def train_epoch(self, epochs):
if not self.init_train:
raise Exception('Train graph is not inited')
with self.train_graph.as_default():
if os.path.isfile(cfg.save_path + '.meta'):
print("##########################")
print('# Model restore.. #')
print("##########################")
self.train_saver.restore(self.train_session, cfg.save_path)
else:
self.train_session.run(self.train_init)
total_loss = 0
total_step = 0
start_time =time.time()
for e in range(epochs):
for step in range(self.total_iter// self.batch_size):
data = next(self.train_data)
week_stock = data['week_stock']
month_stock = data['month_stock']
t_month_stock = data['t_month_stock']
decoder_input = data['decoder_input']
decoder_target = data['decoder_target']
batch_seq = batch_seq_len(data['decoder_target'])
_, loss, sample_id = self.train_session.run([self.optimizer, self.loss, self.sample_id],
feed_dict = {self.week_input : week_stock,
self.month_input : month_stock,
self.t_month_input : t_month_stock,
self.decoder_input : decoder_input,
self.decoder_target : decoder_target,
self.decoder_targets_length : batch_seq})
# total_loss += loss
# total_step += self.total_iter
# loss = total_loss/total_step
end = time.time()
print('epoch: {}|{} minibatch loss: {:.6f} Time: {:.1f} min'.format(e+1, epochs, loss, (end-start_time)/60 ))
if e % 50 ==0:
self.train_saver.save(self.train_session, cfg.save_path)
#랜덤 sid 선택, training output_text
sid = random.randint(0, self.batch_size-1)
target_text = decode_text(decoder_target[sid],self.idx2word_dict)
output_text = decode_text(sample_id[sid],self.idx2word_dict)
print('============ training sample text =============')
print('training_target :' + target_text)
print('training_output :' + output_text)
print('===============================================')
self.eval()
def get_output():
data = request.get_json()
input_sample = data["sample"]
model_name = data["model"]
sup_ids = input_sample["sup_ids"] if "sup_ids" in input_sample else None
answer_text = decode_text(input_sample["answer"])
context = decode_text(input_sample["context"])
answer_start = context.lower().find(answer_text.lower())
answer_dict = {"text": answer_text, "answer_start": answer_start}
print(sup_ids)
if not sup_ids:
if answer_start != -1 and answer_text != "":
sup_ids = [find_sup_char_ids(context, answer_start)]
sample = {
"id": decode_text(input_sample["id"]),
"question": decode_text(input_sample["question"]),
"context": context,
"answer": answer_dict,
"sup_ids": sup_ids
}
prediction, layers, token_indices = generate_model_output(
sample, model_name)
output = {
'hidden_states': layers,
'prediction': prediction,
'token_indices': token_indices
}
return jsonify(output)
def eval(self):
with self.eval_graph.as_default():
self.eval_saver.restore(self.eval_session, cfg.save_path)
all_bleu = [0] * 4
eval_mask_weights = tf.ones(shape=[self.batch_size, self.max_length],dtype=tf.float64)
for step in range(self.val_total_iter//self.batch_size):
data = next(self.val_data)
week_stock = data['week_stock']
month_stock = data['month_stock']
t_month_stock = data['t_month_stock']
batch_seq = batch_seq_len(data['decoder_target'])
#beam search_output
beam_output = self.eval_session.run([self.predicted_ids],
feed_dict = {self.eval_week_input : week_stock,
self.eval_month_input : month_stock,
self.eval_t_month_input : t_month_stock,
self.eval_decoder_targets_length : batch_seq
})
target_text = idx_to_text(data['decoder_input'][:,1:],self.idx2word_dict)
target_text = remove_sent_pad(target_text)
beam_output = np.squeeze(np.array(beam_output),axis=0)
output_text = idx_to_text(beam_output[:,:,0], self.idx2word_dict)
bleu_score = self.cal_metrics(target_text, output_text)
for idx,score in enumerate(bleu_score):
all_bleu[idx] += score
print('================ BLEU score ================')
for idx, bleu in enumerate(bleu_score):#2,3,4,5
print('BLEU-{} : {}'.format(idx+2, bleu))
sid = random.randint(0, self.batch_size-1)
target_text = decode_text(data['decoder_target'][sid],self.idx2word_dict)
output_text = decode_text(beam_output[sid,:,0],self.idx2word_dict)
print('============= Beam search text =============')
print('infer_target : ' + target_text)
print('beam_search : ' + output_text)
print('============================================')
def search(entry, level):
entry = dict_value(entry)
if 'Title' in entry:
if 'A' in entry or 'Dest' in entry:
title = decode_text(str_value(entry['Title']))
dest = entry.get('Dest')
action = entry.get('A')
se = entry.get('SE')
yield (level, title, dest, action, se)
if 'First' in entry and 'Last' in entry:
for x in search(entry['First'], level + 1):
yield x
if 'Next' in entry:
for x in search(entry['Next'], level):
yield x
return
def search(entry, level):
entry = dict_value(entry)
if 'Title' in entry:
if 'A' in entry or 'Dest' in entry:
title = decode_text(str_value(entry['Title']))
dest = entry.get('Dest')
action = entry.get('A')
se = entry.get('SE')
yield (level, title, dest, action, se)
if 'First' in entry and 'Last' in entry:
for x in search(entry['First'], level+1):
yield x
if 'Next' in entry:
for x in search(entry['Next'], level):
yield x
return
