14.9. The Dataset for Pretraining BERT
Open the notebook in Colab
Open the notebook in Colab
Open the notebook in Colab

To pretrain the BERT model as implemented in Section 14.8, we need to generate the dataset in the ideal format to facilitate the two pretraining tasks: masked language modeling and next sentence prediction. On one hand, the original BERT model is pretrained on the concatenation of two huge corpora BookCorpus and English Wikipedia (see Section 14.8.5), making it hard to run for most readers of this book. On the other hand, the off-the-shelf pretrained BERT model may not fit for applications from specific domains like medicine. Thus, it is getting popular to pretrain BERT on a customized dataset. To facilitate the demonstration of BERT pretraining, we use a smaller corpus WikiText-2 [Merity et al., 2016].

Comparing with the PTB dataset used for pretraining word2vec in Section 14.3, WikiText-2 i) retains the original punctuation, making it suitable for next sentence prediction; ii) retains the original case and numbers; iii) is over twice larger.

import collections
from d2l import mxnet as d2l
import mxnet as mx
from mxnet import autograd, gluon, init, np, npx
import os
import random
import time
import zipfile


In the WikiText-2 dataset, each line represents a paragraph where space is inserted between any punctuation and its preceding token. Paragraphs with at least two sentences are retained. To split sentences, we only use the period as the delimiter for simplicity. We leave discussions of more complex sentence splitting techniques in the exercises at the end of this section.

d2l.DATA_HUB['wikitext-2'] = (
    'wikitext-2-v1.zip', '3c914d17d80b1459be871a5039ac23e752a53cbe')

def _read_wiki(data_dir):
    file_name = os.path.join(data_dir, 'wiki.train.tokens')
    with open(file_name, 'r') as f:
        lines = f.readlines()
    # Uppercase letters are converted to lowercase ones
    paragraphs = [line.strip().lower().split(' . ')
                  for line in lines if len(line.split(' . ')) >= 2]
    return paragraphs

14.9.1. Defining Helper Functions for Pretraining Tasks

In the following, we begin by implementing helper functions for the two BERT pretraining tasks: next sentence prediction and masked language modeling. These helper functions will be invoked later when transforming the raw text corpus into the dataset of the ideal format to pretrain BERT. Generating the Next Sentence Prediction Task

According to descriptions of Section, the _get_next_sentence function generates a training example for the binary classification task.

def _get_next_sentence(sentence, next_sentence, paragraphs):
    if random.random() < 0.5:
        is_next = True
        # `paragraphs` is a list of lists of lists
        next_sentence = random.choice(random.choice(paragraphs))
        is_next = False
    return sentence, next_sentence, is_next

The following function generates training examples for next sentence prediction from the input paragraph by invoking the _get_next_sentence function. Here paragraph is a list of sentences, where each sentence is a list of tokens. The argument max_len specifies the maximum length of a BERT input sequence during pretraining.

def _get_nsp_data_from_paragraph(paragraph, paragraphs, vocab, max_len):
    nsp_data_from_paragraph = []
    for i in range(len(paragraph) - 1):
        tokens_a, tokens_b, is_next = _get_next_sentence(
            paragraph[i], paragraph[i + 1], paragraphs)
        # Consider 1 '<cls>' token and 2 '<sep>' tokens
        if len(tokens_a) + len(tokens_b) + 3 > max_len:
        tokens, segments = d2l.get_tokens_and_segments(tokens_a, tokens_b)
        nsp_data_from_paragraph.append((tokens, segments, is_next))
    return nsp_data_from_paragraph Generating the Masked Language Modeling Task

In order to generate training examples for the masked language modeling task from a BERT input sequence, we define the following _replace_mlm_tokens function. In its inputs, tokens is a list of tokens representing a BERT input sequence, candidate_pred_positions is a list of token indices of the BERT input sequence excluding those of special tokens (special tokens are not predicted in the masked language modeling task), and num_mlm_preds indicates the number of predictions (recall 15% random tokens to predict). Following the definition of the masked language modeling task in Section, at each prediction position, the input may be replaced by a special “<mask>” token or a random token, or remain unchanged. In the end, the function returns the input tokens after possible replacement, the token indices where predictions take place and labels for these predictions.

