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

Author's PyTorch implementation of TD3 for OpenAI gym tasks

Pytorch implementation of VAEs for heterogeneous likelihoods.

A large-scale video dataset for the training and evaluation of 3D human pose estimation models

Discriminative Region Suppression for Weakly-Supervised Semantic Segmentation

Database Reasoning Over Text project for ACL paper

Weight initialization schemes for PyTorch nn.Modules

[NeurIPS'21] Shape As Points: A Differentiable Poisson Solver

3D-Transformer: Molecular Representation with Transformer in 3D Space

ESPNet: Efficient Spatial Pyramid of Dilated Convolutions for Semantic Segmentation

A crash course in six episodes for software developers who want to become machine learning practitioners.

Facial expression detector

InferPy: Deep Probabilistic Modeling with Tensorflow Made Easy

Time should be taken seer-iously

Code for the AI lab course 2021/2022 of the University of Verona

Seq2seq - Sequence to Sequence Learning with Keras

Deep learning with dynamic computation graphs in TensorFlow

LUKE -- Language Understanding with Knowledge-based Embeddings

MOpt-AFL provided by the paper "MOPT: Optimized Mutation Scheduling for Fuzzers"

This is the official Pytorch implementation of "Lung Segmentation from Chest X-rays using Variational Data Imputation", Raghavendra Selvan et al. 2020

Framework web SnakeServer.