder = PositionalEncoding(d_model, dropout)encoder_layers = TransformerEncoderLayer(d_model, nhead, d_hid, dropout)self.transformer_encoder = TransformerEncoder(encoder_layers, nlayers)self.embedding = nn.Embedding(ntoken, d_model)self.d_model = d_modelself.linear = nn.Linear(d_model, ntoken)self.init_weights()def init_weights(self) -> None:initrange = 0.1self.embedding.weight.data.uniform_(-initrange, initrange)self.linear.bias.data.zero_()self.linear.weight.data.uniform_(-initrange, initrange)def forward(self, src: Tensor, src_mask: Tensor = None) -> Tensor:"""Arguments:src: Tensor, shape ``[seq_len, batch_size]``src_mask: Tensor, shape ``[seq_len, seq_len]``Returns:output Tensor of shape ``[seq_len, batch_size, ntoken]``"""src = self.embedding(src) * math.sqrt(self.d_model)src = self.pos_encoder(src)output = self.transformer_encoder(src, src_mask)output = self.linear(output)return output

PositionalEncoding模块注入一些关于序列中记号的相对或绝对位置的信息。位置编码器与文本嵌入层具有相同的维数，因此两者可以相加。这里，我们使用不同频率的正弦和余弦函数。

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

加载和批处理数据

本教程使用torchtext生成Wikitext-2数据集。要访问torchtext数据集，请按照GitHub - pytorch/data: A PyTorch repo for data loading and utilities to be shared by the PyTorch domain libraries. 的说明安装torchdata.

%%bash
pip install portalocker
pip install torchdata

词汇对象是基于训练数据集构建的，用于将令牌数值化为张量。Wikitext-2将稀疏令牌表示为unk。给定顺序数据的一维向量，batchify() 方法将数据排列到batch_size列中。如果数据没有均匀地分成batch_size列，那么数据将被裁剪。例如，以字母表为数据(总长度为26)，batch_size=4，我们将字母表分成长度为6的序列，得到4个这样的序列。

批处理支持更多的并行处理。然而，批处理意味着模型独立处理每一列。在上面的例子中，G和F的依赖关系是无法学习的。

from torchtext.datasets import WikiText2
from torchtext.data.utils import get_tokenizer
from torchtext.vocab import build_vocab_from_iteratortrain_iter = WikiText2(split='train')
tokenizer = get_tokenizer('basic_english')
vocab = build_vocab_from_iterator(map(tokenizer, train_iter), specials=['<unk>'])
vocab.set_default_index(vocab['<unk>'])def data_process(raw_text_iter: dataset.IterableDataset) -> Tensor:"""Converts raw text into a flat Tensor."""data = [torch.tensor(vocab(tokenizer(item)), dtype=torch.long) for item in raw_text_iter]return torch.cat(tuple(filter(lambda t: t.numel() > 0, data)))# ``train_iter`` was "consumed" by the process of building the vocab,
# so we have to create it again
train_iter, val_iter, test_iter = WikiText2()
train_data = data_process(train_iter)
val_data = data_process(val_iter)
test_data = data_process(test_iter)device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')def batchify(data: Tensor, bsz: int) -> Tensor:"""Divides the data into ``bsz`` separate sequences, removing extra elementsthat wouldn't cleanly fit.Arguments:data: Tensor, shape ``[N]``bsz: int, batch sizeReturns:Tensor of shape ``[N // bsz, bsz]``"""seq_len = data.size(0) // bszdata = data[:seq_len * bsz]data = data.view(bsz, seq_len).t().contiguous()return data.to(device)batch_size = 20
eval_batch_size = 10
train_data = batchify(train_data, batch_size)  # shape ``[seq_len, batch_size]``
val_data = batchify(val_data, eval_batch_size)
test_data = batchify(test_data, eval_batch_size)

函数生成输入和目标序列

get_batch() 为transformer模型生成一对输入-目标序列。它将源数据细分为长度为bptt的块。对于语言建模任务，模型需要以下单词作为目标。例如，如果bptt值为2，那么当i = 0时，我们将得到以下两个变量:

应该注意的是，数据块的维度是0，与Transformer模型中的S维度一致。批次维度N 的维度是1.

bptt = 35
def get_batch(source: Tensor, i: int) -> Tuple[Tensor, Tensor]:"""Args:source: Tensor, shape ``[full_seq_len, batch_size]``i: intReturns:tuple (data, target), where data has shape ``[seq_len, batch_size]`` andtarget has shape ``[seq_len * batch_size]``"""seq_len = min(bptt, len(source) - 1 - i)data = source[i:i+seq_len]target = source[i+1:i+1+seq_len].reshape(-1)return data, target