def _replace_mlm_tokens(tokens, candidate_pred_positions, num_mlm_preds,
    # Make a new copy of tokens for the input of a masked language model,
    # where the input may contain replaced '<mask>' or random tokens
    mlm_input_tokens = [token for token in tokens]
    pred_positions_and_labels = []
    # Shuffle for getting 15% random tokens for prediction in the masked
    # language modeling task
    for mlm_pred_position in candidate_pred_positions:
        if len(pred_positions_and_labels) >= num_mlm_preds:
        masked_token = None
        # 80% of the time: replace the word with the '<mask>' token
        if random.random() < 0.8:
            masked_token = '<mask>'
            # 10% of the time: keep the word unchanged
            if random.random() < 0.5:
                masked_token = tokens[mlm_pred_position]
            # 10% of the time: replace the word with a random word
                masked_token = random.randint(0, len(vocab) - 1)
        mlm_input_tokens[mlm_pred_position] = masked_token
            (mlm_pred_position, tokens[mlm_pred_position]))
    return mlm_input_tokens, pred_positions_and_labels

By invoking the aforementioned _replace_mlm_tokens function, the following function takes a BERT input sequence (tokens) as an input and returns indices of the input tokens (after possible token replacement as described in Section, the token indices where predictions take place, and label indices for these predictions.

def _get_mlm_data_from_tokens(tokens, vocab):
    candidate_pred_positions = []
    # `tokens` is a list of strings
    for i, token in enumerate(tokens):
        # Special tokens are not predicted in the masked language modeling
        # task
        if token in ['<cls>', '<sep>']:
    # 15% of random tokens are predicted in the masked language modeling task
    num_mlm_preds = max(1, round(len(tokens) * 0.15))
    mlm_input_tokens, pred_positions_and_labels = _replace_mlm_tokens(
        tokens, candidate_pred_positions, num_mlm_preds, vocab)
    pred_positions_and_labels = sorted(pred_positions_and_labels,
                                       key=lambda x: x[0])
    pred_positions = [v[0] for v in pred_positions_and_labels]
    mlm_pred_labels = [v[1] for v in pred_positions_and_labels]
    return vocab[mlm_input_tokens], pred_positions, vocab[mlm_pred_labels]

14.9.2. Transforming Text into the Pretraining Dataset

Now we are almost ready to customize a Dataset class for pretraining BERT. Before that, we still need to define a helper function _pad_bert_inputs to append the special “<mask>” tokens to the inputs. Its argument examples contain the outputs from the helper functions _get_nsp_data_from_paragraph and _get_mlm_data_from_tokens for the two pretraining tasks.

def _pad_bert_inputs(examples, max_len, vocab):
    max_num_mlm_preds = round(max_len * 0.15)
    all_token_ids, all_segments, valid_lens,  = [], [], []
    all_pred_positions, all_mlm_weights, all_mlm_labels = [], [], []
    nsp_labels = []
    for (token_ids, pred_positions, mlm_pred_label_ids, segments,
         is_next) in examples:
        all_token_ids.append(np.array(token_ids + [vocab['<pad>']] * (
            max_len - len(token_ids)), dtype='int32'))
        all_segments.append(np.array(segments + [0] * (
            max_len - len(segments)), dtype='int32'))
        # `valid_lens` excludes count of '<pad>' tokens
        valid_lens.append(np.array(len(token_ids), dtype='float32'))
        all_pred_positions.append(np.array(pred_positions + [0] * (
            max_num_mlm_preds - len(pred_positions)), dtype='int32'))
        # Predictions of padded tokens will be filtered out in the loss via
        # multiplication of 0 weights
            np.array([1.0] * len(mlm_pred_label_ids) + [0.0] * (
                max_num_mlm_preds - len(pred_positions)), dtype='float32'))
        all_mlm_labels.append(np.array(mlm_pred_label_ids + [0] * (
            max_num_mlm_preds - len(mlm_pred_label_ids)), dtype='int32'))
    return (all_token_ids, all_segments, valid_lens, all_pred_positions,
            all_mlm_weights, all_mlm_labels, nsp_labels)

Putting the helper functions for generating training examples of the two pretraining tasks, and the helper function for padding inputs together, we customize the following _WikiTextDataset class as the WikiText-2 dataset for pretraining BERT. By implementing the __getitem__function, we can arbitrarily access the pretraining (masked language modeling and next sentence prediction) examples generated from a pair of sentences from the WikiText-2 corpus.

