A plug-and-play library for neural networks written in Python

Last update: Jul 16, 2022

Overview

Synapses

A plug-and-play library for neural networks written in Python!

# run
pip install synapses-py==7.4.1
# in the directory of your project

Neural Network

Create a neural network

Import Synapses, call NeuralNetwork.init and provide the size of each layer.

from synapses_py import NeuralNetwork, ActivationFunction, DataPreprocessor, Statistics
layers = [4, 6, 5, 3]
neuralNetwork = NeuralNetwork.init(layers)

neuralNetwork has 4 layers. The first layer has 4 input nodes and the last layer has 3 output nodes. There are 2 hidden layers with 6 and 5 neurons respectively.

Get a prediction

inputValues = [1.0, 0.5625, 0.511111, 0.47619]
prediction = \
        NeuralNetwork.prediction(neuralNetwork, inputValues)

prediction should be something like [ 0.8296, 0.6996, 0.4541 ].

Note that the lengths of inputValues and prediction equal to the sizes of input and output layers respectively.

Fit network

learningRate = 0.5
expectedOutput = [0.0, 1.0, 0.0]
fitNetwork = \
        NeuralNetwork.fit(
            neuralNetwork,
            learningRate,
            inputValues,
            expectedOutput
        )

fitNetwork is a new neural network trained with a single observation.

To train a neural network, you should fit with multiple datapoints

Create a customized neural network

The activation function of the neurons created with NeuralNetwork.init, is a sigmoid one. If you want to customize the activation functions and the weight distribution, call NeuralNetwork.customizedInit.

def activationF(layerIndex):
    if layerIndex == 0:
        return ActivationFunction.sigmoid
    elif layerIndex == 1:
        return ActivationFunction.identity
    elif layerIndex == 2:
        return ActivationFunction.leakyReLU
    else:
        return ActivationFunction.tanh

def weightInitF(_layerIndex):
    return 1.0 - 2.0 * random()

customizedNetwork = \
        NeuralNetwork.customizedInit(
            layers,
            activationF,
            weightInitF
        )

Visualization

Call NeuralNetwork.toSvg to take a brief look at its svg drawing.

The color of each neuron depends on its activation function while the transparency of the synapses depends on their weight.

svg = NeuralNetwork.toSvg(customizedNetwork)

Save and load a neural network

JSON instances are compatible across platforms! We can generate, train and save a neural network in Python and then load and make predictions in Javascript!

toJson

Call NeuralNetwork.toJson on a neural network and get a string representation of it. Use it as you like. Save json in the file system or insert into a database table.

json = NeuralNetwork.toJson(customizedNetwork)

ofJson

loadedNetwork = NeuralNetwork.ofJson(json)

As the name suggests, NeuralNetwork.ofJson turns a json string into a neural network.

Encoding and decoding

One hot encoding is a process that turns discrete attributes into a list of 0.0 and 1.0. Minmax normalization scales continuous attributes into values between 0.0 and 1.0. You can use DataPreprocessor for datapoint encoding and decoding.

The first parameter of DataPreprocessor.init is a list of tuples (attributeName, discreteOrNot).

setosaDatapoint = {
    "petal_length": "1.5",
    "petal_width": "0.1",
    "sepal_length": "4.9",
    "sepal_width": "3.1",
    "species": "setosa"
}

versicolorDatapoint = {
    "petal_length": "3.8",
    "petal_width": "1.1",
    "sepal_length": "5.5",
    "sepal_width": "2.4",
    "species": "versicolor"
}

virginicaDatapoint = {
    "petal_length": "6.0",
    "petal_width": "2.2",
    "sepal_length": "5.0",
    "sepal_width": "1.5",
    "species": "virginica"
}

datasetList = [ setosaDatapoint,
                versicolorDatapoint,
                virginicaDatapoint ]

dataPreprocessor = \
        DataPreprocessor.init(
             [ ("petal_length", False),
               ("petal_width", False),
               ("sepal_length", False),
               ("sepal_width", False),
               ("species", True) ],
             iter(datasetList)
        )

encodedDatapoints = map(lambda x:
        DataPreprocessor.encodedDatapoint(dataPreprocessor, x),
        datasetList
)

encodedDatapoints equals to:

[ [ 0.0     , 0.0     , 0.0     , 1.0     , 0.0, 0.0, 1.0 ],
  [ 0.511111, 0.476190, 1.0     , 0.562500, 0.0, 1.0, 0.0 ],
  [ 1.0     , 1.0     , 0.166667, 0.0     , 1.0, 0.0, 0.0 ] ]

Save and load the preprocessor by calling DataPreprocessor.toJson and DataPreprocessor.ofJson.

Evaluation

To evaluate a neural network, you can call Statistics.rootMeanSquareError and provide the expected and predicted values.

expectedWithOutputValuesList = \
        [ ( [ 0.0, 0.0, 1.0], [ 0.0, 0.0, 1.0] ),
          ( [ 0.0, 0.0, 1.0], [ 0.0, 1.0, 1.0] ) ]

expectedWithOutputValuesIter = \
        iter(expectedWithOutputValuesList)

rmse = Statistics.rootMeanSquareError(
                        expectedWithOutputValuesIter
)

A plug-and-play library for neural networks written in Python

Related tags

Overview

Synapses

Neural Network

Create a neural network

Get a prediction

Fit network

Create a customized neural network

Visualization

Save and load a neural network

toJson

ofJson

Encoding and decoding

Evaluation

Owner

Dimos Michailidis

Code artifacts for the submission "Mind the Gap! A Study on the Transferability of Virtual vs Physical-world Testing of Autonomous Driving Systems"

DL course co-developed by YSDA, HSE and Skoltech

PerfFuzz: Automatically Generate Pathological Inputs for C/C++ programs

Python TFLite scripts for detecting objects of any class in an image without knowing their label.

PyTorch Implementation of the SuRP algorithm by the authors of the AISTATS 2022 paper "An Information-Theoretic Justification for Model Pruning"

Astrostatistics class for the MSc degree in Astrophysics at the University of Milan-Bicocca (Italy)

This repository provides a basic implementation of our GCPR 2021 paper "Learning Conditional Invariance through Cycle Consistency"

Lucid Sonic Dreams syncs GAN-generated visuals to music.

Fastshap: A fast, approximate shap kernel

Self-Supervised Learning for Domain Adaptation on Point-Clouds

A PyTorch re-implementation of Neural Radiance Fields

SpeechBrain is an open-source and all-in-one speech toolkit based on PyTorch.

Audio Visual Emotion Recognition using TDA

Unified Pre-training for Self-Supervised Learning and Supervised Learning for ASR

The pyrelational package offers a flexible workflow to enable active learning with as little change to the models and datasets as possible

Deep Illuminator is a data augmentation tool designed for image relighting. It can be used to easily and efficiently generate a wide range of illumination variants of a single image.

Low Complexity Channel estimation with Neural Network Solutions

Hummingbird compiles trained ML models into tensor computation for faster inference.

implementation of the paper "MarginGAN: Adversarial Training in Semi-Supervised Learning"

Code for this paper The Lottery Ticket Hypothesis for Pre-trained BERT Networks.