Groma: Grounded Multimodal Assistant

Groma: Localized Visual Tokenization for Grounding Multimodal Large Language Models
Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, Xiaojuan Qi

Groma is an MLLM with exceptional region understanding and visual grounding capabilities. It can take user-defined region inputs (boxes) as well as generate long-form responses that are grounded to visual context.

Groma presents a novel paradigm of grounded MLLMs. (a) LLM for localization (e.g., Kosmos-2, Shikra); (b) External modules for localization (e.g., Lisa); and (c) Visual tokenier for localization (Groma).

Install
Model
Data
Training
Inference
Evaluation

Performance

State-of-the-art performance on referring expression comprehension (REC) benchmarks among multimodal large language models.

Method	RefCOCO			RefCOCO+			RefCOCOg		Avergae
Method	val	testA	testB	val	testA	testB	val	test	Avergae
Shikra	87.01	90.61	80.24	81.60	87.36	72.12	82.27	82.19	82.93
Ferret	87.49	91.35	82.45	80.78	87.38	73.14	83.93	84.76	83.91
MiniGPT-v2	88.69	91.65	85.33	79.97	85.12	74.45	84.44	84.66	84.29
Qwen-VL	89.36	92.26	85.34	83.12	88.25	77.21	85.58	85.48	85.83
Groma	89.53	92.09	86.26	83.90	88.91	78.05	86.37	87.01	86.52

Installation

Clone the repository

git clone https://github.com/FoundationVision/Groma.git
cd Groma

Create the conda environment and install dependencies

conda create -n groma python=3.9 -y
conda activate groma
conda install pytorch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 pytorch-cuda=11.8 -c pytorch -c nvidia
pip install --upgrade pip  # enable PEP 660 support
pip install -e .

cd mmcv
MMCV_WITH_OPS=1 pip install -e .
cd ..

Install falsh-attention for training

pip install ninja
pip install flash-attn --no-build-isolation

Model Weights

To play with Groma, please download the model weights from huggingface.

We additionally provide pretrained checkpoints from intermediate training stages. You can start from any point to customize training.

Training stage	Required checkpoints
Detection pretraining	DINOv2-L
Alignment pretraining	Vicuna-7b-v1.5, Groma-det-pretrain
Instruction finetuning	Groma-7b-pretrain

Prepare Data

We provide instructions to download datasets used at different training stages of Groma, including Groma Instruct, a 30k viusally grounded conversation dataset constructed with GPT-4V. You don’t have to download all of them unless you want to train Groma from scratch. Please follow instructions in DATA.md to prepare datasets.

Training stage	Data types	Datasets
Detection pretraining	Detection	COCO, Objects365, OpenImages, V3Det, SA1B
Alignment pretraining	Image caption	ShareGPT-4V-PT
	Grounded caption	Flickr30k Entities
	Region caption	Visual Genome, RefCOCOg
	REC	COCO, RefCOCO/g/+, Grit-20m
Instruction finetuning	Grounded caption	Flickr30k Entities
	Region caption	Visual Genome, RefCOCOg
	REC	COCO, RefCOCO/g/+
	Instruction following	Groma Instruct, LLaVA Instruct, ShareGPT-4V

Training

For detection pretraining, please run

bash scripts/det_pretrain.sh {path_to_dinov2_ckpt} {output_dir}

For alignment pretraining, please run

bash scripts/vl_pretrain.sh {path_to_vicuna_ckpt} {path_to_groma_det_pretrain_ckpt} {output_dir}

For instruction finetuning, please run

bash scripts/vl_finetune.sh {path_to_groma_7b_pretrain_ckpt} {output_dir}

Inference

To test on single image, you can run

python -m groma.eval.run_groma \
    --model-name {path_to_groma_7b_finetune} \
    --image-file {path_to_img} \
    --query {user_query} \
    --quant_type 'none' # support ['none', 'fp16', '8bit', '4bit'] for inference

Evaluation

For evaluation, please refer to EVAL.md for more details.

Citation

If you find this repo useful for your research, feel free to give us a star ⭐ or cite our paper:

@article{ma2024groma,
  title={Groma: Localized Visual Tokenization for Grounding Multimodal Large Language Models},
  author={Ma, Chuofan and Jiang, Yi and Wu, Jiannan and Yuan, Zehuan and Qi, Xiaojuan},
  journal={arXiv preprint arXiv:2404.13013},
  year={2024}
}