CoNAL: Learning Conflict-Noticed Architecture for Multi-Task Learning

This repository includes the pytorch implementation of "Learning Conflict-Noticed Architecture for Multi-Task Learning".

Citation

If you find CoNAL useful for research or development, please cite our paper:

@article{CoNAL,
  title={Learning Conflict-Noticed Architecture for Multi-Task Learning},
  author={Zhixiong Yue, Yu Zhang, Jie Liang},
  journal={Proceedings of the 2023 National Conference of the American Association for Artificial Intelligence (AAAI2023)},
  year={2023}
}

Setup Environment

Please install following python packages:

- python
- numpy
- pytorch
- torchvision
- tensorboard

or install from requirements:

pip install -r requirements.txt

Example Usage

Learn MTL architecture in NYUv2 dataset:

python CoNAL/search.py

Then retrain the learned architecture:

python CoNAL/train.py

Visualization

We use tensorboard to visualize the architecture learning process

tensorboard --logdir logs/logdir

The learned architecture can be found in logs/logdir/arch.json