14.10. 预训练BERT¶ Open the notebook in SageMaker Studio Lab
利用 14.8节中实现的BERT模型和 14.9节中从WikiText-2数据集生成的预训练样本，我们将在本节中在WikiText-2数据集上对BERT进行预训练。
from mxnet import autograd, gluon, init, np, npx
from d2l import mxnet as d2l
npx.set_np()
import torch
from torch import nn
from d2l import torch as d2l
import warnings
from d2l import paddle as d2l
warnings.filterwarnings("ignore")
import paddle
from paddle import nn
首先，我们加载WikiText-2数据集作为小批量的预训练样本，用于遮蔽语言模型和下一句预测。批量大小是512，BERT输入序列的最大长度是64。注意，在原始BERT模型中，最大长度是512。
batch_size, max_len = 512, 64
train_iter, vocab = d2l.load_data_wiki(batch_size, max_len)
Downloading ../data/wikitext-2-v1.zip from https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-v1.zip...
[07:00:25] ../src/storage/storage.cc:196: Using Pooled (Naive) StorageManager for CPU
batch_size, max_len = 512, 64
train_iter, vocab = d2l.load_data_wiki(batch_size, max_len)
def load_data_wiki(batch_size, max_len):
"""加载WikiText-2数据集
Defined in :numref:`subsec_prepare_mlm_data`"""
data_dir = d2l.download_extract('wikitext-2', 'wikitext-2')
paragraphs = d2l._read_wiki(data_dir)
train_set = d2l._WikiTextDataset(paragraphs, max_len)
train_iter = paddle.io.DataLoader(dataset=train_set, batch_size=batch_size, return_list=True,
shuffle=True, num_workers=0)
return train_iter, train_set.vocab
batch_size, max_len = 512, 64
train_iter, vocab = load_data_wiki(batch_size, max_len)
14.10.1. 预训练BERT¶
原始BERT (Devlin et al., 2018)有两个不同模型尺寸的版本。基本模型（\(\text{BERT}_{\text{BASE}}\)）使用12层（Transformer编码器块），768个隐藏单元（隐藏大小）和12个自注意头。大模型（\(\text{BERT}_{\text{LARGE}}\)）使用24层，1024个隐藏单元和16个自注意头。值得注意的是，前者有1.1亿个参数，后者有3.4亿个参数。为了便于演示，我们定义了一个小的BERT，使用了2层、128个隐藏单元和2个自注意头。
net = d2l.BERTModel(len(vocab), num_hiddens=128, ffn_num_hiddens=256,
num_heads=2, num_layers=2, dropout=0.2)
devices = d2l.try_all_gpus()
net.initialize(init.Xavier(), ctx=devices)
loss = gluon.loss.SoftmaxCELoss()
[07:01:34] ../src/storage/storage.cc:196: Using Pooled (Naive) StorageManager for GPU
[07:01:34] ../src/storage/storage.cc:196: Using Pooled (Naive) StorageManager for GPU
net = d2l.BERTModel(len(vocab), num_hiddens=128, norm_shape=[128],
ffn_num_input=128, ffn_num_hiddens=256, num_heads=2,
num_layers=2, dropout=0.2, key_size=128, query_size=128,
value_size=128, hid_in_features=128, mlm_in_features=128,
nsp_in_features=128)
devices = d2l.try_all_gpus()
loss = nn.CrossEntropyLoss()
net = d2l.BERTModel(len(vocab), num_hiddens=128, norm_shape=[128],
ffn_num_input=128, ffn_num_hiddens=256, num_heads=2,
num_layers=2, dropout=0.2, key_size=128, query_size=128,
value_size=128, hid_in_features=128, mlm_in_features=128,
nsp_in_features=128)
devices = d2l.try_all_gpus()
loss = nn.CrossEntropyLoss()
W0818 09:27:23.462936 94778 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.8, Runtime API Version: 11.8
W0818 09:27:23.494377 94778 gpu_resources.cc:91] device: 0, cuDNN Version: 8.7.
