search

lzhengning / SubdivNet

Subdivision-based Mesh Convolutional Networks.
MIT License
251 stars 34 forks source link

try to support more faces or verts #18

Closed Yannnnnnnnnnnn closed 3 years ago

Yannnnnnnnnnnn commented 3 years ago

Thanks for your tips!

I replaced the code of calculating E_hash and everything seems okay.

I run bash scripts/coseg-aliens/test.sh twice and the results are below.

before modification

name: coseg-alien
37/37 [00:21<00:00,  1.68it/s]
test acc =  0.9854864712419181
test acc [original] =  0.9638948133961672
test acc [original] [voted] =  0.9732876916279148
test acc per label =  [0.98391198 0.95061119 0.98919777 0.98311102]

after modification

name: coseg-alien
37/37 [00:28<00:00,  1.31it/s]
test acc =  0.9854868921740302
test acc [original] =  0.9638949857770487
test acc [original] [voted] =  0.9732914332466888
test acc per label =  [0.98391648 0.95060499 0.98919779 0.98310933]
lzhengning commented 3 years ago

Hi, thank you for this pull request. Now the latest jittor supports argsort int64 arrays. So it is more efficient to direct convert data type on GPU. I have updated this file. This pr will be closed.