.. _sec_bert-pretraining:
BERTの事前学習
==============
:numref:`sec_bert` で実装した BERT モデルと、
:numref:`sec_bert-dataset` で生成した WikiText-2
データセット由来の事前学習用サンプルを用いて、この節では WikiText-2
データセット上で BERT を事前学習する。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
from d2l import torch as d2l
import torch
from torch import nn
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
from d2l import mxnet as d2l
from mxnet import autograd, gluon, init, np, npx
npx.set_np()
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
batch_size, max_len = 512, 64
train_iter, vocab = d2l.load_data_wiki(batch_size, max_len)
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
batch_size, max_len = 512, 64
train_iter, vocab = d2l.load_data_wiki(batch_size, max_len)
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
batch_size, max_len = 512, 64
train_iter, vocab = d2l.load_data_wiki(batch_size, max_len)
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
batch_size, max_len = 512, 64
train_iter, vocab = d2l.load_data_wiki(batch_size, max_len)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
Downloading ../data/wikitext-2-v1.zip from https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-v1.zip...
[06:58:35] ../src/storage/storage.cc:196: Using Pooled (Naive) StorageManager for CPU
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_bert(train_iter, net, loss, len(vocab), devices, 50)
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_bert(train_iter, net, loss, len(vocab), devices, 50)
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
net = d2l.BERTModel(len(vocab), num_hiddens=128,
ffn_num_hiddens=256, num_heads=2, num_blks=2, dropout=0.2)
devices = d2l.try_all_gpus()
loss = nn.CrossEntropyLoss()
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
net = d2l.BERTModel(len(vocab), num_hiddens=128, ffn_num_hiddens=256,
num_heads=2, num_blks=2, dropout=0.2)
devices = d2l.try_all_gpus()
net.initialize(init.Xavier(), ctx=devices)
loss = gluon.loss.SoftmaxCELoss()
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
[06:59:57] ../src/storage/storage.cc:196: Using Pooled (Naive) StorageManager for GPU
[06:59:57] ../src/storage/storage.cc:196: Using Pooled (Naive) StorageManager for GPU
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
tokens_a = ['a', 'crane', 'is', 'flying']
encoded_text = get_bert_encoding(net, tokens_a)
# Tokens: '', '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:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
tokens_a = ['a', 'crane', 'is', 'flying']
encoded_text = get_bert_encoding(net, tokens_a)
# Tokens: '', '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:: 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):
# Forward pass
_, mlm_Y_hat, nsp_Y_hat = net(tokens_X, segments_X,
valid_lens_x.reshape(-1),
pred_positions_X)
# Compute masked language model loss
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)
# Compute next sentence prediction loss
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_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):
# Forward pass
_, mlm_Y_hat, nsp_Y_hat = net(
tokens_X_shard, segments_X_shard, valid_lens_x_shard.reshape(-1),
pred_positions_X_shard)
# Compute masked language model loss
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)
# Compute next sentence prediction loss
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
.. 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)
# Tokens: '', '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:: 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)
# Tokens: '', '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:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
def train_bert(train_iter, net, loss, vocab_size, devices, num_steps):
net(*next(iter(train_iter))[:4])
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'])
# Sum of masked language modeling losses, sum of next sentence prediction
# losses, no. of sentence pairs, count
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:: latex
\diilbookstyleinputcell
.. code:: python
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'])
# Sum of masked language modeling losses, sum of next sentence prediction
# losses, no. of sentence pairs, count
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)}')
.. 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 = 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:: 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 = 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
.. raw:: html
.. raw:: html