Op Support: torchvision.roi_align.default

[shivadbhavsar]

No description provided.

[shivadbhavsar]

Here is some sample test code to help with the acc_op implementation

import torch
from torchvision.ops import RoIAlign

roi_mod = RoIAlign(output_size=[2, 2], spatial_scale=1.0, sampling_ratio=-1, aligned=False)

inp = torch.randn(2, 3, 4, 4)
boxes = torch.tensor([[0, 0, 0, 0.5, 0.7]])

torch_out = roi_mod(inp, boxes)
print(torch_out)


import torch_migraphx.fx.tracer.acc_tracer.acc_tracer as acc_tracer
from torch_migraphx.fx.fx2mgx import MGXInterpreter
from torch_migraphx.fx.mgx_module import MGXModule
def convert_to_mgx(mod, inp):
    traced = acc_tracer.trace(mod.eval(), inp)
    traced.graph.print_tabular()
    interp = MGXInterpreter(traced, inp)
    interp.run()
    return MGXModule(interp.program, interp.get_input_names())

mgx_mod = convert_to_mgx(roi_mod, [inp, boxes])

mgx_out = mgx_mod(inp, boxes)
print(mgx_out)

Currently, this will fail with RuntimeError: Conversion of function torchvision.ops.roi_align.roi_align not supported.. Do one of the following:
a. Create a acc_op wrapper for torchvision.ops.roi_align.roi_align and implement a converter for it in acc_ops_converters.
b. Dont bother with acc wrapper, just create a torchvision_converters file and directly implement a converter for torchvision.ops.roi_align.roi_align

[bpickrel]

Current status: Converter has been coded for this op, branch aten_roi_align but I learned that the MigraphX implementation of the ROIAlign op. doesn't give correct results. The op. has two coordinate transformation modes 'half_pixel' and 'output_half_pixel', and only the legacy mode 'output_half_pixel' gives correct results.

To debug the 'half_pixel' mode calculations, I'm comparing them in detail with the implementation found in Onnxruntime. This was used as the model for the MigraphX version but the code is just different enough to make checking a big job.

There isn't any PR for this task at this time.

[bpickrel]

Confirmed that the MIGraphX ROIAlign op is buggy and created PR 3482 to fix it. This PR will be a prerequisite for the torch-migraphx code work.

The Migraphx implementation of ROIAlign did not have a test for half_pixel case at all, and contains bugs that weren't revealed by the existing test such as inverting X and Y coordinates in the input array. Thus, I keep finding more problems as I write more tests.

[bpickrel]

PR 3482 is now coded and in review. The torch-migraphx changes, which depend on the MIGraphx PR, are still in progress.