Bridge between Pharo and Keras

KerasBridge gives Pharo developers the capability of using Keras neural network python library directly from Pharo.

Getting Started

First, you need to install Python, we recommend you to look at Python bridge. In particular, you need Python3.6 (Python3.7 is not compatible with Tensorflow yet), Pipenv and a Pharo6.1 or newer image.

Second, you need to run in your image:

Metacello new
    baseline: 'KerasBridge';
    repository: 'github://ObjectProfile/KerasBridge/src';
    load.

If this fails refer to the troubleshooting section of the PythonBridge documentation. Though remember that the pipenv environment of the KerasBridge is different from the base PythonBridge, therfore, all commands should be performed on Keras class and KerasBridge repository.

First neural network

The following code snippet creates a neural network that learns to compute the XOR operation.

Keras do: [ 
    x := #(#(0 0) #(0 1) #(1 0) #(1 1)).
    y := #((0) (1) (1) (0)).

    model := KSequentialModel inputs: 2.
    model addLayer: (KDenseLayer neurons: 8 activation: KTanh).
    model addLayer: (KDenseLayer neurons: 1).
    model addLayer: (KActivationLayer activation: KSigmoid).
    model 
        compileLoss: KBinaryCrossentropy 
        optimizer: KAdam default 
        metrics: { KBinaryAccuracy }.
    (model fit: x labels: y epochs: 700) waitForValue ]

The first line Keras do: [ ... ] initializes the Python process, executes the code in the block and then ensures the destruction of the python process to prevent having zombies. This could be replaced by using start and stop methods.

Keras start.
...
Keras stop.

The x and y variables corresponds to the training data and the labels.

model := KSequentialModel inputs: 2. creates a Keras Sequential model.

model addLayer: (KDenseLayer neurons: 8 activation: KTanh). creates a new Dense layer with 8 neurons and a Tanh activation function, and adds it to the model.

model addLayer: (KActivationLayer activation: KSigmoid). creates a new Activation layer with a Sigmoid activation function. Then it adds the layer to the model.

model compileLoss: KBinaryCrossentropy optimizer: KAdam default metrics: { KBinaryAccuracy }. compiles the model using BinaryCrossentropy loss function, the Adam optimizer and the BinaryAccuracy metric.

model fit: x labels: y epochs: 700 train the model using x data with y labels for 700 epochs. This method returns a promise expecting the training history, if you want to wait for the result of the computation send the promise the message waitForValue.