def _preprocess_fn(self, features, labels, mode):
"""The preprocessing function which will be executed prior to the model_fn.
Args:
features: The input features extracted from a single example in our
in_feature_specification format.
labels: (Optional) The input labels extracted from a single example in our
in_label_specification format.
mode: (ModeKeys) Specifies if this is training, evaluation or prediction.
Returns:
features: The preprocessed features, potentially adding
additional tensors derived from the input features.
labels: (Optional) The preprocessed labels, potentially
adding additional tensors derived from the input features and labels.
"""
if mode == TRAIN:
image = image_transformations.RandomCropImages(
[features.state.image], INPUT_SHAPE, TARGET_SHAPE)[0]
else:
image = image_transformations.CenterCropImages(
[features.state.image], INPUT_SHAPE, TARGET_SHAPE)[0]
image = tf.image.convert_image_dtype(image, tf.float32)
if mode == TRAIN:
image = (image_transformations.ApplyPhotometricImageDistortions(
[image])[0])
features.state.image = image
return features, labels
def testDeterminism(self):
"""Get deterministic distortion by setting global seed and op-level seed."""
input_shape = [32, 32, 3]
gin.bind_parameter('tf.random.uniform.seed', 0)
gin.bind_parameter('tf.random.normal.seed', 0)
with tf.Graph().as_default():
tf.set_random_seed(123)
batch_size = 4
images = self._CreateRampTestImages(batch_size, input_shape[0],
input_shape[1])
tensor_list = []
for i in range(batch_size):
tensor_list.append(images[i])
distorted_1 = image_transformations.ApplyPhotometricImageDistortions(
tensor_list, random_noise_apply_probability=1.0)
distorted_2 = image_transformations.ApplyPhotometricImageDistortions(
tensor_list, random_noise_apply_probability=1.0)
with tf.Session() as sess:
images_1 = sess.run(distorted_1)
images_2 = sess.run(distorted_2)
for image_1, image_2 in zip(images_1, images_2):
self.assertAllClose(image_1, image_2)
def maybe_distort_image_batch(images, mode):
"""Applies data augmentation to given images.
Args:
images: 4D Tensor (batch images) or 5D Tensor (batch of image sequences).
mode: (ModeKeys) Specifies if this is training, evaluation or prediction.
Returns:
Distorted images. Image distortion is identical for every image in the
batch.
"""
if mode == tf.estimator.ModeKeys.TRAIN:
images = image_transformations.ApplyPhotometricImageDistortions(
[images])[0]
return images
def testPhotometricImageDistortions(self, input_shape):
input_shape = input_shape + [3]
with tf.Graph().as_default():
batch_size = 4
images = self._CreateRampTestImages(batch_size, input_shape[0],
input_shape[1])
tensor_list = []
for i in range(batch_size):
tensor_list.append(images[i])
distorted = image_transformations.ApplyPhotometricImageDistortions(
tensor_list, random_noise_apply_probability=1.0)
delta = tf.reduce_sum(tf.square(images - distorted))
with tf.Session() as sess:
images_delta = sess.run(delta)
# Check if any distortion applied.
self.assertGreater(images_delta, 0)
