Skip to content

Image classification using reinforcement learning and multi-agent system

License

Notifications You must be signed in to change notification settings

Ipsedo/MARLClassification

Repository files navigation

MARLClassification

Multi-Agent Image Classification via Reinforcement Learning

TODO

  • train on AID
  • docstring

Results

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

Loss Train Eval
Epoch 0 1.5161 41%, 42% 44%, 44%
Epoch 20 0.5385 80%, 80% 80%, 79%
Epoch 39 0.5218 81%, 81% 82%, 81%

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

Loss Train Eval
prec, rec prec, rec
Epoch 0 2.6236 21%, 26% 25%, 26%
Epoch 20 1.3262 57%, 57% 58%, 57%
Epoch 49 0.9263 68%, 68% 66%, 65%

Installation

$ cd /path/to/MARLClassification
$ # create and activate your virtual env
$ python -m venv venv
$ ./venv/bin/activate
$ # install requirements
$ pip install -r requirements.txt
$ # download datasets using sh scripts in resources folder, ex : MNIST
$ ./resources/download_mnist.sh

You may download datasets with bash scripts in res 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 --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 --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 --nlb 320 --nla 320 --nb-epoch 50 --lr 1e-4 -o ./out/aid_actor_critic

Reference

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