dev(narugo): add those code (92cfd18a) · Commits · git-mirror / Imgutils

imgutils/preprocess/torchvision.py

+8 −3

Original line number	Diff line number	Diff line
		@@ -446,12 +446,17 @@ if _HAS_TORCHVISION:
		# Resize tensor
		mode = _get_interpolation_str_from_mode(self.interpolation)
		resized = F.interpolate(
		tensor,
		tensor.type(torch.float32),
		size=(new_h, new_w),
		mode=mode,
		align_corners=None if mode == 'nearest' or mode == 'area' else False,
		align_corners=None if mode in {'nearest', 'area'} else False,
		antialias=True if mode in {'bicubic', 'bilinear'} else False,
		)
		if tensor.dtype.is_floating_point:
		resized = torch.clip(resized, min=0.0, max=1.0)
		else:
		resized = torch.clip(resized, min=0, max=255)
		resized = resized.to(tensor.device).type(tensor.dtype)

		# Create padded tensor with background color
		# noinspection PyTypeChecker

test/preprocess/test_torchvision.py

+140 −0

Original line number	Diff line number	Diff line
		@@ -13,6 +13,7 @@ from imgutils.preprocess.torchvision import _get_interpolation_mode, create_torc
		from test.testings import get_testfile

		try:
		import torch
		import torchvision
		except (ImportError, ModuleNotFoundError):
		_TORCHVISION_AVAILABLE = False
		@@ -355,6 +356,145 @@ class TestPreprocessPillow:
		def test_pad_to_size_repr_text(self, json_, repr_text):
		assert repr(create_torchvision_transforms(json_)) == repr_text

		@skipUnless(_TORCHVISION_AVAILABLE, 'Torchvision required.')
		def test_pad_to_size_error(self):
		with pytest.raises(TypeError):
		PadToSize((512, 512))(np.random.randn(1, 3, 384, 384))

		@skipUnless(_TORCHVISION_AVAILABLE, 'Torchvision required.')
		@pytest.mark.parametrize(*tmatrix({
		'json_': [
		# {'type': 'pad_to_size', 'size': [512, 768], 'background_color': 'white', 'interpolation': 'nearest'},
		{'type': 'pad_to_size', 'size': [768, 512], 'background_color': 'gray', 'interpolation': 'lanczos'},
		{'type': 'pad_to_size', 'size': [512, 512], 'background_color': (0, 0, 255), 'interpolation': 'bicubic'},
		{'type': 'pad_to_size', 'size': [384, 512], 'background_color': 'blue', 'interpolation': 'box'},
		{'type': 'pad_to_size', 'size': [512, 512], 'background_color': (255, 255, 255),
		'interpolation': 'bilinear'},
		{'type': 'pad_to_size', 'size': [512, 512], 'background_color': 'blue', 'interpolation': 'lanczos'},
		# {'type': 'pad_to_size', 'size': [768, 512], 'background_color': 'red', 'interpolation': 'nearest'},
		{'type': 'pad_to_size', 'size': [512, 512], 'background_color': 'gray', 'interpolation': 'hamming'},
		{'type': 'pad_to_size', 'size': [384, 384], 'background_color': (255, 0, 0, 128), 'interpolation': 'box'},
		{'type': 'pad_to_size', 'size': [384, 384], 'background_color': (0, 0, 255, 128),
		'interpolation': 'hamming'},
		# {'type': 'pad_to_size', 'size': [512, 512], 'background_color': (0, 0, 255, 128),
		# 'interpolation': 'nearest'},
		{'type': 'pad_to_size', 'size': [384, 384], 'background_color': (0, 0, 255), 'interpolation': 'hamming'},
		{'type': 'pad_to_size', 'size': [384, 384], 'background_color': (255, 255, 255),
		'interpolation': 'hamming'},
		{'type': 'pad_to_size', 'size': [512, 512], 'background_color': (255, 0, 0, 128),
		'interpolation': 'hamming'},
		{'type': 'pad_to_size', 'size': [512, 512], 'background_color': 'white', 'interpolation': 'bilinear'},
		],
		'filename': [
		'nian_640_L.png',
		'nian_640_LA.png',
		'nian_640_RGB.png',
		'nian_640_RGBA.png',
		'png_640_m90.png',
		'png_640.png',
		'dori_640.png',
		]
		}, mode='matrix'))
		def test_pad_to_size_tensor_and_image_float(self, json_, filename, image_diff):
		import torch
		p = create_torchvision_transforms(json_)
		image = load_image(get_testfile(filename), force_background=None, mode=None)

