def train(self, train_words, train_chars, train_labels, test_words,
test_chars, test_labels):
global_test_acc, global_step, lr = 0.0, 0, self.cfg.lr
num_batches = math.ceil(float(len(train_words) / self.cfg.batch_size))
self.logger.info("start training...")
for epoch in range(1, self.cfg.epochs + 1):
self.logger.info("Epoch {}/{}:".format(epoch, self.cfg.epochs))
train_words, train_chars, train_labels = sklearn.utils.shuffle(
train_words, train_chars, train_labels)
prog = Progbar(target=num_batches)
for i, (b_words, b_seq_len, b_chars, b_char_seq_len,
b_labels) in enumerate(
batch_iter(train_words, train_chars, train_labels,
self.cfg.batch_size)):
global_step += 1
batch_labels = []
for j in range(self.num_classifier):
ecoc_array = self.nary_ecoc[:, j]
b_lbs = remap_labels(b_labels.copy(), ecoc_array)
b_lbs = dense_to_one_hot(b_lbs, self.num_class)
batch_labels.append(b_lbs)
feed_dict = self.get_feed_dict(b_words,
b_seq_len,
b_chars,
b_char_seq_len,
batch_labels,
lr=lr,
training=True)
_, pred_labels, loss = self.sess.run(
[self.train_op, self.pred_labels, self.loss],
feed_dict=feed_dict)
acc = compute_ensemble_accuracy(pred_labels, self.nary_ecoc,
b_labels)
prog.update(i + 1, [("Global Step", global_step),
("Train Loss", loss),
("Train Acc", acc * 100)])
accuracy, _ = self.test(test_words,
test_chars,
test_labels,
batch_size=200,
print_info=True,
restore=False)
if accuracy > global_test_acc:
global_test_acc = accuracy
self.save_session(epoch)
lr = self.cfg.lr / (1 + epoch * self.cfg.lr_decay)
def test(self, test_dataset, print_info=True):
self.restore_last_session()
test_imgs, test_labels = test_dataset.images, test_dataset.labels
y_labels = []
for j in range(self.num_classifier):
ecoc_array = self.nary_ecoc[:, j]
b_lbs = remap_labels(test_labels.copy(), ecoc_array)
b_lbs = dense_to_one_hot(b_lbs, self.num_class)
y_labels.append(b_lbs)
feed_dict = self._get_feed_dict(test_imgs, y_labels)
pred_labels, test_loss = self.sess.run([self.pred_labels, self.cost],
feed_dict=feed_dict)
test_acc = compute_ensemble_accuracy(pred_labels, self.nary_ecoc,
test_dataset.labels)
if print_info:
self.logger.info(" -- Test Loss: {}, Test Accuracy: {}".format(
test_loss, test_acc))
return pred_labels
def train(self, train_dataset, test_dataset):
global_test_acc = 0.0
global_step = 0
test_imgs, test_labels = test_dataset.images, test_dataset.labels
self.logger.info("start training...")
for epoch in range(1, self.epochs + 1):
self.logger.info("Epoch {}/{}:".format(epoch, self.epochs))
num_batches = train_dataset.num_examples // self.batch_size
prog = Progbar(target=num_batches)
prog.update(0, [("Global Step", 0), ("Train Loss", 0.0),
("Train Acc", 0.0), ("Test Loss", 0.0),
("Test Acc", 0.0)])
for i in range(num_batches):
global_step += 1
train_imgs, train_labels = train_dataset.next_batch(
self.batch_size)
b_labels = []
for j in range(self.num_classifier):
ecoc_array = self.nary_ecoc[:, j]
b_lbs = remap_labels(train_labels.copy(), ecoc_array)
b_lbs = dense_to_one_hot(b_lbs, self.num_class)
b_labels.append(b_lbs)
feed_dict = self._get_feed_dict(train_imgs, b_labels, True)
_, pred_labels, loss = self.sess.run(
[self.train_op, self.pred_labels, self.cost],
feed_dict=feed_dict)
acc = compute_ensemble_accuracy(pred_labels, self.nary_ecoc,
train_labels)
if global_step % 100 == 0:
y_labels = []
for j in range(self.num_classifier):
ecoc_array = self.nary_ecoc[:, j]
b_lbs = remap_labels(test_labels.copy(), ecoc_array)
b_lbs = dense_to_one_hot(b_lbs, self.num_class)
y_labels.append(b_lbs)
feed_dict = self._get_feed_dict(test_imgs, y_labels)
test_pred_labels, test_loss = self.sess.run(
[self.pred_labels, self.cost], feed_dict=feed_dict)
test_acc = compute_ensemble_accuracy(
test_pred_labels, self.nary_ecoc, test_labels)
prog.update(i + 1, [("Global Step", int(global_step)),
("Train Loss", loss),
("Train Acc", acc),
("Test Loss", test_loss),
("Test Acc", test_acc)])
if test_acc > global_test_acc:
global_test_acc = test_acc
self.save_session(global_step)
else:
prog.update(i + 1, [("Global Step", int(global_step)),
("Train Loss", loss),
("Train Acc", acc)])
y_labels = []
for j in range(self.num_classifier):
ecoc_array = self.nary_ecoc[:, j]
b_lbs = remap_labels(test_labels.copy(), ecoc_array)
b_lbs = dense_to_one_hot(b_lbs, self.num_class)
y_labels.append(b_lbs)
feed_dict = self._get_feed_dict(test_imgs, y_labels)
test_pred_labels, test_loss = self.sess.run(
[self.pred_labels, self.cost], feed_dict=feed_dict)
test_acc = compute_ensemble_accuracy(test_pred_labels,
self.nary_ecoc, test_labels)
self.logger.info(
"Epoch: {}, Global Step: {}, Test Loss: {}, Test Accuracy: {}".
format(epoch, global_step, test_loss, test_acc))