The original BERT model uses WordPiece embeddings whose vocabulary size is 30,000 [Wu et al., 2016]. The tokenization method of WordPiece is a slight modification of the original byte pair encoding algorithm in Section 14.6.2. For simplicity, we use the d2l.tokenize function for tokenization. Infrequent tokens that appear less than five times are filtered out.

class _WikiTextDataset(gluon.data.Dataset):
    def __init__(self, paragraphs, max_len):
        # Input `paragraphs[i]` is a list of sentence strings representing a
        # paragraph; while output `paragraphs[i]` is a list of sentences
        # representing a paragraph, where each sentence is a list of tokens
        paragraphs = [d2l.tokenize(
            paragraph, token='word') for paragraph in paragraphs]
        sentences = [sentence for paragraph in paragraphs
                     for sentence in paragraph]
        self.vocab = d2l.Vocab(sentences, min_freq=5, reserved_tokens=[
            '<pad>', '<mask>', '<cls>', '<sep>'])
        # Get data for the next sentence prediction task
        examples = []
        for paragraph in paragraphs:
                paragraph, paragraphs, self.vocab, max_len))
        # Get data for the masked language model task
        examples = [(_get_mlm_data_from_tokens(tokens, self.vocab)
                      + (segments, is_next))
                     for tokens, segments, is_next in examples]
        # Pad inputs
        (self.all_token_ids, self.all_segments, self.valid_lens,
         self.all_pred_positions, self.all_mlm_weights,
         self.all_mlm_labels, self.nsp_labels) = _pad_bert_inputs(
            examples, max_len, self.vocab)

    def __getitem__(self, idx):
        return (self.all_token_ids[idx], self.all_segments[idx],
                self.valid_lens[idx], self.all_pred_positions[idx],
                self.all_mlm_weights[idx], self.all_mlm_labels[idx],

    def __len__(self):
        return len(self.all_token_ids)

By using the _read_wiki function and the _WikiTextDataset class, we define the following load_data_wiki to download and WikiText-2 dataset and generate pretraining examples from it.

def load_data_wiki(batch_size, max_len):
    num_workers = d2l.get_dataloader_workers()
    data_dir = d2l.download_extract('wikitext-2', 'wikitext-2')
    paragraphs = _read_wiki(data_dir)
    train_set = _WikiTextDataset(paragraphs, max_len)
    train_iter = gluon.data.DataLoader(train_set, batch_size, shuffle=True,
    return train_iter, train_set.vocab

Setting the batch size to 512 and the maximum length of a BERT input sequence to be 64, we print out the shapes of a minibatch of BERT pretraining examples. Note that in each BERT input sequence, \(10\) (\(64 \times 0.15\)) positions are predicted for the masked language modeling task.

batch_size, max_len = 512, 64
train_iter, vocab = load_data_wiki(batch_size, max_len)

for (tokens_X, segments_X, valid_lens_x, pred_positions_X, mlm_weights_X,
     mlm_Y, nsp_y) in train_iter:
    print(tokens_X.shape, segments_X.shape, valid_lens_x.shape,
          pred_positions_X.shape, mlm_weights_X.shape, mlm_Y.shape,
(512, 64) (512, 64) (512,) (512, 10) (512, 10) (512, 10) (512,)

In the end, let us take a look at the vocabulary size. Even after filtering out infrequent tokens, it is still over twice larger than that of the PTB dataset.


14.9.3. Summary

  • Comparing with the PTB dataset, the WikiText-2 dateset retains the original punctuation, case and numbers, and is over twice larger.

  • We can arbitrarily access the pretraining (masked language modeling and next sentence prediction) examples generated from a pair of sentences from the WikiText-2 corpus.

14.9.4. Exercises

  1. For simplicity, the period is used as the only delimiter for splitting sentences. Try other sentence splitting techniques, such as the spaCy and NLTK. Take NLTK as an example. You need to install NLTK first: pip install nltk. In the code, first import nltk. Then, download the Punkt sentence tokenizer: nltk.download('punkt'). To split sentences such as sentences = 'This is great ! Why not ?', invoking nltk.tokenize.sent_tokenize(sentences) will return a list of two sentence strings: ['This is great !', 'Why not ?'].

  2. What is the vocabulary size if we do not filter out any infrequent token?