Experiments of ML with Tensorflow

These are the experiments that I replicated:

• https://github.com/googlecodelabs/tensorflow-for-poets-2
• https://github.com/devsisters/DQN-tensorflow

Base image:

• docker pull nvidia/cuda:8.0-cudnn5-runtime-ubuntu16.04

Note:

• I have Ubuntu 16.04 and an Nvidia card gtx 1050
• If you don't have a GPU graphic card, you can start from the ubuntu 16.04 image and use docker instead of nvidia-docker. Not much else has to be changed

Start a container:

• nvidia-docker run -it --name ai-dqn -p 6006:6006 -v /tmp/.X11-unix:/tmp/.X11-unix -v /docker-mounts:/docker-mounts -e DISPLAY=unix$DISPLAY --device /dev/dri --device /dev/snd nvidia/cuda:8.0-cudnn5-runtime-ubuntu16.04 bash

• nvidia-docker run -it --name ai-dqn -p 6006:6006 -v /tmp/.X11-unix:/tmp/.X11-unix -v /docker-mounts:/docker-mounts -e DISPLAY=unix$DISPLAY --device /dev/dri --device /dev/snd nvidia/cuda:9.1-cudnn7-runtime-ubuntu16.04 bash

• port 6006 is the standard Tensorboard portport 6006 is the standard Tensorboard port

Install - utilities

• apt-get update
• apt-get install vim
• apt-get install git
• apt-get upgrade

Install - Tensorflow (https://www.tensorflow.org/install/install_linux)

• apt-get install python-pip python-dev
• pip install --upgrade pip
• apt-get install libcupti-dev (only for GPU)
• pip install tensorflow or pip install tensorflow-gpu (for GPU)
  • pypa/pip#5240

Install - DQN (https://github.com/devsisters/DQN-tensorflow.git)

• Read the installation info from here if you need additional info: https://github.com/devsisters/DQN-tensorflow. Next are the commands I have executed
  • cd  - mkdir prj  - cd prj  - git clone https://github.com/devsisters/DQN-tensorflow.git  - pip install tqdm gym[all]    - If ERROR: openai/gym#218
    • apt-get install xvfb libav-tools xorg-dev libsdl2-dev swig cmake
    • re-run: pip install tqdm gym[all] // Just to be sure
  • execution of "python main.py --env_name=Breakout-v0 --is_train=True --display=False"    - If error: GPU (default is GPU on)      - Disable GPU in main.py - flags.DEFINE_boolean('use_gpu', False, 'Whether to use gpu or not')
    • If error: Memory
      • Change memory usage in config.py
    • If error: AttributeError: 'TimeLimit' object has no attribute 'ale'
      • pip install gym==0.7.3 (Note: downgrade)
    • Now the command runs OK
  • execution of "python main.py --env_name=Breakout-v0 --is_train=True --display=True"
    • If error: pyglet.canvas.xlib.NoSuchDisplayException: Cannot connect to "None"
      • Set these params to docker
        • "docker run ... -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY=unix$DISPLAY --device /dev/dri --device /dev/snd ..."
    • If error: Visualization problem - devsisters/DQN-tensorflow#35
      • pip install atari-py==0.0.21 (Note: downgrade)
    • Now the command runs OK

Save an image of this container:

• nvidia-docker commit ai-dqn stefanutti/cuda:8.0-cudnn5-runtime-ubuntu16.04-tf-dqn-1.2
• nvidia-docker rm ai-dqn
• nvidia-docker run -it --name ai-dqn -p 6006:6006 -v /tmp/.X11-unix:/tmp/.X11-unix -v /docker-mounts:/docker-mounts -e DISPLAY=unix$DISPLAY --device /dev/dri --device /dev/snd stefanutti/cuda:8.0-cudnn5-runtime-ubuntu16.04-tf-dqn-1.2 bash

Other useful commands:

• nvidia-docker start ai-dqn
• nvidia-docker exec -it ai-dqn bash
• nvidia-docker stop ai-dqn

Extra

• tensorboard --logdir training_summaries &
• python main.py --env_name=Breakout-v0 --is_train=True --display=True

Bye