chriswu1997/singleshapegen

code for paper "Learning to Generate 3D Shapes from a Single Example", SIGGRAPH Asia 2022

singleshapegen's Issues

training data prepare

Thanks for your great work.

BTW, I'm not sure how to preprocess the data. I got an STL format file from your link, but

To construct the training data (voxel pyramid) from a mesh, we rely on binvox. After downloading it, make sure you change BINVOX_PATH in voxelization/voxelize.py to your path to excetuable binvox.

this part how to make it to binvox format?

Or, could you share any processed .h5 file?

thanks

One discriminator for all scales?

Hi, if I understand correctly, only one discriminator is used to train all the scales right? It is reused to train the next scale?

Train using 128 resolution

Hello, thanks for the repository

I'm trying to train the model using binvox files with res 128 as you did in your paper.
I used the code python main.py train --tag {your-experiment-tag} -s {path-to-processed-h5-data} -g {gpu-id} but the processing time is the same and I dont see any diferent results compare to res 64.

Is there any parameter I'm missing?

Thanks again

Only updating the parameters of current scale?

Hi, when training the current scale, are all the parameters of previous scales of the generator frozen? and only parameters of current scale are updated?

`flip` when sampling features from plane

Hi, thanks for the great work, I'm wondering why flip(-1) is used when sampling the features from the triplane?

SingleShapeGen/model/networks.py

Line 129 in a9fed7f

sample_yz_feat = F.grid_sample(yz_feat, coords_[..., [1, 2]].flip(-1), align_corners=True).permute(0, 2, 3, 1).squeeze(0).squeeze(0)