.. _sec_bert-pretraining:
预训练BERT
==========
利用 :numref:`sec_bert`\ 中实现的BERT模型和
:numref:`sec_bert-dataset`\ 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。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
import mindspore
import mindspore.nn as nn
from d2l import mindspore as d2l
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
import torch
from torch import nn
from d2l import torch as d2l
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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)
loss = nn.CrossEntropyLoss()
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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()
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
#@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_Y_hat = mlm_Y_hat.astype("")
mlm_Y = mlm_Y.astype("int32")
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_y = nsp_y.astype("int32")
nsp_l = loss(nsp_Y_hat, nsp_y)
l = mlm_l + nsp_l
return mlm_l, nsp_l, l
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
#@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
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
def train_bert(train_dataset, net, loss, vocab_size, num_steps):
trainer = nn.Adam(net.trainable_params(), 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
def forward_fn(vocab_size, tokens_X, segments_X, valid_lens_x,
pred_positions_X, mlm_weights_X, mlm_Y, nsp_y):
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)
return l, mlm_l, nsp_l
grad_fn = mindspore.value_and_grad(forward_fn, None, weights=net.trainable_params(), has_aux=True)
while step < num_steps and not num_steps_reached:
net.set_train()
for tokens_X, segments_X, valid_lens_x, pred_positions_X,\
mlm_weights_X, mlm_Y, nsp_y in train_dataset.create_tuple_iterator():
(l, mlm_l, nsp_l), grads = grad_fn(
vocab_size, tokens_X, segments_X, valid_lens_x,
pred_positions_X, mlm_weights_X, mlm_Y, nsp_y)
trainer(grads)
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 ')
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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)}')
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_bert(train_iter, net, loss, len(vocab), 50)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
MLM loss 5.784, NSP loss 0.773
145.6 sentence pairs/sec on
.. figure:: output_bert-pretraining_41429c_41_1.svg
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_bert(train_iter, net, loss, len(vocab), devices, 50)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
MLM loss 5.546, NSP loss 0.775
33157.7 sentence pairs/sec on [device(type='cuda', index=0)]
.. figure:: output_bert-pretraining_41429c_44_1.svg
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
def get_bert_encoding(net, tokens_a, tokens_b=None):
tokens, segments = d2l.get_tokens_and_segments(tokens_a, tokens_b)
token_ids = mindspore.Tensor(vocab[tokens]).unsqueeze(0)
segments = mindspore.Tensor(segments).unsqueeze(0)
valid_len = mindspore.Tensor(len(tokens)).unsqueeze(0)
encoded_X, _, _ = net(token_ids, segments, valid_len)
return encoded_X
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
tokens_a = ['a', 'crane', 'is', 'flying']
encoded_text = get_bert_encoding(net, tokens_a)
# 词元：'','a','crane','is','flying',''
encoded_text_cls = encoded_text[:, 0, :]
encoded_text_crane = encoded_text[:, 2, :]
encoded_text.shape, encoded_text_cls.shape, encoded_text_crane[0][:3]
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
((1, 6, 128),
(1, 128),
Tensor(shape=[3], dtype=Float32, value= [ 1.01091897e+00, -1.08268477e-01, -1.73249785e-02]))
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
tokens_a = ['a', 'crane', 'is', 'flying']
encoded_text = get_bert_encoding(net, tokens_a)
# 词元：'','a','crane','is','flying',''
encoded_text_cls = encoded_text[:, 0, :]
encoded_text_crane = encoded_text[:, 2, :]
encoded_text.shape, encoded_text_cls.shape, encoded_text_crane[0][:3]
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
(torch.Size([1, 6, 128]),
torch.Size([1, 128]),
tensor([-2.4664, 0.2748, 0.8431], device='cuda:0', grad_fn=))
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
tokens_a, tokens_b = ['a', 'crane', 'driver', 'came'], ['he', 'just', 'left']
encoded_pair = get_bert_encoding(net, tokens_a, tokens_b)
# 词元：'','a','crane','driver','came','','he','just',
# 'left',''
encoded_pair_cls = encoded_pair[:, 0, :]
encoded_pair_crane = encoded_pair[:, 2, :]
encoded_pair.shape, encoded_pair_cls.shape, encoded_pair_crane[0][:3]
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
((1, 10, 128),
(1, 128),
Tensor(shape=[3], dtype=Float32, value= [ 1.16757222e-01, -1.34207364e-02, 1.65760005e+00]))
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
tokens_a, tokens_b = ['a', 'crane', 'driver', 'came'], ['he', 'just', 'left']
encoded_pair = get_bert_encoding(net, tokens_a, tokens_b)
# 词元：'','a','crane','driver','came','','he','just',
# 'left',''
encoded_pair_cls = encoded_pair[:, 0, :]
encoded_pair_crane = encoded_pair[:, 2, :]
encoded_pair.shape, encoded_pair_cls.shape, encoded_pair_crane[0][:3]
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
(torch.Size([1, 10, 128]),
torch.Size([1, 128]),
tensor([-2.3270, 0.2519, 0.9314], device='cuda:0', grad_fn=))
.. raw:: html
.. raw:: html
.. raw:: html
`讨论 `__
.. raw:: html
.. raw:: html
`讨论 `__
.. raw:: html
.. raw:: html