DRLib：A concise deep reinforcement learning library which integrats amost all of off policy RL algos with HER and PER.

A concise deep reinforcement learning library which integrats amost all of off policy RL algos with HER and PER. The library is written based on the code in https://github.com/openai/spinningup, and can be achieved with tensorflow or pytorch. Compared with spinning up, the multi-process and experimental grid wrapper have been deleted for easy application. Besides, the code in our library is convenient to debug with pycharm~

欢迎大家关注我的最新工作D2SR，基于切换的密集转稀疏的高效奖励函数设计方法：

核心代码已经开源在本仓库：D2SSR

对应的中文版论文也在这个仓库中：D2SR: Transferring Dense Reward Function to Sparse by Network Resetting

欢迎大家关注我的最新工作RHER，简洁高效的HER变体： https://github.com/kaixindelele/RHER

最新的、全面的实验结果：

4种tf，3种torch的HER算法在三个操作任务的测试结果。

画图脚本示例：

python spinup_utils/plot.py HER_DRLib_mpi1/2 --select Push

#如果是Windows建议用绝对路径，否则找不到文件

保存一个训练好的模型：net/replay_buffer/norm!

python train_torch_mpi_norm_save.py

重载测试一个训练好的模型：net/replay_buffer/norm!

python train_torch_mpi_norm_load.py

项目特点：

1. tf1和pytorch两个版本的算法，前者快，后者新，任君选择；

2. 在spinup的基础上，封装了DDPG, TD3, SAC等主流强化算法，相比原来的函数形式的封装，调用更方便，且加了pytorch的GPU调用；

3. 添加了HER和PER功能，非常适合做机器人相关任务的同学们；

4. 实现了最简单的并行自动调参（ExperimentGrid）和多进程（MPI_fork-实现了，没有完全实现）部分，适合新手在pycharm中debug，原版的直接调试经常会报错~

教程链接：【Spinning Up】四、python同时启动多个不同参数脚本

多进程教程：没写~

我终于把tf版本-基于mpi的多进程调好了~

torch版本的没有测试完毕，有报错！

如果大家的CPU核心足够多的情况下，试试mpi多进程，性能会提升比较大的。

目前测试的结果是，tf-DDPG的性能最佳，TD3的结果竟然会比ddpg的差，简直了~

5. 最后，全网最详细的环境配置教程！亲测两个小时内，从零配置完全套环境！

6. 求三连，不行的话，求个star！

1. Installation

1. Clone the repo and cd into it:

   git clone https://github.com/kaixindelele/DRLib.git
   cd DRLib

2. Create anaconda DRLib_env env:

   conda create -n DRLib_env python=3.6.9
   source activate DRLib_env

3. Install pip_requirement.txt:

   pip install -r pip_requirement.txt

   If installation of mpi4py fails, try the following command(Only this one can be installed successfully!):

   conda install mpi4py

   或者直接看下面的链接： ubuntu-windows-install-mpi4py-亲测好使！

   conda install seaborn==0.8.1 scipy -y

4. Install tensorflow-gpu=1.14.0

   conda install tensorflow-gpu==1.14.0 # if you have a CUDA-compatible gpu and proper drivers

5. Install torch torchvision

   # CUDA 9.2
   conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=9.2 -c pytorch
   
   # CUDA 10.1
   conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch
   
   # CUDA 10.2
   conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.2 -c pytorch
   
   # CPU Only
   conda install pytorch==1.6.0 torchvision==0.7.0 cpuonly -c pytorch
   
   # or pip install    
   pip --default-timeout=100 install torch -i  http://pypi.douban.com/simple  --trusted-host pypi.douban.com
   [pip install torch 在线安装!非离线!](https://blog.csdn.net/hehedadaq/article/details/111480313)

6. Install mujoco and mujoco-py

   refer to: https://blog.csdn.net/hehedadaq/article/details/109012048

7. Install gym[all]

   refer to https://blog.csdn.net/hehedadaq/article/details/110423154

2. Training models

• Example 1. SAC-tf1-HER-PER with FetchPush-v1:

1. modify params in arguments.py, choose env, RL-algorithm, use PER and HER or not, gpu-id, and so on.

