.. _sec_natural-language-inference-bert:
Suy luận ngôn ngữ tự nhiên: Tinh chỉnh BERT
===========================================
Trong các phần trước của chương này, chúng tôi đã thiết kế một kiến trúc
dựa trên sự chú ý (trong
:numref:`sec_natural-language-inference-attention`) cho nhiệm vụ suy
luận ngôn ngữ tự nhiên trên tập dữ liệu SNLI (như được mô tả trong
:numref:`sec_natural-language-inference-and-dataset`). Bây giờ chúng
tôi xem lại nhiệm vụ này bằng cách tinh chỉnh BERT. Như đã thảo luận
trong :numref:`sec_finetuning-bert`, suy luận ngôn ngữ tự nhiên là một
bài toán phân loại cặp văn bản cấp trình tự, và tinh chỉnh BERT chỉ đòi
hỏi một kiến trúc dựa trên MLP bổ sung, như minh họa trong
:numref:`fig_nlp-map-nli-bert`.
.. _fig_nlp-map-nli-bert:
.. figure:: ../img/nlp-map-nli-bert.svg
This section feeds pretrained BERT to an MLP-based architecture for
natural language inference.
Trong phần này, chúng tôi sẽ tải xuống một phiên bản nhỏ được đào tạo
trước của BERT, sau đó tinh chỉnh nó để suy luận ngôn ngữ tự nhiên trên
bộ dữ liệu SNLI.
.. raw:: html
.. raw:: html
.. code:: python
import json
import multiprocessing
import os
from mxnet import gluon, np, npx
from mxnet.gluon import nn
from d2l import mxnet as d2l
npx.set_np()
.. raw:: html
.. raw:: html
.. code:: python
import json
import multiprocessing
import os
import torch
from torch import nn
from d2l import torch as d2l
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
d2l.DATA_HUB['bert.base'] = (d2l.DATA_URL + 'bert.base.zip',
'7b3820b35da691042e5d34c0971ac3edbd80d3f4')
d2l.DATA_HUB['bert.small'] = (d2l.DATA_URL + 'bert.small.zip',
'a4e718a47137ccd1809c9107ab4f5edd317bae2c')
.. raw:: html
.. raw:: html
.. code:: python
d2l.DATA_HUB['bert.base'] = (d2l.DATA_URL + 'bert.base.torch.zip',
'225d66f04cae318b841a13d32af3acc165f253ac')
d2l.DATA_HUB['bert.small'] = (d2l.DATA_URL + 'bert.small.torch.zip',
'c72329e68a732bef0452e4b96a1c341c8910f81f')
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
def load_pretrained_model(pretrained_model, num_hiddens, ffn_num_hiddens,
num_heads, num_layers, dropout, max_len, devices):
data_dir = d2l.download_extract(pretrained_model)
# Define an empty vocabulary to load the predefined vocabulary
vocab = d2l.Vocab()
vocab.idx_to_token = json.load(open(os.path.join(data_dir, 'vocab.json')))
vocab.token_to_idx = {token: idx for idx, token in enumerate(
vocab.idx_to_token)}
bert = d2l.BERTModel(len(vocab), num_hiddens, ffn_num_hiddens, num_heads,
num_layers, dropout, max_len)
# Load pretrained BERT parameters
bert.load_parameters(os.path.join(data_dir, 'pretrained.params'),
ctx=devices)
return bert, vocab
.. raw:: html
.. raw:: html
.. code:: python
def load_pretrained_model(pretrained_model, num_hiddens, ffn_num_hiddens,
num_heads, num_layers, dropout, max_len, devices):
data_dir = d2l.download_extract(pretrained_model)
# Define an empty vocabulary to load the predefined vocabulary
vocab = d2l.Vocab()
vocab.idx_to_token = json.load(open(os.path.join(data_dir, 'vocab.json')))
vocab.token_to_idx = {token: idx for idx, token in enumerate(
vocab.idx_to_token)}
bert = d2l.BERTModel(len(vocab), num_hiddens, norm_shape=[256],
ffn_num_input=256, ffn_num_hiddens=ffn_num_hiddens,
num_heads=4, num_layers=2, dropout=0.2,
max_len=max_len, key_size=256, query_size=256,
value_size=256, hid_in_features=256,
mlm_in_features=256, nsp_in_features=256)
# Load pretrained BERT parameters
bert.load_state_dict(torch.load(os.path.join(data_dir,
'pretrained.params')))
return bert, vocab
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
devices = d2l.try_all_gpus()
bert, vocab = load_pretrained_model(
'bert.small', num_hiddens=256, ffn_num_hiddens=512, num_heads=4,
num_layers=2, dropout=0.1, max_len=512, devices=devices)
.. parsed-literal::
:class: output
Downloading ../data/bert.small.zip from http://d2l-data.s3-accelerate.amazonaws.com/bert.small.zip...
