The code repository for ["Few-Shot Class-Incremental Learning by Sampling Multi-Phase Tasks"](https://arxiv.org/abs/2203.17030) (TPAMI 2022) in PyTorch. If you use any content of this repo for your work, please cite the following bib entry:

@article{zhou2022few,

title={Few-Shot Class-Incremental Learning by Sampling Multi-Phase Tasks},

author={Zhou, Da-Wei and Ye, Han-Jia and Zhan, De-Chuan},

journal={arXiv preprint arXiv:2203.17030},

year={2022}

}

New classes arise frequently in our ever-changing world, e.g., emerging topics in social media and new types of products in e-commerce. A model should recognize new classes and meanwhile maintain discriminability over old classes. Under severe

circumstances, only limited novel instances are available to incrementally update the model. The task of recognizing few-shot new classes without forgetting old classes is called few-shot class-incremental learning (FSCIL). In this work, we propose a new paradigm for FSCIL based on meta-learning by LearnIng Multi-phase Incremental Tasks (LIMIT), which synthesizes fake FSCIL tasks from the base dataset. The data format of fake tasks is consistent with the ‘real’ incremental tasks, and we can build a generalizable feature space for the unseen tasks through meta-learning. Besides, LIMIT also constructs a calibration module based on transformer, which calibrates the old class classifiers and new class prototypes into the same scale and fills in the semantic gap. The calibration module also adaptively

contextualizes the instance-specific embedding with a set-to-set function. LIMIT efficiently adapts to new classes and meanwhile resists forgetting over old classes. Experiments on three benchmark datasets (CIFAR100, miniImageNet, and CUB200) and large-scale dataset, i.e., ImageNet ILSVRC2012 validate that LIMIT achieves state-of-the-art performance.

<img src='imgs/teaser.png' width='1000' height='300'>

<img src='imgs/result.png' width='900' height='778'>

Please refer to our [paper](https://arxiv.org/abs/2203.17030) for detailed values.

The following packages are required to run the scripts:

- [PyTorch-1.4 and torchvision](https://pytorch.org)

- tqdm

- Download the [pretrained models](https://drive.google.com/drive/folders/11liacLA6F4uz_SWAPLYQshlzLMXMzDkB?usp=sharing) and put them in ./params

We provide the source code on three benchmark datasets, i.e., CIFAR100, CUB200 and miniImageNet. Please follow the guidelines in [CEC](https://github.com/icoz69/CEC-CVPR2021) to prepare them.

There are four parts in the code.

- `models`: It contains the backbone network and training protocols for the experiment.

- `data`: Images and splits for the data sets.

- `dataloader`: Dataloader of different datasets.

- `checkpoint`: The weights and logs of the experiment.

- `params`: Pretrained model weights.

- Train CIFAR100

```

```

- Train CUB200

```

The index files are named like "session_x.txt", where x indicates the session number. Each index file stores the indexes of the images that are selected for the session.

"session_1.txt" stores all the base class training images. Each "session_t.txt" (t>1) stores the 25 (5 classes and 5 shots per class) few-shot new class training images.

You may adopt the following steps to perform the experiments.

First, at session 1, train a base model using the images in session_1.txt;

Then, at session t (t>1), finetune the model trained at the previous session (t-1), only using the images in session_t.txt.

For evaluating the model at session t, first joint all the encountered test sets as a single test set. Then test the current model using all the test images and compute the recognition accuracy.