def _unl_parser(value):
"""Cifar10 parser."""
image_size = params.image_size
value = tf.io.decode_raw(value, tf.uint8)
image = tf.reshape(value[1:], [3, 32, 32]) # uint8
image = tf.transpose(image, [1, 2, 0])
if image_size != 32:
image = tf.image.resize_bicubic([image],
[image_size, image_size])[0]
image.set_shape([image_size, image_size, 3])
image = _flip_and_jitter(image, replace_value=128)
ori_image = image
aug = augment.RandAugment(cutout_const=image_size // 8,
translate_const=image_size // 8,
magnitude=params.augment_magnitude)
aug_image = aug.distort(image)
aug_image = augment.cutout(aug_image,
pad_size=image_size // 4,
replace=128)
aug_image = _flip_and_jitter(aug_image, replace_value=128)
ori_image = convert_and_normalize(params, ori_image)
aug_image = convert_and_normalize(params, aug_image)
return {'ori_images': ori_image, 'aug_images': aug_image}
def _imagenet_parser(params, value, training):
"""ImageNet parser."""
if training:
image_size = params.image_size
else:
if 'eval_image_size' in params:
image_size = max(params.image_size, params.eval_image_size)
else:
image_size = params.image_size
logging.info(f'eval_image_size={image_size}')
keys_to_features = {
'image/encoded': tf.io.FixedLenFeature((), tf.string),
'image/class/label': tf.io.FixedLenFeature([], tf.int64),
}
features = tf.io.parse_single_example(value, keys_to_features)
label = tf.cast(features['image/class/label'], tf.int32) - 1
label = tf.one_hot(label, depth=params.num_classes, dtype=tf.float32)
if training:
image = _decode_and_random_crop(features['image/encoded'], image_size)
image = tf.image.random_flip_left_right(image)
if params.use_augment:
aug = augment.RandAugment(magnitude=params.augment_magnitude)
image = aug.distort(image)
else:
image = _decode_and_center_crop(features['image/encoded'], image_size)
image.set_shape([image_size, image_size, 3])
image = convert_and_normalize(params, image)
return image, label
def _cifar10_parser(params, value, training):
"""Cifar10 parser."""
image_size = params.image_size
value = tf.io.decode_raw(value, tf.uint8)
label = tf.cast(value[0], tf.int32)
label = tf.one_hot(label, depth=params.num_classes, dtype=tf.float32)
image = tf.reshape(value[1:], [3, 32, 32]) # uint8
image = tf.transpose(image, [1, 2, 0])
if image_size != 32:
image = tf.image.resize_bicubic([image], [image_size, image_size])[0]
image.set_shape([image_size, image_size, 3])
if training:
if params.use_augment:
aug = augment.RandAugment(cutout_const=image_size // 8,
translate_const=image_size // 8,
magnitude=params.augment_magnitude)
image = _flip_and_jitter(image, 128)
image = aug.distort(image)
image = augment.cutout(image,
pad_size=image_size // 4,
replace=128)
else:
image = _flip_and_jitter(image, 128)
image = convert_and_normalize(params, image)
return image, label
def _unl_parser(value):
"""ImageNet parser."""
image_size = params.image_size
keys_to_features = {
'image/encoded': tf.io.FixedLenFeature((), tf.string),
}
features = tf.io.parse_single_example(value, keys_to_features)
image = _decode_and_random_crop(features['image/encoded'], image_size)
ori_image = tf.image.random_flip_left_right(image)
ori_image.set_shape([image_size, image_size, 3])
aug_image = tf.image.random_flip_left_right(image)
aug = augment.RandAugment(magnitude=params.augment_magnitude)
aug_image = aug.distort(aug_image)
aug_image.set_shape([image_size, image_size, 3])
ori_image = convert_and_normalize(params, ori_image)
aug_image = convert_and_normalize(params, aug_image)
return {'ori_images': ori_image, 'aug_images': aug_image}
