def random_augmentation(x):
aug_policies = found_policies.randaug_policies()
x_augm = torch.zeros_like(x)
for i in range(x.size()[0]):
chosen_policy = aug_policies[np.random.choice(len(aug_policies))]
aug_image = augmentation_transforms.apply_policy(
chosen_policy, x[i, :, :, :].permute(1, 2, 0).cpu().numpy())
aug_image = augmentation_transforms.cutout_numpy(aug_image)
tmp = torch.tensor(aug_image).permute(2, 0, 1)
x_augm[i, :, :, :] = torch.mean(tmp, dim=0).unsqueeze(0)
return x_augm
def proc_and_dump_unsup_data(sub_set_data, aug_copy_num, used_labels):
ori_images = sub_set_data["images"].copy()
# print('used_labels', used_labels)
ori_images = resampling_unsup_by_labels(ori_images,
sub_set_data["labels"].copy(),
used_labels)
print('unsup ori_images', ori_images.shape)
image_idx = np.arange(len(ori_images))
np.random.shuffle(image_idx)
ori_images = ori_images[image_idx]
# tf.logging.info("first 5 indexes after shuffling: {}".format(
# str(image_idx[:5])))
ori_images = ori_images / 255.0
mean, std = augmentation_transforms.get_mean_and_std()
ori_images = (ori_images - mean) / std
if FLAGS.task_name == "cifar10":
aug_policies = found_policies.randaug_policies()
elif FLAGS.task_name == "svhn":
aug_policies = found_policies.randaug_policies()
example_list = []
for image in ori_images:
chosen_policy = aug_policies[np.random.choice(len(aug_policies))]
aug_image = augmentation_transforms.apply_policy(chosen_policy, image)
aug_image = augmentation_transforms.cutout_numpy(aug_image)
# Write example to the tfrecord file
example = tf.train.Example(features=tf.train.Features(
feature={
"ori_image": _float_feature(image.reshape(-1)),
"aug_image": _float_feature(aug_image.reshape(-1)),
}))
example_list += [example]
out_path = os.path.join(
FLAGS.output_base_dir,
format_unsup_filename(aug_copy_num),
)
save_tfrecord(example_list, out_path)
def __call__(self, img):
policy = self.policies[np.random.choice(len(self.policies))]
final_img = augmentation_transforms.apply_policy(policy, img)
return final_img