Hub Applications

Overview

This shows how to use TensorFlow Hub and Keras to build maximally reusable ML.

This repo builds and shares a pretrained DenseNet121 for feature extraction (forked from DenseNet code in keras-applications).

Use cases

There are two high-level ways to use a Hub Application:

graph+weights: no need to clone any code, no need to use Keras. Just use hub.load and call the module.
code+weights: original Keras Applications method, use this if you're using Keras by cloning the repo and importing the python.

Try it!

Play around with the code:

Run with TensorFlow Serving:

Usage: graph+weights

In this case we use TF-Hub and SavedModel directly - no need to clone any code or even be using Keras:

import tensorflow_hub as hub

module = hub.load('https://github.com/jharmsen/hub-application/releases/download/v1/densenet121_weights_tf_dim_ordering_tf_kernels_notop.tar.gz')

output = module(tf.random.normal(1, 32, 32, 3))

Pros

No model code needed
Can be used across TF ecosystem (e.g., Sonnet, other languages, etc...)
Can be easily used in Keras with hub.KerasLayer

Cons

Less flexibility without full model code

Usage: code+weights

In this case the Keras model code is cloned and produces a keras.Model whose weights are loaded from the SavedModel.

$ pip install git+https://github.com/jharmsen/hub-application.git

from hub_application import densenet
...
model = densenet.DenseNet121()

Pros

Full flexibility in modifying code
Produces a complete keras.Model

Cons

Only applicable if user is using python & Keras

Building a Hub Application

Export your keras.Model with tf.saved_model.save to produce the SavedModel .tar.gz (example notebook)
Upload SavedModel as a release binary file
Add functionality in your keras.Model constructor to
1. Download SavedModel with hub.resolve
2. Load weights using keras.Model.load_weights

See an example constructor here.

Testing

python -m pytest tests/