2D linear upsample with align_corners=False doesn't match PyTorch.

[ysiraichi]

🐛 Bug

Running a 2D linear upsample yields in different results when comparing with PyTorch (both CPU and CUDA). That, however, only happens when align_corners=False.

def run(device):
    a = torch.arange(9, dtype=torch.float32, device=device).view(1, 1, 3, 3)
    m = Upsample(a, scale_factor=1.5, mode="bilinear", align_corners=False)
    return m(a)

>>> run("cpu")
tensor([[[[0.0000, 0.5000, 1.1667, 1.8333],
          [1.5000, 2.0000, 2.6667, 3.3333],
          [3.5000, 4.0000, 4.6667, 5.3333],
          [5.5000, 6.0000, 6.6667, 7.3333]]]],
       grad_fn=<UpsampleBilinear2DBackward0>)

>>> run("xla")
tensor([[[[0.0000, 0.6250, 1.3750, 2.0000],
          [1.8750, 2.5000, 3.2500, 3.8750],
          [4.1250, 4.7500, 5.5000, 6.1250],
          [6.0000, 6.6250, 7.3750, 8.0000]]]], device='xla:0',
       grad_fn=<UpsampleBilinear2DBackward0>)

PyTorch/XLA result seems to align with tensorflow (makes sense, since we took it from there), and OpenCV. However, PyTorch does seem to be the different one here.

Environment

• Reproducible on XLA backend [CPU/TPU/CUDA]: CUDA
• torch_xla version: 3808df0

Additional context

Found this when lowering the backwards (issue #8754). I tried taking the lowering from the same place and it looks like the gradient also matches only when align_corners=True.

This might not affect XLA:TPU, since it uses a custom lowering.

[ysiraichi]

Question: are all of these correct results of a different interpretation?

[ysiraichi]

Apparently, this mismatch is due to PyTorch using the scale_factor, while tensorflow (and OpenCV) uses the ratio between output and input sizes. If we call Upsample(..., recompute_scale_factor=True) we get the same result.

Thank you @vfdev-5 for helping me with this.

[ysiraichi]

I'm opening this issue, since with recompute_scale_factor=False, we still get different results.