.. raw:: html
.. raw:: html
.. code:: python
devices = d2l.try_all_gpus()
bert, vocab = load_pretrained_model(
'bert.small', num_hiddens=256, ffn_num_hiddens=512, num_heads=4,
num_layers=2, dropout=0.1, max_len=512, devices=devices)
.. parsed-literal::
:class: output
Downloading ../data/bert.small.torch.zip from http://d2l-data.s3-accelerate.amazonaws.com/bert.small.torch.zip...
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
class SNLIBERTDataset(gluon.data.Dataset):
def __init__(self, dataset, max_len, vocab=None):
all_premise_hypothesis_tokens = [[
p_tokens, h_tokens] for p_tokens, h_tokens in zip(
*[d2l.tokenize([s.lower() for s in sentences])
for sentences in dataset[:2]])]
self.labels = np.array(dataset[2])
self.vocab = vocab
self.max_len = max_len
(self.all_token_ids, self.all_segments,
self.valid_lens) = self._preprocess(all_premise_hypothesis_tokens)
print('read ' + str(len(self.all_token_ids)) + ' examples')
def _preprocess(self, all_premise_hypothesis_tokens):
pool = multiprocessing.Pool(4) # Use 4 worker processes
out = pool.map(self._mp_worker, all_premise_hypothesis_tokens)
all_token_ids = [
token_ids for token_ids, segments, valid_len in out]
all_segments = [segments for token_ids, segments, valid_len in out]
valid_lens = [valid_len for token_ids, segments, valid_len in out]
return (np.array(all_token_ids, dtype='int32'),
np.array(all_segments, dtype='int32'),
np.array(valid_lens))
def _mp_worker(self, premise_hypothesis_tokens):
p_tokens, h_tokens = premise_hypothesis_tokens
self._truncate_pair_of_tokens(p_tokens, h_tokens)
tokens, segments = d2l.get_tokens_and_segments(p_tokens, h_tokens)
token_ids = self.vocab[tokens] + [self.vocab['
']] \
* (self.max_len - len(tokens))
segments = segments + [0] * (self.max_len - len(segments))
valid_len = len(tokens)
return token_ids, segments, valid_len
def _truncate_pair_of_tokens(self, p_tokens, h_tokens):
# Reserve slots for '', '', and '' tokens for the BERT
# input
while len(p_tokens) + len(h_tokens) > self.max_len - 3:
if len(p_tokens) > len(h_tokens):
p_tokens.pop()
else:
h_tokens.pop()
def __getitem__(self, idx):
return (self.all_token_ids[idx], self.all_segments[idx],
self.valid_lens[idx]), self.labels[idx]
def __len__(self):
return len(self.all_token_ids)
.. raw:: html
.. raw:: html
.. code:: python
class SNLIBERTDataset(torch.utils.data.Dataset):
def __init__(self, dataset, max_len, vocab=None):
all_premise_hypothesis_tokens = [[
p_tokens, h_tokens] for p_tokens, h_tokens in zip(
*[d2l.tokenize([s.lower() for s in sentences])
for sentences in dataset[:2]])]
self.labels = torch.tensor(dataset[2])
self.vocab = vocab
self.max_len = max_len
(self.all_token_ids, self.all_segments,
self.valid_lens) = self._preprocess(all_premise_hypothesis_tokens)
print('read ' + str(len(self.all_token_ids)) + ' examples')
def _preprocess(self, all_premise_hypothesis_tokens):
pool = multiprocessing.Pool(4) # Use 4 worker processes
out = pool.map(self._mp_worker, all_premise_hypothesis_tokens)
all_token_ids = [
token_ids for token_ids, segments, valid_len in out]
all_segments = [segments for token_ids, segments, valid_len in out]
valid_lens = [valid_len for token_ids, segments, valid_len in out]
return (torch.tensor(all_token_ids, dtype=torch.long),
torch.tensor(all_segments, dtype=torch.long),
torch.tensor(valid_lens))
def _mp_worker(self, premise_hypothesis_tokens):
p_tokens, h_tokens = premise_hypothesis_tokens
self._truncate_pair_of_tokens(p_tokens, h_tokens)
tokens, segments = d2l.get_tokens_and_segments(p_tokens, h_tokens)
token_ids = self.vocab[tokens] + [self.vocab['
']] \
* (self.max_len - len(tokens))
segments = segments + [0] * (self.max_len - len(segments))
valid_len = len(tokens)
return token_ids, segments, valid_len
def _truncate_pair_of_tokens(self, p_tokens, h_tokens):
