Evaluation

src/benchmark.py evaluates the performance of pretrained models, including those trained with OBC and Arnold, as well as the expert policies.

Evaluating OBC and Arnold models

Specify the path to the saved model (.zip file) and the task to evaluate:

python src/benchmark.py \
    --load path/to/your/model.zip \
    --task <task_names> \
    --arnold \
    --num_episodes <number_of_episodes> \
    --deterministic \
    --device <cpu_or_cuda> \
    --render

Flag	Meaning
--load	Path to your trained model checkpoint.
--task	One or many of the available tasks.
--arnold	Include if the model was trained with Arnold.
--num_episodes	How many episodes to run for evaluation.
--deterministic	Take deterministic actions from the policy.
--device	cpu or cuda.
--render	Optionally render to video.

Rendering on macOS

--render requires running mjpython instead of python.

Example: an Arnold model

python src/benchmark.py \
    --load data/student_policies/arnold/rl_model_64670238_steps.zip \
    --task kinesis \
    --arnold \
    --num_episodes 10 \
    --deterministic \
    --device cpu

Example: an OBC model

python src/benchmark.py \
    --load data/student_policies/obc/rl_model_54974700_steps.zip \
    --task relocate \
    --arnold \
    --num_episodes 10 \
    --deterministic \
    --device cpu

Evaluating expert policies

For an expert policy, only the task needs to be specified — the checkpoint is resolved from data/expert_policies/:

python src/benchmark.py \
    --task <task_name> \
    --expert \
    --num_episodes <number_of_episodes> \
    --deterministic \
    --device <cpu_or_cuda>

Include the --expert flag to indicate you are testing an expert policy, and set <task_name> to one of the available tasks.

Example

python src/benchmark.py \
    --task relocate \
    --expert \
    --num_episodes 10 \
    --deterministic \
    --device cpu \
    --render

Saving results for the plotting scripts

The plotting scripts read benchmark result JSONs from data/final_benchmarks/ and data/final_benchmarks_extra/. To regenerate them rather than using the released ones, add --save_results and point --out_dir at the directory the relevant plot expects — see CSI-Finetuning and Multi-task RL baselines for the per-arm destinations.