def _example_parser(example: Dict[str, tf.Tensor]) -> Dict[str, tf.Tensor]:
"""Preprocesses Places-365 image Tensors using inception_preprocessing."""
# `preprocess_image` returns images in [-1, 1].
image = inception_preprocessing.preprocess_image(
example['image'],
height=224,
width=224,
is_training=self._is_training)
# Rescale to [0, 1].
image = (image + 1.0) / 2.0
label = tf.cast(example['label'], tf.int32)
return {'features': image, 'labels': label}
def _example_parser(
example: Dict[str, tf.Tensor]) -> Dict[str, tf.Tensor]:
"""Preprocesses ImageNet image Tensors."""
per_example_step_seed = tf.random.experimental.stateless_fold_in(
self._seed, example[self._enumerate_id_key])
if self._preprocessing_type == 'inception':
# `inception_preprocessing.preprocess_image` returns images in [-1, 1].
image = inception_preprocessing.preprocess_image(
example['image'],
height=self._image_size,
width=self._image_size,
seed=per_example_step_seed,
is_training=self._is_training)
# Rescale to [0, 1].
image = (image + 1.0) / 2.0
elif self._preprocessing_type == 'resnet':
# `resnet_preprocessing.preprocess_image` returns images in [0, 1].
image = resnet_preprocessing.preprocess_image(
image_bytes=example['image'],
is_training=self._is_training,
use_bfloat16=self._use_bfloat16,
image_size=self._image_size,
seed=per_example_step_seed,
resize_method=self._resnet_preprocessing_resize_method)
else:
raise ValueError(
'Invalid preprocessing type, must be one of "inception" or '
'"resnet", received {}.'.format(self._preprocessing_type))
if self._normalize_input:
image = (tf.cast(image, tf.float32) -
IMAGENET_MEAN) / IMAGENET_STDDEV
if self._use_bfloat16:
image = tf.cast(image, tf.bfloat16)
# Note that labels are always float32, even when images are bfloat16.
if self._one_hot:
label = tf.one_hot(example['label'], 1000, dtype=tf.float32)
else:
label = tf.cast(example['label'], tf.float32)
parsed_example = {
'features': image,
'labels': label,
}
if 'file_name' in example:
parsed_example['file_name'] = example['file_name']
return parsed_example
def _example_parser(
example: Dict[str, tf.Tensor]) -> Dict[str, tf.Tensor]:
"""Preprocesses Places-365 image Tensors using inception_preprocessing."""
per_example_step_seed = tf.random.experimental.stateless_fold_in(
self._seed, example[self._enumerate_id_key])
# `preprocess_image` returns images in [-1, 1].
image = inception_preprocessing.preprocess_image(
example['image'],
height=224,
width=224,
seed=per_example_step_seed,
is_training=self._is_training)
# Rescale to [0, 1].
image = (image + 1.0) / 2.0
label = tf.cast(example['label'], tf.int32)
return {'features': image, 'labels': label}
def _example_parser(example: Dict[str, tf.Tensor]) -> Dict[str, tf.Tensor]:
"""Preprocesses ImageNet image Tensors using inception_preprocessing."""
# `preprocess_image` returns images in [-1, 1].
image = inception_preprocessing.preprocess_image(
example['image'],
height=224,
width=224,
is_training=self._is_training(split))
# Rescale to [0, 1].
image = (image + 1.0) / 2.0
label = tf.cast(example['label'], tf.int32)
parsed_example = {
'features': image,
'labels': label,
}
if 'file_name' in example:
parsed_example['file_name'] = example['file_name']
return parsed_example