File size: 9,378 Bytes
32b542e |
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 |
# Uni-Perceiver
This repository contains training (pre-training, fine-tuning, prompt-tuning), evaluation code and pretrained models for the following papers:
> [Uni-Perceiver](https://arxiv.org/abs/2112.01522): Pre-training Unified Architecture for Generic Perception for Zero-shot and Few-shot Tasks, CVPR 2022.
> [Uni-Perceiver-MoE](https://arxiv.org/abs/2206.04674): Learning Sparse Generalist Models with Conditional MoEs, NeurIPS 2022.
## Introduction
__Uni-Perceiver__ is a generalist model (generic perception model) that can process a variety of modalities and tasks with unified modeling and shared
parameters. Different perception tasks are modeled as the same formulation, that is, finding the maximum likelihood target for each input through the similarity of their representations. Meanwhile, Uni-Perceiver is pre-trained on several uni-modal and multi-modal tasks, and evaluated on a variety of downstream tasks, including novel tasks that did not appear in the pre-training stage.
Thanks to the unified formulation, it shows the ability of zero-shot inference on novel tasks, and shows promising performance close to or on par with SOTA results by prompt tuning or finetuning.

In __Uni-Perceiver-MoE__, we found that the interference among different tasks and modalities can lead to performance degradation of generalist models on some tasks compared with task-specialized models. We introduce the Conditional Mixture-of-Experts (Conditional MoEs) to mitigate such interference. By incorporating the proposed Conditional MoEs, Uni-Perceiver-MoE can effectively mitigate the interference across tasks and modalities, and achieves state-of-the-art results on a series of downstream tasks via prompt tuning on 1% of downstream data. Moreover, the introduction of Conditional MoEs still holds the generalization ability of generalist models to conduct zero-shot inference on new tasks,

## Main Results and Pretrained Models
### Base Models
<table border="1" width="100%">
<tr align="center">
<th>Task</th>
<th>Image Classification</th>
<th colspan="2">Image Caption</th>
<th colspan="4">Image Retrieval</th>
<th>Video Classification</th><th>Video Caption</th><th colspan="2">Video Retrieval</th>
</tr>
<tr align="center">
<td>Dataset</td><td>ImageNet-1k</td><td>MSCOCO</td><td>Flickr30k</td><td colspan="2">MSCOCO</td><td colspan="2">Flickr30k</td><td>Kinetics-400</td><td>MSVD</td><td colspan="2">MSVD</td>
</tr>
<tr align="center">
<td>Split</td><td>ILSVRC 2012 val</td><td>Karpathy test</td><td>test</td><td colspan="2">Karpathy test</td><td colspan="2">test</td><td>test-dev</td><td>val</td><td>val</td><td colspan="2">val</td>
</tr>
<tr align="center">
<td>Metric</td><td>Acc@1</td><td>BLEU-4</td><td>BLEU-4</td><td>R@1 i2t</td><td>R@1 t2i</td><td>R@1 i2t</td><td>R@1 t2i</td><td>Acc@1</td><td>BLEU-4</td><td>R@1 v2t</td><td>R@1 t2v</td>
</tr>
</tr>
<tr align="center">