初始化实例

模型超参数定义如下。词汇表大小等于词汇表对象的长度。

ntokens = len(vocab)  # size of vocabulary
emsize = 200  # embedding dimension
d_hid = 200  # dimension of the feedforward network model in ``nn.TransformerEncoder``
nlayers = 2  # number of ``nn.TransformerEncoderLayer`` in ``nn.TransformerEncoder``
nhead = 2  # number of heads in ``nn.MultiheadAttention``
dropout = 0.2  # dropout probability
model = TransformerModel(ntokens, emsize, nhead, d_hid, nlayers, dropout).to(device)

运行模型

我们将CrossEntropyLoss与 SGD (随机梯度下降)优化器一起使用。学习率最初设置为5.0，并遵循StepLR 计划。在训练期间，我们使用nn.utils.clip_grad_norm_ 防止梯度爆炸。

import timecriterion = 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: nn.Module) -> None:model.train()  # turn on train modetotal_loss = 0.log_interval = 200start_time = time.time()num_batches = len(train_data) // bpttfor batch, i in enumerate(range(0, train_data.size(0) - 1, bptt)):data, targets = get_batch(train_data, i)output = model(data)output_flat = output.view(-1, ntokens)loss = criterion(output_flat, targets)optimizer.zero_grad()loss.backward()torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5)optimizer.step()total_loss += loss.item()if batch % log_interval == 0 and batch > 0:lr = scheduler.get_last_lr()[0]ms_per_batch = (time.time() - start_time) * 1000 / log_intervalcur_loss = total_loss / log_intervalppl = math.exp(cur_loss)print(f'| epoch {epoch:3d} | {batch:5d}/{num_batches:5d} batches | 'f'lr {lr:02.2f} | ms/batch {ms_per_batch:5.2f} | 'f'loss {cur_loss:5.2f} | ppl {ppl:8.2f}')total_loss = 0start_time = time.time()def evaluate(model: nn.Module, eval_data: Tensor) -> float:model.eval()  # turn on evaluation modetotal_loss = 0.with torch.no_grad():for i in range(0, eval_data.size(0) - 1, bptt):data, targets = get_batch(eval_data, i)seq_len = data.size(0)output = model(data)output_flat = output.view(-1, ntokens)total_loss += seq_len * criterion(output_flat, targets).item()return total_loss / (len(eval_data) - 1)

循环epochs，如果验证损失是目前为止我们看到的最好的，那么保存模型。在每个epoch之后调整学习速率。

best_val_loss = float('inf')
epochs = 3with TemporaryDirectory() as tempdir:best_model_params_path = os.path.join(tempdir, "best_model_params.pt")for epoch in range(1, epochs + 1):epoch_start_time = time.time()train(model)val_loss = evaluate(model, val_data)val_ppl = math.exp(val_loss)elapsed = time.time() - epoch_start_timeprint('-' * 89)print(f'| end of epoch {epoch:3d} | time: {elapsed:5.2f}s | 'f'valid loss {val_loss:5.2f} | valid ppl {val_ppl:8.2f}')print('-' * 89)if val_loss < best_val_loss:best_val_loss = val_losstorch.save(model.state_dict(), best_model_params_path)scheduler.step()model.load_state_dict(torch.load(best_model_params_path)) # load best model states

