def _rotate_pp(data):
# Create labels in the same structure as images!
if create_labels:
data["label"] = utils.tf_apply_to_image_or_images(
lambda _: tf.constant([0, 1, 2, 3]), data["image"], dtype=tf.int32)
data["image"] = utils.tf_apply_to_image_or_images(_four_rots, data["image"])
return data
def _crop_pp(data):
crop_fn = functools.partial(__crop,
is_training=is_training,
crop_size=crop_size)
data["image"] = utils.tf_apply_to_image_or_images(
crop_fn, data["image"])
return data
def _rotation_pp(data):
if is_training:
data['image'] = utils.tf_apply_to_image_or_images(
lambda img: tf.image.rot90(
img, k=tf.random_uniform([], dtype=tf.int32, maxval=4)),
data['image'])
return data
def _crop_patches_pp(data):
image = data["image"]
image_to_patches_fn = functools.partial(_image_to_patches,
is_training=is_training,
split_per_side=split_per_side,
channel_jitter=channel_jitter)
image = utils.tf_apply_to_image_or_images(image_to_patches_fn, image)
data["image"] = image
return data
def _random_flip_lr_pp(data):
if is_training:
data["image"] = utils.tf_apply_to_image_or_images(
tf.image.random_flip_left_right, data["image"])
return data
def _to_gray_pp(data):
data["image"] = utils.tf_apply_to_image_or_images(
lambda img: utils.tf_apply_with_probability(
grayscale_probability, _to_gray, img), data["image"])
return data
def _standardization_pp(data):
# Trick: normalize each patch to avoid low level statistics.
data["image"] = utils.tf_apply_to_image_or_images(
tf.image.per_image_standardization, data["image"])
return data
def _to_gray_pp(data):
data["image"] = utils.tf_apply_to_image_or_images(
lambda img: utils.tf_apply_with_probability( # pylint:disable=g-long-lambda
grayscale_probability, _to_gray, img),
data["image"])
return data
def _inception_crop_patches_pp(data):
# The output becomes float32 because of the tf.image.resize.
data["image"] = utils.tf_apply_to_image_or_images(
_inception_crop_patches, data["image"], dtype=tf.float32)
return data
def _hsvnoise_pp(data): data["image"] = utils.tf_apply_to_image_or_images(_hsvnoise, data["image"]) return data
