---

🧐 About

A simple and modular reinforcement learning library based on PyTorch.

🏁 Getting Started

pip install pyforce-rl

🎈 Usage

from pyforce.env import wrap_openai_gym
from pyforce.nn import default_network_components
from pyforce.agents import PPOAgent
import gym
import torch

device="cuda:0" if torch.cuda.is_available() else "cpu"

env=wrap_openai_gym(gym.make("LunarLanderContinuous-v2"))

observation_processor,hidden_layers,action_mapper=default_network_components(env)

agent=PPOAgent(
    observation_processor,
    hidden_layers,
    action_mapper,
    save_path="./evals/ppo_example",
    value_lr=5e-4,
    policy_lr=5e-4
).to(device)

agent.train(env,episodes=1000,train_freq=2048,eval_freq=50,render=True, batch_size=128,gamma=.99,tau=.95,clip=.2,n_steps=32,entropy_coef=.01)

🚀 Implement custom RL Agents

from pyforce.agents.base import BaseAgent
from torch import nn
  
class  MyAgent(BaseAgent):

def  __init__(self,observationprocessor,hiddenlayers,actionmapper,save_path=None):

	super().__init__(save_path)

	self.policy_net = nn.Sequential(observationprocessor, hiddenlayers, actionmapper)
	self.value_net = ...

def  forward(self, state):
	return  self.policy_net(state)

def  get_action(self, state, eval, args):
	#return action + possible additional information to be stored in the memory
	return  self(state).sample(), {} 

def  after_step(self, done, eval, args):
	if  not  eval:
		if  self.env_steps % args["train_freq"] == 0 and len(self.memory) > 0:
			#do training

	if done and eval:
		#do evaluation

⛏️ Built Using

• PyTorch - ML Framework

• OpenAI Gym - Environment API

• NumPy - Numerical calculations outside PyTorch