paul007pl
Training time and pre-trained models
Hello Liang, thanks for sharing the code of your interesting work. I have some questions.
- I I train your model for 16384 points on RTX 3090 with a batch size of 18. I trained it with cd loss instead of emd loss. It took around 1 hour to train one epoch. How long does it take for you to train the model?
- If possible, could you provide the pre-trained model with 16384 points ? Thanks in advance.
Thank you for asking. Actually, I have been asked many times about the training time with 16384 points as well as the pre-trained model. Unfortunately, it is indeed time-consuming to train the model with 16384 points, and my pre-trained models are lost due to changing the workstation cluster nodes.
When I was doing this task (2 years ago), I mostly conduct experiments with 2048 points or 4096 points, which could finish the training in around one day with a V100 GPU. The results with 2048 points are usually positively related to the results with 16384 points. I did not train many models with 16384 points due to its time-consuming fact (a few days). For the same reason, the following works all prefer to use 2048 points in their experiments. You could also check:
https://github.com/wutong16/Density_aware_Chamfer_Distance
https://github.com/ZhaoyangLyu/Point_Diffusion_Refinement
Perhaps, you could focus on 2048 points first, or you could consider designing efficient operations to ease the training (this is a true research gap).
However, in mechanical engineering, we should have more points to recover CAD model with high quality.Too small points can do nothing.
