nn-wtf - Neural Networks Wrapper for TensorFlow

nn-wtf aims at providing a convenience wrapper to Google's TensorFlow machine learning library. Its focus is on making neural networks easy to set up, train and use.

The library is in pre-alpha right now and does not do anything seriously useful yet.

Installation

nn-wtf runs under Python3.4 and above.

Dependencies

Depending on whether you want GPU support or not, you need to do either

$ pip install tensorflow

$ pip install tensorflow-gpu

See https://www.tensorflow.org/get_started/os_setup for details.

The GPU version of TensorFlow needs the Nvidia CUDA libraries and the cuDNN Framework installed. See TensorFlow-GPU dependencies for instructions.

NN-WTF itself

Simple:

$ pip install nn_wtf

TensorFlow-GPU dependencies

The GPU version of TensorFlow needs the CUDA Toolkit 8.0 and cuDNN 5. There are instructions on how to do this at Install CUDA (GPUs on Linux)

A brief summary for Ubuntu Linux:

Download the CUDA 8 .deb from https://developer.nvidia.com/cuda-downloads and follow the installation instructions.
Download the cuDNN 5 archive from https://developer.nvidia.com/rdp/cudnn-download.
- You will have to register and set up an NVIDIA Developer account and apply for membership in the Accelerated Computing Developer Program (https://developer.nvidia.com/rdp/cuda-registered-developer-program) to download cuDNN.
Install cuDNN by extracting the downloaded archive and copying the extracted files to /usr/local/cuda/include and /usr/local/cuda/lib64.
Verify the installation by opening a Python interpreter and typing

>>> import tensorflow as tf

If no errors occur, your installation is successful.

Documentation

Sorry the documentation is absolutely minimal at this point. More useful documentation will be ready by the time this package reaches alpha status.

List of useful classes and methods

NeuralNetworkGraph: Base class for defining and training neural networks
- __init__(self, input_size, layer_sizes, output_size, learning_rate)
- set_session(self, session=None)
- train(self, data_sets, max_steps, precision, steps_between_checks, run_as_check, batch_size)
- get_predictor().predict(input_data)
MNISTGraph: Subclass of NeuralNetworkGraph suitable for working on MNIST data
NeuralNetworkOptimizer: Optimize geometry of a neural network for fast training
- __init__( self, tested_network, input_size, output_size, training_precision, layer_sizes, learning_rate, verbose, batch_size)
- brute_force_optimal_network_geometry(self, data_sets, max_steps)

Usage example

If you want to try it on MNIST data, try this sample program:

from nn_wtf.mnist_data_sets import MNISTDataSets
from nn_wtf.mnist_graph import MNISTGraph

import tensorflow as tf

data_sets = MNISTDataSets('.')
graph = MNISTGraph(
    learning_rate=0.1, layer_sizes=(64, 64, 16), train_dir='.'
)
graph.train(data_sets, max_steps=5000, precision=0.95)

# verify the training worked by testing if a handwritten number is recognized
image_data = MNISTDataSets.read_one_image_from_url(
    'http://github.com/lene/nn-wtf/blob/master/nn_wtf/data/7_from_test_set.raw?raw=true'
)
prediction = graph.get_predictor().predict(image_data)
assert prediction == 7

If you have cloned or downloaded the git repository at https://github.com/lene/nn-wtf, you can also have a look at main.py in the root directory for some example usage.

From there on, you are on your own for now. More functionality and documentation to come.