Guide on tflite quantization and converting

Repository contains examples on how to do quantization aware training, and converting to tflite freezed graphdef, keras model, or checkpoint file.

Requirements

tensorflow 1.4 or higher python 3

pip3 install -r requirements.txt

Running quantization aware training

Script contains training/exporting/testing of the dummy model for the image-to-image transformation

python3 quantization_aware_training.py

for the export add --mode export, and for the testing --mode test

Converting to tflite trained model

there is a 3 different options for obtaining tflite model - from graphdef, from the keras model and from the checkpoint. Also you can choose to do post training quantization for reducing the model size, which will convert weights to uint8.

python3 convert_tf_model.py --model_type "type of the trained model" --model_path "path to the trained model"

Useful resources

Official documentation of the Tensorflow Lite Blog post with explanation of examples