def build_model(word_vocab):
if FLAGS.model_choice == 'bilstm':
valid_model = model.cnn_sen_enc(
word_vocab_size=word_vocab.size,
word_embed_size=FLAGS.word_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
max_sen_length=FLAGS.max_sen_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
max_doc_length=FLAGS.max_doc_length)
valid_model.update(
model.bilstm_doc_enc(valid_model.input_cnn,
batch_size=FLAGS.batch_size,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_doc_length=FLAGS.max_doc_length,
dropout=0.0))
valid_model.update(model.label_prediction(valid_model.enc_outputs))
valid_model.update(
model.loss_extraction(valid_model.logits, FLAGS.batch_size,
FLAGS.max_doc_length))
elif FLAGS.model_choice == 'lstm':
valid_model = model.cnn_sen_enc(
word_vocab_size=word_vocab.size,
word_embed_size=FLAGS.word_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
max_sen_length=FLAGS.max_sen_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
max_doc_length=FLAGS.max_doc_length)
valid_model.update(
model.lstm_doc_enc(valid_model.input_cnn,
batch_size=FLAGS.batch_size,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_doc_length=FLAGS.max_doc_length,
dropout=0.0))
valid_model.update(
model.lstm_doc_dec(valid_model.input_cnn,
valid_model.final_enc_state,
batch_size=FLAGS.batch_size,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_doc_length=FLAGS.max_doc_length,
dropout=0.0))
valid_model.update(
model.label_prediction_att(valid_model.enc_outputs,
valid_model.dec_outputs))
valid_model.update(
model.loss_extraction(valid_model.logits, FLAGS.batch_size,
FLAGS.max_doc_length))
return valid_model
def build_model(word_vocab, max_doc_length, train):
'''build a training or inference graph, based on the model choice'''
my_model = None
if train:
pretrained_emb = None
embedding_path = FLAGS.embedding_path
if FLAGS.load_model is None and embedding_path != None and os.path.exists(
embedding_path):
pretrained_emb = load_wordvec(embedding_path, word_vocab)
if FLAGS.model_choice == 'bilstm':
my_model = model.cnn_sen_enc(
word_vocab_size=word_vocab.size,
word_embed_size=FLAGS.word_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
max_sen_length=FLAGS.max_sen_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
max_doc_length=max_doc_length,
pretrained=pretrained_emb)
my_model.update(
model.bilstm_doc_enc(my_model.input_cnn,
batch_size=FLAGS.batch_size,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_doc_length=max_doc_length,
dropout=FLAGS.dropout))
my_model.update(model.label_prediction(my_model.enc_outputs))
my_model.update(
model.loss_extraction(my_model.logits, FLAGS.batch_size,
max_doc_length))
my_model.update(
model.training_graph(my_model.loss * max_doc_length,
FLAGS.learning_rate, FLAGS.max_grad_norm))
elif FLAGS.model_choice == 'lstm':
my_model = model.cnn_sen_enc(
word_vocab_size=word_vocab.size,
word_embed_size=FLAGS.word_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
max_sen_length=FLAGS.max_sen_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
max_doc_length=max_doc_length,
pretrained=pretrained_emb)
my_model.update(
model.lstm_doc_enc(my_model.input_cnn,
batch_size=FLAGS.batch_size,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_doc_length=max_doc_length,
dropout=FLAGS.dropout))
my_model.update(
model.lstm_doc_dec(my_model.input_cnn,
my_model.final_enc_state,
batch_size=FLAGS.batch_size,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_doc_length=max_doc_length,
dropout=FLAGS.dropout))
my_model.update(
model.label_prediction_att(my_model.enc_outputs,
my_model.dec_outputs))
my_model.update(
model.loss_extraction(my_model.logits, FLAGS.batch_size,
max_doc_length))
my_model.update(
model.training_graph(my_model.loss * max_doc_length,
FLAGS.learning_rate, FLAGS.max_grad_norm))
else:
if FLAGS.model_choice == 'bilstm':
my_model = model.cnn_sen_enc(
word_vocab_size=word_vocab.size,
word_embed_size=FLAGS.word_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
max_sen_length=FLAGS.max_sen_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
max_doc_length=max_doc_length)
my_model.update(
model.bilstm_doc_enc(my_model.input_cnn,
batch_size=FLAGS.batch_size,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_doc_length=max_doc_length,
dropout=FLAGS.dropout))
my_model.update(model.label_prediction(my_model.enc_outputs))
my_model.update(
model.loss_extraction(my_model.logits, FLAGS.batch_size,
max_doc_length))
elif FLAGS.model_choice == 'lstm':
my_model = model.cnn_sen_enc(
word_vocab_size=word_vocab.size,
word_embed_size=FLAGS.word_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
max_sen_length=FLAGS.max_sen_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
max_doc_length=max_doc_length)
my_model.update(
model.lstm_doc_enc(my_model.input_cnn,
batch_size=FLAGS.batch_size,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_doc_length=max_doc_length,
dropout=FLAGS.dropout))
my_model.update(
model.lstm_doc_dec(my_model.input_cnn,
my_model.final_enc_state,
batch_size=FLAGS.batch_size,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_doc_length=max_doc_length,
dropout=FLAGS.dropout))
my_model.update(
model.label_prediction_att(my_model.enc_outputs,
my_model.dec_outputs))
my_model.update(
model.loss_extraction(my_model.logits, FLAGS.batch_size,
max_doc_length))
return my_model