huggingbutt 0.0.5a0

Client library to download and publish environments, agents on the huggingbutt.com hub

HuggingButt

Client library to download and publish reinforcement learning environments/agents on the huggingbutt.com hub

Installation

The base dependency mlagents-envs has been update to version 1.0.0. However, this version may cause error(on Windows 10). This is due to the package specifying numpy version 1.21.2, which might fail to build on Python 3.10 (Windows 10). I have had two colleagues test it and they encountered the same issues. So, to avoid conflicts, mlagents-envs==0.30.0, NumPy==1.21.2 in this project. MacOS is OK. For the reason above, to ensure safety, we will use Python 3.9.

Create a new python environment using anaconda/miniconda.

conda create -n hb python=3.9

activate the new python environment.

conda activate hb

install huggingbutt from pypi

pip install huggingbutt==0.0.5

or from source code

git clone https://github.com/huggingbutt/huggingbutt.git
cd huggingbutt
python -m pip install .

If there is no error message printed during the installation, congratulations, you have successfully installed this package. Next, you need to apply an access token from the official website https://huggingbutt.com.

Register an account and login, just do as shown in the image below.

Click new token button generate a new token. This access token is mainly used to restrict the download times of each user, as the server cost is relatively high.

Congratulations, you now have an access token!

Just put the generated token in the task code and you're gooooood to go.

Here is a simple training code:

from huggingbutt import Env, Agent, set_access_token

if __name__ == '__main__':
    set_access_token('YOUR_TOKEN')
    env = Env.get("huggingbutt/juggle", 'mac', startup_args=['--time_scale', '10'])

    agent = Agent(
        env=env,
        algorithm='PPO',
        policy='MlpPolicy',
        batch_size=64
    )
    agent.learn(total_timesteps=10000)
    agent.save()
    env.close()

Inference:

from huggingbutt import Env, Agent, set_access_token

if __name__ == '__main__':
    set_access_token('YOUR_TOKEN')
    env = Env.get("huggingbutt/juggle", 'mac', startup_args=['--time_scale', '1'])
    agent = Agent.get(15, env)

    obs = env.reset()
    for i in range(100):
        act, _status_ = agent.predict(obs)
        obs, reward, done, info = env.step(act)
        if done:
            obs = env.reset()
    env.close()

todo

1. Support more types learning environment, such as native game wrapped by python, pygame, class gym...
2. Develop a framework, user can customize the observation, action and reward of the environment developed under this framework. We hope this makes it easier for everyone to iterate the environment's agent.
3. There are still many ideas, I will add them later when I think about them...