def ScaleTransform_img_gt(imgs, *args) -> Tuple[Any, Any]:
"""
Given the input array, return the expected output array and shape after
applying the resize transformation.
Args:
imgs (array): image(s) array before the transform.
args (list): list of arguments. Details can be found in test case.
Returns:
img (array): expected output array after apply the transformation.
None means does not have expected output array for sanity check.
(list): expected shape of the output array. None means does not have
expected output shape for sanity check.
"""
h, w, new_h, new_w, interp = args
float_tensor = to_float_tensor(imgs)
if interp == "nearest":
if float_tensor.dim() == 3:
float_tensor = torch._C._nn.upsample_nearest1d(
float_tensor, (new_h, new_w))
elif float_tensor.dim() == 4:
float_tensor = torch._C._nn.upsample_nearest2d(
float_tensor, (new_h, new_w))
elif float_tensor.dim() == 5:
float_tensor = torch._C._nn.upsample_nearest3d(
float_tensor, (new_h, new_w))
else:
return None, None
elif interp == "bilinear":
if float_tensor.dim() == 4:
float_tensor = torch._C._nn.upsample_bilinear2d(
float_tensor, (new_h, new_w), False)
else:
return None, None
numpy_tensor = to_numpy(float_tensor, imgs.shape, imgs.dtype)
return numpy_tensor, numpy_tensor.shape
def apply_image(self, img: np.ndarray, interp: str = None) -> np.ndarray:
"""
Resize the image(s).
Args:
img (ndarray): of shape NxHxWxC, or HxWxC or HxW. The array can be
of type uint8 in range [0, 255], or floating point in range
[0, 1] or [0, 255].
interp (str): interpolation methods. Options includes `nearest`, `linear`
(3D-only), `bilinear`, `bicubic` (4D-only), and `area`.
Details can be found in:
https://pytorch.org/docs/stable/nn.functional.html
Returns:
ndarray: resized image(s).
"""
if len(img.shape) == 4:
h, w = img.shape[1:3]
elif len(img.shape) in (2, 3):
h, w = img.shape[:2]
else:
raise ("Unsupported input with shape of {}".format(img.shape))
assert (self.h == h and self.w
== w), "Input size mismatch h w {}:{} -> {}:{}".format(
self.h, self.w, h, w)
interp_method = interp if interp is not None else self.interp
# Option of align_corners is only supported for linear, bilinear,
# and bicubic.
if interp_method in ["linear", "bilinear", "bicubic"]:
align_corners = False
else:
align_corners = None
# note: this is quite slow for int8 images because torch does not
# support it https://github.com/pytorch/pytorch/issues/5580
float_tensor = torch.nn.functional.interpolate(
to_float_tensor(img),
size=(self.new_h, self.new_w),
mode=interp_method,
align_corners=align_corners,
)
return to_numpy(float_tensor, img.shape, img.dtype)
def test_convert(self) -> None:
N, C, H, W = 4, 64, 14, 14
np.random.seed(0)
array_HW: np.ndarray = np.random.rand(H, W)
array_HWC: np.ndarray = np.random.rand(H, W, C)
array_NHWC: np.ndarray = np.random.rand(N, H, W, C)
arrays = [
array_HW,
(array_HW * 255).astype(np.uint8),
array_HWC,
(array_HWC * 255).astype(np.uint8),
array_NHWC,
(array_NHWC * 255).astype(np.uint8),
]
for array in arrays:
converted_tensor = to_float_tensor(array)
converted_array = to_numpy(converted_tensor, array.shape,
array.dtype)
self.assertTrue(np.allclose(array, converted_array))
def ScaleTransform_seg_gt(seg, *args) -> Tuple[np.ndarray, list]:
"""
Given the input segmentation, return the expected output array and shape
after applying the blend transformation.
Args:
seg (array): segmentation before the transform.
args (list): list of arguments. Details can be found in test case.
Returns:
seg (array): expected output segmentation after apply the
transformation.
(list): expected shape of the output array.
"""
h, w, new_h, new_w = args
float_tensor = torch.nn.functional.interpolate(
to_float_tensor(seg),
size=(new_h, new_w),
mode="nearest",
align_corners=None,
)
numpy_tensor = to_numpy(float_tensor, seg.shape, seg.dtype)
return numpy_tensor, numpy_tensor.shape
def apply_image(self, img: np.ndarray, interp: str = None) -> np.ndarray:
"""
Apply grid sampling on the image(s).
Args:
img (ndarray): of shape NxHxWxC, or HxWxC or HxW. The array can be
of type uint8 in range [0, 255], or floating point in range
[0, 1] or [0, 255].
interp (str): interpolation methods. Options include `nearest` and
`bilinear`.
Returns:
ndarray: grid sampled image(s).
"""
interp_method = interp if interp is not None else self.interp
float_tensor = torch.nn.functional.grid_sample(
to_float_tensor(img), # NxHxWxC -> NxCxHxW.
torch.from_numpy(self.grid),
mode=interp_method,
padding_mode="border",
align_corners=False,
)
return to_numpy(float_tensor, img.shape, img.dtype)