https://gijskoning.github.io/ReproducingCURL/
This repository is an assignment of the Deep Learning course given by the TUDelft.
It reproduces the paper "Influence-aware Memory Architectures for Deep Reinforcement Learning" written by Miguel Suau (link).
The code for the original paper was developed in Tensorflow (link). This project recreates it in Pytorch and has some additional experiments and evaluations.
The model will be evaluated using the PPO algorithm retrieved from: https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail
Currently, the code is a work in progress and is only a not working copy of the tensorflow one.
The Warehouse environment included in the code of the paper. Todo the hyperparameter are not yet correct! Currently these parameters are correct:
- learning rate, num epoch/num-steps, value coeff/value-loss-coef, entropy/entropy-coef, clip/clip-param, batch size/num-mini-batch.
- The default of GAE delta is 0.95 and thus the same as in the paper.
- The default for gamma/discount is 0.99 which is also the same as in the paper.
Also not sure if value coeff is the same as value-loss-coef. Previously in the PPO algo it was set to 0.5 as default. The paper does 1. Next to that I am not sure if num workers is equal to num-steps. num-steps equal to 3 doesn't actually work for 8 workers.
Use --num-processes 2 --num-steps 16 when low on memory Warehouse with FNN
python main.py --env-name Warehouse --yaml-file FNN --fnn-hidden-sizes 640,256 --num-processes 8 --num-steps 8 --num-mini-batch 32 --log-interval 100 --use-linear-lr-decay --entropy-coef 0.01 --algo ppo --use-gae --lr 2.5e-4 --clip-param 0.1 --value-loss-coef 1
Warehouse with IAM network:
python main.py --env-name Warehouse --yaml-file IAM_static --fnn-hidden-sizes 512,256 --rec-hidden-size 128 --recurrent-policy --num-processes 8 --num-steps 8 --num-mini-batch 32 --log-interval 100 --use-linear-lr-decay --entropy-coef 0.01 --algo ppo --use-gae --lr 2.5e-4 --clip-param 0.1 --value-loss-coef 1
Preferably you use Pytorch with Cuda enabled but this requires a special version. To check your Cuda version (if installed) execute
this command in cmd.exe: nvcc --version
You can find the correct command of Pytorch with the specific Cuda version here: https://pytorch.org/
The command to be executed for Pip packages could look like this (only torch is needed): pip install torch==1.8.0+cu111 -f https://download.pytorch.org/whl/torch_stable.html
There is a bug in atari_py. If you already installed it remove by doing pip uninstall atari_py' and install this version: pip install -f https://github.com/Kojoley/atari-py/releases atari_py`