search

chrischoy / SpatioTemporalSegmentation

4D Spatio-Temporal Semantic Segmentation on a 3D video (a sequence of 3D scans)
MIT License
289 stars 51 forks source link

Model Zoo conv1_kernel_size mismatch #7

Open 96lives opened 4 years ago

96lives commented 4 years ago

Hello,

I noticed that you have put a pertained Mink16UNet34C model. However, I don't think this is a correct pretrained model. If you load the model of the model zoo with model.load_state_dict(torch.load(file_path))['state_dict'] you can see that conv0p1s1 has kernel size of 3, while your implementation of MinkUNet34 has kernel size of 5 in the Minkowski Engine

If you look at the indoor.py and get the pretrained model of via link that you've uploaded, I think you can get the correct pretrained model.

Please let me know if I'm mistaken. Thanks,

chrischoy commented 4 years ago

Ah thanks for letting me know. It is the same weight, but you have to use the argument python main.py --conv1_kernel_size 5 ... for the weights. I'll update the ModelZoo description.

chrischoy commented 4 years ago

It is 0.2% mIoU lower than the current one I put on the Model Zoo in this repository. But for all purposes, they are pretty state of the art.

96lives commented 4 years ago

Thanks for the quick response!!