Best results in ModelNet40

[zeal-up]

Thanks for your job! Pytorch is more elegant for me.
I want to asks that what's your best result of classification trained on ModelNet40 using the default hyper-parameters? Or, what's the best accuracy when you tune the hyper-parameters appropriate?
I'm training the model using your code and I will be appreciated if you post the best results.

[erikwijmans]

This repo matches the performance from the paper. If I recall correctly, the only hyper-parameter you need to change is the number of points (to 10k) for that.

[zeal-up]

@erikwijmans Thanks for your reply! My training results do match the paper, but it is still 0.x% accuracy gap with the paper. And I find that, in your implementation, the architecture seems to be pruned.

        self.SA_modules.append(
            PointnetSAModuleMSG(
                npoint=512,
                radii=[0.1, 0.2, 0.4],
                nsamples=[32, 64, 128],
                `mlps=[[input_channels, 64], [input_channels, 128],` 
################### in the paper, it seem to be a three layer mlp for per sub-pointnet
                      [input_channels, 128]],
                use_xyz=use_xyz
            )
        )

I want to know if you do this on purpose?
Thanks!

[erikwijmans]

I have played around with the architectures a fair amount. Makes sense to change them back to the ones given in Charles' repo, I will make that change.

[mingminzhen]

@erikwijmans How to test the model for modelnet40? Could you give some tips to re-implement the experiment result in the paper?

[erikwijmans]

The default parameters should do that. train/train_cls.py will train an MSG model on modelnet40.

[LiuNull]

@erikwijmans Thanks for your reply! My training results do match the paper, but it is still 0.x% accuracy gap with the paper. And I find that, in your implementation, the architecture seems to be pruned.

        self.SA_modules.append(
            PointnetSAModuleMSG(
                npoint=512,
                radii=[0.1, 0.2, 0.4],
                nsamples=[32, 64, 128],
                `mlps=[[input_channels, 64], [input_channels, 128],` 
################### in the paper, it seem to be a three layer mlp for per sub-pointnet
                      [input_channels, 128]],
                use_xyz=use_xyz
            )
        )

I want to know if you do this on purpose?
Thanks!

Hi, sorry to bother. Did you match the paper's accuracy with the newest arch? I run the code and only got 0.9023 with the default hyper-parameters.