Skip to content
/ imp Public
forked from MILVLG/imp

a family of multimodal small language models

License

Notifications You must be signed in to change notification settings

bruceisme/imp

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

78 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

😈 Imp

[📖 Technical report]  [😈Demo]  [🤗Huggingface]

This repository contains the official training/evaluation code of the Imp project, which aims to provide a family of highly capable yet efficient large multimodal models (LMMs). Imp-v1-3B is our attempt with only 3B parameters, which is build upon a small yet powerful SLM Phi-2 (2.7B) and a powerful visual encoder SigLIP (0.4B), and trained on the LLaVA-v1.5 training set. After that, we further investigate different LLMs, training strategies, and traning data, and obtain a series of models termed Imp-v1.5-2B/3B/4B.

As shown in the Evaluation, our Imp model significantly outperforms the counterparts of similar model sizes, and even achieves slightly better performance than the strong LLaVA-7B model on various multimodal benchmarks.

Now the training and evaluation codes of Imp-v1-3B and Imp-v1.5-2B/3B/4B are released. And the Multimodal instruction tuning data will be released as soon as possible.

Updates

  • July 4, 2024: Training and evaluation codes of Imp-v1.5-2B/3B/4B are released.
  • May 21, 2024: The technical report and corresponding Imp-v1.5-2B/3B/4B model series are released on 🤗HuggingFace.
  • February 9, 2024: Training and evaluation codes of the Imp-v1-3B model are released.

Table of Contents

Prerequisites

  1. Clone this repository and navigate to the folder
git clone https://github.com/MILVLG/imp.git
cd imp
  1. Install Package

We recommend using Anaconda to create a new environment for the project, and install the requirements with the following commands:

conda create -n imp python=3.10 -y
conda activate imp
pip install -r requirements.txt
pip install flash-attn==2.4.2 --no-build-isolation

Model-zoo

All models of Imp family can be checked on Model_Zoo.md.

Customized Continue-Finetuning

You can directly finetune different versions of Imp using your own custom dataset use scripts/finetune_lora_custom.sh. The custom dataset should be in the LLaVA-1.5 format.

If you want to train an imp from scratch, please refers to Training.

#imp_model should be hg repository or local model path with imp
#version should be the same as imp, only in phi2/qwen1.5/phi3
bash scripts/finetune_lora_custom.sh -imp_model "path/to/imp" -version "phi2"

You will get a trained lora model imp-{version}-custom-lora under ./checkpoints/. And you can use the following script to merge lora model and base imp into a single one for release and evaluation. Our evaluation script supports both the lora model and merged model checkpoints.

#version should be the same as imp, only in phi2/qwen1.5/phi3
bash scripts/merge.sh -imp_model "path/to/imp"  -lora "path/to/lora" -version "phi2"

After that, a checkpoint file will be stored in ./checkpoints/imp-{version}-merged.

Training

The training pipeline and datasets of our Imp models are directly inherited from LLaVA-v1.5. The training

  • Multimodal pretraining: train a projector on a subset of ∼558K image-text pairs to connect a frozen pretrained vision encoder and a frozen LLM.
  • Multimodal instruction tuning: fine-tune the projector and LoRA in the LLM with multimodal instruction data and VQA-formatted data to empower the MLSM the ability of multimodal instruction following.
  • Note: For Imp-v1.5-2B/3B/4B, we append OCR&chart-oriented datasets and GPT4V-annotated datasets in the base of llava 1.5's Multimodal instruction tuning datasets. The augmented data will be released as soon as possible. More information about the data refers to imp

Imp is trained on 8 A100 (40G) GPUs. You can reduce the per_device_train_batch_size and increase the gradient_accumulation_steps to match your resources. .But always keep the global batch size the same: global_batch_size = per_device_train_batch_size $\times$ gradient_accumulation_steps $\times$ num_gpus.

Training scripts

Stage-1: Multimodal pretraining

Take Imp-v1-3B as an example, all stages' checkpoint can be downloaded in Model_Zoo.md. Please download the caption annotations blip_laion_cc_sbu_558k.json and images from here. Move the downloaded files to the ./datasets folder, with image folder unzipped and renamed to pretrain_images. Then run the following command to start the training process:

bash scripts/pretrain.sh 

After that, a checkpoint file will be stored in ./checkpoints/imp-v1-3b-stage1.

