Skip to content

Running your first experiment#

Assuming you have installed the full library, you can run your first experiment by calling

PYTHONPATH=. python allenact/main.py minigrid_tutorial -b projects/tutorials -m 8 -o experiment_output/minigrid -s 12345

from the allenact root directory.

  • With -b projects/tutorials we tell allenact that minigrid_tutorial experiment config file will be found in the projects/tutorials directory.
  • With -m 8 we limit the number of subprocesses to 8 (each subprocess will run 16 of the 128 training task samplers).
  • With -o experiment_output/minigrid we set the output folder into which results and logs will be saved.
  • With -s 12345 we set the random seed.

If everything was installed correctly, a simple model will be trained (and validated) in the MiniGrid environment and a new folder experiment_output/minigrid will be created containing:

  • a checkpoints/MiniGridTutorial/LOCAL_TIME_STR/ subfolder with model weight checkpoints,
  • a used_configs/MiniGridTutorial/LOCAL_TIME_STR/ subfolder with all used configuration files,
  • and a tensorboard log file under tb/MiniGridTutorial/LOCAL_TIME_STR/.

Here LOCAL_TIME_STR is a string that records the time when the experiment was started (e.g. the string "2020-08-21_18-19-47" corresponds to an experiment started on August 21st 2020, 47 seconds past 6:19pm.

If we have Tensorboard installed, we can track training progress with

tensorboard --logdir experiment_output/minigrid/tb

which will default to the URL http://localhost:6006/.

After 150,000 steps, the script will terminate and several checkpoints will be saved in the output folder. The training curves should look similar to:

training curves

If everything went well, the valid success rate should converge to 1 and the mean episode length to a value below 4. (For perfectly uniform sampling and complete observation, the expectation for the optimal policy is 3.75 steps.) In the not-so-unlikely event of the run failing to converge to a near-optimal policy, we can just try to re-run (for example with a different random seed). The validation curves should look similar to:

validation curves

A detailed tutorial describing how the minigrid_tutorial experiment configuration was created can be found here.

To run your own custom experiment simply define a new experiment configuration in a file projects/YOUR_PROJECT_NAME/experiments/my_custom_experiment.py after which you may run it with PYTHONPATH=. python allenact/main.py my_custom_experiment -b projects/YOUR_PROJECT_NAME/experiments.