search

NVlabs / handover-sim2real

Official code for CVPR'23 paper: Learning Human-to-Robot Handovers from Point Clouds
https://handover-sim2real.github.io
Other
65 stars 12 forks source link

Training time? #5

Open LyuJZ opened 8 months ago

LyuJZ commented 8 months ago

Hi, if I would like to reproduce the results with A6000 from Nvidia, how long will it take to train the model from scratch?

christsa commented 8 months ago

The training time largely depends on the hardware used.

On our hardware, pretraining takes around 22 hours and finetuning around 13 hours.

Best, Sammy

LyuJZ commented 8 months ago

Thanks for your reply. So PointNet++ is only used to encode point clouds and not trained during reinforcement learning?

christsa commented 8 months ago

No, the PointNet++ is being finetuned as well during training.

ychao-nvidia commented 7 months ago

For reference, here is our benchmark for training time (in hours) on some GPU models.

V100 A100 RTX 3090
Pretraining 48.8 22.2 21.5
Fintuning 30.9 13.5 12.7

Obviously, there could be many other hardware/system factors in play for time performance.