文章目录
简介
本文使用PyTorch自带的transformer层进行机器翻译：从德语翻译为英语。从零开始实现Transformer请参阅PyTorch从零开始实现Transformer，以便于获得对Transfomer更深的理解。
数据集
Multi30k
环境要求
使用torch， torchtext，spacy，其中spacy用来分词。另外，spacy要求在虚拟环境中下载语言模型，以便于进行tokenize（分词）
python
# To install spacy languages do:
python -m spacy download en_core_web_sm
python -m spacy download de_core_news_sm实验代码
代码来源请参考下方的GitHub链接
transformer_translation.py文件
python
# Bleu score 32.02
import torch
import torch.nn as nn
import torch.optim as optim
import spacy
from utils import translate_sentence, bleu, save_checkpoint, load_checkpoint
from torch.utils.tensorboard import SummaryWriter
from torchtext.datasets import Multi30k
from torchtext.data import Field, BucketIterator
"""
To install spacy languages do:
python -m spacy download en_core_web_sm
python -m spacy download de_core_news_sm
"""
spacy_ger = spacy.load("de_core_news_sm")
spacy_eng = spacy.load("en_core_web_sm")
def tokenize_ger(text):
return [tok.text for tok in spacy_ger.tokenizer(text)]
# 将英语进行分词
def tokenize_eng(text):
return [tok.text for tok in spacy_eng.tokenizer(text)]
german = Field(tokenize=tokenize_ger, lower=True, init_token="<sos>", eos_token="<eos>")
english = Field(
tokenize=tokenize_eng, lower=True, init_token="<sos>", eos_token="<eos>"
)
train_data, valid_data, test_data = Multi30k.splits(
exts=(".de", ".en"), fields=(german, english)
)
german.build_vocab(train_data, max_size=10000, min_freq=2)
english.build_vocab(train_data, max_size=10000, min_freq=2)
class Transformer(nn.Module):
def __init__(
self,
embedding_size,
src_vocab_size,
trg_vocab_size,
src_pad_idx,
num_heads,
num_encoder_layers,
num_decoder_layers,
forward_expansion,
dropout,
max_len,
device,
):
super(Transformer, self).__init__()
self.src_word_embedding = nn.Embedding(src_vocab_size, embedding_size)
self.src_position_embedding = nn.Embedding(max_len, embedding_size)
self.trg_word_embedding = nn.Embedding(trg_vocab_size, embedding_size)
self.trg_position_embedding = nn.Embedding(max_len, embedding_size)
self.device = device
self.transformer = nn.Transformer(
embedding_size,
num_heads,
num_encoder_layers,
num_decoder_layers,
forward_expansion,
dropout,
)
self.fc_out = nn.Linear(embedding_size, trg_vocab_size)
self.dropout = nn.Dropout(dropout)
self.src_pad_idx = src_pad_idx
def make_src_mask(self, src):
src_mask = src.transpose(0, 1) == self.src_pad_idx
# (N, src_len)
return src_mask.to(self.device)
def forward(self, src, trg):
src_seq_length, N = src.shape
trg_seq_length, N = trg.shape
src_positions = (
torch.arange(0, src_seq_length)
.unsqueeze(1)
.expand(src_seq_length, N)
.to(self.device)
)
trg_positions = (
torch.arange(0, trg_seq_length)
.unsqueeze(1)
.expand(trg_seq_length, N)
.to(self.device)
)
embed_src = self.dropout(
(self.src_word_embedding(src) + self.src_position_embedding(src_positions))
)
embed_trg = self.dropout(
(self.trg_word_embedding(trg) + self.trg_position_embedding(trg_positions))
)
src_padding_mask = self.make_src_mask(src)
trg_mask = self.transformer.generate_square_subsequent_mask(trg_seq_length).to(
self.device
)
out = self.transformer(
embed_src,
embed_trg,
src_key_padding_mask=src_padding_mask,
tgt_mask=trg_mask,
)
out = self.fc_out(out)
return out
# We're ready to define everything we need for training our Seq2Seq model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
load_model = False
save_model = True
# Training hyperparameters
num_epochs = 10
