search

tusen-ai / SST

Code for a series of work in LiDAR perception, including SST (CVPR 22), FSD (NeurIPS 22), FSD++ (TPAMI 23), FSDv2, and CTRL (ICCV 23, oral).
Apache License 2.0
801 stars 102 forks source link

Pruning CTRL (lower compute/memory usage) #194

Closed Kampffussel03 closed 5 months ago

Kampffussel03 commented 5 months ago

Hi Lue,

many thanks again for your great work. As CTRL_FSD was trained with much compute, what would be the best cuts for a still high-performing, pruned version? Do you have any quick hints or ideas about that @Abyssaledge ?

I noticed that FrameDropout is not used and was wondering why. I was looking for the implementation of the following modifications 2) and 3) mentioned here in your paper: image

Maybe a quick modification to test would be using less frames, e.g. 5 instead of 9, or a higher dropout if that's worth in terms of compute. Have you tried something already in this regard?

Best regards

Kampffussel03 commented 5 months ago

The number of frame for actual training is set in the dataloader class.

For pruning CTRL, you can just use a small backbone, which does not decrease the performance too much.