Auto-Lambda

This repository contains the source code of Auto-Lambda and baselines from the paper, Auto-Lambda: Disentangling Dynamic Task Relationships.

We encourage readers to check out our project page, including more interesting discussions and insights which are not covered in our technical paper.

Multi-task Methods

We implemented all weighting and gradient-based baselines presented in the paper for computer vision tasks: Dense Prediction Tasks (for NYUv2 and CityScapes) and Multi-domain Classification Tasks (for CIFAR-100).

Specifically, we have covered the implementation of these following multi-task optimisation methods:

Weighting-based:

• Equal - All task weightings are 1. --weight equal
• Uncertainty - https://arxiv.org/abs/1705.07115 --weight uncert
• Dynamic Weight Average - https://arxiv.org/abs/1803.10704 --weight dwa
• Auto-Lambda - Our approach. --weight autol

Gradient-based:

• GradDrop - https://arxiv.org/abs/2010.06808 --grad_method graddrop
• PCGrad - https://arxiv.org/abs/2001.06782 --grad_method pcgrad
• CAGrad - https://arxiv.org/abs/2110.14048 --grad_method cagrad

Note: Applying a combination of both weighting and gradient-based methods can further improve performance.

Datasets

We applied the same data pre-processing following our previous project: MTAN which experimented on:

• NYUv2 [3 Tasks] - 13 Class Segmentation + Depth Estimation + Surface Normal. [288 x 384] Resolution.
• CityScapes [3 Tasks] - 19 Class Segmentation + 10 Class Part Segmentation + Disparity (Inverse Depth) Estimation. [256 x 512] Resolution.

Note: We have included a new task: Part Segmentation for CityScapes dataset. The pre-processing file for CityScapes has also been included in the dataset folder.

Experiments

All experiments were written in PyTorch 1.7 and can be trained with different flags (hyper-parameters) when running each training script. We briefly introduce some important flags below.

Training Auto-Lambda in Multi-task / Auxiliary Learning Mode:

python trainer_dense.py --dataset [nyuv2, cityscapes] --task [PRIMARY_TASK] --weight autol --gpu 0   # for NYUv2 or CityScapes dataset
python trainer_cifar.py --subset_id [PRIMARY_DOMAIN_ID] --weight autol --gpu 0   # for CIFAR-100 dataset

Training in Single-task Learning Mode:

python trainer_dense_single.py --dataset [nyuv2, cityscapes] --task [PRIMARY_TASK]  --gpu 0   # for NYUv2 or CityScapes dataset
python trainer_cifar_single.py --subset_id [PRIMARY_DOMAIN_ID] --gpu 0   # for CIFAR-100 dataset

Note: All experiments in the original paper were trained from scratch without pre-training.

Benchmark

For standard 3 tasks in NYUv2 (without dense prediction task) in the multi-task learning setting with Split architecture, please follow the results below.

Note: The results were averaged across three random seeds. You should expect the error range less than +/-1%.

Citation

If you found this code/work to be useful in your own research, please considering citing the following:

@article{liu2022auto-lambda,
  title={Auto-Lambda: Disentangling Dynamic Task Relationships},
  author={Liu, Shikun and James, Stephen and Davison, Andrew J and Johns, Edward},
  journal={arXiv preprint arXiv:2202.03091},
  year={2022}
}

Acknowledgement

We would like to thank @Cranial-XIX for his clean implementation for gradient-based optimisation methods.

Contact

If you have any questions, please contact [email protected].