learning_rate = 3e-4
batch_size = 32
# Model hyperparameters
src_vocab_size = len(german.vocab)
trg_vocab_size = len(english.vocab)
embedding_size = 512
num_heads = 8
num_encoder_layers = 3
num_decoder_layers = 3
dropout = 0.10
max_len = 100
forward_expansion = 4
src_pad_idx = english.vocab.stoi["<pad>"]
# Tensorboard to get nice loss plot
writer = SummaryWriter("runs/loss_plot")
step = 0
train_iterator, valid_iterator, test_iterator = BucketIterator.splits(
(train_data, valid_data, test_data),
batch_size=batch_size,
sort_within_batch=True,
sort_key=lambda x: len(x.src),
device=device,
)
model = Transformer(
embedding_size,
src_vocab_size,
trg_vocab_size,
src_pad_idx,
num_heads,
num_encoder_layers,
num_decoder_layers,
forward_expansion,
dropout,
max_len,
device,
).to(device)
optimizer = optim.Adam(model.parameters(), lr=learning_rate)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer, factor=0.1, patience=10, verbose=True
)
pad_idx = english.vocab.stoi["<pad>"]
criterion = nn.CrossEntropyLoss(ignore_index=pad_idx)
if load_model:
load_checkpoint(torch.load("my_checkpoint.pth.tar"), model, optimizer)
# 'a', 'horse', 'is', 'walking', 'under', 'a', 'bridge', 'next', 'to', 'a', 'boat', '.
sentence = "ein pferd geht unter einer brücke neben einem boot."
for epoch in range(num_epochs):
print(f"[Epoch {epoch} / {num_epochs}]")
if save_model:
checkpoint = {
"state_dict": model.state_dict(),
"optimizer": optimizer.state_dict(),
}
save_checkpoint(checkpoint)
model.eval()
translated_sentence = translate_sentence(
model, sentence, german, english, device, max_length=50
)
print(f"Translated example sentence: \n {translated_sentence}")
model.train()
losses = []
for batch_idx, batch in enumerate(train_iterator):
# Get input and targets and get to cuda
inp_data = batch.src.to(device)
target = batch.trg.to(device)
# Forward prop
output = model(inp_data, target[:-1, :])
# Output is of shape (trg_len, batch_size, output_dim) but Cross Entropy Loss
# doesn't take input in that form. For example if we have MNIST we want to have
# output to be: (N, 10) and targets just (N). Here we can view it in a similar
# way that we have output_words * batch_size that we want to send in into
# our cost function, so we need to do some reshapin.
# Let's also remove the start token while we're at it
output = output.reshape(-1, output.shape[2])
target = target[1:].reshape(-1)
optimizer.zero_grad()
loss = criterion(output, target)
losses.append(loss.item())
# Back prop
loss.backward()
# Clip to avoid exploding gradient issues, makes sure grads are
# within a healthy range
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1)
# Gradient descent step
optimizer.step()
# plot to tensorboard
writer.add_scalar("Training loss", loss, global_step=step)
step += 1
mean_loss = sum(losses) / len(losses)
scheduler.step(mean_loss)
# running on entire test data takes a while
score = bleu(test_data[1:100], model, german, english, device)
print(f"Bleu score {score * 100:.2f}")utils.py文件
python
import torch
import spacy
from torchtext.data.metrics import bleu_score
import sys
def translate_sentence(model, sentence, german, english, device, max_length=50):
# Load german tokenizer
spacy_ger = spacy.load("de_core_news_sm")
# Create tokens using spacy and everything in lower case (which is what our vocab is)
if type(sentence) == str:
tokens = [token.text.lower() for token in spacy_ger(sentence)]
