search

hongfz16 / DS-Net

[CVPR 2021/TPAMI 2023] Rank 1st in the public leaderboard of SemanticKITTI Panoptic Segmentation (2020-11-16)
MIT License
243 stars 29 forks source link

Number of GPUs for training of 4D-DS-Net #14

Open LarsKreuzberg opened 2 years ago

LarsKreuzberg commented 2 years ago

Hi! Thanks for sharing your great project! I have a question regarding 4D-DS-Net. I tried to finetune the backbone of 4D-DS-Net. I have the possibility to train on 4 Geforce GTX 1080Ti. Unfortunately even on 4 Geforce GTX 1080Ti I am getting a "CUDA OUT OF MEMORY ERROR". On how many GPUs you trained the 4D-DS-Net? Is there a possibility that you share your pretrained models for 4D-DS-Net? Thanks in advance!

hongfz16 commented 2 years ago

Thank you for your interest. We trained the 4D-DS-Net using 4 NVIDIA V100 GPUs which have 32GB memory each. So unfortunately, GTX 1080Ti will not be enough to train our recommended setting. One possible solution is to tune down the number of points using in dynamic shifting module. You can adjust the number in the config file here and here. But from my experiment, tuning down the number of points will affect the final results.

As for the pre-trained model for 4D-DS-Net, I will upload it soon.

hongfz16 commented 2 years ago

Uploaded the pre-train model. Go check it out in README!

LarsKreuzberg commented 2 years ago

Many thanks for sharing your final model for 4D-DS-Net! Could you maybe also share your finetuned backbone for 4D-DS-Net? This pretrained model is called "checkpoint_epoch_20_0.622_0.572_0.620.pth" in ./scripts/release/4d-dsnet/train_dsnet_multi_frames_2.sh.