Hello

[molla1998]

Hi

TResNet can be easily converted to onnx. i have converted them in the past. follow the steps listed below, don't reinvent the wheel, or do other stuff

1) update to latest ASL repo 2) edit the class "SpaceToDepthModule" to be:

class SpaceToDepthModule(nn.Module):
    def __init__(self, remove_model_jit=True): # True instead of false. JIT function cannot be converted.
        super().__init__()
        if not remove_model_jit:
            self.op = SpaceToDepthJit()
        else:
            self.op = SpaceToDepth()

    def forward(self, x):
        return self.op(x)

(you can pass the 'remove_model_jit' as an argument for something more modular) 3) After you create the model and load the proper weights, do:

model=InplacABN_to_ABN(model)

This will transfrom inplace bn to equivilent standart bn

If after these steps you encounter problems, let me know

Tal

Originally posted by @mrT23 in https://github.com/Alibaba-MIIL/ASL/issues/73#issuecomment-975339046