LLaVA¶

Visual Instruction Tuning

摘要¶

Instruction tuning large language models (LLMs) using machine-generated instruction-following data has improved zero-shot capabilities on new tasks, but the idea is less explored in the multimodal field. In this paper, we present the first attempt to use language-only GPT-4 to generate multimodal language-image instruction-following data. By instruction tuning on such generated data, we introduce LLaVA: Large Language and Vision Assistant, an end-to-end trained large multimodal model that connects a vision encoder and LLM for general-purpose visual and language understanding.Our early experiments show that LLaVA demonstrates impressive multimodel chat abilities, sometimes exhibiting the behaviors of multimodal GPT-4 on unseen images/instructions, and yields a 85.1% relative score compared with GPT-4 on a synthetic multimodal instruction-following dataset. When fine-tuned on Science QA, the synergy of LLaVA and GPT-4 achieves a new state-of-the-art accuracy of 92.53%. We make GPT-4 generated visual instruction tuning data, our model and code base publicly available.

使用方式¶

According to the license of LLaMA, we cannot provide the merged checkpoint directly. Please use the below script to download and get the merged the checkpoint.

python tools/model_converters/llava-delta2mmpre.py huggyllama/llama-7b liuhaotian/LLaVA-Lightning-7B-delta-v1-1 ./LLaVA-Lightning-7B-delta-v1-1.pth

import torch
from mmpretrain import get_model, inference_model

model = get_model('llava-7b-v1_caption', pretrained='MERGED_CHECKPOINT_PATH', device='cuda')
out = inference_model(model, 'demo/cat-dog.png')
print(out)
# {'pred_caption': 'In the image, there are two cats sitting on a blanket.'}

Prepare your dataset according to the docs.

测试：

python tools/test.py configs/llava/llava-7b-v1_caption.py MERGED_CHECKPOINT_PATH

Models and results¶

Image Caption on COCO¶

模型	Params (M)	BLEU-4	CIDER	配置文件	下载
`llava-7b-v1_caption`	7045.82	Upcoming	Upcoming	config	See the above tutorial

引用¶

@misc{liu2023llava,
      title={Visual Instruction Tuning},
      author={Liu, Haotian and Li, Chunyuan and Wu, Qingyang and Lee, Yong Jae},
      publisher={arXiv:2304.08485},
      year={2023},
}