PyTorch 使用 nn.Transformer 和 TorchText 進(jìn)行序列到序列建模

2020-09-07 17:25 更新
原文： https://pytorch.org/tutorials/beginner/transformer_tutorial.html

這是一個(gè)有關(guān)如何訓(xùn)練使用 nn.Transformer 模塊的序列到序列模型的教程。

PyTorch 1.2 版本包括一個(gè)基于紙張的標(biāo)準(zhǔn)變壓器模塊。 事實(shí)證明，該變壓器模型在許多序列間問(wèn)題上具有較高的質(zhì)量，同時(shí)具有更高的可并行性。 nn.Transformer模塊完全依賴于注意力機(jī)制(另一個(gè)最近實(shí)現(xiàn)為 nn.MultiheadAttention 的模塊）來(lái)繪制輸入和輸出之間的全局依存關(guān)系。 nn.Transformer模塊現(xiàn)在已高度模塊化，因此可以輕松地修改/組成單個(gè)組件(例如本教程中的 nn.TransformerEncoder)。

../_images/transformer_architecture.jpg

定義模型

在本教程中，我們?cè)谡Z(yǔ)言建模任務(wù)上訓(xùn)練nn.TransformerEncoder模型。 語(yǔ)言建模任務(wù)是為給定單詞(或單詞序列）遵循單詞序列的可能性分配概率。 令牌序列首先傳遞到嵌入層，然后傳遞到位置編碼層以說(shuō)明單詞的順序(有關(guān)更多詳細(xì)信息，請(qǐng)參見(jiàn)下一段）。 nn.TransformerEncoder由多層 nn.TransformerEncoderLayer 組成。 與輸入序列一起，還需要一個(gè)正方形的注意掩碼，因?yàn)?code>nn.TransformerEncoder中的自注意層只允許出現(xiàn)在該序列中的較早位置。 對(duì)于語(yǔ)言建模任務(wù)，應(yīng)屏蔽將來(lái)頭寸上的所有標(biāo)記。 為了獲得實(shí)際的單詞，nn.TransformerEncoder模型的輸出將發(fā)送到最終的 Linear 層，然后是 log-Softmax 函數(shù)。

import math
import torch
import torch.nn as nn
import torch.nn.functional as F


class TransformerModel(nn.Module):


    def __init__(self, ntoken, ninp, nhead, nhid, nlayers, dropout=0.5):
        super(TransformerModel, self).__init__()
        from torch.nn import TransformerEncoder, TransformerEncoderLayer
        self.model_type = 'Transformer'
        self.src_mask = None
        self.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 = ninp
        self.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 mask


    def init_weights(self):
        initrange = 0.1
        self.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.device
            mask = self._generate_square_subsequent_mask(len(src)).to(device)
            self.src_mask = mask


        src = 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模塊注入一些有關(guān)令牌在序列中的相對(duì)或絕對(duì)位置的信息。 位置編碼的尺寸與嵌入的尺寸相同，因此可以將兩者相加。 在這里，我們使用不同頻率的sine和cosine功能。

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)

加載和批處理數(shù)據(jù)

訓(xùn)練過(guò)程使用torchtext中的 Wikitext-2 數(shù)據(jù)集。 vocab 對(duì)象是基于訓(xùn)練數(shù)據(jù)集構(gòu)建的，用于將標(biāo)記數(shù)字化為張量。 從順序數(shù)據(jù)開(kāi)始，batchify()函數(shù)將數(shù)據(jù)集排列為列，以修剪掉數(shù)據(jù)分成大小為batch_size的批次后剩余的所有令牌。 例如，以字母為序列(總長(zhǎng)度為 26）并且批大小為 4，我們將字母分為 4 個(gè)長(zhǎng)度為 6 的序列：

這些列被模型視為獨(dú)立的，這意味著無(wú)法了解G和F的依賴性，但可以進(jìn)行更有效的批處理。

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


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)

得出：

downloading wikitext-2-v1.zip
extracting

產(chǎn)生輸入和目標(biāo)序列的功能

get_batch()功能為變壓器模型生成輸入和目標(biāo)序列。 它將源數(shù)據(jù)細(xì)分為長(zhǎng)度為bptt的塊。 對(duì)于語(yǔ)言建模任務(wù)，模型需要以下單詞作為Target。 例如，如果bptt值為 2，則i = 0 時(shí)，我們將獲得以下兩個(gè)變量：

