A3C
My daily used A3C implementaion. Based on openai's A3C implementation(https://github.com/openai/universe-starter-agent.git). I make some tiny changes to fit my working preferences.
environment
To run this algo, I use tmux. I use python3.
coding
- Typically I don't change
a3c.pyandworker.py. - If I want to change the model, I change
model.py. - If I want to handle the env stuff, I change
envs.py. - If I want to handle the hyper-parameters, I also change
envs.py. - I hard-code almost every argument in
envs.py, so that I can easily start the program in CLI without too much typing.
running
To start training with default num-workers(=4), just run below in CLI:
./do.a3c.sh
To start training with 32 workers:
./do.a3c.sh -w 32
To stop training:
./no.a3c.sh
log archiving
python ./ziplog.py
Versioning when tuning
I use VSTR in envs.py to identify my hyper-parameter and algo version info. I use git so I can easily check out
detail of every version. I use ziplog.py to archive tuning result.