Agents Bar Python Client

This package is a python client for Agents Bar service. It isn't supposed to be used in an isolation and you are supposed to have an existing account. Check the website for more information about the service, or check the Agents Bar Docs to learn more how to use this client.

Quick start

The client allows to communicate with service by wrapping around APIs and coveraging common use patterns.

A common usage is replacing your deep reinforcement learning agent with an entity that uses step and act APIs for progressing agent and infering action, respectively. For usage examples as Google Colab check Doc's quick start link and for scripts check examples directory.

For a minimal (almost) working example check this code snippet:

from agents_bar import Client, RemoteAgent
from agents_bar import environments

# Define client to communicate with https://agents.bar. Make sure it's authenticated.
client = Client()

# Create an environment. Simple one is "CartPole-v1" from OpenAI gym repo.
env_name = "CartPole"
environments.create(client, config={"name": env_name, "config": {"gym_name": "CartPole-v1"}})

# Create an agent. Since environment is discrete we use DQN.
agent = RemoteAgent(client, agent_name="CartPoleAgent")
agent.create_agent(obs_size=4, action_size=2, agent_model="DQN")

# Initiat learning loop. Observe env's state, pass to agent, make a decision (action), execute on env. Repeat.
obs = environments.reset(client, env_name)
for iteration in range(10):
    action = agent.act(obs)

    out = environments.step(client, env_name, step={"actions": [action], "commit": True})
    next_obs, reward, done = out.get("observation"), out.get("reward"), done.get("done")

    agent.step(obs, action, reward, next_obs, done)
    obs = next_obs

Support

Agents Bar Client currently supports manipulation of agents, environments and experimnets. We also provide an abstraction over agent which allows you to use the agent as an object, the same as you are already using it.

The client is intended to be used for easy communication. Check documentation for all available APIs. In most cases they should be the same as you see in https://agents.bar/docs.

Installation

Pip (Recommended)

The latest stable version should always be accessible through pip as agents-bar. To install locally add agents-bar to your dependency file, e.g. requirements.txt, or install it directly using

pip install agents-bar

GitHub source

Checkout this package using git clone git@github.com:agents-bar/agents-bar-client-python. This will create a new directory agents-bar-client-python. Go ahead, enter the directory and install the package via pip install -e ..

Note we recommend having a separate python environment for standalone projects, e.g. using python -m venv command.

Authentication

To use the client you need to be pass Agents Bar credentials or some proof that you're a user, e.g. access_token. There are a few ways how to authenticate your client.

Note: Never store your credentials in places easy accessible by others. This includes git repositories that have the slightest chance to leave your computer. Definitely nothing that goes to the GitHub/GitLab.

Environment variables (suggested)

Currently suggested approach for authentication is to set your token or credentials as environment variables. The client looks first for AGENTS_BAR_ACCESS_TOKEN and uses that as its access token. You can use this approach if you want to login using a different application with securely stored credentials and temporarily set the access token. Otherwise, you can also set your username and password in AGENTS_BAR_USERNAME and AGENTS_BAR_PASSWORD, respectively.

As an example, in unix, you can set environment variables by using export command in shell

export AGENTS_BAR_ACCESS_TOKEN=<access_token>
... or ...
export AGENTS_BAR_USERNAME=<username>
export AGENTS_BAR_PASSWORD=<password>

Instantiating with credentials

The RemoteClient can authenticate using access_token or credentials (username and password) provided when instantiating the agent. Only one of these is required and the access_token has priority over credentials pair. Also, note that directly passed variables have priority over the environment variables.

access_token = "<access_token>"
username = "<username>"
password = "<password>"

client = RemoteClient(..., access_token=access_token, username=username, password=password)