# Reserve slots for '', '', and '' tokens for the BERT
# input
while len(p_tokens) + len(h_tokens) > self.max_len - 3:
if len(p_tokens) > len(h_tokens):
p_tokens.pop()
else:
h_tokens.pop()
def __getitem__(self, idx):
return (self.all_token_ids[idx], self.all_segments[idx],
self.valid_lens[idx]), self.labels[idx]
def __len__(self):
return len(self.all_token_ids)
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
# Reduce `batch_size` if there is an out of memory error. In the original BERT
# model, `max_len` = 512
batch_size, max_len, num_workers = 512, 128, d2l.get_dataloader_workers()
data_dir = d2l.download_extract('SNLI')
train_set = SNLIBERTDataset(d2l.read_snli(data_dir, True), max_len, vocab)
test_set = SNLIBERTDataset(d2l.read_snli(data_dir, False), max_len, vocab)
train_iter = gluon.data.DataLoader(train_set, batch_size, shuffle=True,
num_workers=num_workers)
test_iter = gluon.data.DataLoader(test_set, batch_size,
num_workers=num_workers)
.. parsed-literal::
:class: output
read 549367 examples
read 9824 examples
.. raw:: html
.. raw:: html
.. code:: python
# Reduce `batch_size` if there is an out of memory error. In the original BERT
# model, `max_len` = 512
batch_size, max_len, num_workers = 512, 128, d2l.get_dataloader_workers()
data_dir = d2l.download_extract('SNLI')
train_set = SNLIBERTDataset(d2l.read_snli(data_dir, True), max_len, vocab)
test_set = SNLIBERTDataset(d2l.read_snli(data_dir, False), max_len, vocab)
train_iter = torch.utils.data.DataLoader(train_set, batch_size, shuffle=True,
num_workers=num_workers)
test_iter = torch.utils.data.DataLoader(test_set, batch_size,
num_workers=num_workers)
.. parsed-literal::
:class: output
read 549367 examples
read 9824 examples
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
class BERTClassifier(nn.Block):
def __init__(self, bert):
super(BERTClassifier, self).__init__()
self.encoder = bert.encoder
self.hidden = bert.hidden
self.output = nn.Dense(3)
def forward(self, inputs):
tokens_X, segments_X, valid_lens_x = inputs
encoded_X = self.encoder(tokens_X, segments_X, valid_lens_x)
return self.output(self.hidden(encoded_X[:, 0, :]))
.. raw:: html
.. raw:: html
.. code:: python
class BERTClassifier(nn.Module):
def __init__(self, bert):
super(BERTClassifier, self).__init__()
self.encoder = bert.encoder
self.hidden = bert.hidden
self.output = nn.Linear(256, 3)
def forward(self, inputs):
tokens_X, segments_X, valid_lens_x = inputs
encoded_X = self.encoder(tokens_X, segments_X, valid_lens_x)
return self.output(self.hidden(encoded_X[:, 0, :]))
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
net = BERTClassifier(bert)
net.output.initialize(ctx=devices)
.. raw:: html
.. raw:: html
.. code:: python
net = BERTClassifier(bert)
.. raw:: html
.. raw:: html
.. raw:: html
.. code:: python
lr, num_epochs = 1e-4, 5
trainer = gluon.Trainer(net.collect_params(), 'adam', {'learning_rate': lr})
loss = gluon.loss.SoftmaxCrossEntropyLoss()
d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs, devices,
d2l.split_batch_multi_inputs)
.. parsed-literal::
:class: output
loss 0.480, train acc 0.810, test acc 0.785
6981.9 examples/sec on [gpu(0), gpu(1)]
.. figure:: output_natural-language-inference-bert_1857e6_75_1.svg
.. raw:: html
.. raw:: html
.. code:: python
lr, num_epochs = 1e-4, 5
trainer = torch.optim.Adam(net.parameters(), lr=lr)
loss = nn.CrossEntropyLoss(reduction='none')
d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs, devices)
.. parsed-literal::
:class: output
loss 0.518, train acc 0.791, test acc 0.779
10236.7 examples/sec on [device(type='cuda', index=0), device(type='cuda', index=1)]
.. figure:: output_natural-language-inference-bert_1857e6_78_1.svg
.. raw:: html
.. raw:: html
.. raw:: html
`Discussions `__
.. raw:: html
.. raw:: html
`Discussions `__
.. raw:: html
.. raw:: html