def get_model(learning_rate=config.LEARNING_RATE, name='model_v1'):
"""
create model.
:return: model
"""
num_class = 4
embedding_matrix = load_embedding_matrix()
inputs = tf.keras.layers.Input(shape=(config.MAX_SEQUENCE_LENGTH,), name='input')
embedding = tf.keras.layers.Embedding(embedding_matrix.shape[0], embedding_matrix.shape[1],
embeddings_initializer=Constant(embedding_matrix),
input_length=config.MAX_SEQUENCE_LENGTH,
trainable=False)(inputs)
share_hidden = tf.keras.layers.GRU(64, activation='relu', return_sequences=True, reset_after=True)(embedding)
share_hidden = tf.keras.layers.GRU(32, activation='relu', reset_after=True)(share_hidden)
outputs = []
for col in constants.COLS:
# outputs.append(fc_nn(share_hidden, name=col))
outputs.append(tf.keras.layers.Dense(num_class, activation='softmax', name=col)(share_hidden))
model = tf.keras.Model(inputs=inputs, outputs=outputs, name=name)
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
loss='sparse_categorical_crossentropy',
metrics=[tf.keras.metrics.SparseCategoricalAccuracy(name='acc'),
tfa.metrics.F1Score(num_class, average='micro')],)
return model
def train(epochs=config.EPOCHS, learning_rate=config.LEARNING_RATE):
# service waiters attitude classification.
_logger.info('load data.')
x_train, y_train, x_val, y_val = train_data()
_logger.info('load embedding matrix')
embedding_matrix = load_embedding_matrix()
_logger.info('get and compile model')
model = get_model(embedding_matrix)
model.compile(
optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
loss='sparse_categorical_crossentropy',
metrics=['acc'])
log_dir = os.path.join(
config.LOG_DIR,
'fit/{}/{}'.format(model.name,
datetime.datetime.now().strftime("%Y%m%d-%H%M%S")))
tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir,
histogram_freq=1,
update_freq='batch')
checkpoint_path = os.path.join(config.MODEL_CHECKPOINT_PATH,
'{}'.format(model.name))
cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_path,
save_weights_only=True,
verbose=1)
_logger.info('training model')
history = model.fit(x_train,
y_train,
batch_size=config.BATCH_SIZE,
epochs=epochs,
verbose=1,
validation_data=(x_val, y_val),
callbacks=[tensorboard_callback, cp_callback],
workers=config.WORKER_NUM)
_logger.info('save model')
model_path = os.path.join(config.MODEL_PATH, model.name)
# model.save(model_path)
model.save_weights(model_path)
_logger.info('done')
return history
def init_embedding_matrix():
# wv_model = KeyedVectors.load(tencent_pretrained_word_embedding)
wv_model = Word2Vec.load(corpus_word_embdding)
tokenizer = load_tokenizer()
load_embedding_matrix(tokenizer.word_index, wv_model)
# encoding: utf8
import env
from gensim.models import Word2Vec
from gensim.models import KeyedVectors
from senti_analysis import config
from senti_analysis.preprocess import load_embedding_matrix, load_tokenizer
tencent_pretrained_word_embedding = '/Users/hotbaby/Datasets/Tencent_AILab_ChineseEmbedding.txt'
corpus_word_embedding = '/Users/hotbaby/code/github/sentiment-analysis/notebooks/w2v.model'
if __name__ == '__main__':
wv_model = KeyedVectors.load_word2vec_format(
tencent_pretrained_word_embedding, binary=False)
# wv_model = Word2Vec.load(corpus_word_embedding)
wv_model.save(config.W2V_MODEL_PATH)
tokenizer = load_tokenizer()
load_embedding_matrix(tokenizer.word_index, wv_model)