Skip to content

Latest commit

 

History

History
32 lines (25 loc) · 2.52 KB

RUN_MODEL.md

File metadata and controls

32 lines (25 loc) · 2.52 KB

Training and Evaluating SSDA3D

Training Scripts

SSDA3D is a two-stage framework, which are Inter-domain Adaptation stage and Intra-domain Generalization stage, between which is generating pseudo labels for unlabeled target data. So there are three steps to train SSDA3D. Here is an example for training with 1% labeled nuScenes data (and 99% unlabeled nuScenes data). For different amount of labeled nuScenes settings, you can specify different config files, such as 5% labeled (and 95% unlabeled), 10% labeled(and 90% unlabeled) and 20% labeled (and 80% unlabeled).

  • Inter-domain Adaptation.. You can run the following command to conduct STAGE_I training.
bash scripts/stage1_cutmix_dist_train.sh ${NUM_GPUS} --cfg_file cfgs/stage1_cutmix/centerpoint_20_waymo_1_nus_frames_cutmix.yaml --fix_random_seed
  • Generating Pseudo Labeles. After STAGE_I training, you can run the following command to generate pseudo labels for the left unlabeled target data with the model from STAGE_I. Note that, you need to specify a thresholds score to filter the predicted boxes. Specifically, we select 0.2 for 99% unlabeled target data.
bash scripts/generate_pseudo_labels_dish.sh ${NUM_GPUS} --cfg_file cfgs/pseudo_labels/centerpoint_generate_99_pseudo_nus_frames.yaml --ckpt ../output/stage1_cutmix/centerpoint_20_waymo_1_nus_frames_cutmix/default/ckpt/checkpoint_epoch_20.pth --pseudo_thresh 0.2
  • Intra-domain Generalization. After generating pseudo labels, you can run the following command to conduct STAGE_II training.
bash scripts/stage2_mixup_dist_train.sh ${NUM_GPUS} --cfg_file cfgs/stage2_mixup/centerpoint_1_lab_99_unlab_nus_frames_mixup.yaml --pretrained_model ../output/stage1_cutmix centerpoint_20_waymo_1_nus_frames_cutmix/default/ckpt/checkpoint_epoch_20.pth --pseudo_info_path ../output/pseudo_labels/centerpoint_generate_99_pseudo_nus_frames/default/score_0.2_last_80_percent_frames_nuscenes_infos_10sweeps_train.pkl --fix_random_seed

In our experiments, we set different thresholds scores for different setting, which can be found in the following table.

Labeled Target 1% 5% 10% 20%
Thresh score 0.2 0.2 0.2 0.3

Evaluation Scripts

You can run this command to evaluate a model trained by SSDA3D on nuScenes dataset.

bash scripts/dist_test.sh ${NUM_GPUS} --cfg_file cfgs/nuscenes_models/ssda3d_centerpoint.yaml --ckpt ../../OwnOpenPCDet/output/xxx/xxx/xxx/ckpt/checkpoint_epoch.pth