输出

| epoch   1 |   200/ 2928 batches | lr 5.00 | ms/batch 31.00 | loss  8.15 | ppl  3449.06
| epoch   1 |   400/ 2928 batches | lr 5.00 | ms/batch 28.73 | loss  6.25 | ppl   517.05
| epoch   1 |   600/ 2928 batches | lr 5.00 | ms/batch 28.56 | loss  5.61 | ppl   274.25
| epoch   1 |   800/ 2928 batches | lr 5.00 | ms/batch 28.42 | loss  5.31 | ppl   202.30
| epoch   1 |  1000/ 2928 batches | lr 5.00 | ms/batch 28.33 | loss  4.95 | ppl   140.81
| epoch   1 |  1200/ 2928 batches | lr 5.00 | ms/batch 28.28 | loss  4.55 | ppl    94.20
| epoch   1 |  1400/ 2928 batches | lr 5.00 | ms/batch 28.36 | loss  4.21 | ppl    67.25
| epoch   1 |  1600/ 2928 batches | lr 5.00 | ms/batch 28.45 | loss  3.99 | ppl    54.28
| epoch   1 |  1800/ 2928 batches | lr 5.00 | ms/batch 28.65 | loss  3.74 | ppl    41.89
| epoch   1 |  2000/ 2928 batches | lr 5.00 | ms/batch 28.56 | loss  3.66 | ppl    38.71
| epoch   1 |  2200/ 2928 batches | lr 5.00 | ms/batch 28.67 | loss  3.48 | ppl    32.44
| epoch   1 |  2400/ 2928 batches | lr 5.00 | ms/batch 28.74 | loss  3.49 | ppl    32.78
| epoch   1 |  2600/ 2928 batches | lr 5.00 | ms/batch 28.60 | loss  3.38 | ppl    29.50
| epoch   1 |  2800/ 2928 batches | lr 5.00 | ms/batch 28.46 | loss  3.29 | ppl    26.94
-----------------------------------------------------------------------------------------
| end of epoch   1 | time: 86.92s | valid loss  2.06 | valid ppl     7.88
-----------------------------------------------------------------------------------------
| epoch   2 |   200/ 2928 batches | lr 4.75 | ms/batch 28.88 | loss  3.10 | ppl    22.18
| epoch   2 |   400/ 2928 batches | lr 4.75 | ms/batch 28.50 | loss  3.02 | ppl    20.55
| epoch   2 |   600/ 2928 batches | lr 4.75 | ms/batch 28.65 | loss  2.86 | ppl    17.50
| epoch   2 |   800/ 2928 batches | lr 4.75 | ms/batch 28.68 | loss  2.85 | ppl    17.28
| epoch   2 |  1000/ 2928 batches | lr 4.75 | ms/batch 28.59 | loss  2.67 | ppl    14.43
| epoch   2 |  1200/ 2928 batches | lr 4.75 | ms/batch 28.55 | loss  2.68 | ppl    14.57
| epoch   2 |  1400/ 2928 batches | lr 4.75 | ms/batch 28.51 | loss  2.72 | ppl    15.13
| epoch   2 |  1600/ 2928 batches | lr 4.75 | ms/batch 28.44 | loss  2.69 | ppl    14.71
| epoch   2 |  1800/ 2928 batches | lr 4.75 | ms/batch 28.46 | loss  2.60 | ppl    13.51
| epoch   2 |  2000/ 2928 batches | lr 4.75 | ms/batch 28.51 | loss  2.61 | ppl    13.60
| epoch   2 |  2200/ 2928 batches | lr 4.75 | ms/batch 28.64 | loss  2.57 | ppl    13.04
| epoch   2 |  2400/ 2928 batches | lr 4.75 | ms/batch 28.67 | loss  2.57 | ppl    13.08
| epoch   2 |  2600/ 2928 batches | lr 4.75 | ms/batch 28.56 | loss  2.57 | ppl    13.05
| epoch   2 |  2800/ 2928 batches | lr 4.75 | ms/batch 28.61 | loss  2.55 | ppl    12.81
-----------------------------------------------------------------------------------------
| end of epoch   2 | time: 86.63s | valid loss  1.83 | valid ppl     6.24
-----------------------------------------------------------------------------------------
| epoch   3 |   200/ 2928 batches | lr 4.51 | ms/batch 28.71 | loss  2.43 | ppl    11.35
| epoch   3 |   400/ 2928 batches | lr 4.51 | ms/batch 28.82 | loss  2.37 | ppl    10.65
| epoch   3 |   600/ 2928 batches | lr 4.51 | ms/batch 28.67 | loss  2.27 | ppl     9.64
| epoch   3 |   800/ 2928 batches | lr 4.51 | ms/batch 28.74 | loss  2.29 | ppl     9.83
| epoch   3 |  1000/ 2928 batches | lr 4.51 | ms/batch 28.55 | loss  2.22 | ppl     9.22
| epoch   3 |  1200/ 2928 batches | lr 4.51 | ms/batch 28.73 | loss  2.25 | ppl     9.48
| epoch   3 |  1400/ 2928 batches | lr 4.51 | ms/batch 28.57 | loss  2.29 | ppl     9.89
| epoch   3 |  1600/ 2928 batches | lr 4.51 | ms/batch 28.73 | loss  2.36 | ppl    10.62
| epoch   3 |  1800/ 2928 batches | lr 4.51 | ms/batch 28.52 | loss  2.20 | ppl     9.07
| epoch   3 |  2000/ 2928 batches | lr 4.51 | ms/batch 28.61 | loss  2.26 | ppl     9.57
| epoch   3 |  2200/ 2928 batches | lr 4.51 | ms/batch 28.53 | loss  2.20 | ppl     9.03
| epoch   3 |  2400/ 2928 batches | lr 4.51 | ms/batch 28.45 | loss  2.23 | ppl     9.26
| epoch   3 |  2600/ 2928 batches | lr 4.51 | ms/batch 28.56 | loss  2.21 | ppl     9.13
| epoch   3 |  2800/ 2928 batches | lr 4.51 | ms/batch 28.54 | loss  2.31 | ppl    10.03
-----------------------------------------------------------------------------------------
| end of epoch   3 | time: 86.63s | valid loss  1.28 | valid ppl     3.60
-----------------------------------------------------------------------------------------

在测试数据集上评估最佳模型

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

输出

=========================================================================================
| End of training | test loss  1.26 | test ppl     3.54
=========================================================================================