search

lizhaoliu-Lec / CPCM

This is the official repo for Contextual Point Cloud Modeling for Weakly-supervised Point Cloud Semantic Segmentation (ICCV 23).
MIT License
34 stars 3 forks source link

Question about the training script for the 0.02% (S3DIS) #4

Closed jihun1998 closed 12 months ago

jihun1998 commented 1 year ago

Hi. Thank you for releasing the code. However, I cannot find the training script for the 0.02% setting on the S3DIS dataset. I modified the script for 0.01% as 0.02% and tried it. Unfortunately, this setting fails to reproduce the reported performance. Could you provide the official training script for it? Any helpful comment will be greatly appreciated.

lizhaoliu-Lec commented 1 year ago

The hyper-parameters are almost the same for 0.01% and 0.02%. Which dataset are you using and what's the result you get?

jihun1998 commented 1 year ago

Thank you for replying. I am using s3dis dataset with configs for 0.01%. I got about 42 mIoU in validation set (area 5) after 180 epoch training.

lizhaoliu-Lec commented 1 year ago

Maybe provide your training script and config file for me to further diagnosis?

jihun1998 commented 1 year ago

I used same script which is provided from your repository.

CUDA_VISIBLE_DEVICES=0 python launch.py ddp_train.py --config config/default.yaml \ GENERAL.exp_name 1e-4_percentage_consis_weight2_maskGrid075GridSize8_extraMaskStreamSelfCorr_weight10 \ TRAINER.name TwoStreamTrainer \ MODEL.out_channels 13 \ DATA.name StanfordDataLoader \ DATA.dataset StanfordArea5Dataset \ DATA.voxel_size 0.05 \ DATA.batch_size 2 \ DATA.train_limit_numpoints 1000000 \ OPTIMIZER.lr 0.01 \ OPTIMIZER.weight_decay 0.001 \ SCHEDULER.name PolyLR \ TRAINER.epochs 180 \ EVALUATOR.iou_num_class 13 \ DATA.stanford3d_path /mnt/jihun3/s3dis_cpcm \ DATA.stanford3d_sampled_inds /mnt/jihun3/CPCM/prepare_dataset/stanford/points/percentage0.0002evenc \ DATA.sparse_label False \ DATA.two_stream True \ MODEL.two_stream_model_apply True \ TRAINER.two_stream_feats_key semantic_scores \ TRAINER.two_stream_loss_mode js_divergence_v2 \ TRAINER.two_stream_seg_both True \ TRAINER.two_stream_loss_weight 2.0 \ AUGMENTATION.use_color_jitter False \ TRAINER.two_stream_mask_grid_size 8 \ TRAINER.two_stream_loss_mask_mode js_divergence_v2 \ TRAINER.two_stream_mask_ratio 0.75 \ TRAINER.two_stream_mask_mode grid \ TRAINER.two_stream_mask_extra_stream True \ TRAINER.two_stream_mask_feats_key semantic_scores \ TRAINER.two_stream_mask_corr_loss True \ TRAINER.two_stream_mask_self_loss True \ TRAINER.two_stream_loss_mask_weight 10. \ TRAINER.two_stream_mask_loss_threshold -1. \ TRAINER.empty_cache_every 1

lizhaoliu-Lec commented 1 year ago

Thanks for your feedback. I will rerun the exp tonight and get back to you first thing after I get the results.

lizhaoliu-Lec commented 1 year ago

image I have run the experiments and my results are fine. With consis weight = 2, mask weight = 10 (the script you provided), I get the mIoU of 61.16. With consis weight = 5, mask weight = 5, I get the mIoU of 62.34. The exp scripts are shared here: https://codefile.io/f/BP8u0JuRw9.

jihun1998 commented 1 year ago

Thank you for your prompt response. :D