Keras_to_TFLite_for_Android

How to convert a trained Keras neural network to TensorFlow Lite, that can be applied in an Android application

I will summarize this procedure here, as it was a bit time intensive for me to pull together its parts from across the internet.

Starting point: Keras trained model .hdf5 file

The starting point of this conversion is a Keras for Tensorflow - trained neural network model file. A trained Keras model file usually is of format .hdf5 or .h5 in-short.

Step 1: Convert model.hdf5 to protobuf Tensorflow format .pb

Use the attached tool keras_to_tensorflow.py:

--> python keras_to_tensorflow.py -input_model model.h5 –output_model outputmodel_path.pb

Now we have the Keras-trained model in a frozen Tensorflow format .pb.

Step 2: Converting TF protobuf model.pb to TFLite model.tflite

I could not get it to run under Windows. Therefore run it on Linux.

My setup:

• Ubuntu 18.04
• python 3.6
  --> sudo apt-get install python3.6 --> sudo apt-get install python3-pip python3-dev
• Tensorflow 1.6 (this version includes toco) --> pip3 install tensorflow==1.60

Conversion using Toco

For this we assume that the input to our model was a 299x299 pixel RGB image. (Dimensions 299x299x3)

Below command should be one not-interrupted command!

/home/moritz/.local/bin/toco --input_file=Desktop/model.pb --output_file=Desktop/model.tflite --input_format=TENSORFLOW_GRAPHDEF --output_format=TFLITE --input_shape=1,299,299,3
--input_array=input_1 --output_array=output_node
--inference_type=FLOAT

Note how the input_shape has to be defined according to the expected input of your model, as well as the name of the input layer. (input_1 in my case) Also the name of the output layer has to be defined correctly. (output_node in my case)

Finding the output_array name

This can pose a difficulty, as when using standard layers in Keras, the layers names are not obvious. Therefore i have added a line in the code of keras_to_tensorflow.py that will print out the name of the output layer:

Model in TFLite format

If everything went well, you now should have converted the Keras model.hdf5 file to a TFLite model.tflite

Please note: NOT every TF layer is available in TFLite. Please refrain from using complex layers such as GlobalMaxPooling2D().

Supported layers of TFLite can be found here:

https://www.tensorflow.org/lite/guide/ops_compatibility