DRLib：A concise deep reinforcement learning library, integrating HER and PER for almost off policy RL algos.

Last update: Jan 3, 2023

Related tags

Deep Learning DRLib

Overview

DRLib：A concise deep reinforcement learning library, integrating HER and PER for almost off policy RL algos

A concise deep reinforcement learning library, integrating HER and PER for almost off policy RL algos. With tensorflow1.14 and pytorch, add HER and PER, core codes based on https://github.com/openai/spinningup

Compared with spinning up, I delete multi-process and experimental grid wrapper, and our advantage is that it is convenient to debug with pycharm~

项目特点：

tf1和pytorch两个版本的算法，前者快，后者新，任君选择；
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；
添加了HER和PER功能，非常适合做机器人相关任务的同学们；
去除了自动调参（ExperimentGrid）和多进程（MPI_fork）部分，适合新手在pycharm中debug，前者直接跑经常会报错~ 等我熟练了这两个，我再加上去，并附上详细教程；
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！
求三连，不行求个star！

1. Installation

Clone the repo and cd into it:

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

Create anaconda DRLib_env env:

conda create -n DRLib_env python=3.6.9
source activate DRLib_env

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
```

Install tensorflow-gpu=1.14.0

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

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)

Install mujoco and mujoco-py

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

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:

modify params in arguments.py, choose env, RL-algorithm, use PER and HER or not, gpu-id, and so on.
run with train_tf.py or train_torch.py
```
python train_tf.py
```
exp results to local:https://blog.csdn.net/hehedadaq/article/details/114045615
plot results:https://blog.csdn.net/hehedadaq/article/details/114044217

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, can choose whether or not HER and PER
│   │   │   ├── ddpg.py-------------It's copied directly from spinup
│   │   │   ├── __init__.py
│   │   ├── __init__.py
│   │   ├── offPolicy
│   │   │   ├── baseOffPolicy.py----baseOffPolicy, can be used to 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:

Refer to these code bases:

It can be converged, but this code is too difficult. https://github.com/openai/baselines
It can also converged, but only for DDPG-torch-cpu. https://github.com/sush1996/DDPG_Fetch
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。

https://www.bilibili.com/video/BV1BA411x7Wm

4.2. Key tricks for HER:

state-normalize: success rate from 0 to 1 for FetchPush-v1 task.
Q-clip: success rate from 0.5 to 0.7 for FetchPickAndPlace-v1 task.
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：

PPO的封装；
DQN的封装；
多进程的封装；
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

Comments

plot

could you upload some data to plot figure below(especially shaded regions)? the installation takes too much time, it is hard for me to get that license

opened by tinmodeHuang 4
训练结果显示

您好，感谢您的代码分享！我在学习您的代码，正在学习代码中spinningup_utils中绘制RL仿真结果的部分，我跑完了三个随机种子的实验，然后运行命令： python spinning_utils/plot.py ./HER_DRLib_exps --smooth=3 --count，结果在一个图中分别出现了三个随机种子的曲线；

然后我运行命令python spinning_utils/plot.py ./HER_DRLib_exps --smooth=3 想让代码绘制出三个随机种子的平均曲线（有阴影效果的那种），曲线却显示不出来，请问这是怎么回事啊？

opened by ZHQ-air 3
请问td3多进程中的代码

DRLib/algos/pytorch/td3_sp/MPI_td3_per_her.py

line 146 mpi_avg_grads(self.ac.v) # 请问这个代码是正确的嘛 ac.q1 , ac.q2 , q_params

或许说self.ac.v 在那里定义了嘛

谢谢

opened by namjiwon1023 2

当选用HER时，在存储轨迹的时候会出现维度不匹配的错误

十分感谢您的工作和无私的分享，代码非常棒！不过在运行代码时出现如题描述的错误，具体如下：

 File "D:\theory\experiment\DRLib-main\algos\pytorch\offPolicy\baseOffPolicy.py", line 114, in save_episode
    self.store_transition(transition=(obs_arr, action, reward, next_obs_arr, done))
  File "D:\theory\experiment\DRLib-main\algos\pytorch\offPolicy\baseOffPolicy.py", line 90, in store_transition
    self.replay_buffer.store(s, a, r, s_, done)