../_images/transformer_input_target.png

應(yīng)該注意的是，這些塊沿維度 0，與 Transformer 模型中的S維度一致。 批次尺寸N沿尺寸 1。

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

啟動(dòng)實(shí)例

使用下面的超參數(shù)建立模型。 vocab 的大小等于 vocab 對(duì)象的長(zhǎng)度。

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)

運(yùn)行模型

CrossEntropyLoss 用于跟蹤損失， SGD 實(shí)現(xiàn)隨機(jī)梯度下降法作為優(yōu)化器。 初始學(xué)習(xí)率設(shè)置為 5.0。 StepLR 用于通過(guò)歷時(shí)調(diào)整學(xué)習(xí)速率。 在訓(xùn)練過(guò)程中，我們使用 nn.utils.clipgrad_norm 函數(shù)將所有梯度縮放在一起，以防止爆炸。

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)


import time
def train():
    model.train() # Turn on the train mode
    total_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 = 200
        if batch % log_interval == 0 and batch > 0:
            cur_loss = total_loss / log_interval
            elapsed = time.time() - start_time
            print('| 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 = 0
            start_time = time.time()


def evaluate(eval_model, data_source):
    eval_model.eval() # Turn on the evaluation mode
    total_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)

循環(huán)遍歷。 如果驗(yàn)證損失是迄今為止迄今為止最好的，請(qǐng)保存模型。 在每個(gè)時(shí)期之后調(diào)整學(xué)習(xí)率。

best_val_loss = float("inf")
epochs = 3 # The number of epochs
best_model = None


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('-' * 89)


    if val_loss < best_val_loss:
        best_val_loss = val_loss
        best_model = model


    scheduler.step()

得出:

