Open a new kernel, connect to GPU runtime, and run:
!pip install https://github.com/openai/baselines/archive/9ee399f5b20cd70ac0a871927a6cf043b478193f.zip
!pip install tensorflow==1.15.0 mpi4py==3.0.3 gym==0.15.4
!git clone https://github.com/openai/procgen.git
!git clone https://github.com/MandiZhao/train-procgen.git
%cd procgen
!pip install -e .
%cd ..
%cd train-procgen
!pip install -e .
Then all scripts can be run from here.
Need to export variables:
first confirm /usr/local/cuda-10.0/lib64 exists (or other versions of cuda, e.g. cuda-5.0), then
export LD_LIBRARY_PATH=/usr/local/cuda-10.0/lib64:/usr/local/cuda-10.0/lib
if the desired GPU has index 0, do:
export CUDA_VISIBLE_DEVICES=0
As a reference, output from current conda env that uses GPU:
(train-procgen)$ echo $LD_LIBRARY_PATH
/home/mandi/.local/bin:/home/mandi/bin:/home/mandi/miniconda3/envs/train-procgen/bin:/home/mandi/miniconda3/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/ussr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin:/home/mandi/bin:/home/mandi/bin:/home/mandi/lib:/usr/local/cuda-10.0/lib64:/usr/local/cuda-10.0/lib
(train-procgen) $ echo $CUDA_VISIBLE_DEVICES
0
- To run a training experiments with 50 avaliable levels for 1M steps and log every 30 steps, use cross-cut method for input processing on top of the baseline PPO algorithm, and saved the model with idx 1:
$ conda activate train-procgen
$(train-procgen) python train_procgen/train_crop.py -id 1 --use "cross" --num_levels 50 --total_tsteps 1000000 \
--log_interval 30
## If you want to use taskset to limit CPU usage, say only use CPU index 0-5, do
## taskset -c 0-5 python train_procgen/train_recenter.py --nupdates 0 -id 0
## same for commands below
- To load a partially trained random-crop agent at index 4, continue training it for 2M steps, and save at idx 5:
$(train-procgen) python train_procgen/train_crop.py -id 5 --load_id 4 --use "randcrop" --num_levels 50 --total_tsteps 2000000
- To test a trained random-random-cut agent at index 3 on a set of {100, 1000, 2000, ..., 95000} level intervals: (it's easier to put these blocks of commands in a bash .sh file and run from command line)
LOAD=3
USE="randcuts"
taskset -c 0-20 python train_procgen/test_select.py --start_level 100 -id 0 --load_id ${LOAD} --use ${USE}
taskset -c 0-20 python train_procgen/test_select.py --start_level 1000 -id 1 --load_id ${LOAD} --use ${USE}
taskset -c 0-20 python train_procgen/test_select.py --start_level 5000 -id 2 --load_id ${LOAD} --use ${USE}
taskset -c 0-20 python train_procgen/test_select.py --start_level 10000 -id 3 --load_id ${LOAD} --use ${USE}
taskset -c 0-20 python train_procgen/test_select.py --start_level 20000 -id 4 --load_id ${LOAD} --use ${USE}
taskset -c 0-20 python train_procgen/test_select.py --start_level 30000 -id 5 --load_id ${LOAD} --use ${USE}
taskset -c 0-20 python train_procgen/test_select.py --start_level 40000 -id 6 --load_id ${LOAD} --use ${USE}
taskset -c 0-20 python train_procgen/test_select.py --start_level 50000 -id 7 --load_id ${LOAD} --use ${USE}
taskset -c 0-20 python train_procgen/test_select.py --start_level 60000 -id 8 --load_id ${LOAD} --use ${USE}
taskset -c 0-20 python train_procgen/test_select.py --start_level 70000 -id 9 --load_id ${LOAD} --use ${USE}
taskset -c 0-20 python train_procgen/test_select.py --start_level 80000 -id 10 --load_id ${LOAD} --use ${USE}
taskset -c 0-20 python train_procgen/test_select.py --start_level 90000 -id 11 --load_id ${LOAD} --use ${USE}
taskset -c 0-20 python train_procgen/test_select.py --start_level 95000 -id 12 --load_id ${LOAD} --use ${USE}
Original train-procgen README:
Status: Archive (code is provided as-is, no updates expected)
This is code for training agents for some of the experiments in Leveraging Procedural Generation to Benchmark Reinforcement Learning (citation). The code for the environments is in the Procgen Benchmark repo.
Supported platforms:
- macOS 10.14 (Mojave)
- Ubuntu 16.04
Supported Pythons:
- 3.7 64-bit
You can get miniconda from https://docs.conda.io/en/latest/miniconda.html if you don't have it, or install the dependencies from environment.yml manually.
git clone https://github.com/openai/train-procgen.git
conda env update --name train-procgen --file train-procgen/environment.yml
conda activate train-procgen
pip install https://github.com/openai/baselines/archive/9ee399f5b20cd70ac0a871927a6cf043b478193f.zip
pip install -e train-procgen
Please cite using the following bibtex entry:
@article{cobbe2019procgen,
title={Leveraging Procedural Generation to Benchmark Reinforcement Learning},
author={Cobbe, Karl and Hesse, Christopher and Hilton, Jacob and Schulman, John},
journal={arXiv preprint arXiv:1912.01588},
year={2019}
}