Stationary Representations: Optimally Approximating Compatibility and Implications for Improved Model Replacements

This repo contains the source code of the Python package iamcl2r and it is the official implementation of:
Stationary Representations: Optimally Approximating Compatibility and Implications for Improved Model Replacements, Niccolò Biondi, Federico Pernici, Simone Ricci, Alberto Del Bimbo at CVPR2024.
Our paper has been selected as Poster Highlight (notable top 2.8%).
Arxiv link: https://arxiv.org/abs/2405.02581

The repo also contains several examples on Google Colab and IPython Notebook.
See NOTEBOOKS.md for more details. We only support PyTorch for now.

See our paper for a detailed description of $d$-Simplex-HOC and the Improved Asynchronous Model Compatible Lifelong Learning Representation ($\text{IAM-CL}^2\text{R}$) scenario.

Repository Overview

There are several directories in this repo:

• notebooks/ contains the code for the Google Colab and IPython Notebook examples we provide;
• src/iamcl2r/ contains the source code for the package iamcl2r;
• configs/ contains the configuration files for the examples;

Quickstart

1. Installation
   Installing iamcl2r is simply

   git clone https://github.com/miccunifi/iamcl2r
   cd iamcl2r
   make install

   This creates a virtual enviroment and installs all the required dependencies.

   Create the .env file (Optional but suggested)

   Create an .env file to automatically export your env vars before launching training. An example of an .env file is

   WANDB_USERNAME=<your wandb username>
   WANDB_API_KEY=<your wandb private api key> 
   WANDB_ENTITY=<wandb entity>
   WANDB_PROJECT=<wandb project name>

   The .env file is not released for security reasons.

2. Download the pretrained models
   These models are used to replace the fine-tuned model in a $\text{IAM-CL}^2\text{R}$ training.

   make download-pretrained-models

3. Run $\text{IAM-CL}^2\text{R}$ training
   We only support $d$-Simplex-HOC and ER baseline for now.

   # d-Simplex-HOC training
   make run CONFIG=configs/hoc_7tasks.yaml
   make run CONFIG=configs/hoc_31tasks.yaml
   
   # ER baseline training
   make run CONFIG=configs/er_7tasks.yaml
   make run CONFIG=configs/er_31tasks.yaml

4. Run a compatibility evaluation
   Customize the configuration file eval.yaml by changing the checkpoint_path line

   ...
   checkpoint_path: <FOLDER_PATH>
   ...

   After that, you can evaluate the compatibility of the models which checkpoints are in the <FOLDER_PATH> you have specified above.

   make run CONFIG=configs/eval.yaml

(back to top)

Released results

We have released some experimental results obtained using the code in the repo. See NOTEBOOKS.md for more details.

(back to top)

Citation

If you use this code in your research, please kindly cite the following paper:

@inproceedings{biondi2024stationary,
  title={Stationary Representations: Optimally Approximating Compatibility and Implications for Improved Model Replacements},
  author={Biondi, Niccolò and Pernici, Federico and Ricci, Simone and Del Bimbo, Alberto},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2024},
}

(back to top)

Contact

For help, new features, or reporting bugs associated with this repo, please open a GitHub issue or contact us if you have any questions.

• Niccolò Biondi <niccolo.biondi (at) unifi.it>
• Federico Pernici <federico.pernici (at) unifi.it>
• Simone Ricci <simone.ricci (at) unifi.it>
• Alberto Del Bimbo <alberto.delbimbo (at) unifi.it>

(back to top)