Image Classification
vision
self-supervised
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

[![License](https://img.shields.io/badge/License-Apache%202.0-blue?style=for-the-badge)](https://opensource.org/licenses/Apache-2.0)
[![Framework](https://img.shields.io/badge/Framework-ONNX-orange?style=for-the-badge)](https://onnx.ai/)
[![Task](https://img.shields.io/badge/Task-Classification-green?style=for-the-badge)](https://github.com/TexasInstruments/edgeai)
[![Dataset](https://img.shields.io/badge/Dataset-ImageNet--1K-blueviolet?style=for-the-badge)](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>