TexasInstruments-EdgeAI/ViT-Classification

ViT for TI EdgeAI

Pure Transformer for Image Classification at Scale

---

Overview

ViT (Vision Transformer) applies the standard Transformer architecture directly to sequences of non-overlapping image patches — no convolutions. Introduced in An Image is Worth 16x16 Words (Dosovitskiy et al., ICLR 2021), ViT demonstrates that a pure transformer pre-trained on large data transfers strongly to standard image recognition benchmarks.

Pretrained weights are sourced from torchvision (BSD-3-Clause), trained on ImageNet-1K using a DeiT-style recipe. Each exported ONNX model is a single-input classification graph that outputs 1000-class ImageNet logits [1, 1000].

---

Model Variants

Model	Architecture	Params	Top-1 Acc	Validated Devices	Config
vit_b_16	ViT-Base/16	86.6M	81.1%	TDA4VH	vit_b_16_config.yaml
vit_b_32	ViT-Base/32	88.2M	75.9%	TDA4VH	vit_b_32_config.yaml
vit_l_16	ViT-Large/16	304.3M	79.7%	TDA4VH	vit_l_16_config.yaml
vit_l_32	ViT-Large/32	306.5M	77.0%	TDA4VH	vit_l_32_config.yaml

Recommended for edge deployment: vit_b_16 (best accuracy/compute trade-off)

---

Quick Start

Prerequisites

pip install torch torchvision onnx>=1.22.0 onnxruntime>=1.23.2
# Optional but recommended for model optimization:
pip install onnx-simplifier

Export the Model

# Export the default model (vit_b_16)
python prepare_model.py

# Export a specific model variant
python prepare_model.py --model vit_b_32

# Export all supported models
python prepare_model.py --model all

# List all available variants
python prepare_model.py --list-models

The script automatically:

• Downloads pretrained ImageNet-1K weights from torchvision (first run only)
• Exports the model to ONNX (opset 17) with static input shape [1, 3, 224, 224]
• Runs ONNX shape inference
• Optionally simplifies the graph with onnx-simplifier

Compile and Infer uing edgeai-tidlrunner

Note: Run the commands below from inside the tidlrunner directory (the cloned edgeai-tidlrunner repository), with --config_path pointing to this model's config file.

Compile using edgeai-tidlrunner - on PC

cd /path/to/edgeai-tidlrunner
tidlrunner-cli compile --target_device J784S4 \
  --config_path /path/to/vit_b_16_config.yaml

Run Inference Benchmark - on device

cd /path/to/edgeai-tidlrunner
tidlrunner-cli infer --target_device J784S4 \
  --config_path /path/to/vit_b_16_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. For ONNX models, onnxruntime-tidl with TIDL acceleration can be used. Consult the documentation of edgeai-tidl-tools for more details.

---

Citation

@inproceedings{dosovitskiy2021image,
  title     = {An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale},
  author    = {Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and
               Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and
               Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and
               Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2021},
  url       = {https://arxiv.org/abs/2010.11929}
}

---

🔗 Resources

Resource	Link
Paper	arXiv:2010.11929
Source Code	pytorch/vision
Torchvision Docs	VisionTransformer
edgeai-tidl-tools	GitHub
edgeai-tidlrunner	GitHub
EdgeAI SDK	Documentation

---

Related Models

DINOv2 Self-supervised ViT Higher accuracy

DINO Self-supervised ViT Linear classification head

ResNet CNN baseline Lower compute

MobileNetV3 Lightweight CNN Built for edge

---

Maintained by: Texas Instruments EdgeAI Team
Last Updated: August 2026