ProbVLM: Probabilistic Adapter for Frozen Vision-Language Models

Arxiv paper || Blog

Introduction

Abstract. Large-scale vision-language models (VLMs) like CLIP successfully find correspondences between images and text. Through the standard deterministic mapping process, an image or a text sample is mapped to a single vector in the embedding space. This is problematic: as multiple samples (images or text) can abstract the same concept in the physical world, deterministic embeddings do not reflect the inherent ambiguity in the embedding space. We propose ProbVLM, a probabilistic adapter that estimates probability distributions for the embeddings of pre-trained VLMs via inter/intra-modal alignment in a post-hoc manner without needing large-scale datasets or computing. On four challenging datasets, i.e., COCO, Flickr, CUB, and Oxford-flowers, we estimate the multi-modal embedding uncertainties for two VLMs, i.e., CLIP and BLIP, quantify the calibration of embedding uncertainties in retrieval tasks and show that ProbVLM outperforms other methods. Furthermore, we propose active learning and model selection as two real-world downstream tasks for VLMs and show that the estimated uncertainty aids both tasks. Lastly, we present a novel technique for visualizing the embedding distributions using a large-scale pre-trained latent diffusion model.

TLDR: This is the official PyTorch implementation of ProbVLM (from ICCV 2023) that allows estimating calibrated uncertainty for pre-trained (frozen) vision-langugae models in fast and efficient manner.

The structure of the repository is as follows:

ProbVLM
|-src/ (has the relevant code to train ProbVLM for a pretrained vision-language model, e.g., CLIP)
|-requirements.txt (the environment dependencies)
|-figs/ (has some example images)

Getting Started

The src/ directory already provides all the code to load, train, and evaluate the CLIP checkpoints along with the ProbVLM checkpoint. In order to get started, first ensure that you have all the requirements as provided in requirements.txt, this can be setup by running

conda create --name <env> --file requirements.txt

The notebook src/train_ProbVLM_CLIP.ipynb shows how to load the CLIP as the base model and how to train ProbVLM for the same using the function,

train_ProbVLM(
    CLIP_Net,
    ProbVLM_Net,
    cub_train_loader,
    cub_valid_loader,
    Cri = TempCombLoss(),
    device='cuda',
    dtype=torch.cuda.FloatTensor,
    init_lr=8e-5,
    num_epochs=500,
    eval_every=5,
    ckpt_path='../ckpt/ProbVLM_Net',
    T1=1e0,
    T2=1e-4
)

Dataset Setup

To setup the COCO, CUB, Flickr and the Oxford-Flowers dataset, please follow the below instructions:

COCO: Download the 2014 data here and setup the directory in the following way

coco
|-images/
|--train2014 
|--val2014 
|-captions_train2014.json 
|-captions_val2014.json

CUB: Download the CUB-200-2011 dataset here and the captions from https://github.com/reedscot/cvpr2016.

Flowers: Download the images here and the captions from https://github.com/reedscot/cvpr2016.

Flickr: Dwonload the images and captions from here.

Resources

We use the following resources to use as the base models (on top of which we train different ProbVLM models)

CLIP: https://github.com/openai/CLIP
BLIP: https://github.com/salesforce/LAVIS
BayesCap: https://github.com/ExplainableML/BayesCap

BibTex

Please cite the following works

@inproceedings{Upa_probvlm,
  title = {ProbVLM: Probabilistic Adapter for Frozen Vision-Language Models},
  author = {Upadhyay, U. and Karthik, S. and Mancini, M. and Akata, Z.},
  booktitle = {International Conference on Computer Vision (ICCV 2023)},
  year = {2023}
}

@inproceedings{Upa_bayescap,
  title = {BayesCap: Bayesian Identity Cap for Calibrated Uncertainty in Frozen Neural Networks},
  author = {Upadhyay, U. and Karthik, S. and Chen, Y. and Mancini, M. and Akata, Z.},
  booktitle = {European Conference on Computer Vision (ECCV 2022)},
  year = {2022}
}

@inproceedings{upadhyay2021uncerguidedi2i,
  title={Uncertainty Guided Progressive GANs for Medical Image Translation},
  author={Upadhyay, Uddeshya and Chen, Yanbei and Hebb, Tobias and Gatidis, Sergios and Akata, Zeynep},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI)},
  year={2021},
  organization={Springer}
}

@inproceedings{UpaCheAka21,
  title = {Robustness via Uncertainty-aware Cycle Consistency},
  author = {Upadhyay, U. and Chen, Y. and Akata, Z.},
  booktitle = {Advances in Neural Information Processing Systems 34 (NeurIPS 2021)},
  year = {2021}
}

ExplainableML/ProbVLM