script_usage.md

Script usage

Experiment Groups

The script structures all runs into experiment groups. To create a new experiment group, you should create a subdirectory of models_dir, and create a file config.yml in that subdirectory. See the config usage document for futher details.

Training

To run a new experiment, simply run

python -m script \
    --mode=train \
    --experiment_group=EXPERIMENT_GROUP_NAME \
    --env_name=ENV_NAME \
    --seed=RNG_SEED_INT

Checkpointing

By default, the script looks for saved checkpoints, and will use them if you do not delete them.
The checkpoints can be found in

models_dir/EXPERIMENT_GROUP_NAME/ENV_NAME/RNG_SEED_INT/checkpoints

Tensorboard logs

To monitor tensorboard logs, you should type

tensorboard \
    --logdir=models_dir/EXPERIMENT_GROUP_NAME/ENV_NAME/RNG_SEED_INT/tensorboard_logs \
    --host=localhost

And then navigate to http://localhost:6006 in your browser.

Evaluation

Sometimes algorithm evaluation is conducted after training of the RL agent has ended. To evaluate an agent that supports this type of evaluation, you can run

python -m script \
    --mode=evaluate \
    --experiment_group=EXPERIMENT_GROUP_NAME \
    --env_name=ENV_NAME \
    --seed=RNG_SEED_INT

Video

To see video of the agent interacting with the environment, you can run

python -m script \
    --mode=video \
    --experiment_group=EXPERIMENT_GROUP_NAME \
    --env_name=ENV_NAME \
    --seed=RNG_SEED_INT

The video will be saved to

models_dir/EXPERIMENT_GROUP_NAME/ENV_NAME/RNG_SEED_INT/media