def run_epoch(sess, m_train, m_valid, training_data, validation_data):
iters = 0
# TODO: shuffle and fetch data
sf_train = SF(training_data, CONFIG.BATCH_SIZE, is_training=True)
sf_valid = SF(validation_data, CONFIG.BATCH_SIZE, is_training=False)
for tup in sf_train.get_batch():
print('Training Batch: ', iters)
_, doc, summary, docLens, sumLens = tup
doc_batch = _get_doc_batch(doc)
summary_batch = _get_summary_batch(summary)
label_batch = _get_label_batch(summary)
loss = m_train.run_train_step(sess, doc_batch, summary_batch,
label_batch, np.array(docLens),
np.array(sumLens))
print('training loss: ', loss)
it = 0
for tup_valid in sf_valid.get_batch():
if (it > 0):
break
it += 1
_, doc_valid, summary_valid, docLens_valid, sumLens_valid = tup
doc_batch_v = _get_doc_batch(doc_valid)
summary_batch_v = _get_summary_batch(summary_valid)
label_batch_v = _get_label_batch(summary_valid)
m_valid.run_valid_step(sess, doc_batch_v, summary_batch_v,
label_batch_v, np.array(docLens_valid),
np.array(sumLens_valid))
iters += 1
p_valid = tf.nn.softmax(logits = u_valid)
#####################################
# start session and init
sess = tf.Session()
sess.run(init_op)
sess.run(embedding_init, feed_dict = {embedding_placeholder: np.matrix(Mword)})
ac_train_list = []
ac_validation_list = []
outfile = open ('results_rnn_last.txt', 'w')
for i in range(EPOCH):
print ('------------------running ' + str(i)+ 'th epoch --------------------')
print ('shuffling batch')
sf = SF(training_data, BATCH_SIZE)
sf.shuffle()
for tup in sf.get_batch(): # getting a batch of *documents*
j, batch_xs, batch_ys, batch_lens = tup # note: j means the j'th batch
#print ('training the ', str(j), 'th batch')
num_sentences = len(batch_xs[0])
# np_state for initialization at start of doc: [BATCH_SIZE, DIM_WordEmbedding]
np_state = sess.run(cell.zero_state(BATCH_SIZE, dtype=tf.float32))
sess.run(init_assign, feed_dict={init_state_placeholder:np_state})
for ns in range(num_sentences):
# prepare the batch of sentences, label, sentence_lens
sentences_batch = getSentenceBatch(batch_xs, ns)
labels_batch = getSentenceLabels(batch_ys, ns)
lens_batch = getSentenceLength(batch_lens, ns)