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

这是一个mobilenet-yolov4-lite的库，把yolov4主干网络修改成了mobilenet，修改了Panet的卷积组成，使参数量大幅度缩小。

Pytorch implementation of the paper Improving Text-to-Image Synthesis Using Contrastive Learning

A list of Machine Learning Art Colabs

Semi-Autoregressive Transformer for Image Captioning

SubOmiEmbed: Self-supervised Representation Learning of Multi-omics Data for Cancer Type Classification

audioLIME: Listenable Explanations Using Source Separation

💛 Code and Dataset for our EMNLP 2021 paper: "Perspective-taking and Pragmatics for Generating Empathetic Responses Focused on Emotion Causes"

Code and data of the ACL 2021 paper: Few-Shot Text Ranking with Meta Adapted Synthetic Weak Supervision

Unofficial implementation of HiFi-GAN+ from the paper "Bandwidth Extension is All You Need" by Su, et al.

Square Root Bundle Adjustment for Large-Scale Reconstruction

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

Official Tensorflow implementation of "M-LSD: Towards Light-weight and Real-time Line Segment Detection"

Code for Blind Image Decomposition (BID) and Blind Image Decomposition network (BIDeN).

A scikit-learn-compatible module for estimating prediction intervals.

ParaGen is a PyTorch deep learning framework for parallel sequence generation

Bayesian optimisation library developped by Huawei Noah's Ark Library

Simple helper library to convert a collection of numpy data to tfrecord, and build a tensorflow dataset from the tfrecord.

noisy labels; missing labels; semi-supervised learning; entropy; uncertainty; robustness and generalisation.

Improving 3D Object Detection with Channel-wise Transformer

Official PyTorch implementation for FastDPM, a fast sampling algorithm for diffusion probabilistic models