MARLClassification

PyTorch implementation of Multi-Agent Image Classification via Reinforcement Learning paper.

TODO

• train on WorldStrat
• docstring

Installation

You need to have python (at least version 3.14) and uv executables in your PATH.

First you can clone the project :

git clone https://github.com/Ipsedo/MARLClassification.git

Then install dependencies :

cd /path/to/MARLClassification
# install dependencies
uv sync
# download datasets using sh scripts in resources folder, ex : MNIST
./resources/download_mnist.sh

You may download datasets with bash scripts in resources folder.

Usage

To run training :

cd /path/to/MARLClassification
# train on MNIST
python -m marl_classification -a 3 --step 5 --cuda --run-id train_mnist train --action [[1,0],[-1,0],[0,1],[0,-1]] --img-size 28 --nb-class 10 -d 2 --f 6 --ft-extr mnist --nb 64 --na 64 --nm 16 --nmo 24 --nd 8 --nlb 96 --nla 96 --batch-size 32 --lr 1e-3 --nb-epoch 40 -o ./out/mnist_actor_critic
# train on NWPU-RESISC45
python -m marl_classification -a 16 --step 16 --cuda --run-id train_resisc45 train --action [[1,0],[-1,0],[0,1],[0,-1]] --ft-extr resisc45 --batch-size 8 --nb-class 45 --img-size 256 -d 2 --nb 256 --na 256 --nd 16 --f 12 --nm 64 --nmo 96 --nlb 384 --nla 384 --nb-epoch 50 --lr 1e-4 -o ./out/resisc45_actor_critic
# train on AID
python -m marl_classification -a 16 --step 16 --cuda --run-id train_aid train --action [[3,0],[-3,0],[0,3],[0,-3]] --ft-extr aid --batch-size 8 --nb-class 30 --img-size 600 -d 2 --nb 256 --na 256 --nd 16 --f 24 --nm 64 --nmo 96 --nlb 320 --nla 320 --nb-epoch 50 --lr 1e-4 -o ./out/aid_actor_critic

Results

Training on MNIST (see resources/trained_models/mnist) :

Epoch 50
--------
[Train]
precision : 81.6%
recall    : 81.3%
--------
[Eval]
precision : 82.4%
recall    : 81.2%

Training on image dataset NWPU-RESISC45 (see resources/trained_models/resisc45) :

Epoch 50
--------
[Train]
precision : 71.4%
recall    : 71.8%
--------
[Eval]
precision : 68.7%
recall    : 67.8%

Training on image dataset AID (see resources/trained_models/aid) :

Epoch 50
--------
[Train]
precision : 80.9%
recall    : 80.6%
--------
[Eval]
precision : 73.7%
recall    : 72.5%

Training on image dataset World Strat (see resources/trained_models/world_strat) :

Epoch 50
--------
[Train]
precision : 65.8%
recall    : 53.9%
--------
[Eval]
precision : 55.6%
recall    : 55.8%

Reference

[1]: https://arxiv.org/abs/1905.04835, Hossein K. Mousavi, Mohammadreza Nazari, Martin Takáč, Nader Motee - 2019