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

A Unified Generative Framework for Various NER Subtasks.

Differentiable Prompt Makes Pre-trained Language Models Better Few-shot Learners

The implementation of FOLD-R++ algorithm

Code for "FPS-Net: A convolutional fusion network for large-scale LiDAR point cloud segmentation".

Official Matlab Implementation for "Tiny Obstacle Discovery by Occlusion-aware Multilayer Regression", TIP 2020

HTSeq is a Python library to facilitate processing and analysis of data from high-throughput sequencing (HTS) experiments.

Implementation of 🦩 Flamingo, state-of-the-art few-shot visual question answering attention net out of Deepmind, in Pytorch

A Rao-Blackwellized Particle Filter for 6D Object Pose Tracking

P-Tuning v2: Prompt Tuning Can Be Comparable to Finetuning Universally Across Scales and Tasks

AdamW optimizer and cosine learning rate annealing with restarts

This repository contains the map content ontology used in narrative cartography

Multiband spectro-radiometric satellite image analysis with K-means cluster algorithm

Hidden-Fold Networks (HFN): Random Recurrent Residuals Using Sparse Supermasks

Secure Distributed Training at Scale

[ICCV2021] Safety-aware Motion Prediction with Unseen Vehicles for Autonomous Driving

Official implementation of the paper "Topographic VAEs learn Equivariant Capsules"

PyTorch implementation of ICLR 2022 paper PiCO: Contrastive Label Disambiguation for Partial Label Learning

Structural Constraints on Information Content in Human Brain States

Apply our monocular depth boosting to your own network!

Unofficial PyTorch implementation of Masked Autoencoders Are Scalable Vision Learners