Converting the models to Tensorflow Lite

5 July, 2021

Introduction

The goal is to convert the following models to the tflite format that is more suitable for low computational environments:

• SSD MobileNet v2 320x320
• CenterNet MobileNetV2 FPN 512x512
• EfficientDet D0 512x512

As is said in the previous post the only two models that can be converted are SSD MobileNet(using standard Tensorflow Lite) and EfficientDet(using Tensorflow Lite Model Maker), but in the zip file of CenterNet MobileNet appeared a tflite file of the model so is added.

SSD MobileNet

I followed this tutorial.

Firstly, I needed to convert the model to an inference graph of tflite that can be done with this command in the directory research/object_detection in the repository:

python object_detection/export_tflite_graph_tf2.py --pipeline_config_path=object_detection/ssd_mobilenet/pipeline.config 
--trained_checkpoint_dir=object_detection/ssd_mobilenet/checkpoint --output_directory=object_detection/models_tflite/ssd_mobilenet

After that, we can pass to the last part to create the tflite model with all necessary metadata. For that, I wrote a tiny script of Python with the information provided by the tutorial, this script is located in the following directory research/object_detection/models_tflite/ssd_mobilenet.

import tensorflow as tf
from tflite_support.metadata_writers import object_detector
from tflite_support.metadata_writers import writer_utils

if __name__ == "__main__":
    _TFLITE_MODEL_PATH = "model_ssd.tflite"
    _TFLITE_LABEL_PATH = "../../data/mscoco_label_map.pbtxt"
    _SAVED_MODEL_PATH = "saved_model"

    converter = tf.lite.TFLiteConverter.from_saved_model(_SAVED_MODEL_PATH)
    converter.optimizations = [tf.lite.Optimize.DEFAULT]
    tflite_model = converter.convert()

    with open(_TFLITE_MODEL_PATH, 'wb') as f:
        f.write(tflite_model)

    writer = object_detector.MetadataWriter.create_for_inference(
        writer_utils.load_file(_TFLITE_MODEL_PATH), input_norm_mean=[0],
        input_norm_std=[255], label_file_paths=[_TFLITE_LABEL_PATH])
    writer_utils.save_file(writer.populate(), _TFLITE_MODEL_PATH)

The difference respect to the tutorial are that as that I use dynamic quantization I only need to pass to the optimizations the default optimization.

EfficientDet

Firstly, I though about creating a script and following this tutorial, but I wasn´t capable of configure Google Cloud so I finally decided using the Google Colab notebook provided by the same tutorial. I only needed to do a small change in the line with object_detector.create adding to the parameters do_train=False for skipping the train as the model is pretrained with the dataset that we are using and if I train the model will be modified to adapt to the dataset of the notebook.

Alejandro Fernández Camello