| epoch   1 |   200/ 2981 batches | lr 5.00 | ms/batch 29.47 | loss  8.04 | ppl  3112.50
| epoch   1 |   400/ 2981 batches | lr 5.00 | ms/batch 28.38 | loss  6.78 | ppl   882.16
| epoch   1 |   600/ 2981 batches | lr 5.00 | ms/batch 28.38 | loss  6.38 | ppl   589.27
| epoch   1 |   800/ 2981 batches | lr 5.00 | ms/batch 28.40 | loss  6.23 | ppl   508.15
| epoch   1 |  1000/ 2981 batches | lr 5.00 | ms/batch 28.41 | loss  6.12 | ppl   454.63
| epoch   1 |  1200/ 2981 batches | lr 5.00 | ms/batch 28.40 | loss  6.09 | ppl   441.65
| epoch   1 |  1400/ 2981 batches | lr 5.00 | ms/batch 28.42 | loss  6.04 | ppl   418.77
| epoch   1 |  1600/ 2981 batches | lr 5.00 | ms/batch 28.41 | loss  6.04 | ppl   421.53
| epoch   1 |  1800/ 2981 batches | lr 5.00 | ms/batch 28.40 | loss  5.96 | ppl   387.98
| epoch   1 |  2000/ 2981 batches | lr 5.00 | ms/batch 28.41 | loss  5.96 | ppl   386.42
| epoch   1 |  2200/ 2981 batches | lr 5.00 | ms/batch 28.42 | loss  5.85 | ppl   346.77
| epoch   1 |  2400/ 2981 batches | lr 5.00 | ms/batch 28.42 | loss  5.89 | ppl   362.54
| epoch   1 |  2600/ 2981 batches | lr 5.00 | ms/batch 28.42 | loss  5.90 | ppl   364.01
| epoch   1 |  2800/ 2981 batches | lr 5.00 | ms/batch 28.43 | loss  5.80 | ppl   329.20
-----------------------------------------------------------------------------------------
| end of epoch   1 | time: 88.26s | valid loss  5.73 | valid ppl   307.01
-----------------------------------------------------------------------------------------
| epoch   2 |   200/ 2981 batches | lr 4.51 | ms/batch 28.58 | loss  5.79 | ppl   328.13
| epoch   2 |   400/ 2981 batches | lr 4.51 | ms/batch 28.42 | loss  5.77 | ppl   319.25
| epoch   2 |   600/ 2981 batches | lr 4.51 | ms/batch 28.42 | loss  5.60 | ppl   270.79
| epoch   2 |   800/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.63 | ppl   279.91
| epoch   2 |  1000/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.58 | ppl   265.99
| epoch   2 |  1200/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.61 | ppl   273.55
| epoch   2 |  1400/ 2981 batches | lr 4.51 | ms/batch 28.42 | loss  5.63 | ppl   277.59
| epoch   2 |  1600/ 2981 batches | lr 4.51 | ms/batch 28.45 | loss  5.66 | ppl   287.09
| epoch   2 |  1800/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.58 | ppl   266.00
| epoch   2 |  2000/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.61 | ppl   272.58
| epoch   2 |  2200/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.50 | ppl   244.59
| epoch   2 |  2400/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.57 | ppl   262.87
| epoch   2 |  2600/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.58 | ppl   265.65
| epoch   2 |  2800/ 2981 batches | lr 4.51 | ms/batch 28.44 | loss  5.51 | ppl   246.48
-----------------------------------------------------------------------------------------
| end of epoch   2 | time: 88.16s | valid loss  5.53 | valid ppl   253.40
-----------------------------------------------------------------------------------------
| epoch   3 |   200/ 2981 batches | lr 4.29 | ms/batch 28.58 | loss  5.55 | ppl   256.02
| epoch   3 |   400/ 2981 batches | lr 4.29 | ms/batch 28.43 | loss  5.55 | ppl   256.76
| epoch   3 |   600/ 2981 batches | lr 4.29 | ms/batch 28.46 | loss  5.36 | ppl   212.31
| epoch   3 |   800/ 2981 batches | lr 4.29 | ms/batch 28.44 | loss  5.42 | ppl   225.88
| epoch   3 |  1000/ 2981 batches | lr 4.29 | ms/batch 28.46 | loss  5.38 | ppl   217.24
| epoch   3 |  1200/ 2981 batches | lr 4.29 | ms/batch 28.45 | loss  5.41 | ppl   223.82
| epoch   3 |  1400/ 2981 batches | lr 4.29 | ms/batch 28.43 | loss  5.42 | ppl   226.87
| epoch   3 |  1600/ 2981 batches | lr 4.29 | ms/batch 28.44 | loss  5.47 | ppl   238.34
| epoch   3 |  1800/ 2981 batches | lr 4.29 | ms/batch 28.45 | loss  5.41 | ppl   223.13
| epoch   3 |  2000/ 2981 batches | lr 4.29 | ms/batch 28.45 | loss  5.44 | ppl   230.23
| epoch   3 |  2200/ 2981 batches | lr 4.29 | ms/batch 28.44 | loss  5.32 | ppl   205.28
| epoch   3 |  2400/ 2981 batches | lr 4.29 | ms/batch 28.44 | loss  5.40 | ppl   221.60
| epoch   3 |  2600/ 2981 batches | lr 4.29 | ms/batch 28.45 | loss  5.42 | ppl   224.76
| epoch   3 |  2800/ 2981 batches | lr 4.29 | ms/batch 28.44 | loss  5.34 | ppl   209.38
-----------------------------------------------------------------------------------------
| end of epoch   3 | time: 88.18s | valid loss  5.48 | valid ppl   240.75
-----------------------------------------------------------------------------------------

使用測(cè)試數(shù)據(jù)集評(píng)估模型

應(yīng)用最佳模型以檢查測(cè)試數(shù)據(jù)集的結(jié)果。

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)

得出:

=========================================================================================
| End of training | test loss  5.39 | test ppl   219.13
=========================================================================================

腳本的總運(yùn)行時(shí)間：(4 分鐘 39.556 秒）

Download Python source code: transformer_tutorial.py Download Jupyter notebook: transformer_tutorial.ipynb

由獅身人面像畫廊生成的畫廊


以上內(nèi)容是否對(duì)您有幫助：
在線筆記
App下載
App下載

掃描二維碼

下載編程獅App

公眾號(hào)
微信公眾號(hào)

編程獅公眾號(hào)