  File "D:\theory\experiment\DRLib-main\memory\sp_memory_torch.py", line 25, in store
    self.obs_buf[self.ptr] = obs
ValueError: could not broadcast input array from shape (31) into shape (28)

经调试，问题应该在这个函数中：

    def convert_dict_to_array(self, obs_dict,
                              exclude_key=['achieved_goal']):
        obs_array = np.concatenate([obs_dict[key] for key, value in obs_dict.items() if key not in exclude_key])
        return obs_array

经过这个函数后，obs的维度就会扩张，我想是不是exclude_key里是不是还要增加一个要排除掉的字符串，比如"grip_goals"

opened by theorynice 1

UnicodeDecodeError: 'gbk' codec can't decode byte oxad in position 2505 错误修改

File d\DRLib-main1\spinup_utils\logx.py", line 225 with open(root_dir,'r') as file_to_read: 修改为： with open(root_dir,'r',encoding='utf-8' ) as file_to_read:

opened by 1dSHallRed 1

DRLib：A concise deep reinforcement learning library, integrating HER and PER for almost off policy RL algos.

Related tags

Overview

DRLib：A concise deep reinforcement learning library, integrating HER and PER for almost off policy RL algos

项目特点：

1. Installation

2. Training models

3. File tree and introduction:

4. HER introduction:

4.1. My understanding and video:

4.2. Key tricks for HER:

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

5. PER introduction:

6. Summary：

6.1. 下一步添加的功能：

7. Contact：

Comments

plot

训练结果显示

请问td3多进程中的代码

当选用HER时，在存储轨迹的时候会出现维度不匹配的错误

UnicodeDecodeError: 'gbk' codec can't decode byte oxad in position 2505 错误修改

Owner

Pytorch implementations of popular off-policy multi-agent reinforcement learning algorithms, including QMix, VDN, MADDPG, and MATD3.

Implementation of Advantage-Weighted Regression: Simple and Scalable Off-Policy Reinforcement Learning

DiffQ performs differentiable quantization using pseudo quantization noise. It can automatically tune the number of bits used per weight or group of weights, in order to achieve a given trade-off between model size and accuracy.

PyTorch implementation of Advantage Actor Critic (A2C), Proximal Policy Optimization (PPO), Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation (ACKTR) and Generative Adversarial Imitation Learning (GAIL).

PyTorch implementation of Advantage Actor Critic (A2C), Proximal Policy Optimization (PPO), Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation (ACKTR) and Generative Adversarial Imitation Learning (GAIL).

A clear, concise, simple yet powerful and efficient API for deep learning.

PGPortfolio: Policy Gradient Portfolio, the source code of "A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem"(https://arxiv.org/pdf/1706.10059.pdf).

Official Pytorch implementation of ICLR 2018 paper Deep Learning for Physical Processes: Integrating Prior Scientific Knowledge.

Official Implementation of 'UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers' ICLR 2021(spotlight)

A concise but complete implementation of CLIP with various experimental improvements from recent papers

A concise but complete implementation of CLIP with various experimental improvements from recent papers

This is the unofficial code of Deep Dual-resolution Networks for Real-time and Accurate Semantic Segmentation of Road Scenes. which achieve state-of-the-art trade-off between accuracy and speed on cityscapes and camvid, without using inference acceleration and extra data

This project provides a stock market environment using OpenGym with Deep Q-learning and Policy Gradient.

Conservative Q Learning for Offline Reinforcement Reinforcement Learning in JAX

Reinforcement-learning - Repository of the class assignment questions for the course on reinforcement learning

Deep Learning and Reinforcement Learning Library for Scientists and Engineers 🔥

Deep Learning and Reinforcement Learning Library for Scientists and Engineers 🔥

MINIROCKET: A Very Fast (Almost) Deterministic Transform for Time Series Classification

Dark Finix: All in one hacking framework with almost 100 tools