Support coordinate transformation modes for Resample2d

[Honry]

Issue: DeepLabV3 contains Resize node in ONNX model and ResizeBilinear in TFLite model. It's corresponding op in WebNN is Resample2d. Resize node contains coordinate_transformation_mode=align_corners, ResizeBilinear node contains align_corners=true and half_pixel_centers=false. While WebNN doesn't support such coordinate transformation modes and its default behavior equals to align_corners=false and half_pixel_centers=true.

Compares all the behaviors in different backends and frameworks as follows:

TFLite (half_pixel_centers, align_corners)	ONNX (coordinate_transformation_mode enum)	OpenVINO (coordinate_transformation_mode enum)	DML	WebNN
half_pixel_centers=true	half_pixel	half_pixel	Supported	Default behavior
N/A	pytorch_half_pixel	pytorch_half_pixel	Supported	N/A
half_pixel_centers=false align_corners=false	asymmetric	asymmetric	Supported	N/A
N/A	N/A	tf_half_pixel_for_nn	N/A	N/A
align_corners=true	align_corners	align_corners	Supported	N/A
N/A	tf_crop_and_resize	N/A	Supported	N/A

Note: DML backend can support these modes by calculating its InputPixelOffsets and outputPixelOffsets members.
You can refer to its implementation in ONNXRumtime DML backend.

Open: Can we support these coordinate transformation modes in WebNN? At least for asymmetric, half_pixel_centers and align_corners. Or maybe we could refer to DML by defining InputPixelOffsets and outputPixelOffsets options in Resample2d to implement various coordinate transformation modes.

[fdwr]

Most of these resampling transformations were mistakes (yes, I added them to DML for completeness using a generic transform approach, but I don't recommend propagating them all into perpetuity). The correct approach from computer vision/imaging experts is to resample the center point of the pixels, not asymmetrically (which causes image shifting) nor as point samples (which causes misalignment). That approach (half pixel) is used by libraries like OpenCV, SciPy, Matlab. Recent versions of PyTorch and TF do the right thing too, and it's ONNX's default.

Also see: https://medium.com/hackernoon/how-tensorflows-tf-image-resize-stole-60-days-of-my-life-aba5eb093f35

[inexorabletash]

What do folks think about closing this issue? I think @fdwr has given plenty of evidence here and in #358 and #816 that what WebNN currently provides are what graphics experts recommend, and other coordinate transformation modes are a historical mistake.

[fdwr]

What do folks think about closing this issue? I think @fdwr has given plenty of evidence here and in #358 and #816 that what WebNN currently provides are what graphics experts recommend, and other coordinate transformation modes are a historical mistake.

Now mind you, this is from an imaging perspective, which is the primary use for resampling anyway, but if someone was using resampling in some other domain like audio, or for some other clever use (such as quickly initializing a linearly spaced 1D series of numbers 0 to 100 by interpolating a 2-element tensor into a larger target tensor, like numpy.linspace) then we may want to support modes treating the endpoints as point samples rather than pixels. Though, we can always extend WebNN later when we hit the use case. ⌛... Content to close.

[fdwr]

I'd normally leave it to the issue opener to close an issue, but I see @Honry liked Joshua's comment along with Ningxin too. So closing. 📫➡️📪

[Honry]

Thanks @fdwr, @inexorabletash, +1 to close. 😊