def fit(self, sess, input_data, verbose=None):
"""
Runs an epoch of validation or test. return test error
Args:
sess: tf.Session() object
input_data: tuple of (encode_input, decode_input, decode_label)
Returns:
avg_loss: scalar. Average minibatch loss of model on epoch.
"""
total_loss = []
for step, (b_data, b_order) in enumerate(
helper.data_iter(input_data, self.config.batch_size)):
order_indices = b_order
losses = []
for i in range(self.config.processing_step):
(ret_batch, ret_label, sent_num_enc, sent_num_dec,
sent_len) = helper.shuffleData(b_data, order_indices,
self.vocab)
feed_dict = self.create_feed_dict(ret_batch, sent_len,
sent_num_enc, ret_label,
sent_num_dec)
loss, pred = sess.run([self.loss, self.prediction],
feed_dict=feed_dict)
pred = pred.tolist()
order_indices = helper.reorder(order_indices, pred,
sent_num_dec)
losses.append(loss)
total_loss.append(np.mean(losses))
avg_loss = np.mean(total_loss)
return avg_loss
def predict(self, sess, input_data, verbose=None):
preds = []
true_label = []
lengths = []
for _, (b_data, b_order) in enumerate(
helper.data_iter(input_data, self.config.batch_size)):
order_indices = b_order
pred = None
ret_label = None
sent_num_dec = None
for i in range(self.config.processing_step):
(ret_batch, ret_label, sent_num_enc, sent_num_dec,
sent_len) = helper.shuffleData(b_data, order_indices,
self.vocab)
feed_dict = self.create_feed_dict(ret_batch, sent_len,
sent_num_enc, ret_label,
sent_num_dec)
pred = sess.run(self.prediction, feed_dict=feed_dict)
pred = pred.tolist()
order_indices = helper.reorder(order_indices, pred,
sent_num_dec)
preds += pred
true_label += ret_label.tolist()
lengths += sent_num_dec
return preds, true_label, lengths
def run_epoch(self, sess, input_data, verbose=None):
"""
Runs an epoch of training.
Trains the model for one-epoch.
Args:
sess: tf.Session() object
input_data: tuple of (encode_input, decode_input, decode_label)
Returns:
avg_loss: scalar. Average minibatch loss of model on epoch.
"""
data_len = len(input_data)
total_steps =data_len // self.config.batch_size
total_loss = []
for step, (b_data, b_order) in enumerate(
helper.data_iter(input_data, self.config.batch_size)):
order_indices = b_order
losses = []
for i in range(self.config.processing_step):
(ret_batch, ret_label, sent_num_enc, sent_num_dec, sent_len
) = helper.shuffleData(b_data, order_indices, self.vocab)
feed_dict = self.create_feed_dict(ret_batch, sent_len, sent_num_enc, ret_label, sent_num_dec)
_, loss, lr, pred = sess.run([self.train_op, self.loss, self.learning_rate, self.prediction], feed_dict=feed_dict)
pred = pred.tolist()
order_indices = helper.reorder(order_indices, pred, sent_num_dec)
losses.append(loss)
total_loss.append(np.mean(losses))
if verbose and step % verbose == 0:
sys.stdout.write('\r{} / {} : loss = {}, lr = {}'.format(
step, total_steps, np.mean(total_loss[-verbose:]), lr))
sys.stdout.flush()
sys.stdout.write('\n')
avg_loss = np.mean(total_loss)
return avg_loss