Stage-2: Multimodal instruction tuning

Please download the annotation file of the mixed instruction tuning data llava_v1_5_mix665k.json, and download the images from constituting datasets:

After downloading all of them, organize the data as follows:

datasets
├── llava_v1_5_mix665k.json
└── finetune_images
    ├── coco
    │   └── train2017
    ├── gqa
    │   └── images
    ├── ocr_vqa
    │   └── images
    ├── textvqa
    │   └── train_images
    └── vg
        ├── VG_100K
        └── VG_100K_2

Then, you can start the training process by the following script. If you use your custom dataset, you can refer to llava_v1_5_mix665k.json to format your data.

bash scripts/finetune_lora.sh
# bash scripts/finetune.sh # fully finetuning is not recommended

You will get a trained model imp-v1-3b-stage2-lora (a LoRA diff if you use finetune_lora.sh) under ./checkpoints/ when the training is done.

Lora merging

After the above model training, the model checkpoint consists of base model and lora model. You can merge the lora model and base imp model into a single one by scripts merge.sh in merge.Our evaluation script supports both the lora model and merged model checkpoints. However, if you want to fine-tune the model on your own custom dataset, only the merged model is supported.

Evaluation

We follow the evaluation of LLaVA-v1.5 and conduct experiments on 9 commonly-used benchmarks, including 5 academic VQA benchmarks and 4 popular MLLM benchmarks. All evaluation scripts are placed in the scripts/eval folder.

Before preparing task-specific data, you should download eval.zip and unzip it to ./playground/data/eval. For more specific instructions, please refer to LLaVA's Evaluation.md.

It is supported to evaluate your reproduced model checkpoints or our released model. For more detailed evaluation scripts, please refer to Evaluation.md.

Using our provided model, you can reproduce the following results. Our model significantly outperforms existing MSLMs of similar model sizes, and is comparable with the strong LLaVA-v1.5-7b model.

Models VQAv2 GQA VizWiz SQA(IMG) TextVQA POPE MME(P) MMB MM-Vet
LLaVA-v1.5-lora (7B) 79.10 63.00 47.80 68.40 58.20 86.40 1476.9 66.10 30.2
TinyGPT-V (3B) - 33.60 24.80 - - - - - -
LLaVA-Phi (3B) 71.40 - 35.90 68.40 48.60 85.00 1335.1 59.80 28.9
MobileVLM (3B) - 59.00 - 61.00 47.50 84.90 1288.9 59.60 -
MC-LLaVA (3B) 64.24 49.60 24.88 - 38.59 80.59 - - -
Imp-v1-3B 79.45 58.55 50.09 69.96 59.38 88.02 1434.0 66.49 33.1
Imp-v1.5-2B-Qwen1.5 79.2 61.93 39.16 66.14 54.52 86.74 1304.8 56.95 33.5
Imp-v1.5-3B-Phi2 81.18 63.54 54.13 72.78 59.84 88.87 1446.4 72.94 43.3
Imp-v1.5-4B-Phi3 81.46 63.51 51.16 78.28 60.16 86.86 1507.7 73.28 44.6

Deployment

Based on MLC-LLM, we provide a lightweight deployment solution so that imp can inference efficiently on the mobile device.

  • After 4-bit quantification, imp only takes up about 1.9G of storage space and is fully capable of running on mobile phones.
  • All Android devices are supported and IOS will come soon.
  • Textual and visual modalities are supported.

More details can be found in MILVLG/mlc-imp.

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

About us

This project is maintained by the MILVLG@Hangzhou Dianzi University (HDU) led by Prof. Zhou Yu and Jun Yu, and is mainly developed by Zhenwei Shao and Xuecheng Ouyang. We hope our model may serve as a strong baseline to inspire future research on MSLM, as well as its derivative applications on mobile devices and robots.

Citation

If you use our model or refer our work in your studies, please cite:

@article{imp2024,
  title={Imp: Highly Capable Large Multimodal Models for Mobile Devices},
  author={Shao, Zhenwei and Yu, Zhou and Yu, Jun and Ouyang, Xuecheng and Lihao, Zheng and Zhenbiao, Gai and Mingyang, Wang and Jiajun, Ding},
  journal={arXiv preprint arXiv:2405.12107},
  year={2024}
}

About

a family of multimodal small language models

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 92.0%
  • Shell 4.7%
  • JavaScript 1.7%
  • HTML 1.3%
  • CSS 0.3%