VL-ICL Bench
收藏资源简介:
一个用于多模态上下文学习的综合基准,涵盖了广泛的涉及图像和文本作为输入和输出的任务,以及从感知到推理和长上下文长度的不同类型的挑战。
A comprehensive benchmark for multimodal in-context learning, covering a wide range of tasks that take images and text as both inputs and outputs, and encompassing diverse challenges spanning from perception to reasoning as well as those involving long context lengths.
VL-ICL 数据集概述
数据准备
- 数据集托管平台: HuggingFace
- 数据集链接: VL-ICL 数据集
- 下载与解压命令: bash git lfs install git clone https://huggingface.co/datasets/ys-zong/VL-ICL cd VL-ICL bash unzip.sh cd ..
环境配置
-
创建环境: bash conda create -n {env_name} python==3.10 -y pip install -r requirements/{model.txt} conda activate {env_name}
-
替换文件:
{model.txt}替换为相应的文件名。
使用说明
图像到文本 (I2T)
-
推理命令: bash CUDA_VISIBLE_DEVICES=0 python I2T_inference.py --engine {model_name} --n_shot {shots} --dataset {dataset_name} --task_description detailed
-
示例: bash CUDA_VISIBLE_DEVICES=0 python I2T_inference.py --engine qwen-vl-chat --n_shot 0 1 2 4 5 --task_description detailed --dataset open_mi
-
评估命令: bash python I2T_evaluate.py --dataset {dataset_name} --engine {model_name} --n_shot {shots}
文本到图像 (T2I)
-
推理命令: bash CUDA_VISIBLE_DEVICES=0 python T2I_inference.py --engine {model_name} --n_shot {shots} --dataset {dataset_name} --task_description detailed
-
示例: bash CUDA_VISIBLE_DEVICES=0 python T2I_inference.py --engine emu1-gen --n_shot 0 1 2 4 5 --task_description detailed --dataset open_t2i_mi
-
评估命令: bash CUDA_VISIBLE_DEVICES=0 python T2I_evaluate.py --dataset open_t2i_mi --engine seed-llama
引用
@article{zong2024vlicl, title={VL-ICL Bench: The Devil in the Details of Benchmarking Multimodal In-Context Learning}, author={Zong, Yongshuo and Bohdal, Ondrej and Hospedales, Timothy}, journal={arXiv preprint arXiv:2403.13164}, year={2024} }




