tensorflow.keras save model in python and loading in Java

If you have a model saved in the SavedModel format (which it appears you do, and things like tf.contrib.saved_model.save_keras_model can help create), then in Java you can use SavedModelBundle.load to load and serve it. You do not need to "freeze" the model.

You may find the following useful:

• Example program using the SavedModel format for an object detection model in Java
• Slide deck about the Java API and the SavedModel format

But the basic idea is that your code will look something like:

try (SavedModelBundle model = SavedModelBundle.load("<directory>", "serve")) {
  try (Tensor<?> input = makeInputTensor();
       Tensor<?> output = model.session().runner().feed("INPUT_TENSOR", input).fetch("OUTPUT_TENSOR", output).run().get(0)) {
  // Use output
  }
}

Where "INPUT_TENSOR" and "OUTPUT_TENSOR" are the names of the input and output nodes in the TensorFlow graph. The saved_model_cli command-line tool installed when you install TensorFlow for Python can show you the names of these tensors in your model.

Note that using the TensorFlow Java API may be more suited to server/desktop applications than using TensorFlow Lite as suggested by another commenter. This is because the TensorFLow Lite runtime, while optimized (in terms of memory footprint etc.) for small devices, cannot export all models yet. While the TensorFlow Java API is using the exact same runtime and thus has the exact same abilities as TensorFlow for Python.

Hope that helps.