def validate(self):
print('\n\nValidating epoch: %d' % self.epoch)
print('-' * 80)
total = len(self.dataset_val)
val_generator = self.dataset_val.generator(self.options.batch_size)
progbar = Progbar(total, width=25)
for input_rgb in val_generator:
originItems = []
sketchItems = []
for ix in range(len(input_rgb)):
originItems.append(input_rgb[ix][0])
sketchItems.append(input_rgb[ix][1])
originItems = np.array(originItems)
sketchItems = np.array(sketchItems)
feed_dic = {
self.input_rgb:
originItems,
self.input_gray:
sketchItems.reshape(self.options.batch_size, 512, 512, 1)
}
self.sess.run([self.dis_loss, self.gen_loss, self.accuracy],
feed_dict=feed_dic)
lossD, lossD_fake, lossD_real, lossG, lossG_l1, lossG_gan, acc, step = self.eval_outputs(
feed_dic=feed_dic)
progbar.add(len(input_rgb),
values=[("D loss", lossD), ("D fake", lossD_fake),
("D real", lossD_real), ("G loss", lossG),
("G L1", lossG_l1), ("G gan", lossG_gan),
("accuracy", acc)])
print('\n')
def validate(self):
print('\n\nValidating epoch: %d' % self.epoch)
total = len(self.dataset_val)
val_generator = self.dataset_val.generator(self.options.batch_size)
progbar = Progbar(total, width=25)
for input_rgb in val_generator:
feed_dic = {self.input_rgb: input_rgb}
self.sess.run([self.dis_loss, self.gen_loss, self.accuracy],
feed_dict=feed_dic)
lossD, lossD_fake, lossD_real, lossG, lossG_l1, lossG_gan, acc, step = self.eval_outputs(
feed_dic=feed_dic)
progbar.add(len(input_rgb),
values=[("D loss", lossD), ("D fake", lossD_fake),
("D real", lossD_real), ("G loss", lossG),
("G L1", lossG_l1), ("G gan", lossG_gan),
("accuracy", acc)])
print('\n')
def train(self):
total = len(self.dataset_train)
for epoch in range(self.options.epochs):
lr_rate = self.sess.run(self.learning_rate)
print('Training epoch: %d' % (epoch + 1) + " - learning rate: " +
str(lr_rate))
print('-' * 80)
print(self.dataset_train.load())
self.epoch = epoch + 1
self.iteration = 0
generator = self.dataset_train.generator(self.options.batch_size)
progbar = Progbar(total,
width=25,
stateful_metrics=['epoch', 'iter', 'step'])
for input_rgb in generator:
print("-" * 80)
#print(len(input_rgb))
#print(input_rgb[0][0].shape)
#print(input_rgb[0][1].shape)
originItems = []
sketchItems = []
for ix in range(len(input_rgb)):
originItems.append(input_rgb[ix][0])
sketchItems.append(input_rgb[ix][1])
originItems = np.array(originItems)
sketchItems = np.array(sketchItems)
feed_dic = {
self.input_rgb:
originItems,
self.input_gray:
sketchItems.reshape(self.options.batch_size, 512, 512, 1)
}
self.iteration = self.iteration + 1
self.sess.run([self.dis_train], feed_dict=feed_dic)
self.sess.run([self.gen_train, self.accuracy],
feed_dict=feed_dic)
self.sess.run([self.gen_train, self.accuracy],
feed_dict=feed_dic)
lossD, lossD_fake, lossD_real, lossG, lossG_l1, lossG_gan, acc, step = self.eval_outputs(
feed_dic=feed_dic)
progbar.add(len(input_rgb),
values=[("epoch", epoch + 1),
("iter", self.iteration), ("step", step),
("D loss", lossD), ("D fake", lossD_fake),
("D real", lossD_real), ("G loss", lossG),
("G L1", lossG_l1), ("G gan", lossG_gan),
("accuracy", acc)])
# log model at checkpoints
if self.options.log and step % self.options.log_interval == 0:
with open(self.train_log_file, 'a') as f:
f.write('%d %d %f %f %f %f %f %f %f\n' %
(self.epoch, step, lossD, lossD_fake,
lossD_real, lossG, lossG_l1, lossG_gan, acc))
if self.options.visualize:
visualize(self.train_log_file, self.test_log_file,
self.options.visualize_window, self.name)
# sample model at checkpoints
if self.options.sample and step % self.options.sample_interval == 0:
self.sample(show=False)
# validate model at checkpoints
if self.options.validate and self.options.validate_interval > 0 and step % self.options.validate_interval == 0:
self.validate()
# save model at checkpoints
if self.options.save and step % self.options.save_interval == 0:
self.save()
if self.options.validate:
self.validate()
BATCH_SIZE = 128
LEARNING_RATE = 0.1
learning_rate.set_value(np.cast['float32'](LEARNING_RATE))
training_history = []
valid_history = []
for epoch in range(NB_EPOCH):
prog = Progbar(target=X_train.shape[0])
n = 0
history = []
while n < X_train.shape[0]:
start = n
end = min(n + BATCH_SIZE, X_train.shape[0])
c = f_train(X_train[start:end], y_train[start:end])
prog.title = 'Epoch: %.2d, Cost: %.4f' % (epoch + 1, c)
prog.add(end - start)
n += BATCH_SIZE
history.append(c)
# end of epoch, start validating
y = np.argmax(f_predict(X_valid), axis=-1)
accuracy = accuracy_score(y_valid, y)
print('Validation accuracy:', accuracy)
# save history
training_history.append(np.mean(history))
valid_history.append(accuracy)
y = np.argmax(f_predict(X_test), axis=-1)
accuracy = accuracy_score(y_test, y)
print('Test accuracy:', accuracy)
print('Classification report:')
print(classification_report(y_test, y))
