File size: 4,858 Bytes
382bd97 cb298cd 382bd97 cb298cd d80fe34 cb298cd d80fe34 cb298cd d80fe34 cb298cd ab5a36d d80fe34 cb298cd ab5a36d cb298cd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 |
---
license: apache-2.0
pipeline_tag: text-generation
datasets:
- liuhaotian/LLaVA-Pretrain
- liuhaotian/LLaVA-Instruct-150K
---
# 😈 Imp
> A very small man can cast a very large shadow.
>
> ——*George R.R. Martin, A Clash of Kings*
\[[Paper](https://arxiv.org/abs/2405.12107)\] [[Demo](https://xmbot.net/imp/)\] [[Github](https://github.com/MILVLG/imp)\]
## Introduction
The Imp project aims to provide a family of a strong multimodal `small` language models (MSLMs). Our `Imp-v1.5-2B-Qwen1.5` is a strong MSLM with only **2B** parameters, which is build upon a small yet powerful SLM [Qwen1.5-1.8B-Chat ](https://huggingface.co/Qwen/Qwen1.5-1.8B-Chat)(1.8B) and a powerful visual encoder [SigLIP ](https://huggingface.co/google/siglip-so400m-patch14-384)(0.4B), and trained on the [LLaVA-v1.5](https://github.com/haotian-liu/LLaVA) training set.
As shown in the Table below, `Imp-v1.5-2B-Qwen1.5` significantly outperforms the counterparts of similar model sizes, and even achieves slightly better performance than the strong LLaVA-7B model on various multimodal benchmarks.
We release our model weights and provide an example below to run our model . Detailed technical report and corresponding training/evaluation code will be released soon on our [GitHub repo](https://github.com/MILVLG/imp). We will persistently improve our model and release the next versions to further improve model performance :)
## How to use
**Install dependencies**
```bash
pip install transformers # latest version is ok, but we recommend v4.36.0
pip install -q pillow accelerate einops
```
You can use the following code for model inference. The format of text instruction is similar to [LLaVA](https://github.com/haotian-liu/LLaVA). A Colab page to run this example is provided [here](https://colab.research.google.com/drive/1EBYky6xIPjnlPppo2gZaiNK6gEsjXgom?usp=drive_link#scrollTo=2-VpU6QzWCVZ). Note that the example can only be run on GPUs currently.
```Python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from PIL import Image
torch.set_default_device("cuda")
#Create model
model = AutoModelForCausalLM.from_pretrained(
"MILVLG/Imp-v1.5-2B-Qwen1.5",
torch_dtype=torch.float16,
device_map="auto",
trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("MILVLG/Imp-v1.5-2B-Qwen1.5", trust_remote_code=True)
#Set inputs
text = "<|im_start|>system\nA chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.<|im_end|>\n<|im_start|>user\n<image>\nWhat are the colors of the bus in the image?<|im_end|>\n<|im_start|>assistant"
image = Image.open("images/bus.jpg")
input_ids = tokenizer(text, return_tensors='pt').input_ids
image_tensor = model.image_preprocess(image)
#Generate the answer
output_ids = model.generate(
input_ids,
max_new_tokens=100,
images=image_tensor,
use_cache=True)[0]
print(tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip())
```
## Model evaluation
We conduct evaluation on 9 commonly-used benchmarks, including 5 academic VQA benchmarks and 4 popular MLLM benchmarks, to compare our Imp model with LLaVA (7B) and existing MSLMs of similar model sizes.
| Models | Size | VQAv2 | GQA | SQA(IMG) | TextVQA | POPE | MME(P) | MMB |MMBCN |MM-Vet|
|:--------:|:-----:|:----:|:-------------:|:--------:|:-----:|:----:|:-------:|:-------:|:-------:|:-------:|
| [Mini-Gemini-2B](https://github.com/dvlab-research/MGM) | 2B |- | -| 56.2 |-| -| 1341 | 59.8 |- | 31.1|
| [Bunny-v1.0-2B-zh](https://huggingface.co/BAAI/Bunny-v1_0-2B-zh) | 2B |76.6 | 59.6| 64.6 |-| 85.8 | 1300.8 | 59.1 |58.5 | 31.1|
| **Imp-v1.5-2B-Qwen1.5** | 3B | 79.2 | 61.93 | 66.14| 54.52 | 86.74| 1304.8 | 63.83| 61.34 |33.5|
## License
This project is licensed under the Apache License 2.0 - see the [LICENSE](https://www.apache.org/licenses/LICENSE-2.0) file for details.
## About us
This project is maintained by the [MILVLG](https://github.com/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:
```bibtex
@misc{imp2024,
author = {Shao, Zhenwei and Ouyang, Xuecheng and Yu, Zhou and Yu, Jun},
title = {Imp: An Emprical Study of Multimodal Small Language Models},
year = {2024},
url = {https://huggingface.co/MILVLG/Imp-v1.5-2B-Qwen1.5}
}
``` |