在定义训练代码实现之前，我们定义了一个辅助函数_get_batch_loss_bert。给定训练样本，该函数计算遮蔽语言模型和下一句子预测任务的损失。请注意，BERT预训练的最终损失是遮蔽语言模型损失和下一句预测损失的和。
#@save
def _get_batch_loss_bert(net, loss, vocab_size, tokens_X_shards,
segments_X_shards, valid_lens_x_shards,
pred_positions_X_shards, mlm_weights_X_shards,
mlm_Y_shards, nsp_y_shards):
mlm_ls, nsp_ls, ls = [], [], []
for (tokens_X_shard, segments_X_shard, valid_lens_x_shard,
pred_positions_X_shard, mlm_weights_X_shard, mlm_Y_shard,
nsp_y_shard) in zip(
tokens_X_shards, segments_X_shards, valid_lens_x_shards,
pred_positions_X_shards, mlm_weights_X_shards, mlm_Y_shards,
nsp_y_shards):
# 前向传播
_, mlm_Y_hat, nsp_Y_hat = net(
tokens_X_shard, segments_X_shard, valid_lens_x_shard.reshape(-1),
pred_positions_X_shard)
# 计算遮蔽语言模型损失
mlm_l = loss(
mlm_Y_hat.reshape((-1, vocab_size)), mlm_Y_shard.reshape(-1),
mlm_weights_X_shard.reshape((-1, 1)))
mlm_l = mlm_l.sum() / (mlm_weights_X_shard.sum() + 1e-8)
# 计算下一句子预测任务的损失
nsp_l = loss(nsp_Y_hat, nsp_y_shard)
nsp_l = nsp_l.mean()
mlm_ls.append(mlm_l)
nsp_ls.append(nsp_l)
ls.append(mlm_l + nsp_l)
npx.waitall()
return mlm_ls, nsp_ls, ls
#@save
def _get_batch_loss_bert(net, loss, vocab_size, tokens_X,
segments_X, valid_lens_x,
pred_positions_X, mlm_weights_X,
mlm_Y, nsp_y):
# 前向传播
_, mlm_Y_hat, nsp_Y_hat = net(tokens_X, segments_X,
valid_lens_x.reshape(-1),
pred_positions_X)
# 计算遮蔽语言模型损失
mlm_l = loss(mlm_Y_hat.reshape(-1, vocab_size), mlm_Y.reshape(-1)) *\
mlm_weights_X.reshape(-1, 1)
mlm_l = mlm_l.sum() / (mlm_weights_X.sum() + 1e-8)
# 计算下一句子预测任务的损失
nsp_l = loss(nsp_Y_hat, nsp_y)
l = mlm_l + nsp_l
return mlm_l, nsp_l, l
#@save
def _get_batch_loss_bert(net, loss, vocab_size, tokens_X,
segments_X, valid_lens_x,
pred_positions_X, mlm_weights_X,
mlm_Y, nsp_y):
# 前向传播
_, mlm_Y_hat, nsp_Y_hat = net(tokens_X, segments_X,
valid_lens_x.reshape([-1]),
pred_positions_X)
# 计算遮蔽语言模型损失
mlm_l = loss(mlm_Y_hat.reshape([-1, vocab_size]), mlm_Y.reshape([-1])) *\
mlm_weights_X.reshape([-1, 1])
mlm_l = mlm_l.sum() / (mlm_weights_X.sum() + 1e-8)
# 计算下一句子预测任务的损失
nsp_l = loss(nsp_Y_hat, nsp_y)
l = mlm_l + nsp_l
return mlm_l, nsp_l, l
通过调用上述两个辅助函数，下面的train_bert函数定义了在WikiText-2（train_iter）数据集上预训练BERT（net）的过程。训练BERT可能需要很长时间。以下函数的输入num_steps指定了训练的迭代步数，而不是像train_ch13函数那样指定训练的轮数（参见
13.1节）。
def train_bert(train_iter, net, loss, vocab_size, devices, num_steps):
trainer = gluon.Trainer(net.collect_params(), 'adam',
{'learning_rate': 0.01})
step, timer = 0, d2l.Timer()
animator = d2l.Animator(xlabel='step', ylabel='loss',
xlim=[1, num_steps], legend=['mlm', 'nsp'])
# 遮蔽语言模型损失的和，下一句预测任务损失的和，句子对的数量，计数
metric = d2l.Accumulator(4)
num_steps_reached = False
while step < num_steps and not num_steps_reached:
for batch in train_iter:
(tokens_X_shards, segments_X_shards, valid_lens_x_shards,
pred_positions_X_shards, mlm_weights_X_shards,
