def make_augmentation_config(data_config, num_labels):
"""Returns dataset augmentation configuration."""
random_config = datasets.RandomizedAugmentationConfig(
rotation_probability=data_config.rotation_probability,
smooth_probability=data_config.smooth_probability,
contrast_probability=data_config.contrast_probability,
resize_probability=data_config.resize_probability,
negate_probability=data_config.negate_probability,
roll_probability=data_config.roll_probability,
angle_range=data_config.angle_range,
rotate_by_90=data_config.rotate_by_90)
if data_config.per_label_augmentation:
with tf.variable_scope('augmentations'):
return datasets.AugmentationConfig(children=[
datasets.AugmentationConfig(random_config=random_config)
for _ in range(num_labels)
])
else:
return datasets.AugmentationConfig(random_config=random_config)
def test_full_augmentations(self):
"""Testing image processing with all augmentations."""
aug_config = datasets.AugmentationConfig(
random_config=datasets.RandomizedAugmentationConfig(
rotation_probability=1.0,
smooth_probability=1.0,
contrast_probability=1.0,
resize_probability=1.0,
negate_probability=1.0))
self._test_augmentations(aug_config)
def test_augmentation_config_randomization(self):
"""Testing randomization in AugmentationConfig."""
aug_config = datasets.AugmentationConfig(
random_config=datasets.RandomizedAugmentationConfig())
rand_op = aug_config.randomize_op()
angle = aug_config.angle.value
with self.session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(rand_op)
v1 = sess.run(angle)
v2 = sess.run(angle)
sess.run(rand_op)
v3 = sess.run(angle)
self.assertAlmostEqual(v1, v2)
self.assertNotAlmostEqual(v1, v3)
def test_get_batch(self):
"""Tests image and label generation in the `TaskGenerator`."""
batch_size, image_size = 8, 4
data = self._make_data(batch_size=batch_size, image_size=image_size)
gen = datasets.TaskGenerator(data, num_labels=4, image_size=image_size)
aug_config = datasets.AugmentationConfig(
random_config=datasets.RandomizedAugmentationConfig(
rotation_probability=0.0,
smooth_probability=0.0,
contrast_probability=0.0))
images, labels, classes = gen.get_batch(batch_size=batch_size,
config=aug_config)
with self.session() as sess:
sess.run(tf.global_variables_initializer())
v_images, v_labels, v_classes = sess.run((images, labels, classes))
self.assertEqual(v_images.shape,
(batch_size, image_size, image_size, 1))
self.assertEqual(v_labels.shape, (batch_size, ))
self.assertEqual(v_classes.shape, (batch_size, ))