size mismatch in PartDecoder, m1: [64 x 50], m2: [100 x 100]

daerduoCarey / structurenet

StructureNet: Hierarchical Graph Networks for 3D Shape Generation

https://cs.stanford.edu/~kaichun/structurenet/

Other

258 stars 45 forks source link

size mismatch in PartDecoder, m1: [64 x 50], m2: [100 x 100] #2

Closed Hydrogenion closed 4 years ago

Hydrogenion commented 4 years ago

We meet problem that PartEncoder give a feature [64 x 50], but the input of PartDecoder is [100 x100]. It cause the RuntimeError: size mismatch, m1: [64 x 50], m2: [100 x 100] at /tmp/pip-req-build-9xcrj8au/aten/src/THC/generic/THCTensorMathBlas.cu:268

Please help me, thanks!

daerduoCarey commented 4 years ago

feat_len should be a parameter you can set in config.py. I think we are using 100 for all exps? Can you provide more details how do you get the 50-dim feature?

Hydrogenion commented 4 years ago

Why the program always be killed? Some op always in cuda:0

Hydrogenion commented 4 years ago

I find the reason. The cd loss will always run in cuda:0, if you just change device in config.py, it may cause some problem. use os.environ["CUDA_VISIBLE_DEVICES"] = "2" will solve the problem.

BTW, my computer will kill the process which in cuda:0, so i find this problem.