mlm_Y_shards, nsp_y_shards) = [gluon.utils.split_and_load(
elem, devices, even_split=False) for elem in batch]
timer.start()
with autograd.record():
mlm_ls, nsp_ls, ls = _get_batch_loss_bert(
net, loss, vocab_size, tokens_X_shards, segments_X_shards,
valid_lens_x_shards, pred_positions_X_shards,
mlm_weights_X_shards, mlm_Y_shards, nsp_y_shards)
for l in ls:
l.backward()
trainer.step(1)
mlm_l_mean = sum([float(l) for l in mlm_ls]) / len(mlm_ls)
nsp_l_mean = sum([float(l) for l in nsp_ls]) / len(nsp_ls)
metric.add(mlm_l_mean, nsp_l_mean, batch[0].shape[0], 1)
timer.stop()
animator.add(step + 1,
(metric[0] / metric[3], metric[1] / metric[3]))
step += 1
if step == num_steps:
num_steps_reached = True
break
print(f'MLM loss {metric[0] / metric[3]:.3f}, '
f'NSP loss {metric[1] / metric[3]:.3f}')
print(f'{metric[2] / timer.sum():.1f} sentence pairs/sec on '
f'{str(devices)}')
def train_bert(train_iter, net, loss, vocab_size, devices, num_steps):
net = nn.DataParallel(net, device_ids=devices).to(devices[0])
trainer = torch.optim.Adam(net.parameters(), lr=0.01)
step, timer = 0, d2l.Timer()
animator = d2l.Animator(xlabel='step', ylabel='loss',
xlim=[1, num_steps], legend=['mlm', 'nsp'])
# 遮蔽语言模型损失的和，下一句预测任务损失的和，句子对的数量，计数
metric = d2l.Accumulator(4)
num_steps_reached = False
while step < num_steps and not num_steps_reached:
for tokens_X, segments_X, valid_lens_x, pred_positions_X,\
mlm_weights_X, mlm_Y, nsp_y in train_iter:
tokens_X = tokens_X.to(devices[0])
segments_X = segments_X.to(devices[0])
valid_lens_x = valid_lens_x.to(devices[0])
pred_positions_X = pred_positions_X.to(devices[0])
mlm_weights_X = mlm_weights_X.to(devices[0])
mlm_Y, nsp_y = mlm_Y.to(devices[0]), nsp_y.to(devices[0])
trainer.zero_grad()
timer.start()
mlm_l, nsp_l, l = _get_batch_loss_bert(
net, loss, vocab_size, tokens_X, segments_X, valid_lens_x,
pred_positions_X, mlm_weights_X, mlm_Y, nsp_y)
l.backward()
trainer.step()
metric.add(mlm_l, nsp_l, tokens_X.shape[0], 1)
timer.stop()
animator.add(step + 1,
(metric[0] / metric[3], metric[1] / metric[3]))
step += 1
if step == num_steps:
num_steps_reached = True
break
print(f'MLM loss {metric[0] / metric[3]:.3f}, '
f'NSP loss {metric[1] / metric[3]:.3f}')
print(f'{metric[2] / timer.sum():.1f} sentence pairs/sec on '
f'{str(devices)}')
def train_bert(train_iter, net, loss, vocab_size, devices, num_steps):
trainer = paddle.optimizer.Adam(parameters=net.parameters(), learning_rate=0.01)
step, timer = 0, d2l.Timer()
animator = d2l.Animator(xlabel='step', ylabel='loss',
xlim=[1, num_steps], legend=['mlm', 'nsp'])
# 遮蔽语言模型损失的和，下一句预测任务损失的和，句子对的数量，计数
metric = d2l.Accumulator(4)
num_steps_reached = False
while step < num_steps and not num_steps_reached:
for tokens_X, segments_X, valid_lens_x, pred_positions_X,\
mlm_weights_X, mlm_Y, nsp_y in train_iter:
trainer.clear_grad()
timer.start()
