File size: 8,573 Bytes
06dd81f 6c6c140 76038a7 f805e81 76038a7 f805e81 76038a7 bfd5ad3 76038a7 bfd5ad3 76038a7 |
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 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 |
---
license: apache-2.0
---
This is the official repo for paper [Supervised Fine-tuning *in turn* Improves Visual Foundation Models](https://arxiv.org/abs/2401.10222).
<div align="center">
📃[**Paper (ArXiv)**](https://arxiv.org/abs/2401.10222) **|** [**Code**](https://github.com/TencentARC/ViSFT) **|** 🤗[**Huggingface**](https://huggingface.co/TencentARC/ViSFT)
</div>
## News
* [2024/01/19] We open source the [ViSFT]() including training scripts and weights. Evaluation codes will be released soon.
## Introduction
Image-text training like CLIP has dominated the pretraining of vision foundation models in recent years. Subsequent efforts have been made to introduce region-level visual learning into CLIP’s pretraining but face scalability challenges due to the lack of large-scale region-level datasets. Drawing inspiration from supervised fine-tuning (SFT) in natural language processing such as instruction tuning, we explore the potential of fine-grained SFT in enhancing the generation of vision foundation models after their pretraining. Thus a two-stage method **ViSFT** (**Vi**sion **SFT**) is proposed to unleash the fine-grained knowledge of vision foundation models. In ViSFT, the vision foundation model is enhanced by performing visual joint learning on some in-domain tasks and then tested on out-of-domain benchmarks. With updating using ViSFT on 8 V100 GPUs in less than 2 days, a vision transformer with over 4.4B parameters shows improvements across various out-of-domain benchmarks including vision and vision-linguistic scenarios.
## Installation
### creating a conda environment
```
conda create -n ViSFT python=3.8
conda activate ViSFT
```
### Install pytorch
we use torch1.12 with CUDA11.3 on 8 NVIDIA Volta V100- SXM2-32GB GPUs
```
pip install --extra-index-url https://download.pytorch.org/whl/cu113 torch==1.12.0
pip install --extra-index-url https://download.pytorch.org/whl/cu113 torchvision==0.13.0
pip install --extra-index-url https://download.pytorch.org/whl/cu113 torchaudio==0.12.0
```
### xformers installation
Flash attention is required for running EVA-ViT-E.
please refer to [xformers](https://github.com/facebookresearch/xformers)
### loralib installation
```
pip install --user git+https://github.com/microsoft/LoRA
```
### compile MSDeform for Mask2former head
```
cd ./mmf/models/visft/ops
sudo sh make.sh
# back to root dir
cd ../../../../
```
### Other packages installation
```
pip install -r requirements.txt
```
## Dataset Preparation
export DATA_PATH=your_data_path
### image caption
Generating hdf5 files for image caption following [hdf5](https://github.com/sgrvinod/a-PyTorch-Tutorial-to-Image-Captioning/blob/master/create_input_files.py)
file strcture:
```
DATA_PATH/
└── processed_datasets/
└─── coco_caption_hdf5_files
├──TEST_CAPLENS_coco_5_cap_per_img_5_min_word_freq.json
├──TEST_CAPTIONS_coco_5_cap_per_img_5_min_word_freq.json
├──TEST_IMAGES_coco_5_cap_per_img_5_min_word_freq.hdf5
├──TRAIN_CAPLENS_coco_5_cap_per_img_5_min_word_freq.json
├──TRAIN_CAPTIONS_coco_5_cap_per_img_5_min_word_freq.json
├──TRAIN_IMAGES_coco_5_cap_per_img_5_min_word_freq.hdf5
├──VAL_CAPLENS_coco_5_cap_per_img_5_min_word_freq.json
├──VAL_CAPTIONS_coco_5_cap_per_img_5_min_word_freq.json
├──VAL_IMAGES_coco_5_cap_per_img_5_min_word_freq.hdf5
└───WORDMAP_coco_5_cap_per_img_5_min_word_freq.json
```
### Detection & Segmentation
file strcture:
```
DATA_PATH/
└── public_datasets/
└─── coco
├──train2017
├──val2017
├──test2017
└───annotations
├──instances_train2017.json
├──instances_val2017.json
└───image_info_test-dev2017.json
```
## Training
### Stage1
To get compatible in-domain task heads. Using 8 NVIDIA Volta V100-SXM2-32GB GPUs for every in-domain task head.
