Can LION support various number of point instead of freezing the number to 2048

nv-tlabs / LION

Latent Point Diffusion Models for 3D Shape Generation

Other

735 stars 57 forks source link

Can LION support various number of point instead of freezing the number to 2048 #50

Closed OswaldoBornemann closed 1 year ago

OswaldoBornemann commented 1 year ago

Can LION support various number of point instead of freezing the number to 2048

ZENGXH commented 1 year ago

yes. You can set it throughdata.tr_max_sample_points and data.te_max_sample_points;

if the model is trained 2048, you can use different number of points in inference (I tried ~4k and it's working)
you can also train the model with different number of points, I tried a much larger number before

OswaldoBornemann commented 1 year ago

@ZENGXH Thank you very much for your response. What I mean by "various number of point" here is that the number of points in each batch can vary. For example, there might be 2048 points in some batches and 1564 points in others. I've roughly identified the code for the decoder, and during the decoding process, I'm attempting to pass in a dynamically adjustable "num_points".