Rainy tutorial

Run CLI examples

Though this project is still WIP, all examples are verified to work.

First, install pipenv. E.g. you can install it via

pip3 install pipenv --user

Then, clone this repository and create a virtual environment in it.

git clone https://github.com/kngwyu/Rainy.git
cd Rainy
pipenv --site-packages --three install

Now you are ready to start!

pipenv run python examples/acktr_cart_pole.py train

After training, you can run learned agents.

Please replace (log-directory) in the below command with your real log directory. It should be named like Results/acktr_cart_pole/191119-143319.

pipenv run python acktr_cart_pole.py eval (log-directory) --render

You can also plot training results in your log directory. This command opens an ipython shell with your log file.

pipenv run python -m rainy.ipython

Then you can plot training rewards via

log = open_log('log-directory')
log.plot_reward(12 * 20, max_steps=int(4e5), title='ACKTR cart pole')

Distributed training is supported via horovod.

E.g., if you want to use it with NCLL, you can install it via

sudo env HOROVOD_GPU_ALLREDUCE=NCCL pip3 install --no-cache-dir horovod

You can run the training script using horovodrun command.

E.g., if you want to use two hosts(localhost and anotherhost) and run ppo_atari.py, use

horovodrun -np 2 -H localhost:1,anotherhost:1 pipenv run python examples/ppo_atari.py train