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
)
52 Dec 04, 2022
1.8k Dec 28, 2022
13 Mar 05, 2022
67 Dec 20, 2022
100 Dec 25, 2022
1 Jan 14, 2022
6 Nov 17, 2022
83 Dec 26, 2022
60 Dec 31, 2022
28 Dec 09, 2022
2.3k Jan 05, 2023
7 Aug 07, 2022
1 Nov 12, 2021
293 Dec 27, 2022
3.9k Dec 23, 2022
123 Dec 28, 2022
0 Nov 02, 2021
1k Jan 01, 2023
19 Dec 13, 2022
76 Sep 30, 2022