1. Conx Neural Networks

1.1. Deep Learning for Simple Folk

Built in Python 3 on Keras 2.

CircleCI codecov Documentation Status PyPI version

Read the documentation at conx.readthedocs.io

Ask questions on the mailing list: conx-users

Implements Deep Learning neural network algorithms using a simple interface with easy visualizations and useful analytical. Built on top of Keras, which can use either TensorFlow, Theano, or CNTK.

The network is specified to the constructor by providing sizes. For example, Network("XOR", 2, 5, 1) specifies a network named "XOR" with a 2-node input layer, 5-unit hidden layer, and a 1-unit output layer.

1.2. Example

Computing XOR via a target function:

from conx import Network, SGD

dataset = [[[0, 0], [0]],
           [[0, 1], [1]],
           [[1, 0], [1]],
           [[1, 1], [0]]]

net = Network("XOR", 2, 5, 1, activation="sigmoid")
            optimizer=SGD(lr=0.3, momentum=0.9))
net.train(2000, report_rate=10, accuracy=1)

Creates dynamic, rendered visualizations like this:

1.3. Install

conx requires Python3, Keras version 2.0.8 or greater, and some other Python modules that are installed automatically with pip.

Note: you may need to use pip3, or admin privileges (eg, sudo), or a user environment.

pip install conx -U

You will need to decide whether to use Theano, TensorFlow, or CNTK. Pick one. See docs.microsoft.com for installing CNTK on Windows or Linux. All platforms can also install either of the others using pip:

pip install theano


pip install tensorflow

On MacOS, you may also need to render the SVG visualizations:

brew install cairo

1.3.1. Use with Jupyter Notebooks

To use the Network.dashboard() and camera functions, you will need to install and enable ipywidgets:

With pip:

pip install ipywidgets
jupyter nbextension enable --py widgetsnbextension

With conda

conda install -c conda-forge ipywidgets

Installing ipywidgets with conda will also enable the extension for you.

1.3.2. Changing Keras Backends

To use a Keras backend other than TensorFlow, edit (or create) ~/.keras/kerson.json, like:

    "backend": "theano",
    "image_data_format": "channels_last",
    "epsilon": 1e-07,
    "floatx": "float32"

1.4. Examples

See the notebooks folder and the documentation for additional examples.

1.5. Differences with Keras

  1. Conx does not allow targets to be a single value. Keras will automatically turn single values into a onehot encoded vectors. In conx, you should just convert such "labels" into their encodings before training.