TFlite Optimization

Deep learning model optimization demo using tensorflow lite

Setup

Make sure you have Python 3.10+ installed. Install Poetry: pip install poetry Clone this repository and navigate to the project directory. Run poetry install to install the dependencies.

Alternatively, run pip install -r requirements.txt

Usage

Current configurations are for (movenet)[https://www.kaggle.com/models/google/movenet/], a pose landmark estimation model. Some changes might be required for other models.

Model quantization

model_optimization/quantize_model.py: quantie a tensorflow model in from SavedModel format to tflite format. Dynamic range quantization (float to int8) is used, covered by tf.lite.Optimize.DEFAULT. This optimization is suitable for cpu/edge devices and does not require any representative dataset.

Example:

 python3 model_optimization/quantize_model.py --input_path ./models/movenet-tensorflow2-multipose-lightning-v1/ --output_path ./models/optimized_movenet.tflite

Model inference

model_optimization/infer.py: use a tflite model to infer pose landmarks from an input image. The script displays a visualization and saves it as output.jpg.

Example:

python3 model_optimization/infer.py --model_path ./optimized_movenet.tflite --input_image /home/srg/Downloads/jumping_jack.jpg