Skip to content

Latest commit

 

History

History
53 lines (36 loc) · 2.97 KB

training.md

File metadata and controls

53 lines (36 loc) · 2.97 KB
Raw

Training

NOTE: training is not needed if you only wish to evaluate PF-Track.

As described in the paper, training PF-Track involves two steps: (1) training a single-frame detector; (2) training a multi-frame tracker initialized from the detector. This procedure saves time and accelerates convergence.

If you have downloaded our checkpoints, you can directly use our detector checkpoints and advance to step two. However, please make sure the load_from=xxx fields in the configuration files point to the correct paths of checkpoints.

1. Training Procedure

1.1 Step One: Single-frame Training

You can skip this step if you have downloaded our pretrained single-frame model (see "pretrained models" for details).

  • CONFIG_PATH will be your desired model configurations in configs. In this repository, these are the files starting with f1.
  • GPU_NUM is the number of GPUs. By default, I use 8.
  • PATH_TO_SAVE is the directory to save the checkpoints.

The training command is as below:

bash tools/dist_train.sh CONFIG_PATH GPU_NUM --work-dir PATH_TO_SAVE

For example, if you want to train a small-resolution model, the command will be:

bash tools/dist_train.sh projects/configs/tracking/petr/f1_q5_800x320.py 8 --work-dir work_dirs/f1_pf_track/

The results should be similar to the single-frame detector checkpoints provided by us (see "pretrained models" for details).

1.2 Step Two: Multi-frame Training

  • CONFIG_PATH will be your desired model configurations in configs. In this repository, these are the files starting with f3.
  • GPU_NUM is the number of GPUs. By default, I use 8.
  • PATH_TO_SAVE is the directory to save the checkpoints.

The training command is as below:

bash tools/dist_train.sh CONFIG_PATH GPU_NUM --work-dir PATH_TO_SAVE

For example, if you want to train a small-resolution model, the command will be:

bash tools/dist_train.sh projects/configs/tracking/petr/f3_q5_800x320.py 8 --work-dir work_dirs/f3_pf_track/

The results should be similar to the f3_all/final.pth pretrained model provided by us.

2. Notes for Developers

  • Best practice for debugging? We recommand setting GPU_NUM=1 when running the above commands, and set pdb.set_trace() at the place you want to probe. Meanwhile, you can also switch the infos file to tracking_forecasting-mini_infos_val.pkl to save the time of data loading.
  • Core functions? The core functions of tracking happens in [code link], the function forward_tracking. If you have any difficulty understanding my implementation, please read "my designs," from which you will learn about how to build an end-to-end tracking system.
  • Names for loss functions? If you check the log, we use loss_for to represent motion forecasting losses in future reasoning, and loss_mem to represent past reasoning losses.