def test_conversion_image_to_array(self):
feature_extractor = ImageFeatureExtractionMixin()
image = get_random_image(16, 32)
# Conversion with defaults (rescale + channel first)
array1 = feature_extractor.to_numpy_array(image)
self.assertTrue(array1.dtype, np.float32)
self.assertEqual(array1.shape, (3, 16, 32))
# Conversion with rescale and not channel first
array2 = feature_extractor.to_numpy_array(image, channel_first=False)
self.assertTrue(array2.dtype, np.float32)
self.assertEqual(array2.shape, (16, 32, 3))
self.assertTrue(np.array_equal(array1, array2.transpose(2, 0, 1)))
# Conversion with no rescale and channel first
array3 = feature_extractor.to_numpy_array(image, rescale=False)
self.assertTrue(array3.dtype, np.uint8)
self.assertEqual(array3.shape, (3, 16, 32))
self.assertTrue(
np.array_equal(array1,
array3.astype(np.float32) / 255.0))
# Conversion with no rescale and not channel first
array4 = feature_extractor.to_numpy_array(image,
rescale=False,
channel_first=False)
self.assertTrue(array4.dtype, np.uint8)
self.assertEqual(array4.shape, (16, 32, 3))
self.assertTrue(
np.array_equal(array2,
array4.astype(np.float32) / 255.0))
def test_conversion_array_to_array(self):
feature_extractor = ImageFeatureExtractionMixin()
array = np.random.randint(0, 256, (16, 32, 3), dtype=np.uint8)
# By default, rescale (for an array of ints) and channel permute
array1 = feature_extractor.to_numpy_array(array)
self.assertTrue(array1.dtype, np.float32)
self.assertEqual(array1.shape, (3, 16, 32))
self.assertTrue(np.array_equal(array1, array.transpose(2, 0, 1).astype(np.float32) / 255.0))
# Same with no permute
array2 = feature_extractor.to_numpy_array(array, channel_first=False)
self.assertTrue(array2.dtype, np.float32)
self.assertEqual(array2.shape, (16, 32, 3))
self.assertTrue(np.array_equal(array2, array.astype(np.float32) / 255.0))
# Force rescale to False
array3 = feature_extractor.to_numpy_array(array, rescale=False)
self.assertTrue(array3.dtype, np.uint8)
self.assertEqual(array3.shape, (3, 16, 32))
self.assertTrue(np.array_equal(array3, array.transpose(2, 0, 1)))
# Force rescale to False and no channel permute
array4 = feature_extractor.to_numpy_array(array, rescale=False, channel_first=False)
self.assertTrue(array4.dtype, np.uint8)
self.assertEqual(array4.shape, (16, 32, 3))
self.assertTrue(np.array_equal(array4, array))
# Now test the default rescale for a float array (defaults to False)
array5 = feature_extractor.to_numpy_array(array2)
self.assertTrue(array5.dtype, np.float32)
self.assertEqual(array5.shape, (3, 16, 32))
self.assertTrue(np.array_equal(array5, array1))
def test_center_crop_array(self):
feature_extractor = ImageFeatureExtractionMixin()
image = get_random_image(16, 32)
array = feature_extractor.to_numpy_array(image)
# Test various crop sizes: bigger on all dimensions, on one of the dimensions only and on both dimensions.
crop_sizes = [8, (8, 64), 20, (32, 64)]
for size in crop_sizes:
cropped_array = feature_extractor.center_crop(array, size)
self.assertTrue(isinstance(cropped_array, np.ndarray))
expected_size = (size, size) if isinstance(size, int) else size
self.assertEqual(cropped_array.shape[-2:], expected_size)
# Check result is consistent with PIL.Image.crop
cropped_image = feature_extractor.center_crop(image, size)
self.assertTrue(
np.array_equal(
cropped_array,
feature_extractor.to_numpy_array(cropped_image)))