Skip to content

Implementation of DrQ-v2 with an improved replay buffer and better information tracking.

Notifications You must be signed in to change notification settings

JankowskiChristopher/DrQ-v2plus

Repository files navigation

DrQ-v2 with improved replay buffer

can be run on a 12GB GPU and 16GB of RAM!

Running experiments

Locally run script

rsync.sh

Or with option -n (see description below)

On Entropy cluster start the virtual environment:

source setup.sh

If you need to install (this won't probably activate venv) requirements run:

setup.sh -i 

For some reasons this time without source. And run the job script:

srun_experiment.sh

Important commands

Entropy

To see the current queue of jobs run (either version):

squeue
entropy_queue
entropy_my_queue

To see the stats and memory allocated by jobs run:

sacct --format='JobID,Elapsed,MaxVMSize'

Python path

It's possible that you will have to specify the Python path. cd into the project directory (distributional-sac) and run:

export PYTHONPATH="$PYTHONPATH:$PWD"

This line is also added to the setup.sh script, so probably no need to run it manually.

Linux screen

Srun script uses the screen functionality, to run the jobs in the background even when the terminal is closed. To see the list of screens run:

screen -ls

To attach to a screen run:

screen -r <screen_name>

To detach from a screen press Ctrl+A and then D. To kill all screens see script kill_screens.sh below.

Useful scripts

When running the scripts for the first time remember to run:

chmod u+x <script_name>.sh

setup.sh

Scripts sets up the environment on the cluster and optionally installs the dependencies. Note: in order to work must be used with source

source setup.sh

And to create (if does not already exist) and activate the virtual environment with installation from requirements.txt:

source setup.sh -i

rsync.sh

Rsync sends the code to the cluster. Remember to change the paths to your username in the script When used without flags it copies the whole src to the cluster under directory src:

./rsync.sh

When used with flag -n it copies the src but puts it inside an experiment folder with a timestamp.

./rsync.sh -n

srun_experiment.sh

Script runs the experiment on the cluster by creating a screen and later detaching it. The results of srun are copied from stdout and stderr to a file called results.txt that is stored in the directory of the experiment. Usage with default options (change them for your defaults) is:

./srun_experiment.sh

You can specify the options (see help of the script) like this:

./srun_experiment.sh -n <name e.g. common without ""> -q <qos e.g. student> -g <e.g. titanv:2> -t <time e.g. 1-0 meaning 1 day> -m <path to main.py>

kill_screens.sh

Script kills all the detached screens.

./kill_screens.sh

Singularity

Installation

https://www.linuxwave.info/2022/02/installing-singularity-in-ubuntu-2004.html

Build image

sudo singularity build <image_name.sif> <image_file.def>

About

Implementation of DrQ-v2 with an improved replay buffer and better information tracking.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published