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

Source code for Fathony, Sahu, Willmott, & Kolter, "Multiplicative Filter Networks", ICLR 2021.

LTR_CrossEncoder: Legal Text Retrieval Zalo AI Challenge 2021

3rd Place Solution of the Traffic4Cast Core Challenge @ NeurIPS 2021

Constraint-based geometry sketcher for blender

Learning Dynamic Network Using a Reuse Gate Function in Semi-supervised Video Object Segmentation.

A pre-trained model with multi-exit transformer architecture.

[IJCAI-2021] A benchmark of data-free knowledge distillation from paper "Contrastive Model Inversion for Data-Free Knowledge Distillation"

Is RobustBench/AutoAttack a suitable Benchmark for Adversarial Robustness?

Prometheus Exporter for data scraped from datenplattform.darmstadt.de

🔥🔥High-Performance Face Recognition Library on PaddlePaddle & PyTorch🔥🔥

Neural network chess engine trained on Gary Kasparov's games.

Secure Distributed Training at Scale

Pixray is an image generation system

Official PyTorch implementation of "Evolving Search Space for Neural Architecture Search"

VM3000 Microphones

Pneumonia Detection using machine learning - with PyTorch

Fine-tuning StyleGAN2 for Cartoon Face Generation

Official implementation of Long-Short Transformer in PyTorch.

Code for the paper "Adapting Monolingual Models: Data can be Scarce when Language Similarity is High"

codes for paper Combining Dynamic Local Context Focus and Dependency Cluster Attention for Aspect-level sentiment classification