2. run with train_tf.py or train_torch.py

   python train_tf.py

3. exp results to local:https://blog.csdn.net/hehedadaq/article/details/114045615

4. plot results:https://blog.csdn.net/hehedadaq/article/details/114044217

超强版强化学习画图脚本！

相比于原始的plot.py文件，增加了如下的功能

1.可以直接在pycharm或者vscode执行，也可以用命令行传参；

2.按exp_name排序，而不是按时间排序；

3.固定好每个exp_name的颜色；

4.可以调节曲线的线宽，便于观察；

5.保存图片到本地，便于远程ssh画图~

6.自动显示全屏

7.图片自适应

8.针对颜色不敏感的人群,可以在每条legend上注明性能值,和性能序号

9.对图例legend根据性能从高到低排序，便于分析比较

10.提供clip_xaxis值，对训练程度进行统一截断，图看起来更整洁。 seaborn版本0.8.1

3. File tree and introduction:

.
├── algos
│   ├── pytorch
│   │   ├── ddpg_sp
│   │   │   ├── core.py-------------It's copied directly from spinup, and modified some details.
│   │   │   ├── ddpg_per_her.py-----inherits from offPolicy.baseOffPolicy, where one can choose whether or not HER and PER
│   │   │   ├── ddpg.py-------------It's copied directly from spinup
│   │   │   ├── __init__.py
│   │   ├── __init__.py
│   │   ├── offPolicy
│   │   │   ├── baseOffPolicy.py----baseOffPolicy, DDPG/TD3/SAC and so on.
│   │   │   ├── norm.py-------------state normalizer, update mean/std with training process.
│   │   ├── sac_auto
│   │   ├── sac_sp
│   │   │   ├── core.py-------------likely as before.
│   │   │   ├── __init__.py
│   │   │   ├── sac_per_her.py
│   │   │   └── sac.py
│   │   └── td3_sp
│   │       ├── core.py
│   │       ├── __init__.py
│   │       ├── td3_gpu_class.py----td3_class modified from spinup
│   │       └── td3_per_her.py
│   └── tf1
│       ├── ddpg_sp
│       │   ├── core.py
│       │   ├── DDPG_class.py------------It's copied directly from spinup, and wrap algorithm from function to class.
│       │   ├── DDPG_per_class.py--------Add PER.
│       │   ├── DDPG_per_her_class.py----DDPG with HER and PER without inheriting from offPolicy.
│       │   ├── DDPG_per_her.py----------Add HER and PER.
│       │   ├── DDPG_sp.py---------------It's copied directly from spinup, and modified some details.
│       │   ├── __init__.py
│       ├── __init__.py
│       ├── offPolicy
│       │   ├── baseOffPolicy.py
│       │   ├── core.py
│       │   ├── norm.py
│       ├── sac_auto--------------------SAC with auto adjust alpha parameter version.
│       │   ├── core.py
│       │   ├── __init__.py
│       │   ├── sac_auto_class.py
│       │   ├── sac_auto_per_class.py
│       │   └── sac_auto_per_her.py
│       ├── sac_sp--------------------SAC with alpha=0.2 version.
│       │   ├── core.py
│       │   ├── __init__.py
│       │   ├── SAC_class.py
│       │   ├── SAC_per_class.py
│       │   ├── SAC_per_her.py
│       │   ├── SAC_sp.py
│       └── td3_sp
│           ├── core.py
│           ├── __init__.py
│           ├── TD3_class.py
│           ├── TD3_per_class.py
│           ├── TD3_per_her_class.py
│           ├── TD3_per_her.py
│           ├── TD3_sp.py
├── arguments.py-----------------------hyperparams scripts
├── drlib_tree.txt
├── HER_DRLib_exps---------------------demo exp logs
│   ├── 2021-02-21_HER_TD3_FetchPush-v1
│   │   ├── 2021-02-21_18-26-08-HER_TD3_FetchPush-v1_s123
│   │   │   ├── checkpoint
│   │   │   ├── config.json
│   │   │   ├── params.data-00000-of-00001
│   │   │   ├── params.index
│   │   │   ├── progress.txt
│   │   │   └── Script_backup.py
├── memory
│   ├── __init__.py
│   ├── per_memory.py--------------mofan version
│   ├── simple_memory.py-----------mofan version
│   ├── sp_memory.py---------------spinningup tf1 version, simple uniform buffer memory class.
│   ├── sp_memory_torch.py---------spinningup torch-gpu version, simple uniform buffer memory class.
│   ├── sp_per_memory.py-----------spinningup tf1 version, PER buffer memory class.
│   └── sp_per_memory_torch.py
├── pip_requirement.txt------------pip install requirement, exclude mujoco-py,gym,tf,torch.
├── spinup_utils-------------------some utils from spinningup, about ploting results, logging, and so on.
│   ├── delete_no_checkpoint.py----delete the folder where the experiment did not complete.
│   ├── __init__.py
│   ├── logx.py
│   ├── mpi_tf.py
│   ├── mpi_tools.py
│   ├── plot.py
│   ├── print_logger.py------------save the information printed by the terminal to the local log file。
│   ├── run_utils.py---------------now I haven't used it. I have to learn how to multi-process.
│   ├── serialization_utils.py
│   └── user_config.py
├── train_tf1.py--------------main.py for tf1
└── train_torch.py------------main.py for torch

