mBLIP

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

Updates

02.10.23: Updates for v2 on arxiv:

• Support for training and evaluating pure decoders (BLOOMZ)
• 4bit bnb LLM loading
• Training data: VQA with translated labels, ImageNet, Caption/ImageNet matching task
• Evaluation: Added POPE and CHAIR for object hallucination evaluation

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 using a multilingual task mixture.

<img src="architecture.png" alt="The mBLIP architecture" width="600"/>

This allows the model to be used for tasks like:

• image captioning
• visual question answering (VQA)

in 96 languages.

Checkpoints

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

Demo

We provide a simple Gradio demo. When using 8-bit quantization to load the model, the demo requires ~10GB VRAM (during generation of sequences up to 256 tokens) along with ~12GB memory.

Alternative, use python demo.py --cpu to load and run the model on CPU only. This needs around ~20GB of memory.

Training and Evaluation

Data

We release the data used to train and evaluate the model on Hugging Face. See here for details on how to reproduce our data files and licensing information.

Setup

We recommend Python 3.9 (newer version will likely also work but are untested).

To run our training and evaluation code, please install the following:

pip install -r requirements.txt
pip install -e ./trident

We use trident, a modular framework by Fabian Schmidt that combines pytorch-lightning with hydra configs and Hugging Face datasets (along with some QoL features).

Our code is found in src and the hydra configs in configs.

Training & Testing

The entry point into our code is done by calling python run.py experiment=EXPERIMENT-FILENAME_FROM_configs/experiments as the examples below show.

On blip_checkpoint: We initialize our model with a BLIP-2 model checkpoint but only with the Q-Former and ViT. Loading the model with .from_pretrained("Salesforce/blip2-flan-t5-xl") would load the LLM, too, which we do not want. Instead, we manually downloaded the shard containing the ViT and Q-Former here, remove any LLM weights from the state_dict, and save this state_dict to disk (see this example script). blip_checkpoint points to this state_dict file.

On data_prefix: Our data setup expects all dataset json files and images to be in specific folders relative to this. If your folder structure is different, adapt the following part in the experiment yaml files:

train_data: ${data_prefix}/mblip/data  # <- Root to all dataset jsons, e.g., ${train_data}/xgqa etc.
train_image_root: ${data_prefix}/mblip/data/pretrain/images # Contains both the MSCOCO images and web images from BLIP used in instruction training.
xm3600_image_root: ${data_prefix}/iglue/datasets/Crossmodal3600/images
flickr_image_root: ${data_prefix}/iglue/datasets/flickr30k/flickr_images # Note: This folder also contains all MSCOCO images in our case.
gqa_image_root: ${data_prefix}/iglue/datasets/gqa/images
marvl_img_root: ${data_prefix}/iglue/datasets/marvl/images # Note: This folder also contains the NLVR images in the root (or for train, the numbered folders).

Warm-up

Instruction Training

If you want to merge the LoRA weights after training into the LLM, use the scripts in util.

Evaluate

You can also use module.model.huggingface_checkpoint=Gregor/mblip-mt0-xl to directly load the checkpoint from HuggingFace. Note that 1) 8bit results are slightly worse than in the paper because the ViT and Q-Former is also loaded in 8bit and not just the LLM, and 2) training does not work with this option because the Q-Former is loaded in 8bit.

Finetune

Why test_after_training=False? It was easier for us to run a separate test script than configure lightning to correctly load the best checkpoint and LoRa weights after training.

How to use with Hugging Face

The mBLIP model uses the existing BLIP-2 architecture available in Hugging Face so you can use it right away in your code. For more code examples, we refer to the BLIP-2 documentation.

Running the model on CPU

Running the model on GPU

In full precision

In half precision (bfloat16)

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.

Click to expand

# 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))

License

Our model and code is MIT licensed.

Contact

This repository was created by Gregor Geigle and Abhay Jain. If you have any questions, feel free to open an issue.

Citation

If you use our model or code, 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},
}