Supplementary code for "Sparse Concept Bottleneck Models: Gumbel tricks in Contrastive Learning"

How to install

git clone https://github.com/Andron00e/SparseCBM
cd SparseCBM
pip install -r requirements.txt

Repository structure

• additional_evaluations contains training examples of CBMs, evaluation of CMS, concepts generation and latent space visualization.
• In bottleneck you may find a code for: our model, key objective functions, training utilities and a setup for CMS.
• Use data to look through the concept sets and class labels we use.
• Please, run demo.ipynb in demo_notebooks to train your own CBM with any hyperparameters you need and any of the datasets provided. We suggest you to use CUB200 examples, for simplicity, and train with a small learning rates both of B/32 and L/14 configurations (CBMs are sensitive to lrs). A simple 10 epochs example with CLIP-ViT-B/32 backbone is already presented in demo, but feel free to adjust more changes.
• See experiments to verify our results for the rebuttal. Here we included both Linear Probing and Zero-shot inference of the CLIP-ViT-L/14 model. Along with it, we show an interpretability proxy, by comparing CBL activation with CLIP outputs.

New experimental results

We report with the additional results on interpretability and final classification accuracy.

1. General accuracy tables for CBM and CMS:

Model	CIFAR10	CIFAR100	ImageNet	CUB 200	Places365
Sparse-CBM (Ours)	91.17%	74.88%	71.61%	80.02%	41.34%
$\ell_1$-CBM (Ours)	85.11%	73.24%	71.02%	74.91%	40.87%
Contrastive-CBM	84.75%	68.46%	70.22%	67.04%	40.22%
Label-free CBM	86.40%	65.13%	71.95%	74.31%	43.68%
Post-hoc CBM (CLIP)	83.34%	57.20%	62.57%	63.92%	39.66%
LaBo (full-supervised)	87.90%	69.10%	70.40%	71.80%	39.43%
----------------------	-------	--------	--------	--------	--------
Linear Probing	96.12%	80.03%	83.90%	79.29%	48.33%

Model	CIFAR10	CIFAR100	ImageNet	CUB 200	Places365
Concept Matrix Search (Ours)	85.03%	62.95%	77.82%	65.17%	39.43%
DescriptionCLS	81.61%	68.32%	75.00%	63.46%	40.55%
Zero-shot CLIP-ViT-L/14	81.79%	52.84%	76.20%	62.63%	41.12%

3. For interpretability, we show a comparison between concepts extracted by CBM variant and backbone CLIP model:

Concepts extracted by Sparse-CBM (Ours)

Concepts extracted by $\ell_1$-CBM (Ours)

Concepts extracted by Contrastive-CBM (Ours)

Concepts extracted by CLIP

Concepts extracted by Sparse-CBM (Ours)

Concepts extracted by $\ell_1$-CBM (Ours)

Concepts extracted by Contrastive-CBM (Ours)

Concepts extracted by CLIP

Concepts extracted by Sparse-CBM (Ours)

Concepts extracted by $\ell_1$-CBM (Ours)

Concepts extracted by Contrastive-CBM

Concepts extracted by CLIP

Implementation details

To run our BottleneckTrainer you should create an appropriate CBMConfig instance and prepare dataloaders as in the demo.ipynb. Next, we provide a mock example of how the CBM training starts:

from configs import *
from trainer_utils import *

config = CBMConfig(
    num_nets=3,
    num_concepts=len(concepts),
    num_classes=200,
    run_name="demo_run",
    net_types=["base", "base", "base"],
    backbones=[Constants.clip_large_link, Constants.clip_large_link, Constants.clip_large_link],
    displayed_names=["model no 1", "model no 2", "model no 3"],
    training_methods=["gumbel", "contrastive", "l1"],
    optimizers=["SGD", "Adam", "AdamW"],
    lrs=[1e-3, 3e-4, 3e-4],
    cbl_lrs=[1e-3, 3e-4, 3e-4],
    train_backbones=[False, False, True],
)

trainer = BottleneckTrainer(
    config,
    train_loader,
    val_loader,
    test_loader,
    num_epochs=10,
    device=device,
)

trainer.train()
trainer.test()

If you would like to record the interpretability of the trained CBM, feel free to check the overall metrics and play with several images similarly to interpret.ipynb and cub_interpret.ipynb.

plot_trainer_metrics(trainer.hist)

image = PIL.Image.open("my_image.jpg")
processor = transformers.CLIPProcessor.from_pretrained(Constants.clip_large_link)

scores = cbm_interpretability_scores(trainer.nets[0], processor, concepts, image, device)
draw_bottleneck(image, scores, k=10, concepts=concepts, draw_probs=False)

We believe the details provided are clear enough to reproduce the main findings of our paper.

@misc{semenov2024sparse,
      title={Sparse Concept Bottleneck Models: Gumbel Tricks in Contrastive Learning}, 
      author={Andrei Semenov and Vladimir Ivanov and Aleksandr Beznosikov and Alexander Gasnikov},
      year={2024},
      eprint={2404.03323},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}