def test_mask_applied():
test_image = tf.ones([10, 40, 40, 1], dtype=np.uint8)
result_image = random_cutout(test_image, 20, seed=1234)
total_expected_masked_count = 20 * 20 * test_image.shape[0]
np.testing.assert_equal(
np.sum(result_image) + total_expected_masked_count, np.sum(test_image))
def test_batch_size():
test_image = tf.random.uniform([10, 40, 40, 1],
dtype=np.float32,
seed=1234)
result_image = random_cutout(test_image, 20, seed=1234)
np.testing.assert_allclose(tf.shape(result_image), [10, 40, 40, 1])
means = np.mean(result_image, axis=(1, 2, 3))
np.testing.assert_allclose(len(set(means)), 10)
def test_different_channels():
for channel in [0, 1, 3, 4]:
test_image = tf.ones([1, 40, 40, channel], dtype=np.uint8)
cutout_area = tf.zeros([4, 4], dtype=np.uint8)
cutout_area = tf.pad(cutout_area, ((0, 36), (0, 36)), constant_values=1)
expect_image = to_4D_image(cutout_area)
expect_image = tf.tile(expect_image, [1, 1, 1, channel])
result_image = random_cutout(test_image, 20, seed=1234)
np.testing.assert_allclose(tf.shape(result_image), tf.shape(expect_image))
def test_channel_first():
test_image = tf.ones([10, 1, 40, 40], dtype=np.uint8)
cutout_area = tf.zeros([4, 4], dtype=np.uint8)
cutout_area = tf.pad(cutout_area, ((0, 36), (0, 36)), constant_values=1)
expect_image = tf.expand_dims(cutout_area, 0)
expect_image = tf.expand_dims(expect_image, 0)
expect_image = tf.tile(expect_image, [10, 1, 1, 1])
result_image = random_cutout(test_image,
20,
seed=1234,
data_format="channels_first")
np.testing.assert_allclose(tf.shape(result_image), tf.shape(expect_image))
def test_mask_larger_than_image():
test_image = tf.ones([10, 40, 40, 1], dtype=np.uint8)
result_image = random_cutout(test_image, 60, seed=1234)
np.testing.assert_equal(np.sum(result_image), 0)
