search

nv-tlabs / LION

Latent Point Diffusion Models for 3D Shape Generation
Other
735 stars 57 forks source link

Many-class 3D shape Generation Model #35

Closed yuanzhen2020 closed 1 year ago

yuanzhen2020 commented 1 year ago

Hi @ZENGXH , thanks for your great work! I want to train a many-class unconditional 3D shape generation model. Does this mean I need to use all classes of ShapeNet to train the VAE? I replaced "car" in the example with "all", but found that the loss became NAN at the beginning of training. How can I solve this problem?

ZENGXH commented 1 year ago

Using data.normalize_global False data.normalize_shape_box True in the args is better: I use the shape_box norm for model train on all class in the paper. And it seems to be more stable. I just upload the scripts: scripts/train_vae_all.sh with the arguments. You can try it.

yuanzhen2020 commented 1 year ago

Using data.normalize_global False data.normalize_shape_box True in the args is better: I use the shape_box norm for model train on all class in the paper. And it seems to be more stable. I just upload the scripts: scripts/train_vae_all.sh with the arguments. You can try it.

Thanks for your quick reply!

yuanzhen2020 commented 1 year ago

Hi @ZENGXH, I found that train a multi-class VAE is very time-consuming (4 V100, batchsize = 24, ~1000 hours). Is there a way to speed up the training process?