To convert the Model to TorchScript and Inferencing in C++

[jakeju92]

Hi @erikwijmans ,

Have you tried to convert the pointnet2 model you wrote to Torch script for inferencing in C++?

I've met some issues. One of the issues is about the attribute 'npoint' in _PointnetSAModuleBase class.

class _PointnetSAModuleBase(torch.jit.ScriptModule):
    # __constants__ = ['npoint']

    def __init__(self):
        super(_PointnetSAModuleBase, self).__init__()
        # self.register_buffer('npoint', None)
        self.npoint = None
        self.groupers = None
        self.mlps = None

    @torch.jit.script_method
    def forward(self, xyz, features=None):

It throws out a runtime error:

RuntimeError: 
attribute 'npoint' of type 'int' is not usable in a script method (did you forget to add it __constants__?):
...
            pointnet2_utils.gather_operation(
                xyz_flipped, pointnet2_utils.furthest_point_sample(xyz, self.npoint)
            )
            .transpose(1, 2)
            .contiguous()
            if self.npoint is not None
               ~~~~~~~~~~~ <--- HERE
            else None
        )
...

I tried to put __constants__ = ['npoint'] one line before the def __init__(self):. However, that doesn't work. npoint should be a variable and its default value is None and could be assigned to int eventually.

Do you have any idea how to fix this? Thanks in advance!

[erikwijmans]

I haven't tried torchscript for this repo and I have no idea how c++ extensions interact with torchscript. For npoint, you can try annotating as Optional[int] to see if that makes torchscript happy.

[doublexxking]

@jb892 Do you convert to torchscript successfully ? I meet the similar questions in this process...

[jakeju92]

@doublexxking Nope, I failed to do so. But I rewrite the whole network manually by caffe2 and load weights from a trained PyTorch model. It works perfectly for me. I hope it helps!