def next_batch(self):
"""Return the next minibatch of augmented data."""
next_train_index = self.curr_train_index + self.hparams.batch_size
if next_train_index > self.num_train:
# Increase epoch number
epoch = self.epochs + 1
self.reset()
self.epochs = epoch
batched_data = (
self.train_images[self.curr_train_index:self.curr_train_index +
self.hparams.batch_size],
self.train_labels[self.curr_train_index:self.curr_train_index +
self.hparams.batch_size])
final_imgs = []
images, labels = batched_data
for data in images:
epoch_policy = self.good_policies[np.random.choice(
len(self.good_policies))]
final_img = augmentation_transforms.apply_policy(
epoch_policy, data)
final_img = augmentation_transforms.random_flip(
augmentation_transforms.zero_pad_and_crop(final_img, 4))
# Apply cutout
final_img = augmentation_transforms.cutout_numpy(final_img)
final_imgs.append(final_img)
batched_data = (np.array(final_imgs, np.float32), labels)
self.curr_train_index += self.hparams.batch_size
return batched_data
def func_autoaug(img, selected_policies):
epoch_policy = selected_policies[np.random.choice(len(selected_policies))]
img = np.array(img)
img = AT.apply_policy(epoch_policy, img)
img = np.array(img)
img = AT.random_flip(AT.zero_pad_and_crop(img, 4))
img = AT.cutout_numpy(img)
return img
def adv_random_batch(self):
next_train_index = self.curr_train_index + self.batch_size
if next_train_index > self.fooled_images.shape[0]:
perm = np.arange(self.fooled_images.shape[0])
np.random.shuffle(perm)
self.fooled_images = self.fooled_images[perm]
self.fooled_labels = self.fooled_labels[perm]
self.curr_train_index = 0
batch = (self.fooled_images[self.curr_train_index: self.curr_train_index + self.batch_size], \
self.fooled_labels[self.curr_train_index: self.curr_train_index + self.batch_size])
images, labels = batch
final_imgs = []
for data in images:
final_img = augmentation_transforms.random_flip(
augmentation_transforms.zero_pad_and_crop(data, 4))
final_imgs.append(final_img)
batched_data = (np.array(final_imgs, np.float32), labels)
self.curr_train_index = self.curr_train_index + self.batch_size
return batched_data
def random_batch(self):
"""Return the next minibatch of non-augmented data."""
next_train_index = self.curr_train_index + self.hparams.batch_size
if next_train_index > self.num_train:
# Increase epoch number
epoch = self.epochs + 1
self.reset()
self.epochs = epoch
final_imgs = []
batched_data = (
self.train_images[self.curr_train_index:
self.curr_train_index + self.hparams.batch_size],
self.train_labels[self.curr_train_index:
self.curr_train_index + self.hparams.batch_size])
images, labels = batched_data
for data in images:
final_img = augmentation_transforms.random_flip(
augmentation_transforms.zero_pad_and_crop(data, 4))
final_imgs.append(final_img)
batched_data = (np.array(final_imgs, np.float32), labels)
self.curr_train_index += self.hparams.batch_size
return batched_data
def next_batch(self):
"""Return the next minibatch of augmented data."""
next_train_index = self.curr_train_index + self.hparams.batch_size
if next_train_index > self.num_train:
# Increase epoch number
epoch = self.epochs + 1
self.reset()
self.epochs = epoch
batched_data = (
self.train_images[self.curr_train_index:self.curr_train_index +
self.hparams.batch_size],
self.train_labels[self.curr_train_index:self.curr_train_index +
self.hparams.batch_size])
final_imgs = []
images, labels = batched_data
if self.hparams.augment_type == 'mixup':
images, labels = augmentation_transforms.mixup_batch(
images, labels, self.hparams.mixup_alpha)
elif self.hparams.augment_type == 'image_freq':
images, labels = augmentation_transforms.freq_augment(
images,
labels,
amplitude=self.hparams.freq_augment_amplitude,
magnitude=self.hparams.augmentation_magnitude,
proportion_f=self.hparams.freq_augment_ffrac,
probability=self.hparams.augmentation_probability)
for data in images:
if self.hparams.augment_type == 'autoaugment':
epoch_policy = self.good_policies[np.random.choice(
len(self.good_policies))]
final_img = augmentation_transforms.apply_policy(epoch_policy, data)
elif self.hparams.augment_type == 'random':
epoch_policy = found_policies.random_policy(
self.hparams.num_augmentation_layers,
self.hparams.augmentation_magnitude,
self.hparams.augmentation_probability)
final_img = augmentation_transforms.apply_policy(epoch_policy, data)
else:
final_img = np.copy(data)
if self.hparams.apply_flip_crop:
final_img = augmentation_transforms.random_flip(
augmentation_transforms.zero_pad_and_crop(data, 4))
# Apply cutout
if self.hparams.apply_cutout:
final_img = augmentation_transforms.cutout_numpy(final_img)
final_imgs.append(final_img)
final_imgs = np.array(final_imgs, np.float32)
if self.hparams.noise_type == 'radial':
labels = augmentation_transforms.add_radial_noise(
final_imgs, labels, self.hparams.frequency, self.hparams.amplitude,
self.hparams.noise_class, self.hparams.normalize_amplitude)
elif self.hparams.noise_type == 'random' or self.hparams.noise_type == 'fourier' or self.hparams.noise_type == 'f' or self.hparams.noise_type == '1/f':
labels = augmentation_transforms.add_sinusoidal_noise(
final_imgs, labels, self.hparams.frequency, self.hparams.amplitude,
self.direction, self.hparams.noise_class,
self.hparams.normalize_amplitude)
elif self.hparams.noise_type == 'uniform':
labels = augmentation_transforms.add_uniform_noise(
labels, self.hparams.amplitude, self.hparams.noise_class)
batched_data = (final_imgs, labels)
self.curr_train_index += self.hparams.batch_size
return batched_data
def func_manual(self, img):
img = np.array(img)
img = AT.random_flip(AT.zero_pad_and_crop(img, 4))
img = AT.cutout_numpy(img)
return img