def testWrongRandomCropImages(self):
"""Tests that all ValueErrors are triggered for RandomCropImages."""
with tf.Graph().as_default():
input_shape = [32, 32, 3]
batch_size = 4
images = self._CreateRampTestImages(batch_size, input_shape[0],
input_shape[1])
with self.assertRaises(ValueError):
# The input shape is (height, width) but (height, width, channels) is
# required.
image_transformations.RandomCropImages([images], [32, 32],
[20, 64])
with self.assertRaises(ValueError):
# The input shape is (height, width, channel, random) but
# (height, width, channels) is required.
image_transformations.RandomCropImages([images],
[32, 32, 3, 4],
[20, 64])
with self.assertRaises(ValueError):
# The output shape is (height, ) but (height, width) is required.
image_transformations.RandomCropImages([images], [32, 32, 3],
[20])
with self.assertRaises(ValueError):
# The output shape is (height, width, random) but (height, width) is
# required.
image_transformations.RandomCropImages([images], [32, 32, 3],
[20, 32, 64])
def crop_image(img, mode, input_size=(512, 640), target_size=(472, 472)):
"""Takes a crop of the image, either randomly or from the center.
The crop is consistent across all images given in the batch.
Args:
img: 4D image Tensor [batch, height, width, channels].
mode: (ModeKeys) Specifies if this is training, evaluation or prediction.
input_size: (height, width) of input.
target_size: (height, width) of desired crop.
Returns:
img cropped to the desired size, randomly if mode == TRAIN and from the
center otherwise.
"""
if input_size == target_size:
# Don't even bother adding the ops.
return img
input_height, input_width = input_size
input_shape = (input_height, input_width, 3)
target_shape = target_size
if mode == tf.estimator.ModeKeys.TRAIN:
crops = image_transformations.RandomCropImages(
[img], input_shape=input_shape, target_shape=target_shape)[0]
else:
crops = image_transformations.CenterCropImages(
[img], input_shape=input_shape, target_shape=target_shape)[0]
return crops
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 testFaultyRandomCrop(self):
with tf.Graph().as_default():
input_shape = [32, 32, 3]
output_shape = [20, 64]
batch_size = 4
images = self._CreateRampTestImages(batch_size, input_shape[0],
input_shape[1])
with tf.Session() as sess:
with self.assertRaises(tf.errors.InvalidArgumentError):
cropped = image_transformations.RandomCropImages(
[images], input_shape, output_shape)[0]
sess.run(cropped)
def testRandomCrop(self, output_shape):
with tf.Graph().as_default():
input_shape = [32, 32, 3]
batch_size = 4
images = self._CreateRampTestImages(batch_size, input_shape[0],
input_shape[1])
cropped = image_transformations.RandomCropImages([images],
input_shape,
output_shape)[0]
with tf.Session() as sess:
cropped_image = sess.run(cropped)
self.assertAllEqual(cropped_image.shape,
[batch_size] + output_shape + [3])
self.assertEqual(
cropped_image[0, -1, 0, 1] - cropped_image[0, 0, 0, 1],
output_shape[0] - 1)
self.assertEqual(
cropped_image[0, 0, -1, 0] - cropped_image[0, 0, 0, 0],
output_shape[1] - 1)
