mnli.py

import os

import numpy as np
import pandas as pd

from keras_xlnet.backend import keras
from keras_bert.layers import Extract
from keras_xlnet import PretrainedList, get_pretrained_paths
from keras_xlnet import Tokenizer, load_trained_model_from_checkpoint, ATTENTION_TYPE_BI

EPOCH = 5
BATCH_SIZE = 16
SEQ_LEN = 100
MODEL_NAME = 'MNLI.h5'

CLASSES = {
    'neutral': 0,
    'entailment': 1,
    'contradiction': 2,
}

current_path = os.path.dirname(os.path.abspath(__file__))
train_path = os.path.join(current_path, 'train.tsv')
dev_matched_path = os.path.join(current_path, 'dev_matched.tsv')
dev_mismatched_path = os.path.join(current_path, 'dev_mismatched.tsv')

paths = get_pretrained_paths(PretrainedList.en_cased_base)
tokenizer = Tokenizer(paths.vocab)


# Read data
class DataSequence(keras.utils.Sequence):

    def __init__(self, x, y):
        self.x = x
        self.y = y

    def __len__(self):
        return (len(self.y) + BATCH_SIZE - 1) // BATCH_SIZE

    def __getitem__(self, index):
        s = slice(index * BATCH_SIZE, (index + 1) * BATCH_SIZE)
        return [item[s] for item in self.x], self.y[s]


def generate_sequence(path):
    tokens, classes = [], []
    df = pd.read_csv(path, sep='\t', error_bad_lines=False)
    for _, row in df.iterrows():
        text_a, text_b, cls = row['sentence1'], row['sentence2'], row['gold_label']
        if not isinstance(text_a, str) or not isinstance(text_b, str) or cls not in CLASSES:
            continue
        encoded_a, encoded_b = tokenizer.encode(text_a)[:48], tokenizer.encode(text_b)[:49]
        encoded = encoded_a + [tokenizer.SYM_SEP] + encoded_b + [tokenizer.SYM_SEP]
        encoded = [tokenizer.SYM_PAD] * (SEQ_LEN - 1 - len(encoded)) + encoded + [tokenizer.SYM_CLS]
        tokens.append(encoded)
        classes.append(CLASSES[cls])
    tokens, classes = np.array(tokens), np.array(classes)
    segments = np.zeros_like(tokens)
    segments[:, -1] = 1
    lengths = np.zeros_like(tokens[:, :1])
    return DataSequence([tokens, segments, lengths], classes)


current_path = os.path.dirname(os.path.abspath(__file__))
train_seq = generate_sequence(train_path)
dev_matched_seq = generate_sequence(dev_matched_path)
dev_mismatched_seq = generate_sequence(dev_mismatched_path)


# Load pretrained model
model = load_trained_model_from_checkpoint(
    config_path=paths.config,
    checkpoint_path=paths.model,
    batch_size=BATCH_SIZE,
    memory_len=0,
    target_len=SEQ_LEN,
    in_train_phase=False,
    attention_type=ATTENTION_TYPE_BI,
)


# Build classification model
last = Extract(index=-1, name='Extract')(model.output)
dense = keras.layers.Dense(units=768, activation='tanh', name='Dense')(last)
dropout = keras.layers.Dropout(rate=0.1, name='Dropout')(dense)
output = keras.layers.Dense(units=3, activation='softmax', name='Softmax')(dropout)
model = keras.models.Model(inputs=model.inputs, outputs=output)
model.summary()


# Fit model
if os.path.exists(MODEL_NAME):
    model.load_weights(MODEL_NAME)

model.compile(
    optimizer=keras.optimizers.Adam(lr=3e-5),
    loss='sparse_categorical_crossentropy',
    metrics=['sparse_categorical_accuracy'],
)

model.fit_generator(
    generator=train_seq,
    validation_data=dev_matched_seq,
    epochs=EPOCH,
    callbacks=[keras.callbacks.EarlyStopping(monitor='val_loss', patience=2)],
)

model.save_weights(MODEL_NAME)

# Evaluation
# Use dev set because the results of test set is unknown
for dev_seq in [dev_matched_seq, dev_mismatched_seq]:
    results = model.predict_generator(dev_seq, verbose=True).argmax(axis=-1)
    print('Accuracy: %.2f' % (100.0 * np.sum(results == dev_seq.y) / len(results)))