E5-V：基于多模态大语言模型的通用嵌入方法

E5-V是基于lmms-lab/llama3-llava-next-8b进行微调的模型。

概述

我们提出了名为E5-V的框架，用于适配多模态大语言模型（MLLMs）以实现多模态嵌入。E5-V有效弥合了不同类型输入之间的模态鸿沟，即使未经微调也能在多模态嵌入任务中展现出色性能。我们还提出了针对E5-V的单模态训练方法，该模型仅通过文本对进行训练，其表现优于多模态训练方式。

更多细节请访问：https://github.com/kongds/E5-V

示例

import torch
import torch.nn.functional as F
import requests
from PIL import Image
from transformers import AutoTokenizer
from transformers import LlavaNextProcessor, LlavaNextForConditionalGeneration

llama3_template = '<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n \n'

processor = LlavaNextProcessor.from_pretrained('royokong/e5-v')
model = LlavaNextForConditionalGeneration.from_pretrained('royokong/e5-v', torch_dtype=torch.float16).cuda()

img_prompt = llama3_template.format('<image>\n用一词概括上图：')
text_prompt = llama3_template.format('<sent>\n用一词概括上文：')

urls = ['https://upload.wikimedia.org/wikipedia/commons/thumb/4/47/American_Eskimo_Dog.jpg/360px-American_Eskimo_Dog.jpg',
        'https://upload.wikimedia.org/wikipedia/commons/thumb/b/b6/Felis_catus-cat_on_snow.jpg/179px-Felis_catus-cat_on_snow.jpg']
images = [Image.open(requests.get(url, stream=True).raw) for url in urls]

texts = ['一只坐在草地上的狗',
         '一只站在雪地里的猫']

text_inputs = processor([text_prompt.replace('<sent>', text) for text in texts], return_tensors="pt", padding=True).to('cuda')
img_inputs = processor([img_prompt]*len(images), images, return_tensors="pt", padding=True).to('cuda')

with torch.no_grad():
    text_embs = model(**text_inputs, output_hidden_states=True, return_dict=True).hidden_states[-1][:, -1, :]
    img_embs = model(**img_inputs, output_hidden_states=True, return_dict=True).hidden_states[-1][:, -1, :]

    text_embs = F.normalize(text_embs, dim=-1)
    img_embs = F.normalize(img_embs, dim=-1)

print(text_embs @ img_embs.t())