Keras BERT

[中文|English]

Implementation of the BERT. Official pre-trained models could be loaded for feature extraction and prediction.

Install

pip install keras-bert

Usage

• Load Official Pre-trained Models
• Tokenizer
• Train & Use
• Use Warmup
• Download Pretrained Checkpoints
• Extract Features
• Use Adapter

External Links

• Kashgari is a Production-ready NLP Transfer learning framework for text-labeling and text-classification

Load Official Pre-trained Models

In feature extraction demo, you should be able to get the same extraction results as the official model chinese_L-12_H-768_A-12. And in prediction demo, the missing word in the sentence could be predicted.

Run on TPU

The extraction demo shows how to convert to a model that runs on TPU.

The classification demo shows how to apply the model to simple classification tasks.

Tokenizer

The Tokenizer class is used for splitting texts and generating indices:

from keras_bert import Tokenizer

token_dict = {
    '[CLS]': 0,
    '[SEP]': 1,
    'un': 2,
    '##aff': 3,
    '##able': 4,
    '[UNK]': 5,
}
tokenizer = Tokenizer(token_dict)
print(tokenizer.tokenize('unaffable'))  # The result should be `['[CLS]', 'un', '##aff', '##able', '[SEP]']`
indices, segments = tokenizer.encode('unaffable')
print(indices)  # Should be `[0, 2, 3, 4, 1]`
print(segments)  # Should be `[0, 0, 0, 0, 0]`

print(tokenizer.tokenize(first='unaffable', second='钢'))
# The result should be `['[CLS]', 'un', '##aff', '##able', '[SEP]', '钢', '[SEP]']`
indices, segments = tokenizer.encode(first='unaffable', second='钢', max_len=10)
print(indices)  # Should be `[0, 2, 3, 4, 1, 5, 1, 0, 0, 0]`
print(segments)  # Should be `[0, 0, 0, 0, 0, 1, 1, 0, 0, 0]`

Train & Use

import keras
from keras_bert import get_base_dict, get_model, compile_model, gen_batch_inputs


# A toy input example
sentence_pairs = [
    [['all', 'work', 'and', 'no', 'play'], ['makes', 'jack', 'a', 'dull', 'boy']],
    [['from', 'the', 'day', 'forth'], ['my', 'arm', 'changed']],
    [['and', 'a', 'voice', 'echoed'], ['power', 'give', 'me', 'more', 'power']],
]


# Build token dictionary
token_dict = get_base_dict()  # A dict that contains some special tokens
for pairs in sentence_pairs:
    for token in pairs[0] + pairs[1]:
        if token not in token_dict:
            token_dict[token] = len(token_dict)
token_list = list(token_dict.keys())  # Used for selecting a random word


# Build & train the model
model = get_model(
    token_num=len(token_dict),
    head_num=5,
    transformer_num=12,
    embed_dim=25,
    feed_forward_dim=100,
    seq_len=20,
    pos_num=20,
    dropout_rate=0.05,
)
compile_model(model)
model.summary()

def _generator():
    while True:
        yield gen_batch_inputs(
            sentence_pairs,
            token_dict,
            token_list,
            seq_len=20,
            mask_rate=0.3,
            swap_sentence_rate=1.0,
        )

model.fit_generator(
    generator=_generator(),
    steps_per_epoch=1000,
    epochs=100,
    validation_data=_generator(),
    validation_steps=100,
    callbacks=[
        keras.callbacks.EarlyStopping(monitor='val_loss', patience=5)
    ],
)


# Use the trained model
inputs, output_layer = get_model(
    token_num=len(token_dict),
    head_num=5,
    transformer_num=12,
    embed_dim=25,
    feed_forward_dim=100,
    seq_len=20,
    pos_num=20,
    dropout_rate=0.05,
    training=False,      # The input layers and output layer will be returned if `training` is `False`
    trainable=False,     # Whether the model is trainable. The default value is the same with `training`
    output_layer_num=4,  # The number of layers whose outputs will be concatenated as a single output.
                         # Only available when `training` is `False`.
)

Use Warmup

AdamWarmup optimizer is provided for warmup and decay. The learning rate will reach lr in warmpup_steps steps, and decay to min_lr in decay_steps steps. There is a helper function calc_train_steps for calculating the two steps:

import numpy as np
from keras_bert import AdamWarmup, calc_train_steps

train_x = np.random.standard_normal((1024, 100))

total_steps, warmup_steps = calc_train_steps(
    num_example=train_x.shape[0],
    batch_size=32,
    epochs=10,
    warmup_proportion=0.1,
)

optimizer = AdamWarmup(total_steps, warmup_steps, lr=1e-3, min_lr=1e-5)

Download Pretrained Checkpoints

Several download urls has been added. You can get the downloaded and uncompressed path of a checkpoint by:

from keras_bert import get_pretrained, PretrainedList, get_checkpoint_paths

model_path = get_pretrained(PretrainedList.multi_cased_base)
paths = get_checkpoint_paths(model_path)
print(paths.config, paths.checkpoint, paths.vocab)

Extract Features

You can use helper function extract_embeddings if the features of tokens or sentences (without further tuning) are what you need. To extract the features of all tokens:

from keras_bert import extract_embeddings

model_path = 'xxx/yyy/uncased_L-12_H-768_A-12'
texts = ['all work and no play', 'makes jack a dull boy~']

embeddings = extract_embeddings(model_path, texts)

The returned result is a list with the same length as texts. Each item in the list is a numpy array truncated by the length of the input. The shapes of outputs in this example are (7, 768) and (8, 768).

When the inputs are paired-sentences, and you need the outputs of NSP and max-pooling of the last 4 layers:

from keras_bert import extract_embeddings, POOL_NSP, POOL_MAX

model_path = 'xxx/yyy/uncased_L-12_H-768_A-12'
texts = [
    ('all work and no play', 'makes jack a dull boy'),
    ('makes jack a dull boy', 'all work and no play'),
]

embeddings = extract_embeddings(model_path, texts, output_layer_num=4, poolings=[POOL_NSP, POOL_MAX])

There are no token features in the results. The outputs of NSP and max-pooling will be concatenated with the final shape (768 x 4 x 2,).

The second argument in the helper function is a generator. To extract features from file:

import codecs
from keras_bert import extract_embeddings

model_path = 'xxx/yyy/uncased_L-12_H-768_A-12'

with codecs.open('xxx.txt', 'r', 'utf8') as reader:
    texts = map(lambda x: x.strip(), reader)
    embeddings = extract_embeddings(model_path, texts)

Use Adapter

You can use adapters for fine-tuning:

import os
from keras_bert import load_trained_model_from_checkpoint

layer_num = 12
checkpoint_path = '.../uncased_L-12_H-768_A-12'

config_path = os.path.join(checkpoint_path, 'bert_config.json')
model_path = os.path.join(checkpoint_path, 'bert_model.ckpt')
model = load_trained_model_from_checkpoint(
    config_path,
    model_path,
    training=False,
    use_adapter=True,
    trainable=['Encoder-{}-MultiHeadSelfAttention-Adapter'.format(i + 1) for i in range(layer_num)] +
    ['Encoder-{}-FeedForward-Adapter'.format(i + 1) for i in range(layer_num)] +
    ['Encoder-{}-MultiHeadSelfAttention-Norm'.format(i + 1) for i in range(layer_num)] +
    ['Encoder-{}-FeedForward-Norm'.format(i + 1) for i in range(layer_num)],
)

Use tensorflow.python.keras

Add TF_KERAS=1 to environment variables to use tensorflow.python.keras.

Use theano Backend

Add KERAS_BACKEND=theano to environment variables to enable theano backend.