G-JWLee/COINCIDE_code

COINCIDE

Concept-skill Transferability-based Data Selection for Large Vision-Language Models, EMNLP 2024

• Authors: Jaewoo Lee, Boyang Li^, Sung Ju Hwang^
• Paper

Introduction

Training Large Vision-Langauge Models on extensive visual instruction tuning (VIT) data incurs significant computational cost, making the process infeasible for small academic labs and individual researchers. Additionally, it is not clear if all the VIT data are necessary for good generalization, as different vision-language (VL) tasks have different abilities to transfer to downstream tasks.

We propose COINCIDE (COre-INstruction Concept-skIll Data Election), which identifies VL concept-skill compositions through data clustering using activations from an off-the-shelf, small VLM.

From each cluster, COINCIDE selects training data for a target LVLM by considering transferability (i.e., how well knowledge from each cluster can facilitate LVLM's learning in other clusters) and internal density of clusters.

Install

Since we use different models for 1. clustering and 2. training, different conda environments are required for each process.

Clustering

conda create -n coincide_cluster python=3.10 -y
conda activate coincide_cluster
pip install --upgrade pip  # enable PEP 660 support
pip install -e .
pip install -e ".[train]"

Training

conda create -n coincide_train python=3.10 -y
conda activate coincide_train
cd COINCIDE_train
pip install --upgrade pip  # enable PEP 660 support
pip install -e .
pip install -e ".[train]"
pip install flash-attn --no-build-isolation

Dataset

LLaVA-1.5

Follow the visual instruction tuning dataset download guides in the official Github page Place the downloaded files in COINCIDE_train/playground/data. Also, prepare for the evaluation benchmark datasets by following the instructions in the page.

Vision-Flan

Similarly, download the Vision-Flan visual instruction tuning dataset through the official project page Place the downloaded files in COINCIDE_train/playground/data.

Models

In this paper, we focus on two models: TinyLLavA-2B for clustering and LLaVA-1.5-7B for training. Please download the models in /checkpoints directory.

You can try to use other LVLMs such as TinyLLaVA-OpenELM-450M-SigLIP-0.89B or LLaVA-1.5-13B, as we have done in our ablation studies.

LLaVA-1.5 experiment

Clustering

Change directory cd COINCIDE_cluster.

bash scripts/tiny_llava/eval/llava_score/coincide/extract_embed.sh
bash scripts/tiny_llava/eval/llava_score/coincide/clustering.sh

Training

Change directory cd COINCIDE_train.

bash scripts/v1_5/finetune_lora_coincide.sh

Vision-Flan experiment

Clustering

Change directory cd COINCIDE_cluster.

bash scripts/tiny_llava/eval/vision_flan_score/coincide/extract_embed.sh
bash scripts/tiny_llava/eval/vision_flan_score/coincide/clustering.sh

Training

Change directory cd COINCIDE_train.

bash scripts/v1_5/vision_flan_score/finetune_lora_coincide.sh

Evaluation

MODEL_NAME is the name of the directory that stores the model.
{eval_benchmark} could be gqa, llavabench, mme, etc.

bash scripts/v1_5/eval/{eval_benchmark}.sh MODEL_NAME

Bibtex

@inproceedings{lee2024coincide,
      title={Concept-skill Transferability-based Data Selection for Large Vision-Language Models},
      author={Jaewoo Lee and Boyang Li and Sung Ju Hwang},
      year={2024},
      booktitle={Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP)},
}