README.md

Official Implementation of "Less is more: Summarizing Patch Tokens for efficient Multi-Label Class-Incremental Learning (MULTI-LANE)", published at 3rd Conference on Lifelong Learning Agents (CoLLAs 2024)

🛠️ Installation

Create and activate a conda environment with Python 3.11.17:

$ conda create -n multilane python=3.11.17
$ conda activate multilane

Install Python requirements from file:

$ pip install -r requirements.txt

🏋️ Datasets and Pre-Trained Weights

Both are automatically downloaded at training time.

Datasets are stored in the datasets/ folder (it will be automatically created). To specify a different folder, pass the argument --data_path /path/to/datasets when launching the training script.

Pre-Trained Weights are stored in the default timm directory.

🏃‍♀️ Training

We prepared six training scripts to train MULTI-LANE on each dataset, and configuration, we show in the paper:

• COCO B0-C10 (MLCIL) $ ./train_coco.sh
• COCO B40-C10 (MLCIL) $ ./train_coco_40.sh
• VOC B0-C4 (MLCIL) $ ./train_voc.sh
• VOC B10-C2 (MLCIL) $ ./train_voc_10.sh
• ImageNet-R (CIL) $ ./train_inr.sh
• CIFAR-100 (CIL) $ ./train_c100.sh

🗿 Changing the number of Selectors

To change the number of selectors at training time, pass the argument --num_selectors with the chosen amount of selectors. For example, to train on coco with 1 selector, run:

$ ./train_coco.sh --num_selectors 1

Or edit train_coco.sh and add the argument there.

Similarly, to use token merging instead of selectors pass the argument --tome with the number of tokens to be retained. Additionally, remove selectors from training by setting their quantity to 0 --num_selectors 0. For example:

$ ./train_coco.sh --num_selectors 0 --tome 30

🙅‍♀️Ablations

To run one of the ablations we show in the paper, use the following arguments:

• Disabling Normalization $ ./train_coco.sh --normalize 'none'
• Disabling Parallel Pathways (use query-key selection mechanism) $ ./train_coco.sh --head_model 'task'
• Disabling Drop and Replace $ ./train_coco.sh --disable_dandr

Citing our work

@InProceedings{pmlr-v274-min25a,
  title = {Less is more: Summarizing Patch Tokens for efficient Multi-Label Class-Incremental Learning},
  author = {De Min, Thomas and Mancini, Massimiliano and Lathuili{\`{e}}re, St{\'{e}}phane and Roy, Subhankar and Ricci, Elisa},
  booktitle = {Proceedings of The 3rd Conference on Lifelong Learning Agents (CoLLAs)},
  year = {2024}
}