**For eva-vit-g**
Preparing weights from [LAVIS](https://github.com/salesforce/LAVIS)
```
wget https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/eva_vit_g.pth
```
Adding your weights path to configs under dir:./projects/visft/configs/stage1/eva_g/
```
backbone_dir: path/eva_vit_g.pth
```
Implementing training
```
bash ./scripts/stage1_train/eva_g/caption.sh
bash ./scripts/stage1_train/eva_g/detection.sh
bash ./scripts/stage1_train/eva_g/segment.sh
```
**For eva-vit-e**
Preparing EVA-CLIP weights from [EVA](https://huggingface.co/QuanSun/EVA-CLIP/blob/main/EVA02_CLIP_E_psz14_plus_s9B.pt)
Extact ViT weights
```
python ./scripts/preprocess/extract_eva_e_vit.py
```
Adding your weights path to configs under dir:./projects/visft/configs/stage1/eva_e/
```
backbone_dir: path/EVA02_CLIP_E_psz14_plus_s9B_Visual.pt
```
Implementing training
```
# can be executed in parallel
bash ./scripts/stage1_train/eva_e/caption.sh
bash ./scripts/stage1_train/eva_e/detection.sh
bash ./scripts/stage1_train/eva_e/segment.sh
```
Or you can use the weights we provided.
| In-domain Heads | | |
|----------|:-------------:|:-------------:|
| | EVA-G | EVA-E|
| Caption Head | [weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_g_caption_heads.ckpt) | [weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_e_caption_heads.ckpt)|
| Segment Head | [weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_g_segment_heads.ckpt) |[weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_e_segment_heads.ckpt)|
| Detection Head | [weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_g_detection_heads.ckpt) |[weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_e_detection_heads.ckpt)|
### Stage2
**For eva-vit-g**
Adding your weights path to configs under dir:./projects/visft/configs/stage2/eva_g/stage2.yaml
```
backbone_dir: path/eva_vit_g.pth
caption_ckpt_path: 'path/eva_g_caption_heads.ckpt'
segment_ckpt_path:'path/eva_g_segment_heads.ckpt'
detection_ckpt_path: 'path/eva_g_detection_heads.ckpt'
```
Implementing training
```
bash ./scripts/stage2_train/eva_g/stage2.sh
```
**For eva-vit-e**
Adding your weights path to configs under dir:./projects/visft/configs/stage2/eva_e/stage2.yaml
```
backbone_dir: path/EVA02_CLIP_E_psz14_plus_s9B_Visual.pt
caption_ckpt_path: 'path/eva_e_caption_heads.ckpt'
segment_ckpt_path:'path/eva_e_segment_heads.ckpt'
detection_ckpt_path: 'path/eva_e_detection_heads.ckpt'
```
Implementing training
```
bash ./scripts/stage2_train/eva_e/stage2.sh
```
### Get LoRA Weights
You can extract expected LoRA weights by
```
python ./scripts/postprocess/extract_lora_weights.py
```
Or use the LoRA weights we provide:
| LoRA weights | | |
|----------|:-------------:|:-------------:|
| Iters| EVA-G | EVA-E|
| 5k | [weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_g_lora_5000.pt) | [weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_e_lora_5000.pt)|
| 10k | [weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_g_lora_10000.pt) |[weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_e_lora_10000.pt)|
| 15k | [weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_g_lora_15000.pt) |[weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_e_lora_15000.pt)|
| 20k | [weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_g_lora_20000.pt) |[weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_e_lora_20000.pt)|
| 50k | [weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_g_lora_50000.pt) |[weights](https://huggingface.co/TencentARC/ViSFT/blob/main/eva_e_lora_50000.pt)|
## Evaluation Benchmarks
- [] Zero-shot Image Classification
- [] Zero-shot Image-text Retrieval
- [] OCR
- [] Grounded Object Indentification
- [] VQA
- [] Image Captioning on NoCaps
## Acknowledgement
The code of ViSFT is based on the official implementation of [mmf](https://github.com/facebookresearch/mmf), [EVA](https://github.com/baaivision/EVA/tree/master) and [LAVIS](https://github.com/salesforce/LAVIS/tree/main)
## Citation
If you found our work valuable, please cite:
```
@misc{jiang2024supervised,
title={Supervised Fine-tuning in turn Improves Visual Foundation Models},
author={Xiaohu Jiang and Yixiao Ge and Yuying Ge and Chun Yuan and Ying Shan},
year={2024},
eprint={2401.10222},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
|