def parser(self, serialized_example):
"""Parses a single tf.Example into image and label tensors."""
if self._test_small_sample:
image_raw = serialized_example
image_preprocessing = serialized_example
label = tf.constant(0, tf.int32)
else:
features = tf.parse_single_example(
serialized_example,
features={
'image/encoded':
tf.FixedLenFeature([], tf.string, default_value=''),
'image/class/label': (tf.FixedLenFeature([], tf.int64)),
})
image_raw = tf.reshape(features['image/encoded'], shape=[])
image_preprocessing = tf.image.decode_image(image_raw,
dtype=tf.float32)
image_preprocessing = preprocess_image(image=image_preprocessing,
image_size=224,
is_training=False)
if self._dataset_name == 'imagenet':
# Subtract one so that labels are in [0, 1000).
label = tf.cast(tf.reshape(features['image/class/label'],
shape=[]),
dtype=tf.int32) - 1
else:
label = tf.cast(tf.reshape(features['image/class/label'],
shape=[]),
dtype=tf.int32)
return (image_raw, image_preprocessing, label)
def parser(self, serialized_example):
"""Parses a single tf.Example into image and label tensors."""
if self.test_small_sample:
image = serialized_example
label = tf.constant(0, tf.int32)
else:
features = tf.parse_single_example(
serialized_example,
features={
'raw_image':
tf.FixedLenFeature((), tf.string, default_value=''),
'height':
tf.FixedLenFeature([], dtype=tf.int64, default_value=-1),
'width':
tf.FixedLenFeature([], dtype=tf.int64, default_value=-1),
self.saliency_method:
tf.VarLenFeature(tf.float32),
'label':
tf.FixedLenFeature([], dtype=tf.int64, default_value=-1),
'prediction_class':
tf.FixedLenFeature([], dtype=tf.int64, default_value=-1)
})
image = tf.image.decode_image(features['raw_image'], 3)
image = tf.image.convert_image_dtype(image, dtype=tf.float32)
saliency_heatmap = tf.expand_dims(features[self.saliency_method].values,
0)
saliency_heatmap = tf.reshape(saliency_heatmap, IMAGE_DIMS)
if self.transformation in ['modified_image', 'random_baseline']:
# we apply test_time pre-processing to the raw image before modifying
# according to the estimator ranking.
image_preprocess = preprocess_image(
image, image_size=IMAGE_DIMS[0], is_training=False)
if self.transformation == 'modified_image':
tf.logging.info('Computing feature importance estimate now...')
image = compute_feature_ranking(
input_image=image_preprocess,
saliency_map=saliency_heatmap,
threshold=self.threshold,
global_mean=self.global_mean,
rescale_heatmap=True,
keep_information=self.keep_information,
use_squared_value=self.use_squared_value)
if self.transformation == 'random_baseline':
tf.logging.info('generating a random baseline')
image = random_ranking(
input_image=image_preprocess,
global_mean=self.global_mean,
threshold=self.threshold,
keep_information=self.keep_information)
if self.mode == 'train':
is_training = True
else:
is_training = False
if self.transformation in ['random_baseline', 'modified_image']:
tf.logging.info('starting pre-processing for training/eval')
image = preprocess_image(
image, image_size=IMAGE_DIMS[0], is_training=is_training)
if self.transformation == 'raw_image':
tf.logging.info('starting pre-processing for training/eval')
image = preprocess_image(
image, image_size=IMAGE_DIMS[0], is_training=is_training)
label = tf.cast(tf.reshape(features['label'], shape=[]), dtype=tf.int32)
return image, label