mlm_l, nsp_l, l = _get_batch_loss_bert(
net, loss, vocab_size, tokens_X, segments_X, valid_lens_x,
pred_positions_X, mlm_weights_X, mlm_Y, nsp_y)
l.backward()
trainer.step()
metric.add(mlm_l, nsp_l, tokens_X.shape[0], 1)
timer.stop()
animator.add(step + 1,
(metric[0] / metric[3], metric[1] / metric[3]))
step += 1
if step == num_steps:
num_steps_reached = True
break
print(f'MLM loss {metric[0] / metric[3]:.3f}, '
f'NSP loss {metric[1] / metric[3]:.3f}')
print(f'{metric[2] / timer.sum():.1f} sentence pairs/sec on '
f'{str(devices)}')
在预训练过程中，我们可以绘制出遮蔽语言模型损失和下一句预测损失。
train_bert(train_iter, net, loss, len(vocab), devices, 50)
MLM loss 7.333, NSP loss 0.827
2279.2 sentence pairs/sec on [gpu(0), gpu(1)]
train_bert(train_iter, net, loss, len(vocab), devices, 50)
MLM loss 5.425, NSP loss 0.775
3485.7 sentence pairs/sec on [device(type='cuda', index=0), device(type='cuda', index=1)]
train_bert(train_iter, net, loss, len(vocab), devices[:1], 50)
MLM loss 5.849, NSP loss 0.822
9518.3 sentence pairs/sec on [Place(gpu:0)]
14.10.2. 用BERT表示文本¶
在预训练BERT之后，我们可以用它来表示单个文本、文本对或其中的任何词元。下面的函数返回tokens_a和tokens_b中所有词元的BERT（net）表示。
def get_bert_encoding(net, tokens_a, tokens_b=None):
tokens, segments = d2l.get_tokens_and_segments(tokens_a, tokens_b)
token_ids = np.expand_dims(np.array(vocab[tokens], ctx=devices[0]),
axis=0)
segments = np.expand_dims(np.array(segments, ctx=devices[0]), axis=0)
valid_len = np.expand_dims(np.array(len(tokens), ctx=devices[0]), axis=0)
encoded_X, _, _ = net(token_ids, segments, valid_len)
return encoded_X
def get_bert_encoding(net, tokens_a, tokens_b=None):
tokens, segments = d2l.get_tokens_and_segments(tokens_a, tokens_b)
token_ids = torch.tensor(vocab[tokens], device=devices[0]).unsqueeze(0)
segments = torch.tensor(segments, device=devices[0]).unsqueeze(0)
valid_len = torch.tensor(len(tokens), device=devices[0]).unsqueeze(0)
encoded_X, _, _ = net(token_ids, segments, valid_len)
return encoded_X
def get_bert_encoding(net, tokens_a, tokens_b=None):
tokens, segments = d2l.get_tokens_and_segments(tokens_a, tokens_b)
token_ids = paddle.to_tensor(vocab[tokens]).unsqueeze(0)
segments = paddle.to_tensor(segments).unsqueeze(0)
valid_len = paddle.to_tensor(len(tokens))
encoded_X, _, _ = net(token_ids, segments, valid_len)
return encoded_X
考虑“a crane is flying”这句话。回想一下
14.8.4节中讨论的BERT的输入表示。插入特殊标记“<cls>”（用于分类）和“<sep>”（用于分隔）后，BERT输入序列的长度为6。因为零是“<cls>”词元，encoded_text[:, 0, :]是整个输入语句的BERT表示。为了评估一词多义词元“crane”，我们还打印出了该词元的BERT表示的前三个元素。
tokens_a = ['a', 'crane', 'is', 'flying']
encoded_text = get_bert_encoding(net, tokens_a)
# 词元：'<cls>','a','crane','is','flying','<sep>'
encoded_text_cls = encoded_text[:, 0, :]
encoded_text_crane = encoded_text[:, 2, :]
encoded_text.shape, encoded_text_cls.shape, encoded_text_crane[0][:3]
((1, 6, 128),
(1, 128),
