FLOWER:

This directory contain our implementation of our paper entitled ”Few-Shot Continual Learning via Flat-to-Wide Approaches" (FLOWER) This project is developed based on F2M Project with modification and addition of our methods.

Code:

Pleasee see "code" directory to see the implementation of the algorithms.

Logs:

Pleasee see "logs" directory to see the raw results of our experiments.

System Requirements:

This project is run by using the following environment:

1. NVIDIA NGC container 21.03-py3: The container contains python 3.8, pytorch 1.9.0, torchvision, numpy, and the other libraries. Please read the fllowing link for the complete documentation: https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel_21-03.html#rel_21-03
2. wandb (optional) You can install wandb by using command "pip install wandb" on top of the container

Datasets

We evaluate our system in the benchmark datasets, including CUB-200-2011, CIFAR100, miniImageNet. Please download CUB-200-2011, CIFAR100 and miniImageNet.

Example of Running Script:

Please see file example_running_script.txt in code directory

Contribution (please see code directory):

The main contribution of our work is as follow:

1. Base task learning (please see *_model.py methods directory): F2MJ (Flat minima + projection), F2MMAS (Flat minima + MAS weight importance)+ F2MMASJ (FLOWER base task learning), F2MMASJNOPL (FLOWER base task learning without prototype loss), F2MSI Flat minima + SI weight importance). Those files are developed from F2MModel with necessary modifications.

2. Continual tasks learning (please see *_model.py methods directory): flower1 (FLOWER continual tasks learning), flower_no_psi (FLOWER without projection), flower_no_fm (FLOWER without flat minima), flower_no_ball (FLOWER without ball augmentation) flower_no_mas (FLOWER without MAS weight importance), flower_no_pl (FLOWER without prototype loss)

3. Incremental learning procedure without memory: incremental_procedure_nomem.py The file is developed from F2MModel with necessary modifications.

4. Minor modification: We modify incremental learning procedure script (incremental_procedure.py) for debugging purpose.

LICENSE

This FLOWER project is released under the Apache 2.0 license. Please see code/LICENSE directory.