Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discovery

Paper | Code | Poster | Video | Slides

Setup

Prerequisites

All the dependencies and packages can be installed using pip. The code was tested using Python 3.10.

Installing the Packages

Use:

pip install -r requirements.txt
pip install -e sparse_autoencoder/
pip install -e .

Dataset for training Sparse Autoencoder (CC3M)

Download the CC3M tar file to train the SAE

Note: Number of downloaded paired dataset might be less than we used for our training as we downloaded the dataset in December, 2023. And as of now some more urls might be invalid.

1) Download the ‘Train_GCC-training.tsv’ and ‘Validation_GCC-1.1.0-Validation.tsv’ from https://ai.google.com/research/ConceptualCaptions/download by clicking on training split and validation split.

2) Change their names to cc3m_training.tsv and cc3m_validation.tsv

3) For training dataset:

    sed -i '1s/^/caption\turl\n/' cc3m_training.tsv 
    img2dataset --url_list cc3m_training.tsv --input_format "tsv" --url_col "url" --caption_col "caption" --output_format webdataset --output_folder training --processes_count 16 --thread_count 64 --image_size 256 --enable_wandb True

4) for validation dataset:

    sed -i '1s/^/caption\turl\n/' cc3m_validation.tsv 
    img2dataset --url_list cc3m_validation.tsv --input_format "tsv" --url_col "url" --caption_col "caption" --output_format webdataset --output_folder validation --processes_count 16 --thread_count 64 --image_size 256 --enable_wandb True

Vocabulary for naming concepts

We use the vocabulary of 20k words used by CLIP-Dissect, from here. Download and place the text file named as "clipdissect_20k.txt in vocab_dir specified in config.py. Then compute normalized CLIP embeddings of each text and save them as embeddings_<encoder_name>_clipdissect_20k.pth in vocab_dir. For example, for CLIP ResNet-50, the embedding file should be named embeddings_clip_RN50_clipdissect_20k.pth.

Datasets for training downstream probes

These are the datasets on which linear probes are trained on the learnt concept bottleneck to form a concept bottleneck model (CBM). In our paper, we use four datasets: Places365, ImageNet, CIFAR10, CIFAR100. Instructions for running experiments on these datasets is provided below, for other datasets you may need to define your own utils.

• Download the respective datasets:
  • Places365
  • ImageNet
  • CIFAR10
  • CIFAR100
• Set the paths to the datasets in config.py.

Usage

The following shows example usage with CLIP ResNet-50 as the model, CC3M as the dataset for training the SAE, and Places365 as the dataset for downstream classification.

Training a Sparse Autoencoder (SAE)

Save the CLIP features on CC3M to train the SAE on

python scripts/save_cc3m_features.py --img_enc_name clip_RN50 

Train the SAE

python scripts/train_sae_img.py --lr 5e-4 --l1_coeff 3e-5 --expansion_factor 8 --img_enc_name clip_RN50 --num_epochs 200 --resample_freq 10 --ckpt_freq 0 --val_freq 1 --train_sae_bs 4096

Assigning Names to Concepts

python scripts/assign_names.py --lr 5e-4 --l1_coeff 3e-5 --expansion_factor 8 --img_enc_name clip_RN50 --num_epochs 200 --resample_freq 10 --train_sae_bs 4096

Checkpoints and Assigned names can be found at the following link:

• Checkpoints and Assigned Names

Training a Linear Probe for the Concept Bottleneck Model

Save the CLIP features of probe dataset

python scripts/save_probe_features.py --img_enc_name clip_RN50  --probe_dataset places365

Save concept strengths using the trained SAE

python scripts/save_concept_strengths.py --lr 5e-4 --l1_coeff 3e-5 --expansion_factor 8 --img_enc_name clip_RN50 --num_epochs 200  --resample_freq 10  --train_sae_bs 4096  --probe_dataset places365 --probe_split train

Train the probe on the saved concept strengths

python scripts/train_linear_probe.py --lr 5e-4 --l1_coeff 3e-5 --expansion_factor 8 --img_enc_name clip_RN50 --resample_freq 10 --train_sae_bs 4096 --num_epochs 200 --ckpt_freq 0 --val_freq 1 --probe_lr 1e-2  --probe_sparsity_loss_lambda 0.1 --probe_classification_loss 'CE' --probe_epochs 200 --probe_sparsity_loss L1 --probe_eval_coverage_freq 50 --probe_dataset places365

Visualization plots

Example on the SAE trained above

Task Agnosticity

First, save the required files for plotting:

python scripts/visualization/vis_task_agnosticity_dump.py --img_enc_name clip_RN50 --probe_split val --method_name ours  --device cuda --sae_dataset cc3m

Then, plot the task agnosticity plots:

python scripts/visualization/vis_task_agnosticity_plot.py --img_enc_name clip_RN50 --probe_split val --method_name ours --device cpu --sae_dataset cc3m 

Meta Clusters

First, save the required files for plotting:

python scripts/visualization/vis_meta_clustering_dump.py --img_enc_name clip_RN50 --method_name ours --probe_split val --probe_dataset places365 --sae_dataset cc3m --device cuda

Then, plot the meta clusters:

python scripts/visualizations/vis_meta_clustering_plot.py --img_enc_name clip_RN50 --probe_split val --method_name ours  --probe_dataset places365 --sae_dataset cc3m --device cpu

Local Explanation

First, save the required files for plotting:

python scripts/visualization/vis_local_explanations_dump.py --sae_dataset cc3m --img_enc_name clip_RN50 --method_name ours  --probe_split val --probe_dataset places365 --which_ckpt final

Then, plot the local explanations:

python scripts/visualization/vis_local_explanations_plot.py --sae_dataset cc3m --img_enc_name clip_RN50 --method_name ours --probe_split val --probe_dataset places365  --which_ckpt final --device cpu

Global Explanation

First, save the required files for plotting:

python scripts/visualization/vis_global_explanations_dump.py --sae_dataset cc3m --img_enc_name clip_RN50 --method_name ours --probe_split val   --probe_dataset places365 --which_ckpt final

Then, plot the global explnataions:

python scripts/visualization/vis_global_explanations_plot.py --img_enc_name clip_RN50 --method_name ours --probe_split val   --probe_dataset places365 --which_ckpt final --device cpu

Acknowledgements

This repository uses code from the following repositories:

• openai/CLIP
• ai-safety-foundation/sparse_autoencoder

Citation

Please cite as follows:

@inproceedings{Rao2024Discover,
    author    = {Rao, Sukrut and Mahajan, Sweta and B\"ohle, Moritz and Schiele, Bernt},
    title     = {Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discovery},
    booktitle = {European Conference on Computer Vision},
    year      = {2024}
}