joyrl 0.1.13

A Library for Deep Reinforcement Learning

JoyRL

Install

conda create -n easyrl python=3.7
conda activate easyrl
pip install -r requirements

Torch:

# CPU
conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cpuonly -c pytorch
# GPU
conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=11.3 -c pytorch -c conda-forge
# GPU with mirrors
pip install torch==1.10.0+cu113 torchvision==0.11.0+cu113 torchaudio==0.10.0 --extra-index-url https://download.pytorch.org/whl/cu113

Usage

the following presents a demo to use joyrl, all your need is just to set parameters including GeneralConfig() and AlgoConfig(), which is also shown in examples folder, and well trained results are shown in the benchmarks folder as well.

import joyrl as jj
import gym
class GeneralConfig():
    def __init__(self) -> None:
        self.env_name = "CartPole-v1" # name of environment
        self.algo_name = "DQN" # name of algorithm
        self.mode = "train" # train or test
        self.seed = 0 # random seed
        self.device = "cpu" # device to use
        self.train_eps = 100 # number of episodes for training
        self.test_eps = 20 # number of episodes for testing
        self.eval_eps = 10 # number of episodes for evaluation
        self.eval_per_episode = 5 # evaluation per episode
        self.max_steps = 200 # max steps for each episode
        self.load_checkpoint = False
        self.load_path = "tasks" # path to load model
        self.show_fig = False # show figure or not
        self.save_fig = True # save figure or not

class AlgoConfig():
    def __init__(self) -> None:
        # set epsilon_start=epsilon_end can obtain fixed epsilon=epsilon_end
        self.epsilon_start = 0.95  # epsilon start value
        self.epsilon_end = 0.01  # epsilon end value
        self.epsilon_decay = 500  # epsilon decay rate
        self.gamma = 0.95  # discount factor
        self.lr = 0.0001  # learning rate
        self.buffer_size = 100000  # size of replay buffer
        self.batch_size = 64  # batch size
        self.target_update = 4  # target network update frequency
        self.value_layers = [
            {'layer_type': 'linear', 'layer_dim': ['n_states', 256],
             'activation': 'relu'},
            {'layer_type': 'linear', 'layer_dim': [256, 256],
             'activation': 'relu'},
            {'layer_type': 'linear', 'layer_dim': [256, 'n_actions'],
             'activation': 'none'}]
if __name__ == "__main__":
    general_cfg = GeneralConfig()
    algo_cfg = AlgoConfig()
    jj.run.run(general_cfg,algo_cfg)

Algorithms

Name	Reference	Author	Notes
DQN	DQN Paper	johnjim0816