search

huangjh-pub / synorim

[T-PAMI'22] [Jittor & Pytorch] Multiway Non-rigid Point Cloud Registration via Learned Functional Map Synchronization
https://cg.cs.tsinghua.edu.cn/people/~huangjh/publication/synorim/
118 stars 10 forks source link

KeyError: 'flow_masks is not a file in the archive' #5

Closed elahehomidkhoda closed 1 year ago

elahehomidkhoda commented 1 year ago

Hi, thank you for your excellent work. I want to run synorim project via jittor, but when execute "python train.py configs/train_desc.yaml" command, I got this error raise KeyError("%s is not a file in the archive" % key) KeyError: 'flow_masks is not a file in the archive' could you please guide me?

heiwang1997 commented 1 year ago

Thanks for your interest in our work! This is weird though. Does the pytorch version also give you the same error?

Best, Jiahui

elahehomidkhoda commented 1 year ago

I don't run pytorch version because of Minkowski Engine doesn't build on my system. Can I ask your hardware specification to build Minkowski Engine? My hardware specification are: Ubuntu 22.04 GPU: NVIDIA GEFORCE GTX 960M RAM: 12G Driver Version: 515.65.01 Cuda: 11.7 But I guess, this error will occured in pytorch version too, because npz files in "dataset/mpc-cape/data/" contain just "pcs.npy" and "flows.npy"!!

heiwang1997 commented 1 year ago

Jittor version does not use sparse conv but PointNet++ instead so you will not be able to fully reproduce our results in the paper, unfortunately. My suggestion is still to try building MinkowskiEngine. May I know what is the error when you build ME? My hardware is: Ubuntu 20.04 GPU: nvidia gtx 3090 RAM: 128g Driver version: 510.85.02 Cuda: 11.6

The pytorch version runs with no problem on my side btw :)

elahehomidkhoda commented 1 year ago

Thanks for your quick and kind replies, finally I could build ME on my system, and pytorch version runs with no problem :) One more question about testing on my own data. I have 8 pointcloud that captured from different side of body, is it possible to construct full 3d body model with this project? If I save 8 pointcloud in numpy array (.npy), is it correct?

heiwang1997 commented 1 year ago

Hey @elahehomidkhoda

That's correct! I would recommend putting the body sides into world space before doing our non-rigid registration. It would be good if there could be some overlaps between these scans :)