array([ 0.7835793, 1.1049025, -2.072324 ], ctx=gpu(0)))
tokens_a = ['a', 'crane', 'is', 'flying']
encoded_text = get_bert_encoding(net, tokens_a)
# 词元：'<cls>','a','crane','is','flying','<sep>'
encoded_text_cls = encoded_text[:, 0, :]
encoded_text_crane = encoded_text[:, 2, :]
encoded_text.shape, encoded_text_cls.shape, encoded_text_crane[0][:3]
(torch.Size([1, 6, 128]),
torch.Size([1, 128]),
tensor([-0.5007, -1.0034, 0.8718], device='cuda:0', grad_fn=<SliceBackward0>))
tokens_a = ['a', 'crane', 'is', 'flying']
encoded_text = get_bert_encoding(net, tokens_a)
# 词元：'<cls>','a','crane','is','flying','<sep>'
encoded_text_cls = encoded_text[:, 0, :]
encoded_text_crane = encoded_text[:, 2, :]
encoded_text.shape, encoded_text_cls.shape, encoded_text_crane[0][:3]
([1, 6, 128],
[1, 128],
Tensor(shape=[3], dtype=float32, place=Place(gpu:0), stop_gradient=False,
[ 1.23072958, -0.46575257, -0.91060257]))
现在考虑一个句子“a crane driver came”和“he just
left”。类似地，encoded_pair[:, 0, :]是来自预训练BERT的整个句子对的编码结果。注意，多义词元“crane”的前三个元素与上下文不同时的元素不同。这支持了BERT表示是上下文敏感的。
tokens_a, tokens_b = ['a', 'crane', 'driver', 'came'], ['he', 'just', 'left']
encoded_pair = get_bert_encoding(net, tokens_a, tokens_b)
# 词元：'<cls>','a','crane','driver','came','<sep>','he','just',
# 'left','<sep>'
encoded_pair_cls = encoded_pair[:, 0, :]
encoded_pair_crane = encoded_pair[:, 2, :]
encoded_pair.shape, encoded_pair_cls.shape, encoded_pair_crane[0][:3]
((1, 10, 128),
(1, 128),
array([ 0.7827732, 1.1043007, -2.07267 ], ctx=gpu(0)))
tokens_a, tokens_b = ['a', 'crane', 'driver', 'came'], ['he', 'just', 'left']
encoded_pair = get_bert_encoding(net, tokens_a, tokens_b)
# 词元：'<cls>','a','crane','driver','came','<sep>','he','just',
# 'left','<sep>'
encoded_pair_cls = encoded_pair[:, 0, :]
encoded_pair_crane = encoded_pair[:, 2, :]
encoded_pair.shape, encoded_pair_cls.shape, encoded_pair_crane[0][:3]
(torch.Size([1, 10, 128]),
torch.Size([1, 128]),
tensor([ 0.5101, -0.4041, -1.2749], device='cuda:0', grad_fn=<SliceBackward0>))
tokens_a, tokens_b = ['a', 'crane', 'driver', 'came'], ['he', 'just', 'left']
encoded_pair = get_bert_encoding(net, tokens_a, tokens_b)
# 词元：'<cls>','a','crane','driver','came','<sep>','he','just',
# 'left','<sep>'
encoded_pair_cls = encoded_pair[:, 0, :]
encoded_pair_crane = encoded_pair[:, 2, :]
encoded_pair.shape, encoded_pair_cls.shape, encoded_pair_crane[0][:3]
([1, 10, 128],
[1, 128],
Tensor(shape=[3], dtype=float32, place=Place(gpu:0), stop_gradient=False,
[ 1.19337428, -0.45544022, -0.01078355]))
在 15节中，我们将为下游自然语言处理应用微调预训练的BERT模型。
14.10.3. 小结¶
原始的BERT有两个版本，其中基本模型有1.1亿个参数，大模型有3.4亿个参数。
在预训练BERT之后，我们可以用它来表示单个文本、文本对或其中的任何词元。
在实验中，同一个词元在不同的上下文中具有不同的BERT表示。这支持BERT表示是上下文敏感的。
14.10.4. 练习¶
在实验中，我们可以看到遮蔽语言模型损失明显高于下一句预测损失。为什么？
将BERT输入序列的最大长度设置为512（与原始BERT模型相同）。使用原始BERT模型的配置，如\(\text{BERT}_{\text{LARGE}}\)。运行此部分时是否遇到错误？为什么？