DeepRL

How to create a python working environment:

virtualenv --system-site-packages deeprl-python (not recommented in python 3) source ./deeprl-python/bin/activate pip install --upgrade pip # check if there is a newer version. This link is only for 0.8

CPU

pip install --upgrade https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow-0.8.0-cp27-none-linux_x86_64.whl

GPU

pip install --upgrade https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-0.8.0-cp27-none-linux_x86_64.whl pip install jupyter pip install gym[all]

How to add an RSA-public key to a remote computer:

ssh-copy-id -i id_rsa.pub [email protected]

then add to ~/.ssh/config something like

Host host01 Hostname hostname01.example.com User user21

How to configure CUDA and cuDNN for tensorflow 0.8

1. make sure cuda-7.5 is installed

if cuDNN is already installed

2. add to .bashrc
	 export PATH=/usr/local/cuda-7.5/bin:$PATH
	 export LD_LIBRARY_PATH=/usr/local/cuda-7.5/lib64:$LD_LIBRARY_PATH

else

2. get a free developer account to download cuDNN and download it
3. move it to the remote computer and untar
	$ scp cudnn.tgz host01:~/
	$ tar -xvzf cudnn.tgz
4. add to .bashrc
	 export PATH=/usr/local/cuda-7.5/bin:$PATH
	 export LD_LIBRARY_PATH=~/cuda/lib64:$LD_LIBRARY_PATH

If CUDA is not installed, install it to any directory (no root needed) and add it to the path variables.

How to get remote access to Jupyter notebook and Tensorboard

* get them launched in the host
	jupyter-notebook --no-browser
	tensorboard --logdir=./wherever/your/data/is
* tunnel from the client
	# jupyter
	ssh -NL <local_port>:localhost:8888 host01
	# tensorboard
	ssh -NL <other_local_port>:0.0.0.0:6006
* access through your browser with localhost:<local_port> and localhost:<other_local_port>

How to renew kerberos credentials in one line

while true; do echo "password" | kinit; while true; do krenew; if [ $? -ne 0 ]; then break; fi; sleep 30m; done; done

DQN from_pixels

A small modification of OpenAI's gym is required for efficiently obtaining images from the classic environments avoiding on-screen rendering. The changes can be found in https://github.com/garibarba/gym/commit/69c58d91d64cf3b28c44077bc30c599ed354af1e. Then a minimal change must be done in each environment in order to call the newly defined functions instead of the regular one.

xvfs, GLX and nvidia drivers

These don't play well with each other. A solution is possible but requires reinstalling the nvidia drivers: https://davidsanwald.github.io/ec2-openAI-gym-tensorflow-GPU-cuda-deep-learning.html#ec2-openAI-gym-tensorflow-GPU-cuda-deep-learning