language:

• en
• multilingual license: mit tags:
• vision
• image-to-text
• image-captioning
• visual-question-answering pipeline_tag: image-to-text inference: false datasets:
• Gregor/mblip-train

mBLIP mT0-XL

This is the model checkpoint for our work mBLIP: Efficient Bootstrapping of Multilingual Vision-LLMs.

Model description

mBLIP is a BLIP-2 model which consists of 3 sub-models: a Vision Transformer (ViT), a Query-Transformer (Q-Former) and a large language model (LLM).

The Q-Former and ViT have both been initialized by an English BLIP-2 checkpoint (blip2-flan-t5-xl) and then re-aligned to the multilingual LLM (mt0-xl) using a multilingual task mixture.

This allows the model to be used for tasks like:

• image captioning
• visual question answering (VQA)

in 96 languages.

Languages

mBLIP was trained on the following 96 languages:

af, am, ar, az, be, bg, bn, ca, ceb, cs, cy, da, de, el, en, eo, es, et, eu, fa, fi, fil, fr, ga, gd, gl, gu, ha, hi, ht, hu, hy, id, ig, is, it, iw, ja, jv, ka, kk, km, kn, ko, ku, ky, lb, lo, lt, lv, mg, mi, mk, ml, mn, mr, ms, mt, my, ne, nl, no, ny, pa, pl, ps, pt, ro, ru, sd, si, sk, sl, sm, sn, so, sq, sr, st, su, sv, sw, ta, te, tg, th, tr, uk, ur, uz, vi, xh, yi, yo, zh, zu

Direct Use and Downstream Use

You can use the raw model for conditional text generation given an image and prompt text in a zero-shot setup or alternatively finetune it for downstream applications. We strongly recommend LoRA applied to the LLM when finetuning and to use bf16 as data type - standard fp16 can cause NaN loss.

See our repository for the code used to train and finetune this model.

Bias, Risks, Limitations, and Ethical Considerations

While mBLIP can work in theory with up to 100 languages, in practice, we expect best results when prompted in high-resource languages like English, German, Spanish, etc.

mBLIP inherits the risk, limitations, and biases from the models used to initialize it. mBLIP has not been tested in real world applications. It should not be directly deployed in any applications. Researchers should first carefully assess the safety and fairness of the model in relation to the specific context they’re being deployed within.

How to use

For code examples, we refer to the BLIP-2 documentation.

Running the model on CPU

import requests
from PIL import Image
from transformers import BlipProcessor, Blip2ForConditionalGeneration

processor = BlipProcessor.from_pretrained("Gregor/mblip-mt0-xl")
model = Blip2ForConditionalGeneration.from_pretrained("Gregor/mblip-mt0-xl")

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 = "Describe the image in German."
inputs = processor(raw_image, question, return_tensors="pt")

out = model.generate(**inputs)
print(processor.decode(out[0], skip_special_tokens=True))

Running the model on GPU

In full precision

# pip install accelerate
import requests
from PIL import Image
from transformers import Blip2Processor, Blip2ForConditionalGeneration

processor = Blip2Processor.from_pretrained("Gregor/mblip-mt0-xl")
model = Blip2ForConditionalGeneration.from_pretrained("Gregor/mblip-mt0-xl", device_map="auto")

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 = "Describe the image in German."
inputs = processor(raw_image, question, return_tensors="pt").to("cuda")

out = model.generate(**inputs)
print(processor.decode(out[0], skip_special_tokens=True))

In half precision (bfloat16)

# pip install accelerate
import torch
import requests
from PIL import Image
from transformers import Blip2Processor, Blip2ForConditionalGeneration

processor = Blip2Processor.from_pretrained("Gregor/mblip-mt0-xl")
model = Blip2ForConditionalGeneration.from_pretrained("Gregor/mblip-mt0-xl", torch_dtype=torch.bfloat16, device_map="auto")

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 = "Describe the image in German."
inputs = processor(raw_image, question, return_tensors="pt").to("cuda", torch.bfloat16)

out = model.generate(**inputs)
print(processor.decode(out[0], skip_special_tokens=True))

In 8-bit precision (int8)

Important: Paper results only use int8 for the LLM weights while this loads all weights in int8. We see that this gives slightly worse results but currently int8 for some model parts is not supported by HuggingFace.

# pip install accelerate bitsandbytes
import torch
import requests
from PIL import Image
from transformers import Blip2Processor, Blip2ForConditionalGeneration

processor = Blip2Processor.from_pretrained("Gregor/mblip-mt0-xl")
model = Blip2ForConditionalGeneration.from_pretrained("Gregor/mblip-mt0-xl", load_in_8bit=True, device_map="auto")

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 = "Describe the image in German."
inputs = processor(raw_image, question, return_tensors="pt").to("cuda", torch.bfloat16)

out = model.generate(**inputs)
print(processor.decode(out[0], skip_special_tokens=True))

Citation

If you use our model, please cite the following:

@article{geigle2023mblip,
  author       = {Gregor Geigle and
                  Abhay Jain and
                  Radu Timofte and
                  Goran Glava\v{s}},
  title        = {mBLIP: Efficient Bootstrapping of Multilingual Vision-LLMs},
  journal      = {arXiv},
  volume       = {abs/2307.06930},
  year         = {2023},
  url          = {https://arxiv.org/abs/2307.06930},
  eprinttype    = {arXiv},
  eprint       = {2307.06930},
}