Below is a tutorial on how to evaluate a policy and visualize the results.
cd {'PATH/TO/PROJECT/DIRECTORY'}
python run/scripts/evaluate.py --ckpt {'PATH/TO/CHECKPOINT'}
--num-humans # Number of humans in the scene
--ckpt # Path to the checkpoint
--ckpt-num # Checkpoint number, without the prefix zeros
--env-name # Environment setup configured under activepose/env_config.py
--map-name # Map name
--render-steps # Number of steps to evaluate (or render if --no-render is not specified)
--num-episodes # Number of episodes to evaluate
--no-render # Disable rendering (only evaluate the policy, significantly faster)
--use-gt # Use ground truth 2D pose as input instead of using predicted 2D pose by default
An example of evaluating the policy with checkpoint ray_results/mappo_ctcr_wdl/PPO_1bd01_00000_Feb08/checkpoint_001400/checkpoint_1400
. This is the default pattern generated by Ray RLlib.
python run/scripts/evaluate.py --ckpt ray_results/mappo_ctcr_wdl/PPO_1bd01_00000_Feb08 --ckpt-num 1400
Or you can without specifying the checkpoint number to evaluate the latest checkpoint:
python run/scripts/evaluate.py --ckpt ray_results/mappo_ctcr_wdl/PPO_1bd01_00000_Feb08
If you choose to render the results (that is to not specify --no-render
), the 3D pose estimation sequences and 2D camera recordings will be saved under the folder
render_data
.
Note: make sure to leave render option ON when evaluating the policy.
We develop a interactive tool 3DPoseViewer
to visualize the reconstruction results and learned camera policy. If you would like to use this viewer alone (visualize your own data), please find the instructions here
cd {ROJECT_ROOT}
python -m run.scripts.visualize