def train():
batch_size = 100
seq_len = 128
data_set = DataSet(batch_size, 'data_assistant/vocabs', seq_len)
data_set.load_data('data_assistant/train/')
graph = tf.Graph()
with graph.as_default():
model = MyTextCNN(seq_len, 2, data_set.get_vocab_size())
model.build_model()
with tf.Session(graph=graph) as sess:
tf.global_variables_initializer().run()
print('start!')
epoch_batch_cnt = data_set.get_data_size() // batch_size
print('batch_per_epoch={b}'.format(b=epoch_batch_cnt))
total_step = 0
for epoch in range(1000):
print('epoch {e}'.format(e=epoch))
for ii in range(epoch_batch_cnt + 1):
X, Y = data_set.get_batch()
loss_val, accuracy = model.train(sess, X, Y, total_step)
if total_step % 2 == 0:
print('step {c}, loss={l}, accuracy={a}'.format(
c=total_step, l=loss_val, a=accuracy))
if math.isnan(loss_val):
print('Nan loss!!')
return
if total_step % 100 == 0:
model.save(sess, "data_assistant/model/", total_step)
print('Saved!')
total_step += 1
