Class-Incremental

Introduction

This template aims to create a baseline for handling the image classification using Class-Incremental Method. It also depends on the requirements and datasets to whether use or not use FastSAM or CLIP models to further improve the performance.

Dataset

Now suppose that we have 2 sets in raw: train and val (public_test) corresponding to 2 folders in data/raw. There are 10 phases and each phases we will have 10 different classes so in total we have 100 classes.

Before going to the project experiment, we need to make sure that the data structure follow the below format!

Structure

```
.
├── data

│   ├── raw
│   │   ├── Train
│   │   │   ├── phase 1
│   │   │   │   ├── class_0
│   │   │   │   │   ├── 000.jpg
│   │   │   │   │   ├── ...
│   │   │   │   ├── class_1
│   │   │   │   ├── ...
│   │   │   │   ├── class_10
│   │   │   ├── phase 2
│   │   │   ├── ...
│   │   │   ├── phase 10
│   │   ├── Val
│   │   │   ├── 00.jpg
│   │   │   ├── ...

│   ├── processed
│   │   ├── Train
│   │   ├── Val
```

Set-up

git clone https://github.com/manhph2211/Class-Incremental.git
cd Class-Incremental
conda create -n ci python=3.11
conda activate ci
pip install -r requirements.txt
wget https://huggingface.co/spaces/An-619/FastSAM/resolve/main/weights/FastSAM.pt -O src/models/segmentator/weights/FastSAM.pt

Usage

python src/dataloader/dataset.py # you might wanna see some samples

python src/tools/segment.py --output " " --image_folder " " --text_prompt "eg. a bird" # Optional
 
python src/tools/train.py # Training and validation

python src/tools/classify.py # Testing an image folder and submission

Reference

@misc{zhao2023fast,
      title={Fast Segment Anything},
      author={Xu Zhao and Wenchao Ding and Yongqi An and Yinglong Du and Tao Yu and Min Li and Ming Tang and Jinqiao Wang},
      year={2023},
      eprint={2306.12156},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}