failed to convert SAFMN

[hooke007]

Information:

• Chainner version: Alpha v0.21.2
• OS: win11

Description
failed to convert SAFMN (pth → onnx)

Logs
main.log
renderer.log

[hooke007]

Maybe related https://github.com/sunny2109/SAFMN/tree/main/scripts/to_onnx

[RunDevelopment]

Maybe related https://github.com/sunny2109/SAFMN/tree/main/scripts/to_onnx
[from the readme] "The input shape of the current code is [1, 3, 640, 960]. If you want to support inputs with other resolution, you must adapt the code yourself."

Are they telling us that SAFMN ONNX models only support fixed resolutions? The error message in main.log seems to confirm this: "Unsupported: ONNX export of operator adaptive pooling, since output_size is not constant."

If so, then SAFMN ONNX models are pretty much useless, since they only support input images of a fixed resolution.

[hooke007]

The author said we could use maxpool instead.

[RunDevelopment]

I tried that and it doesn't work. I just get a different error message:

  File "C:\Users\micha\Git\chaiNNer\backend\src\packages\chaiNNer_pytorch\pytorch\utility\convert_to_onnx.py", line 89, in convert_to_onnx_node
    onnx_model_bytes = convert_to_onnx_impl(
  File "C:\Users\micha\Git\chaiNNer\backend\src\nodes\impl\pytorch\convert_to_onnx_impl.py", line 56, in convert_to_onnx_impl
    torch.onnx.export(
  File "C:\Python38\lib\site-packages\torch\onnx\utils.py", line 516, in export
    _export(
  File "C:\Python38\lib\site-packages\torch\onnx\utils.py", line 1596, in _export
    graph, params_dict, torch_out = _model_to_graph(
  File "C:\Python38\lib\site-packages\torch\onnx\utils.py", line 1139, in _model_to_graph
    graph = _optimize_graph(
  File "C:\Python38\lib\site-packages\torch\onnx\utils.py", line 677, in _optimize_graph
    graph = _C._jit_pass_onnx(graph, operator_export_type)
  File "C:\Python38\lib\site-packages\torch\onnx\utils.py", line 1940, in _run_symbolic_function
    return symbolic_fn(graph_context, *inputs, **attrs)
  File "C:\Python38\lib\site-packages\torch\onnx\symbolic_helper.py", line 395, in wrapper
    return fn(g, *args, **kwargs)
  File "C:\Python38\lib\site-packages\torch\onnx\symbolic_helper.py", line 289, in wrapper
    args = [
  File "C:\Python38\lib\site-packages\torch\onnx\symbolic_helper.py", line 290, in <listcomp>
    _parse_arg(arg, arg_desc, arg_name, fn_name)  # type: ignore[assignment]
  File "C:\Python38\lib\site-packages\torch\onnx\symbolic_helper.py", line 107, in _parse_arg
    raise errors.SymbolicValueError(
torch.onnx.errors.SymbolicValueError: Failed to export a node '%143 : Long(requires_grad=0, device=cpu) = onnx::Cast[to=7](%142), scope: nodes.impl.pytorch.convert_to_onnx_impl.convert_to_onnx_impl.<locals>.FakeModel:: # C:\Python38\lib\site-packages\torch\_tensor.py:942:0
' (in list node %147 : int[] = prim::ListConstruct(%143, %146), scope: nodes.impl.pytorch.convert_to_onnx_impl.convert_to_onnx_impl.<locals>.FakeModel::
) because it is not constant. Please try to make things (e.g. kernel sizes) static if possible.  [Caused by the value '147 defined in (%147 : int[] = prim::ListConstruct(%143, %146), scope: nodes.impl.pytorch.convert_to_onnx_impl.convert_to_onnx_impl.<locals>.FakeModel::
)' (type 'List[int]') in the TorchScript graph. The containing node has kind 'prim::ListConstruct'.]

    Inputs:
        #0: 143 defined in (%143 : Long(requires_grad=0, device=cpu) = onnx::Cast[to=7](%142), scope: nodes.impl.pytorch.convert_to_onnx_impl.convert_to_onnx_impl.<locals>.FakeModel:: # C:\Python38\lib\site-packages\torch\_tensor.py:942:0
    )  (type 'Tensor')
        #1: 146 defined in (%146 : Long(requires_grad=0, device=cpu) = onnx::Cast[to=7](%145), scope: nodes.impl.pytorch.convert_to_onnx_impl.convert_to_onnx_impl.<locals>.FakeModel:: # C:\Python38\lib\site-packages\torch\_tensor.py:942:0
    )  (type 'Tensor')
    Outputs:
        #0: 147 defined in (%147 : int[] = prim::ListConstruct(%143, %146), scope: nodes.impl.pytorch.convert_to_onnx_impl.convert_to_onnx_impl.<locals>.FakeModel::
    )  (type 'List[int]')

Here's what I changed:

@@ -126,7 +126,7 @@ class SAFM(nn.Module):
         for i in range(self.n_levels):
             if i > 0:
                 p_size = (h // 2**i, w // 2**i)
-                s = F.adaptive_max_pool2d(xc[i], p_size)
+                s = F.max_pool2d(xc[i], p_size)
                 s = self.mfr[i](s)
                 s = F.interpolate(s, size=(h, w), mode="nearest")
             else: