BLIP：用于统一视觉语言理解和生成的引导式语言图像预训练

BLIP（Bootstrapping Language-Image Pre-training）基于图像-文本匹配进行训练，使用了大规模架构（ViT 大型骨干网络）在 Flickr30k 数据集上进行了训练。

TL;DR

来自 paper 的作者在摘要中写道：

视觉-语言预训练（VLP）已经提升了许多视觉-语言任务的性能。然而，大多数已有的预训练模型只在理解型任务或生成型任务中表现出色。此外，通过扩大使用从网络上收集的嘈杂的图像-文本对数据集来改善性能，这种方式并不是最优的监督来源。在本文中，我们提出了 BLIP，这是一个新的 VLP 框架，可以灵活地转移到视觉-语言理解和生成任务中。BLIP 通过引导标题来高效地利用嘈杂的网络数据，其中一个标题生成器生成合成标题，而一个过滤器则删除噪音标题。我们在广泛的视觉-语言任务上取得了最先进的结果，例如图像-文本检索（平均召回率 @ 1 提高了 2.7%）、图像字幕生成（CIDEr 提高了 2.8%）和视觉问答（VQA 评分提高了 1.6%）。BLIP 还展示了强大的泛化能力，在零-shot 方式下直接转移到视频-语言任务上。代码、模型和数据集已发布。

用法

您可以将此模型用于有条件和无条件的图像字幕生成。

使用 PyTorch 模型

在 CPU 上运行模型

点击扩展

import requests
from PIL import Image
from transformers import BlipProcessor, BlipForImageTextRetrieval

processor = BlipProcessor.from_pretrained("Salesforce/blip-itm-large-flickr")
model = BlipForImageTextRetrieval.from_pretrained("Salesforce/blip-itm-large-flickr")

img_url = 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/demo.jpg' 
raw_image = Image.open(requests.get(img_url, stream=True).raw).convert('RGB')

question = "A woman and a dog sitting together in a beach."
inputs = processor(raw_image, question, return_tensors="pt")

itm_scores = model(**inputs)[0]
cosine_score = model(**inputs, use_itm_head=False)[0]

在 GPU 上运行模型

以全精度运行

点击扩展

import requests
from PIL import Image
from transformers import BlipProcessor, BlipForImageTextRetrieval

processor = BlipProcessor.from_pretrained("Salesforce/blip-itm-large-flickr")
model = BlipForImageTextRetrieval.from_pretrained("Salesforce/blip-itm-large-flickr").to("cuda")

img_url = 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/demo.jpg' 
raw_image = Image.open(requests.get(img_url, stream=True).raw).convert('RGB')

question = "A woman and a dog sitting together in a beach."
inputs = processor(raw_image, question, return_tensors="pt").to("cuda")

itm_scores = model(**inputs)[0]
cosine_score = model(**inputs, use_itm_head=False)[0]

以半精度（float16）运行

点击扩展

import torch
import requests
from PIL import Image
from transformers import BlipProcessor, BlipForImageTextRetrieval

processor = BlipProcessor.from_pretrained("Salesforce/blip-itm-large-flickr")
model = BlipForImageTextRetrieval.from_pretrained("Salesforce/blip-itm-large-flickr", torch_dtype=torch.float16).to("cuda")

img_url = 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/demo.jpg' 
raw_image = Image.open(requests.get(img_url, stream=True).raw).convert('RGB')

question = "A woman and a dog sitting together in a beach."
inputs = processor(raw_image, question, return_tensors="pt").to("cuda", torch.float16)

itm_scores = model(**inputs)[0]
cosine_score = model(**inputs, use_itm_head=False)[0]

BibTex 和引用信息

@misc{https://doi.org/10.48550/arxiv.2201.12086,
  doi = {10.48550/ARXIV.2201.12086},
  
  url = {https://arxiv.org/abs/2201.12086},
  
  author = {Li, Junnan and Li, Dongxu and Xiong, Caiming and Hoi, Steven},
  
  keywords = {Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, FOS: Computer and information sciences},
  
  title = {BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation},
  
  publisher = {arXiv},
  
  year = {2022},
  
  copyright = {Creative Commons Attribution 4.0 International}
}