Open ruiwang2uber opened 5 years ago
Sure. We will have a related diff coming out soon. Once it is done, you'll be able to run ml/rl/test/gym/world_model/mdnrnn_gym.py to get feature importance/sensitivity of any gym environment.
Expect timeline: 2 weeks.
Thank you! Look forward to it!
Any updates on "Data Understanding Tool"? Thanks.
Hi,
Sorry for some delay. Now you can use the following command to check feature importance and feature sensitivity in the cartpole environment:
python mdnrnn_gym.py -- -p ml/rl/test/configs/mdnrnn_cartpole_v0.json -f -s
The output should be something like this:
INFO:ml.rl.evaluation.world_model_evaluator: Debug tool feature importance : tensor([18.9650, 9.1601, 0.1277, 13.4917, 0.7331, 12.0240]) [2019-06-19 13:48:59,942] Debug tool feature importance : tensor([18.9650, 9.1601, 0.1277, 13.4917, 0.7331, 12.0240]) action 0, feature importance: 18.964975357055664 action 1, feature importance: 9.16006088256836 state 0, feature importance: 0.1277211308479309 state 1, feature importance: 13.491726875305176 state 2, feature importance: 0.7331321239471436 state 3, feature importance: 12.024038314819336 INFO:ml.rl.evaluation.world_model_evaluator: Debug tool feature sensitivity : tensor([0.0067, 0.2099, 0.0066, 0.3125]) [2019-06-19 13:49:07,774] Debug tool feature sensitivity : tensor([0.0067, 0.2099, 0.0066, 0.3125]) state 0, feature sensitivity: 0.006689509842544794 state 1, feature sensitivity: 0.2099362462759018 state 2, feature sensitivity: 0.00659541692584753 state 3, feature sensitivity: 0.31252968311309814
Thanks for updating the paper.
In the updated paper, you mentioned you have implemented a "data understanding tool" based on world model. This is super useful and important.
Could you provide some example on how to use the tool?