<td>Uni-Perceiver<sub>BASE</sub> w/o Tuning</td><td>79.2 </td><td>32.0</td><td>14.7 </td><td>64.9 </td><td>50.7 </td><td>82.3 </td><td>71.1</td> <td>74.5 </td><td>22.6 </td><td>50.3</td><td>38.7 </td>
</tr>
<tr align="center">
<td>Uni-Perceiver<sub>BASE</sub> PT (1%)</td><td>80.9 </td><td>35.5</td><td>30.2</td><td>68.4 </td><td>51.9 </td><td>91.0 </td><td>76.0 </td><td>74.8 </td><td>59.5 </td><td>62.7 </td><td>43.8 </td>
</tr>
<tr align="center">
<td>Uni-Perceiver<sub>BASE</sub> FT (100%)</td><td>84.0</td><td>36.4 </td><td>31.2 </td><td>69.8</td><td>53.9 </td><td>92.7</td><td>77.5</td><td>77.7 </td><td>63.3 </td><td>62.8</td><td>45.8 </td>
</tr>
<tr align="center">
<td>Uni-Perceiver-MoE<sub>BASE</sub> w/o Tuning</td><td>80.3 </td><td>33.2</td><td>15.9 </td><td>64.6 </td><td>51.6 </td><td>82.1 </td><td>75.8</td> <td>76.8 </td><td>23.4 </td><td>52.8</td><td>40.0 </td>
</tr>
<tr align="center">
<td>Uni-Perceiver-MoE<sub>BASE</sub> PT (1%)</td><td>82.0 </td><td>36.8</td><td>30.7</td><td>68.9 </td><td>52.6 </td><td>91.3 </td><td>78.5 </td><td>77.2 </td><td>60.0 </td><td>65.6 </td><td>45.3 </td>
</tr>
<tr align="center">
<td>Uni-Perceiver-MoE<sub>BASE</sub> FT (100%)</td><td>84.5</td><td>37.3 </td><td>32.4 </td><td>70.5</td><td>54.1 </td><td>93.6</td><td>79.8</td><td>79.3 </td><td>65.4 </td><td>65.0</td><td>47.8 </td>
</tr>
</table>
### Large Models
<table border="1" width="100%">
<tr align="center">
<th>Task</th>
<th>Image Classification</th>
<th colspan="2">Image Caption</th>
<th colspan="4">Image Retrieval</th>
<th>Video Classification</th><th>Video Caption</th><th colspan="2">Video Retrieval</th>
</tr>
<tr align="center">
<td>Dataset</td><td>ImageNet-1k</td><td>MSCOCO</td><td>Flickr30k</td><td colspan="2">MSCOCO</td><td colspan="2">Flickr30k</td><td>Kinetics-400</td><td>MSVD</td><td colspan="2">MSVD</td>
</tr>
<tr align="center">
<td>Split</td><td>ILSVRC 2012 val</td><td>Karpathy test</td><td>test</td><td colspan="2">Karpathy test</td><td colspan="2">test</td><td>test-dev</td><td>val</td><td>val</td><td colspan="2">val</td>
</tr>
<tr align="center">
<td>Metric</td><td>Acc@1</td><td>BLEU-4</td><td>BLEU-4</td><td>R@1 i2t</td><td>R@1 t2i</td><td>R@1 i2t</td><td>R@1 t2i</td><td>Acc@1</td><td>BLEU-4</td><td>R@1 v2t</td><td>R@1 t2v</td>
</tr>
<tr align="center">
<td>Uni-Perceiver<sub>LARGE</sub> w/o Tuning</td><td>82.7 </td><td> 35.3 </td><td> 15.1 </td><td>67.8 </td><td>54.1 </td><td> 83.7</td><td> 74.2 </td><td> 79.5</td><td>24.7 </td><td> 45.4 </td><td>34.2 </td>
</tr>
<tr align="center">
<td>Uni-Perceiver<sub>LARGE</sub> PT (1%)</td><td>84.2 </td><td>38.6 </td><td> 32.9</td><td> 73.3 </td><td>56.2 </td><td>92.1 </td><td> 80.0</td><td> 80.0</td><td> 67.2</td><td> 65.5 </td><td>48.6 </td>
</tr>
<tr align="center">
<td>Uni-Perceiver<sub>LARGE</sub> FT (100%)</td><td>86.2 </td><td> 39.2 </td><td> 35.5 </td><td>74.4 </td><td>57.9 </td><td>94.7 </td><td> 82.1</td><td>81.9 </td><td>68.3 </td><td> 65.2 </td><td>50.8 </td>
</tr>
<tr align="center">
