Based on this example:

https://github.com/tensorflow/hub/tree/master/tensorflow_hub/tools/make_image_classifier

To install dependencies:

pip install "tensorflow~=2.0"
pip install "tensorflow-hub[make_image_classifier]~=0.6"

To download training data:

curl http://download.tensorflow.org/example_images/flower_photos.tgz -O

To train the classifier:

make_image_classifier   --image_dir flower_photos   --tfhub_module https://tfhub.dev/google/tf2-preview/mobilenet_v2/feature_vector/4   --image_size 224   --saved_model_dir model   --labels_output_file class_labels.txt   --tflite_output_file new_mobile_model.tflite

Then the file https://github.com/amaurs/flowers/blob/master/backend/classifier.py was written inspired in https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/examples/python/label_image.py

With the classifier in place I built a simple backend with chalice. It creates an instance of the classifier and keeps it in memory for inference to start it:

cd backend
pip install -r requirements.txt
chalice local

It will start listening in http://127.0.0.1:8000.

Once the server is up and running you can run the client:

cd frontend
npm install
npm start

It will start running in http://localhost:3000/ and launch a broswer automatically. You can use one of the images to test.