Awesome

POD-AANets

[Paper] [Original GitHub Repository for AANets]

See the original GitHub repository for AANets here: https://git.io/JYHyt

We are still cleaning up the code for PODNet w/ AANets. If you need to use it now, here is a preliminary version.

Installing Requirements

You need to install the same environment as PODNet to run this code. You may see a list of the requirements here.

Preparing Datasets

Our code uses exactly the same dataset splits as PODNet.

You need to put the data of ILSVRC2012 under this folder: https://github.com/yaoyao-liu/POD-AANets/tree/main/data/imagenet

The folders for the training and validation data should be named as train and val, respectively.

Running Experiments

You may run the experiments on ImageNet-Subset using the following command:

python run_exp.py

Results on ImageNet-Subset (5-phase)

You may see the log file here.

	Avg.	Phase 0	Phase 1	Phase 2	Phase 3	Phase 4	Phase 5
`PODNet`	75.9	86.4	80.6	76.3	74.7	69.8	67.6
`POD-AANets (paper)`	77.0	86.1	81.5	77.7	75.5	71.6	69.4
`POD-AANets (this repo)`	77.3	87.0	82.2	78.5	75.5	71.6	69.0

Reporting Issues

If you have any questions, please create an issue in the original repository. Thanks!

Citation

Please cite our paper if it is helpful to your work:

@inproceedings{Liu2020AANets,
  author    = {Liu, Yaoyao and Schiele, Bernt and Sun, Qianru},
  title     = {Adaptive Aggregation Networks for Class-Incremental Learning},
  booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {2544-2553},
  year      = {2021}
}

Acknowledgements

Our implementation uses the source code from the following repository:

PODNet: Pooled Outputs Distillation for Small-Tasks Incremental Learning