else:
tokens = [token.lower() for token in sentence]
# Add <SOS> and <EOS> in beginning and end respectively
tokens.insert(0, german.init_token)
tokens.append(german.eos_token)
# Go through each german token and convert to an index
text_to_indices = [german.vocab.stoi[token] for token in tokens]
# Convert to Tensor
sentence_tensor = torch.LongTensor(text_to_indices).unsqueeze(1).to(device)
outputs = [english.vocab.stoi["<sos>"]]
for i in range(max_length):
trg_tensor = torch.LongTensor(outputs).unsqueeze(1).to(device)
with torch.no_grad():
output = model(sentence_tensor, trg_tensor)
best_guess = output.argmax(2)[-1, :].item()
outputs.append(best_guess)
if best_guess == english.vocab.stoi["<eos>"]:
break
translated_sentence = [english.vocab.itos[idx] for idx in outputs]
# remove start token
return translated_sentence[1:]
def bleu(data, model, german, english, device):
targets = []
outputs = []
for example in data:
src = vars(example)["src"]
trg = vars(example)["trg"]
prediction = translate_sentence(model, src, german, english, device)
prediction = prediction[:-1] # remove <eos> token
targets.append([trg])
outputs.append(prediction)
return bleu_score(outputs, targets)
def save_checkpoint(state, filename="my_checkpoint.pth.tar"):
print("=> Saving checkpoint")
torch.save(state, filename)
def load_checkpoint(checkpoint, model, optimizer):
print("=> Loading checkpoint")
model.load_state_dict(checkpoint["state_dict"])
optimizer.load_state_dict(checkpoint["optimizer"])实验结果
对下面的这句德文进行翻译
python
sentence = "ein pferd geht unter einer brücke neben einem boot."翻译结果为
python
['a', 'horse', 'walks', 'underneath', 'a', 'bridge', 'next', 'to', 'a', 'boat', '.', '<eos>']Bleu score为31.73
跑了10个Epoch，结果如下所示：
python
# Result
=> Loading checkpoint
[Epoch 0 / 10]
=> Saving checkpoint
Translated example sentence:
['a', 'horse', 'walks', 'under', 'a', 'boat', 'next', 'to', 'a', 'boat', '.', '<eos>']
[Epoch 1 / 10]
=> Saving checkpoint
Translated example sentence:
['a', 'horse', 'walks', 'underneath', 'a', 'bridge', 'beside', 'a', 'boat', '.', '<eos>']
[Epoch 2 / 10]
=> Saving checkpoint
Translated example sentence:
['a', 'horse', 'is', 'walking', 'beside', 'a', 'boat', 'under', 'a', 'bridge', '.', '<eos>']
[Epoch 3 / 10]
=> Saving checkpoint
Translated example sentence:
['a', 'horse', 'walks', 'under', 'a', 'bridge', 'next', 'to', 'a', 'boat', '.', '<eos>']
[Epoch 4 / 10]
=> Saving checkpoint
Translated example sentence:
['a', 'horse', 'walks', 'under', 'a', 'bridge', 'next', 'to', 'a', 'boat', '.', '<eos>']
[Epoch 5 / 10]
=> Saving checkpoint
Translated example sentence:
['a', 'horse', 'walks', 'beside', 'a', 'boat', 'next', 'to', 'a', 'boat', '.', '<eos>']
[Epoch 6 / 10]
=> Saving checkpoint
Translated example sentence:
['a', 'horse', 'is', 'walking', 'underneath', 'a', 'bridge', 'under', 'a', 'boat', '.', '<eos>']
[Epoch 7 / 10]
=> Saving checkpoint
Translated example sentence:
['a', 'horse', 'walks', 'under', 'a', 'bridge', 'next', 'to', 'a', 'boat', '.', '<eos>']
[Epoch 8 / 10]
=> Saving checkpoint
Translated example sentence:
['a', 'horse', 'walks', 'beneath', 'a', 'bridge', 'next', 'to', 'a', 'boat', '.', '<eos>']
[Epoch 9 / 10]
=> Saving checkpoint
Translated example sentence:
['a', 'horse', 'walks', 'underneath', 'a', 'bridge', 'next', 'to', 'a', 'boat', '.', '<eos>']
Bleu score 31.73参考来源
[1] https://blog.csdn.net/weixin_43632501/article/details/98731800