jun-hyeok / clip-feat-vis

In-depth Understanding of CLIP Neurons via Faceted Feature Visualization

Hyesu Lim, Changdae Oh, Junhyeok Park, Rohan Prasad

In this project, we reproduce the result of multimodal neuron analysis (OpenAI, 2021) and further investigate the behavior of such neurons after being fine-tuned on diverse downstream task.

Quick Start

1. Prepare the basic experiment pipeline by cloning this repository.

git clone https://github.com/hyesulim/clip-feat-vis.git
cd clip-feat-vis

2. Environment setup (any of python ver >= 3.7, torch ver >= 1.8 should be okay). However, this code has been extensively tested on Python 3.11.6 and PyTorch 2.1.1 and is what we recommend. A requirements.txt file is included to help resolve any version conflicts.

conda create -n mmn python=3.11
conda activate mmn
pip install -r requirements.txt

3. A sample Jupyter notebook is included inside notebooks. This should help you get started.

Module description

1. linear_probe - This folder contains all our code related to training the Linear Probe used for faceted visualization

2. `initial_ablations' - This folder contains the scripts and codes that used Lucent for performing the initial ablations on vanilla feature visualization

3. faceted_visualization - This folder contains all the experiments and code related to faceted visualization

   a.) visualizer - this is a python package containing reusable code for using faceted visualization in your project. It currently only supports CLIP models.

4. finetuning - This folder contains our the code required for finetuning

Ablation Details

You can find the links to our ablations here:

CelebA

Pre-Trained (Rohan)

Pre-Trained (Changdae)

Fine-Tuned

SUN397

Pre-Trained (Rohan)

Pre-Trained (Changdae)

Fine-Tuned

Aircraft

Pre-Trained (Rohan)

Pre-Trained (Changdae)

Fine-Tuned

Acknowledgement

This repository is built on top of Lucent library, and we would like to thank the authors of Lucent for their extensive efforts.!