def test_conversion_array_to_image(self):
feature_extractor = ImageFeatureExtractionMixin()
array = np.random.randint(0, 256, (16, 32, 3), dtype=np.uint8)
# By default, no rescale (for an array of ints)
image1 = feature_extractor.to_pil_image(array)
self.assertTrue(isinstance(image1, PIL.Image.Image))
self.assertTrue(np.array_equal(np.array(image1), array))
# If the array is channel-first, proper reordering of the channels is done.
image2 = feature_extractor.to_pil_image(array.transpose(2, 0, 1))
self.assertTrue(isinstance(image2, PIL.Image.Image))
self.assertTrue(np.array_equal(np.array(image2), array))
# If the array has floating type, it's rescaled by default.
image3 = feature_extractor.to_pil_image(
array.astype(np.float32) / 255.0)
self.assertTrue(isinstance(image3, PIL.Image.Image))
self.assertTrue(np.array_equal(np.array(image3), array))
# You can override the default to rescale.
image4 = feature_extractor.to_pil_image(array.astype(np.float32),
rescale=False)
self.assertTrue(isinstance(image4, PIL.Image.Image))
self.assertTrue(np.array_equal(np.array(image4), array))
# And with floats + channel first.
image5 = feature_extractor.to_pil_image(
array.transpose(2, 0, 1).astype(np.float32) / 255.0)
self.assertTrue(isinstance(image5, PIL.Image.Image))
self.assertTrue(np.array_equal(np.array(image5), array))
def test_conversion_tensor_to_image(self):
feature_extractor = ImageFeatureExtractionMixin()
tensor = torch.randint(0, 256, (16, 32, 3))
array = tensor.numpy()
# By default, no rescale (for a tensor of ints)
image1 = feature_extractor.to_pil_image(tensor)
self.assertTrue(isinstance(image1, PIL.Image.Image))
self.assertTrue(np.array_equal(np.array(image1), array))
# If the tensor is channel-first, proper reordering of the channels is done.
image2 = feature_extractor.to_pil_image(tensor.permute(2, 0, 1))
self.assertTrue(isinstance(image2, PIL.Image.Image))
self.assertTrue(np.array_equal(np.array(image2), array))
# If the tensor has floating type, it's rescaled by default.
image3 = feature_extractor.to_pil_image(tensor.float() / 255.0)
self.assertTrue(isinstance(image3, PIL.Image.Image))
self.assertTrue(np.array_equal(np.array(image3), array))
# You can override the default to rescale.
image4 = feature_extractor.to_pil_image(tensor.float(), rescale=False)
self.assertTrue(isinstance(image4, PIL.Image.Image))
self.assertTrue(np.array_equal(np.array(image4), array))
# And with floats + channel first.
image5 = feature_extractor.to_pil_image(
tensor.permute(2, 0, 1).float() / 255.0)
self.assertTrue(isinstance(image5, PIL.Image.Image))
self.assertTrue(np.array_equal(np.array(image5), array))
def test_conversion_image_to_image(self):
feature_extractor = ImageFeatureExtractionMixin()
image = get_random_image(16, 32)
# On an image, `to_pil_image1` is a noop.
image1 = feature_extractor.to_pil_image(image)
self.assertTrue(isinstance(image, PIL.Image.Image))
self.assertTrue(np.array_equal(np.array(image), np.array(image1)))