def main(_):
word2idx = pickle_load(FLAGS.word2idx_path)
word2vec = pickle_load(FLAGS.word2vec_path)
embedding_size = list(word2vec.values())[0].shape[0]
word_embeddings = np.zeros([len(word2idx) + 1,
embedding_size]) # index 0 for UNK token
for word in word2idx.keys():
word_embeddings[word2idx[word]] = word2vec[word]
FLAGS.n_words = word_embeddings.shape[0]
FLAGS.embedding_size = word_embeddings.shape[1]
train_data = read_data(FLAGS.train_data_path, word2idx,
FLAGS.max_sequence_len)
valid_data = read_data(FLAGS.valid_data_path, word2idx,
FLAGS.max_sequence_len)
graph = tf.Graph()
with tf.Session(graph=graph) as sess:
model = BiLSTM_CRF(FLAGS, sess, word_embeddings)
model.build_model()
model.train(train_data, valid_data)
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
char_embedding_mat = np.load('data/char_embedding_matrix.npy')
X_train = np.load('data/X_train.npy')
X_dev = np.load('data/X_dev.npy')
y_train = np.load('data/y_train.npy')
y_dev = np.load('data/y_dev.npy')
# ner_model = BiLSTM_CRF(n_input=200, n_vocab=char_embedding_mat.shape[0],
# n_embed=100, embedding_mat=char_embedding_mat,
# keep_prob=0.5, n_lstm=100, keep_prob_lstm=0.8,
# n_entity=7, optimizer='adam', batch_size=64, epochs=500)
ner_model = BiLSTM_CRF(n_input=200, n_vocab=char_embedding_mat.shape[0],
n_embed=100, embedding_mat=char_embedding_mat,
keep_prob=0.5, n_lstm=256, keep_prob_lstm=0.6,
n_entity=7, optimizer='adam', batch_size=16, epochs=50)
cp_folder, cp_file = 'checkpoints', 'bilstm_crf_weights_best.hdf5'
log_filepath = 'logs/bilstm_crf_summaries'
cb = [ModelCheckpoint(os.path.join(cp_folder, cp_file), monitor='val_loss',
verbose=1, save_best_only=True, save_weights_only=True, mode='min'),
EarlyStopping(min_delta=1e-8, patience=10, mode='min'),
ReduceLROnPlateau(factor=0.2, patience=6, verbose=0, mode='min',
epsilon=1e-6, cooldown=4, min_lr=1e-8),
TensorBoard(log_dir=log_filepath, write_graph=True, write_images=True,
histogram_freq=1)]
ner_model.train(X_train, y_train, X_dev, y_dev, cb)