Code and Environment for Causal Induction From Visual Observations for Goal-Directed Tasks.

Environment

Consists of the light switch environment for studying visual causal induction, where N switches control N lights, under various causal structures. Includes common cause, common effect, and causal chain relationships. Environment code resides under env/light_env.py.

Induction Models

The different induction models used are located under F_models.py, incuding our proposed iterative attention network, as well as baselines which do not use attention or use temporal convolutions.

Reproducing Experiments

Step 1: Generate Data

python3 collectdata.py --horizon 7 --num 7 --fixed-goal 0 --structure one_to_one --seen 10 --images 1 --data-dir output/

Step 2: Train Induction Model

python3 trainF.py --horizon 7 --num 7 --fixed-goal 0 --structure one_to_one --type iter --images 1 --seen 10 --data-dir output/

Step 3: Eval Induction Model

python3 evalF.py --horizon 7 --num 7 --fixed-goal 0 --structure one_to_one --method trajFi --images 1 --seen 10 --data-dir output/

Step 4: Train Policy via Imitation

python3 learn_planner.py --horizon 7 --num 7 --fixed-goal 0 --structure one_to_one --method trajFi --seen 10 --images 1 --data-dir output/