DQN

This implements training of DQN on the game BreakoutNoFrameskip-v4, mainly modified from pytorch/examples.

DQN Detail

Deep Q-Learning (DQN) combines the method of neural network and Q learning.

Requirements

• Install PyTorch and torchvision(pytorch.org)

• 安装requirements.txt里面要求的依赖

• 配置mujoco

  从https://www.roboti.us/license.html上点击Activation key下载mjkey.txt

  从https://www.roboti.us/download/mjpro150_linux.zip下载mjpro150_linux.zip

  在root目录下创建隐藏文件夹.mujoco(不要忘记带.)，并将mjpro150_linux.zip安装包解压到这个文件夹下

  将mjkey.txt移动到~/.mujoco 和 ~/.mujoco/mjpro150/bin下

  添加环境变量, 打开～/.bashrc文件,添加以下指令

  export LD_LIBRARY_PATH=~/.mujoco/mjpro150/bin${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

  export MUJOCO_KEY_PATH=~/.mujoco${MUJOCO_KEY_PATH}

• baselines安装

  git clone https://github.com/openai/baselines.git

  cd baselines

  pip install -e .

Training

To train a model, run train_dqn.py with the desired model architecture:

# training 1p accuracy
bash test/train_full_1p.sh 

# training 1p performance
bash test/train_performance_1p.sh

# training 1p eval
bash test/train_eval_1p.sh --pth_path=data/DQNAgent-train_full_1p-xx.model ---status_path=data/DQNAgent-train_full_1p-xx.stats

Log path:

test/output/{device_id}/train_full_1p_{device_id}.txt

test/output/{device_id}/train_performance_1p_{device_id}.txt

test/output/{device_id}/train_eval_1p_{device_id}.txt

DQN training result

Acc@1	FPS	Npu_nums	steps	AMP_Type
-	62.9 step/s	1	80000	O1
99.3	-	1	5000000	O1

Statement

For details about the public address of the code in this repository, you can get from the file public_address_statement.md