4. HER introduction:

the achievement of HER is based on the following code :

1. It can be converged, but this code is too difficult. https://github.com/openai/baselines

2. It can also converged, but only for DDPG-torch-cpu. https://github.com/sush1996/DDPG_Fetch

3. It can not be converged, but this code is simpler. https://github.com/Stable-Baselines-Team/stable-baselines

• paper: https://arxiv.org/pdf/1709.10089.pdf

4.1. My understanding and video:

种瓜得豆来解释her: 第一步在春天（state），种瓜（origin-goal）得豆，通过HER，把目标换成种豆，按照之前的操作，可以学会在春天种豆得豆； 第二步种米得瓜，学会种瓜得瓜； 即只要是智能体中间经历过的状态，都可以当做它的目标，进行学会。 即如果智能体能遍历所有的状态空间，那么它就可以学会达到整个状态空间。

论文分析视频：https://www.bilibili.com/video/BV1BA411x7Wm

代码分析文档：https://github.com/kaixindelele/DRLib/blob/main/algos/pytorch/offPolicy/HER_introduction.md

4.2. Key tricks for HER:

1. state-normalize: success rate from 0 to 1 for FetchPush-v1 task.
2. Q-clip: success rate from 0.5 to 0.7 for FetchPickAndPlace-v1 task.
3. action_l2: little effect for Push task.

4.3. Performance about HER-DDPG with FetchPush-v1:

5. PER introduction:

refer to:off-policy全系列（DDPG-TD3-SAC-SAC-auto）+优先经验回放PER-代码-实验结果分析

6. Summary：

这个库我封装了好久，整个代码库简洁、方便、功能比较齐全，在环境配置这块几乎是手把手教程，希望能给大家节省一些时间~

从零开始配置，不到两小时，从下载代码库，到配置环境，到在自己的环境中跑通，全流程非常流畅。

6.1. 下一步添加的功能：

1. PPO的封装；---PPO不封装了！机械臂操作不用PPO~

2. DQN的封装；---这个好像用的人也不多，放弃了~

3. 多进程的封装；

4. ExperimentGrid的封装；

7. Contact：

深度强化学习-DRL：799378128

欢迎关注知乎帐号：未入门的炼丹学徒

CSDN帐号：https://blog.csdn.net/hehedadaq

Star History

欢迎引用！

@inproceedings{luo2023d2sr, title={D2SR: Transferring dense reward function to sparse by network resetting}, author={Luo, Yongle and Wang, Yuxin and Dong, Kun and Liu, Yu and Sun, Zhiyong and Zhang, Qiang and Song, Bo}, booktitle={2023 IEEE International Conference on Real-time Computing and Robotics (RCAR)}, pages={906--911}, year={2023}, organization={IEEE} }