[torchlib] Implement various upsample functions

[justinchuby]

• aten_upsample_bicubic2d AddOp(upsample_bicubic2d) | feat(torchlib) #1208
• aten__upsample_bicubic2d_aa Add Op (_upsample_bilinear2d_aa, _upsample_bicubic2d_aa) | feat(torchlib) #1259
• aten__upsample_bilinear2d_aa Add Op (_upsample_bilinear2d_aa, _upsample_bicubic2d_aa) | feat(torchlib) #1259
• aten_upsample_bilinear2d.vec Add 'aten_upsample_bilinear2d_vec' for unet #1249
• aten_upsample_linear1d Add op(upsample linear1d) | feat(torchlib) #1245
• aten_upsample_nearest1d Add op(upsample nearst1/2/3d) | feat(torchlib) #1246
• aten_upsample_nearest3d Add op(upsample nearst1/2/3d) | feat(torchlib) #1246
• aten_upsample_trilinear3d AddOp(upsample_trilinear3d) | feat(torchlib) #1260

Reference:

onnxscript/onnxscript/function_libs/torch_lib/ops/nn.py

Lines 2121 to 2137 in 24d77d9

	@torch_op("aten::upsample_bilinear2d", trace_only=True)
	def aten_upsample_bilinear2d(
	self: TReal,
	output_size: Optional[INT64] = None,
	scales_h: Optional[float] = None,
	scales_w: Optional[float] = None,
	align_corners: bool = True, # pylint: disable=unused-argument
	) -> TReal:
	"""upsample_bilinear2d(Tensor self, SymInt[2] output_size, bool align_corners, float? scales_h=None, float? scales_w=None) -> Tensor"""
	if output_size is not None:
	result = _aten_upsample_bilinear2d_output_size(self, output_size)
	else:
	assert scales_h is not None
	assert scales_h == scales_w
	result = _aten_upsample_bilinear2d_scales(self, scales_h, scales_w)
	return result

[justinchuby]

• Decomp various private upsample ops pytorch/pytorch#116706

[BowenBao]

How are we validating this works e2e for pytorch export?

I think there might be an issue inside pytorch that decompositions are forced for these ops, similar to the einsum situation. pytorch/pytorch#116684

[BowenBao]

align_corners: bool = True, # pylint: disable=unused-argument

Why is align_corners an unused argument?

[BowenBao]

Regarding Resize.nearest_mode. In torchscript exporter it was set to "floor". In torchlib it is left as default "round_prefer_floor"

Is the change intentional?

[justinchuby]

We do need to implement them and match the behaviors @xiaowuhu please correct me if I’m missing anything