<td>Uni-Perceiver-MoE<sub>LARGE</sub> w/o Tuning</td><td>83.4 </td><td> 35.5 </td><td> 15.8 </td><td>67.9 </td><td>55.3 </td><td> 83.6</td><td> 75.9 </td><td> 82.1</td><td>24.6 </td><td> 45.7 </td><td>41.9 </td>
</tr>
<tr align="center">
<td>Uni-Perceiver-MoE<sub>LARGE</sub> PT (1%)</td><td>84.9 </td><td>39.3 </td><td> 33.7</td><td> 73.3 </td><td>57.1 </td><td>92.4 </td><td> 80.6</td><td> 83.0</td><td> 67.6</td><td> 66.4 </td><td>50.3 </td>
</tr>
<tr align="center">
<td>Uni-Perceiver-MoE<sub>LARGE</sub> FT (100%)</td><td>86.4 </td><td> 40.5 </td><td> 36.2 </td><td>74.7 </td><td>58.3 </td><td>94.1 </td><td> 83.7</td><td>84.2 </td><td>68.9 </td><td> 67.6 </td><td>52.3 </td>
</tr>
</table>
* The numbers are slightly better than the original paper of Uni-Perceiver, which are from the reproduced version of Uni-Perceiver used as the baseline of [Uni-Perceiver-MoE](https://arxiv.org/abs/2206.04674).
* The image resolution for all tasks is `224x224`.
* See [OtherResults.md](data/other_results.md) for results on more tasks and datasets.
## Usage
### Requirements
* Linux, CUDA>=10.1, GCC>=5.4
* Python >=3.7
* pytorch >= 1.8.0
* JAVA >= 1.8 (for caption task evaluation)
### Installation
```bash
git clone https://github.com/fundamentalvision/Uni-Perceiver
cd Uni-Perceiver
pip install -r requirements.txt
```
### Data
See [prepare_data.md](data/prepare_data.md).
### Pre-trained Model Weights
See [checkpoints.md](data/checkpoints.md).
### Pre-training
See [pretraining.md](data/pretraining.md).
### Fine-tuning
See [finetuning.md](data/finetuning.md).
### Prompt-tuning
See [prompt_tuning.md](data/prompt_tuning.md).
### Inference
See [inference.md](data/inference.md).
### TODO
* release more pretrained models
- [ ] Uni-Perceiver Tiny model
- [ ] Uni-Perceiver Small model
- [ ] Uni-Perceiver Huge model
* support more datasets and tasks
## License
Uni-Perceiver is licensed under the [Apache-2.0 License](./LICENSE).
<br></br>
## Citing Uni-Perceiver
If you find Uni-Perceiver useful in your research, please consider giving a star ⭐ and citing:
```bibtex
@article{zhu2021uni,
title={Uni-Perceiver: Pre-training Unified Architecture for Generic Perception for Zero-shot and Few-shot Tasks},
author={Zhu, Xizhou and Zhu, Jinguo and Li, Hao and Wu, Xiaoshi and Wang, Xiaogang and Li, Hongsheng and Wang, Xiaohua and Dai, Jifeng},
journal={arXiv preprint arXiv:2112.01522},
year={2021}
}
```
```bibtex
@article{zhu2022uni,
title={Uni-Perceiver-MoE: Learning Sparse Generalist Models with Conditional MoEs},
author={Zhu, Jinguo and Zhu, Xizhou and Wang, Wenhai and Wang, Xiaohua and Li, Hongsheng and Wang, Xiaogang and Dai, Jifeng},
journal={arXiv preprint arXiv:2206.04674},
year={2022}
}
```
### Acknowledgements
Many thanks to following codes that help us a lot in building this codebase:
* [Detectron2](https://github.com/facebookresearch/detectron2)
* [X-modaler](https://github.com/YehLi/xmodaler)
* [deit](https://github.com/facebookresearch/deit)
* [VL-BERT](https://github.com/jackroos/VL-BERT)
* [TimeSformer](https://github.com/facebookresearch/TimeSformer)
* [CLIP](https://github.com/openai/CLIP)
|