The paper is under review.

conda create -n simignore python=3.10
conda activate simignore
cd src
bash setup.sh

You can download LLaVA1.5-7b from Hugging Face and save it under checkpoint/llava.

You can download ScienceQA from Google Drive and unzip the images under data/scienceqa/images.

We provide the Zero-shot inference procedure for the LLaVA1.5-7b model and the LLaVA1.5-13b model on the ScienceQA(Image) dataset. We conduct the following experiments on one 4090D GPU (24G).

Inference using LLaVA-v1.5-7b model.

bash ./src/Simignore/inference/eval/eval_sqa_latency_inplace.sh

Inference using LLaVA-v1.5-13b model.

bash ./src/Simignore/inference/eval/eval_sqa_latency_inplace_13b.sh 

Evaluation using LLaVA-v1.5-7b model.

bash ./src/Simignore/inference/eval/generate_sqa_results.sh

Evaluation using LLaVA-v1.5-13b model.

bash ./src/Simignore/inference/eval/generate_sqa_results_13b.sh