Awesome

Exemplar-Free Continual Transformer with Convolutions [Paper] [Website]

This repository contains the implementation details of our Exemplar-Free Continual Transformer with Convolutions (ConTraCon) approach for continual learning with transformer backbone.

Anurag Roy, Vinay K. Verma, Sravan Voonna, Kripabandhu Ghosh, Saptarshi Ghosh, Abir Das, "Exemplar-Free Continual Transformer with Convolutions"

If you use the codes and models from this repo, please cite our work. Thanks!

@InProceedings{roy_2023_ICCV,
    author    = {Roy, Anurag and Verma, Vinay and Voonna, Sravan and Ghosh, Kripabandhu and Ghosh, Saptarshi and Das, Abir},
    title     = {Exemplar-Free Continual Transformer with Convolutions},
    booktitle = {International Conference on Computer Vision (ICCV)},
    year      = {2023}
}

Requirements

The code is written for python 3.8.16, but should work for other version with some modifications.

pip install -r requirements.txt

Data Preparation

1. Download the datasets to the root diretory, Datasets.

2. CIFAR100 dataset will be automatically downloaded, while [ImageNet100, TinyImageNet] requires manual download.

3. Overview of dataset root diretory

   ├── cifar100
   │   └── cifar-100-python
   ├── tinyimagenet
   │   ├── tiny-imagenet-200
   │   ├── train
   │   ├── val
   │   └── test
   └── imagenet-100
       ├── imagenet-r
       ├── train_list.txt
       └── val_list.txt
   

NOTE -- After downloading and extracting the tinyimagenet dataset inside the Datasets folder, run

python val_format.py

This is to change the way the test dataset is stored for tinyimagenet.

Python script overview

auto_run.py - Contains the training and the inference code for the ConTraCon approach.

src/* - Contains the source code for the backbone transformer architecture and the convolutional task adaptation mechanisms.

src/utils/model_parts.py - Contains the task specific adaptation classes and functions.

incremental_dataloader.py - Contains the code for the dataloaders for different datasets.

Key Parameters:

ker_sz: kernel size of the convolution kernels which are applied on the key, query and value weight matrices of the MHSA layers
num_tasks : Number of tasks to split the given dataset into. This will split the classes in the datasets equally among the tasks
nepochs: Number of training epochs for each task
is_task0: Denotes whether training the first task. For the first task, the entire backbone transformer is trained from scratch.
use_saved: Use saved weights and resume training from next task. For example, for a 10 task setup, if trained till task 2, you can resume training from task 3 by using this flag. If training on all tasks have completed, then this flag can be used for re-evaluation of the trained model.
dataset: Denotes the dataset.
data_path: The path for the dataset.
scenario: Evaluation scenario. We have evaluated our models in two scenarios -- til (task incremental learning) and cil (class incremental learning).

Training ConTraCon

• For training on x% labeled data scenario, the first task needs to be trained first. This can be done by using the --is_task0 flag.
• For training on subsequent tasks, run without the -is_task0 flag.

Sample Code to train ConTraCon

The code to train ConTraCon on the ImageNet-100 dataset is provided as follows:

1. Training the first task :

python auto_run.py --ker_sz 15 --nepochs 500 --dataset imagenet100 --data_path ./Datasets/imagenet-100/ --num_tasks 10 --is_task0 --scenario til

2. Training the rest of the tasks:

python auto_run.py --ker_sz 15 --nepochs 500 --dataset imagenet100 --data_path ./Datasets/imagenet-100/ --num_tasks 10 --scenario til

Sample Code to Evaluate ConTraCon

• To evaluate ConTraCon in the til setup, run:

python auto_run.py --ker_sz 15 --nepochs 500 --dataset imagenet100 --data_path ./Datasets/imagenet-100/ --num_tasks 10 --use_saved --scenario til

• To evaluate ConTraCon in the cil setup, run:

python auto_run.py --ker_sz 15 --nepochs 500 --dataset imagenet100 --data_path ./Datasets/imagenet-100/ --num_tasks 10 --use_saved --scenario cil

Reference

The implementation reused some portions from CCT[1].

1. Ali Hassani, Steven Walton, Nikhil Shah, Abulikemu Abuduweili, Jiachen Li, Humphrey Shi. "Escaping the Big Data Paradigm with Compact Transformers." Arxiv Preprint. 2021.