TexasInstruments-EdgeAI/MobileNetV3-Classification

MobileNetV3 for TI EdgeAI

Efficient Mobile CNN for Image Classification

---

Overview

MobileNetV3 (Searching for MobileNetV3, Howard et al., 2019) combines hardware-aware Neural Architecture Search (NAS) with NetAdapt and a redesigned last stage to deliver state-of-the-art accuracy for mobile and edge inference. Key improvements over MobileNetV2 include hard-swish activations, squeeze-and-excitation modules in the bottleneck layers, and an optimized final classifier.

Both variants are evaluated at 224×224 input resolution on ImageNet-1K and distributed via torchvision.

---

Model Variants

Model	Architecture	Params	GFLOPs	Top-1 Acc	Top-5 Acc	Validated Devices	Config
mobilenetv3_large	MobileNetV3-Large	5.48M	0.22	75.274%	92.566%	TDA4VH	mobilenetv3_large_config.yaml
mobilenetv3_small	MobileNetV3-Small	2.54M	0.06	67.668%	87.402%	N/A	N/A

mobilenetv3_large uses IMAGENET1K_V2 weights (improved training recipe). mobilenetv3_small is excluded from prepare_model.py's export catalog because it produces poor accuracy under TIDL compilation — the mobilenetv3_small.onnx bundled in this folder is provided for reference only and has no validated TIDL config.

Recommended for edge deployment: mobilenetv3_large (best accuracy/compute trade-off with a validated TIDL config)

---

Quick Start

Prerequisites

pip install torch torchvision onnx>=1.14.0 onnxruntime>=1.16.0
# Optional but recommended for model optimization:
pip install onnx-simplifier

Export the Model

# Export the default model (MobileNetV3-Large)
python prepare_model.py

# Export a specific model variant
python prepare_model.py --model mobilenetv3_large

# Export with a custom input resolution
python prepare_model.py --model mobilenetv3_large --shape 224 224

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

The script automatically:

• Downloads pretrained ImageNet-1K weights from torchvision (MobileNet_V3_Large_Weights.IMAGENET1K_V2)
• Exports to ONNX (opset 17) with a static [1, 3, 224, 224] input shape
• Runs ONNX shape inference across all intermediate tensors
• Optionally simplifies the graph with onnxsim (use --no-simplify to skip)

Note: mobilenetv3_small is currently excluded from the export catalog (poor accuracy under TIDL compilation), so --model mobilenetv3_small and --model all only produce mobilenetv3_large.

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/mobilenetv3_large_config.yaml

Run Inference Benchmark - on device

cd /path/to/edgeai-tidlrunner
tidlrunner-cli infer --target_device J784S4 \
  --config_path /path/to/mobilenetv3_large_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

If you use these models, please cite:

@inproceedings{Howard2019MobileNetV3,
  title     = {Searching for MobileNetV3},
  author    = {Howard, Andrew and Sandler, Mark and Chu, Grace and Chen, Liang-Chieh
               and Chen, Bo and Tan, Mingxing and Wang, Weijun and Zhu, Yukun
               and Pang, Ruoming and Vasudevan, Vijay and Le, Quoc V. and Adam, Hartwig},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2019}
}

---

🔗 Resources

Resource	Link
Paper	arXiv:1905.02244
PyTorch Docs	torchvision MobileNetV3
Source Code	pytorch/vision
edgeai-tidl-tools	GitHub
edgeai-tidlrunner	GitHub
EdgeAI SDK	Documentation

---

Related Models

ResNet Deeper CNN Higher accuracy

ConvNeXt Modern CNN ViT-inspired design

ViT Vision Transformer Attention-based

DINOv2 Self-supervised Rich feature embeddings

---

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