File size: 5,729 Bytes
4f7fd59 e444b91 4f7fd59 e444b91 4f7fd59 667524a 4f7fd59 667524a 4f7fd59 e444b91 4f7fd59 e444b91 4f7fd59 e444b91 4f7fd59 | 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 | ---
license: apache-2.0
tags:
- vision
- image-classification
- self-supervised
datasets:
- imagenet-1k
---
<div align="center">
# DINO for TI EdgeAI
### Self-Supervised Vision Transformer Backbone for Image Classification
[](https://opensource.org/licenses/Apache-2.0)
[](https://onnx.ai/)
[](https://github.com/TexasInstruments/edgeai)
[](http://www.image-net.org/)
</div>
---
## Overview
**DINO** (Self-**Di**stillation with **No** labels) is a self-supervised Vision Transformer pre-training method from Meta AI. The backbone models produce rich feature embeddings that achieve strong performance on ImageNet classification without any labels during pre-training.
These ONNX models include the **full backbone + pretrained linear classification head**, outputting 1000-class ImageNet logits `[1, 1000]`. Feature extraction follows DINO's `eval_linear.py` conventions:
- **ViT-S models**: CLS tokens from last 4 blocks concatenated → `[B, 1536]`
- **ViT-B models**: CLS token + averaged patch tokens (interleaved) → `[B, 1536]`
- **ResNet-50**: avgpool output → `[B, 2048]`
> See [DINOv2](../DINOv2/) for the improved second-generation models.
---
## Model Variants
| Model | Architecture | Params | Linear Top-1 | k-NN Top-1 | Validated Devices | Config |
|-------|-------------|--------|-------------|-----------|----------|--------|
| `dino_vits16` | ViT-S/16 | 21M | 77.0% | 74.5% | TDA4VH | [dino_vits16_config.yaml](dino_vits16_config.yaml) |
| `dino_vits8` | ViT-S/8 | 21M | 79.7% | 78.3% | TDA4VH | [dino_vits8_config.yaml](dino_vits8_config.yaml) |
| `dino_vitb16` | ViT-B/16 | 85M | 78.2% | 76.1% | TDA4VH | [dino_vitb16_config.yaml](dino_vitb16_config.yaml) |
| `dino_vitb8` | ViT-B/8 | 85M | 80.1% | 77.4% | TDA4VH | [dino_vitb8_config.yaml](dino_vitb8_config.yaml) |
| `dino_resnet50` | ResNet-50 | 23M | 75.3% | 67.5% | TDA4VH | [dino_resnet50_config.yaml](dino_resnet50_config.yaml) |
**Recommended for edge deployment:** `dino_vits16` (best accuracy/compute trade-off)
---
## Quick Start
### Prerequisites
```bash
pip install onnx>=1.22.0 onnxruntime>=1.23.2
```
### Export the Model
```bash
# Export the default model (ViT-S/16)
python prepare_model.py
# Export a specific model variant
python prepare_model.py --model dino_vitb16
# Export all supported models
python prepare_model.py --model all
# Re-run shape fixing on an already-exported ONNX
python prepare_model.py --model dino_vits16 --skip-export
```
The script automatically:
- Loads pretrained backbone from PyTorch Hub (`facebookresearch/dino:main`)
- Downloads pretrained linear classification weights from Meta AI
- Combines backbone + linear head into a single classification model
- Exports to ONNX (opset 17) and fixes input shapes to [1, 3, 224, 224]
- Validates the model outputs `[1, 1000]` class logits
### Compile and Infer uing edgeai-tidlrunner
> **Note:** Run the commands below from inside the `tidlrunner` directory (the cloned [edgeai-tidlrunner](https://github.com/TexasInstruments/edgeai-tidlrunner) repository), with `--config_path` pointing to this model's config file.
**Compile using edgeai-tidlrunner - on PC**
```bash
cd /path/to/edgeai-tidlrunner
tidlrunner-cli compile --target_device J784S4 \
--config_path /path/to/dino_vits16_config.yaml
```
**Run Inference Benchmark - on device**
```bash
cd /path/to/edgeai-tidlrunner
tidlrunner-cli infer --target_device J784S4 \
--config_path /path/to/dino_vits16_config.yaml
```
### Compile and Infer using edgeai-tidl-tools (Advanced):
Follow the instructions at https://github.com/TexasInstruments/edgeai-tidl-tools
### Deploy using edgeai-tidl-tools:
Deplyment can be done using **[edgeai-tidl-tools](https://github.com/TexasInstruments/edgeai-tidl-tools)**. For ONNX models, onnxruntime-tidl with TIDL acceleration can be used. Consult the documentation of edgeai-tidl-tools for more details.
---
## Citation
If you use these models, please cite:
```bibtex
@inproceedings{caron2021emerging,
title={Emerging Properties in Self-Supervised Vision Transformers},
author={Caron, Mathilde and Touvron, Hugo and Misra, Ishan and
J{\'e}gou, Herv{\'e} and Mairal, Julien and Bojanowski, Piotr
and Joulin, Armand},
booktitle={Proceedings of the IEEE/CVF International Conference
on Computer Vision (ICCV)},
year={2021}
}
```
---
## 🔗 Resources
| Resource | Link |
|----------|------|
| **Paper** | [arXiv:2104.14294](https://arxiv.org/abs/2104.14294) |
| **Source Code** | [facebookresearch/dino](https://github.com/facebookresearch/dino) |
| **edgeai-tidl-tools** | [GitHub](https://github.com/TexasInstruments/edgeai-tidl-tools) |
| **edgeai-tidlrunner** | [GitHub](https://github.com/TexasInstruments/edgeai-tidlrunner) |
| **EdgeAI SDK** | [Documentation](https://github.com/TexasInstruments/edgeai/blob/main/edgeai-mpu/readme_sdk.md) |
| **DINOv2** | [Improved successor](../DINOv2/) |
---
## Related Models
<table>
<tr>
<td align="center">
**DINOv2**
Improved DINO
Higher accuracy
</td>
<td align="center">
**ViT-S/16**
Recommended
Best edge trade-off
</td>
<td align="center">
**ResNet-50**
CNN backbone
Lower compute
</td>
<td align="center">
**CLIP**
Vision-Language
Zero-shot capable
</td>
</tr>
</table>
---
<div align="center">
**Maintained by:** Texas Instruments EdgeAI Team
**Last Updated:** August 2026
</div>
|