def valid(sess, model):
total_data = dataManager.train(True)
acc_array = np.zeros((2), dtype=np.float32)
count, total_loss, cnt = 0, 0., 0
document, candidates, ori_labels, ori_locs = [], [], [], []
acc_blank = np.zeros((2, 2), dtype=np.float32)
for doc, can, lab, loc, _ in total_data:
document.append(doc)
candidates.append(can)
ori_labels.append(lab)
ori_locs.append(loc)
cnt += 1
if cnt == FLAGS.batch_size * 2:
document, doc_length, candidates, labels, locs, ori_labels, mask = \
prepare_batch_data(document, candidates, ori_labels, ori_locs)
valid_result = model.train_step(sess, document, doc_length,
candidates, labels, locs, mask,
'dev')
count += 1
total_loss += valid_result[0]
caculate_result = caculate_acc(ori_labels, valid_result[1])
acc_array += caculate_result[0]
acc_blank += caculate_result[1]
document, candidates, ori_labels, ori_locs = [], [], [], []
cnt = 0
if cnt > 0:
document, doc_length, candidates, labels, locs, ori_labels, mask = \
prepare_batch_data(document, candidates, ori_labels, ori_locs)
valid_result = model.train_step(sess, document, doc_length, candidates,
labels, locs, mask, 'dev')
count += 1
total_loss += valid_result[0]
caculate_result = caculate_acc(ori_labels, valid_result[1])
acc_array += caculate_result[0]
acc_blank += caculate_result[1]
print(caculate_result)
print(acc_array[0])
print(acc_array[1])
acc = acc_array[0] / acc_array[1]
acc_single = acc_blank[0, 0] / (acc_blank[0, 1] + 1e-12)
acc_multi = acc_blank[1, 0] / (acc_blank[1, 1] + 1e-12)
avg_loss = total_loss / count
print("*** Valid" +
" acc %.5f loss %.3f single_acc %.5f multi_acc %.5f" %
(acc, avg_loss, acc_single, acc_multi))
return acc, avg_loss
def train(sess, model):
acc_array = np.zeros((2), dtype=np.float32)
st_time, count, total_loss, cnt = time.time(), 0, 0., 0
if tf.train.get_checkpoint_state(train_dir):
best_dev_acc = valid(sess, model)[0]
best_iter = model.global_step.eval()
else:
best_dev_acc = 0.1
prev_loss = [1e15, 1e15]
document, candidates, ori_labels, ori_locs = [], [], [], []
for _ in range(10):
total_data = dataManager.train()
for doc, can, lab, loc, _ in total_data:
document.append(doc)
candidates.append(can)
ori_labels.append(lab)
ori_locs.append(loc)
cnt += 1
if cnt == FLAGS.batch_size:
document, doc_length, candidates, labels, locs, ori_labels, mask = \
prepare_batch_data(document, candidates, ori_labels, ori_locs)
train_result = model.train_step(sess, document, doc_length,
candidates, labels, locs, mask)
total_loss += train_result[0]
acc_array += caculate_acc(ori_labels, train_result[1])[0]
count += 1
if model.global_step.eval() % 100 == 0:
temp_avg_time = (time.time() - st_time) / count
temp_avg_loss = total_loss / count
acc = acc_array[0] / acc_array[1]
record_str = "step %d lr %.6f acc %.4f time %.3f loss %.3f" % \
(model.global_step.eval(), model.lr.eval(), acc, temp_avg_time, temp_avg_loss)
print(record_str)
if model.global_step.eval() % 1000 == 0:
dev_acc, dev_loss = valid(sess, model)
if dev_acc > best_dev_acc:
model.saver.save(sess,
"%s/checkpoint" % train_dir,
global_step=model.global_step)
best_dev_acc = dev_acc
best_iter = model.global_step.eval()
print("** New best iteration %d" % best_iter)
print("* Best iteration %d" % best_iter)
if dev_loss > max(prev_loss):
sess.run(model.lr_decay)
prev_loss = [prev_loss[1], dev_loss]
acc_array = np.zeros((2), dtype=np.float32)
st_time, count, total_loss = time.time(), 0, 0.
document, candidates, ori_labels, ori_locs = [], [], [], []
cnt = 0
def valid(sess, model, mode="dev", record=False):
if record:
file = open("pred_result_" + FLAGS.model + "_" + mode + ".txt", "w")
total_data = dataManager.valid(mode)
acc_array = np.zeros((2), dtype=np.float32)
count, total_loss, cnt = 0, 0., 0
document, candidates, ori_labels, ori_locs = [], [], [], []
acc_blank = np.zeros((2, 2), dtype=np.float32)
for doc, can, lab, loc in total_data:
document.append(doc)
candidates.append(can)
ori_labels.append(lab)
ori_locs.append(loc)
cnt += 1
if cnt == FLAGS.batch_size * 2:
document, doc_length, candidates, labels, locs, ori_labels, mask = \
prepare_batch_data(document, candidates, ori_labels, ori_locs)
valid_result = model.train_step(sess, document, doc_length,
candidates, labels, locs, False,
mask)
count += 1
total_loss += valid_result[0]
caculate_result = caculate_acc(ori_labels, valid_result[1])
acc_array += caculate_result[0]
acc_blank += caculate_result[1]
if record:
for id in range(len(ori_labels)):
ori_label = ori_labels[id]
pred_result = valid_result[1][id]
for pred in pred_result[:len(ori_label)]:
file.write(str(int(pred)) + " ")
file.write("\n")
document, candidates, ori_labels, ori_locs = [], [], [], []
cnt = 0
if cnt > 0:
document, doc_length, candidates, labels, locs, ori_labels, mask = \
prepare_batch_data(document, candidates, ori_labels, ori_locs)
valid_result = model.train_step(sess, document, doc_length, candidates,
labels, locs, False, mask)
count += 1
total_loss += valid_result[0]
caculate_result = caculate_acc(ori_labels, valid_result[1])
acc_array += caculate_result[0]
acc_blank += caculate_result[1]
if record:
for id in range(len(ori_labels)):
ori_label = ori_labels[id]
pred_result = valid_result[1][id]
for pred in pred_result[:len(ori_label)]:
file.write(str(int(pred)) + " ")
file.write("\n")
if record:
file.close()
acc = acc_array[0] / acc_array[1]
acc_single = acc_blank[0, 0] / (acc_blank[0, 1] + 1e-12)
acc_multi = acc_blank[1, 0] / (acc_blank[1, 1] + 1e-12)
avg_loss = total_loss / count
print("*** " + mode.upper() +
" acc %.5f loss %.3f single_acc %.5f multi_acc %.5f" %
(acc, avg_loss, acc_single, acc_multi))
return acc, avg_loss