		expected_image = p(image)
		if image.mode == 'L':
		tensor = torch.tensor(np.array(image)[np.newaxis, ...] / 255.0).type(torch.float32)
		else:
		tensor = torch.tensor(np.array(image).transpose(2, 0, 1) / 255.0).type(torch.float32)
		actual_tensor = p(tensor)
		if image.mode == 'L':
		actual_image = Image.fromarray(
		(actual_tensor.numpy()[0] * 255.0).astype(np.uint8),
		mode=image.mode
		)
		else:
		actual_image = Image.fromarray(
		(actual_tensor.numpy().transpose((1, 2, 0)) * 255.0).astype(np.uint8),
		mode=image.mode
		)

		assert tensor.dtype == actual_tensor.dtype
		assert tensor.device == actual_tensor.device

		assert image_diff(
		grid_transparent(expected_image),
		grid_transparent(actual_image),
		throw_exception=False,
		) < 1.5e-2

		@skipUnless(_TORCHVISION_AVAILABLE, 'Torchvision required.')
		@pytest.mark.parametrize(*tmatrix({
		'json_': [
		# {'type': 'pad_to_size', 'size': [512, 768], 'background_color': 'white', 'interpolation': 'nearest'},
		{'type': 'pad_to_size', 'size': [768, 512], 'background_color': 'gray', 'interpolation': 'lanczos'},
		{'type': 'pad_to_size', 'size': [512, 512], 'background_color': (0, 0, 255), 'interpolation': 'bicubic'},
		{'type': 'pad_to_size', 'size': [384, 512], 'background_color': 'blue', 'interpolation': 'box'},
		{'type': 'pad_to_size', 'size': [512, 512], 'background_color': (255, 255, 255),
		'interpolation': 'bilinear'},
		{'type': 'pad_to_size', 'size': [512, 512], 'background_color': 'blue', 'interpolation': 'lanczos'},
		# {'type': 'pad_to_size', 'size': [768, 512], 'background_color': 'red', 'interpolation': 'nearest'},
		{'type': 'pad_to_size', 'size': [512, 512], 'background_color': 'gray', 'interpolation': 'hamming'},
		{'type': 'pad_to_size', 'size': [384, 384], 'background_color': (255, 0, 0, 128), 'interpolation': 'box'},
		{'type': 'pad_to_size', 'size': [384, 384], 'background_color': (0, 0, 255, 128),
		'interpolation': 'hamming'},
		# {'type': 'pad_to_size', 'size': [512, 512], 'background_color': (0, 0, 255, 128),
		# 'interpolation': 'nearest'},
		{'type': 'pad_to_size', 'size': [384, 384], 'background_color': (0, 0, 255), 'interpolation': 'hamming'},
		{'type': 'pad_to_size', 'size': [384, 384], 'background_color': (255, 255, 255),
		'interpolation': 'hamming'},
		{'type': 'pad_to_size', 'size': [512, 512], 'background_color': (255, 0, 0, 128),
		'interpolation': 'hamming'},
		{'type': 'pad_to_size', 'size': [512, 512], 'background_color': 'white', 'interpolation': 'bilinear'},
		],
		'filename': [
		'nian_640_L.png',
		'nian_640_LA.png',
		'nian_640_RGB.png',
		'nian_640_RGBA.png',
		'png_640_m90.png',
		'png_640.png',
		'dori_640.png',
		]
		}, mode='matrix'))
		def test_pad_to_size_tensor_and_image_int8(self, json_, filename, image_diff):
		import torch
		p = create_torchvision_transforms(json_)
		image = load_image(get_testfile(filename), force_background=None, mode=None)

		expected_image = p(image)
		if image.mode == 'L':
		tensor = torch.tensor(np.array(image)[np.newaxis, ...]).type(torch.uint8)
		else:
		tensor = torch.tensor(np.array(image).transpose(2, 0, 1)).type(torch.uint8)
		actual_tensor = p(tensor)
		if image.mode == 'L':
		actual_image = Image.fromarray(
		(actual_tensor.numpy()[0]).astype(np.uint8),
		mode=image.mode
		)
		else:
		actual_image = Image.fromarray(
		(actual_tensor.numpy().transpose((1, 2, 0))).astype(np.uint8),
		mode=image.mode
		)

		assert tensor.dtype == actual_tensor.dtype
		assert tensor.device == actual_tensor.device

		assert image_diff(
		grid_transparent(expected_image),
		grid_transparent(actual_image),
		throw_exception=False,
		) < 1.5e-2

		@skipUnless(_TORCHVISION_AVAILABLE, 'Torchvision required.')
		def test_pad_to_size_dim_error(self):
		import torch
		with pytest.raises(ValueError):
		PadToSize((512, 512))(torch.randn(384, 384))
		with pytest.raises(ValueError):
		PadToSize((512, 512))(torch.randn(1, 1, 1, 384, 384))