def train(n_vocab, labels, embedding, embed, int_to_vocab):
loss_op, train_op = get_loss_and_training_op(n_vocab, labels, embed)
valid_words = sample_eval_data()
with tf.Session() as sess:
saver = tf.train.Saver()
all_losses = []
batch_loss = []
sess.run(tf.global_variables_initializer())
start = time.time()
for i in range(FLAGS.total_iterations):
loss, _ = sess.run([loss_op, train_op])
all_losses.append(loss)
batch_loss.append(loss)
if i % FLAGS.log_every == 0:
end = time.time()
print(
'Iteration {}/{} '.format(i, FLAGS.total_iterations),
'Average Loss: {:.4f}'.format(np.mean(batch_loss)),
'{:.4f} sec/{} iterations'.format((end - start),
FLAGS.log_every))
batch_loss = []
start = time.time()
if i % FLAGS.evaluate_every == 0:
saver.save(sess, 'checkpoint/model-{}.ckpt'.format(i))
pred_op = get_predictions(valid_words, embedding)
predictions = sess.run(pred_op)
words = get_top_10_words(predictions, int_to_vocab)
saver.save(sess, 'checkpoint/model.ckpt')
np.savez('checkpoint/all_losses.npz', all_losses)
def predict(valid_words, estimator, int_to_vocab, vocab_to_int):
pred_generator = estimator.predict(
input_fn=lambda: get_eval_dataset(valid_words))
for predictions in pred_generator:
words = get_top_10_words(np.expand_dims(predictions['similarity'], 0),
int_to_vocab)
return words
def predict(valid_words, embedding, int_to_vocab):
pred_op = get_predictions(valid_words, embedding)
with tf.Session() as sess:
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
saver.restore(sess, 'checkpoint/model.ckpt')
predictions = sess.run(pred_op)
words = get_top_10_words(predictions, int_to_vocab)
return words