Skip to content

Repository files navigation

MobileNetV3-v4: Tiny AI for ARM

A lightweight image classification model with 0.13 MB size, 98x faster than MobileNetV3, fully optimized for ARM architecture.

Features

  • ✅ Ultra-small size (0.13 MB)
  • ✅ High speed (2-5ms per inference)
  • ✅ Trained on CIFAR-100 (100 classes)
  • ✅ ARM-ready via TFLite
  • ✅ Custom architecture built from scratch

Results

Metric Value
Parameters 43,124
Model Size (TFLite) 0.13 MB
Accuracy (CIFAR-100) ~12%
Speed (ARM) 98x faster than MobileNetV3

How to Use

import tensorflow as tf

Load the TFLite model

interpreter = tf.lite.Interpreter(model_path='model_v4_trained.tflite') interpreter.allocate_tensors()

Get input/output details

input_details = interpreter.get_input_details() output_details = interpreter.get_output_details()

Run inference on an image

... (see test_model.py for full example)

Repository Contents

  • model_v4_trained.tflite - The optimized TFLite model
  • model_v4_trained.keras - The full Keras model
  • test_model.py - Testing and benchmarking script
  • requirements.txt - Python dependencies

Competition

This project is submitted to the ARM AI Optimization Challenge.

License

MIT License


Releases

Packages

Contributors

Languages