Movatterモバイル変換

jczic/MicroMLPPublic

NotificationsYou must be signed in to change notification settings
Fork29
Star194

A micro neural network multilayer perceptron for MicroPython (used on ESP32 and Pycom modules)

micromlp.hc2.fr

License

MIT license

194 stars 29 forks Branches Tags Activity

Star

Notifications

You must be signed in to change notification settings

Branches Tags

Folders and files

Name		Name	Last commit message	Last commit date
Latest commit History 53 Commits
CNAME		CNAME
LICENSE.md		LICENSE.md
README.md		README.md
_config.yml		_config.yml
graphe_gaussian.png		graphe_gaussian.png
graphe_heaviside.png		graphe_heaviside.png
graphe_relu.png		graphe_relu.png
graphe_sigmoid.png		graphe_sigmoid.png
graphe_softplus.png		graphe_softplus.png
graphe_tanh.png		graphe_tanh.png
hc2-deep-learning.png		hc2-deep-learning.png
hc2.png		hc2.png
microMLP.py		microMLP.py
test points.py		test points.py
xor.py		xor.py

Repository files navigation

MicroMLP is a micro artificial neural network multilayer perceptron (principally used on ESP32 and Pycom modules)

Very easy to integrate and very light with one file only :

"microMLP.py"

MicroMLP features :

Modifiable multilayer and connections structure
Integrated bias on neurons
Plasticity of the connections included
Activation functions by layer
Parameters Alpha, Eta and Gain
Managing set of examples and learning
QLearning functions to use reinforcement learning
Save and load all structure to/from json file
Various activation functions :
- Heaviside binary step
- Logistic (sigmoid or soft step)
- Hyperbolic tangent
- SoftPlus rectifier
- ReLU (rectified linear unit)
- Gaussian function

Use deep learning for :

Signal processing (speech processing, identification, filtering)
Image processing (compression, recognition, patterns)
Control (diagnosis, quality control, robotics)
Optimization (planning, traffic regulation, finance)
Simulation (black box simulation)
Classification (DNA analysis)
Approximation (unknown function, complex function)

UsingMicroMLP static functions :

Name	Function
Create	`mlp = MicroMLP.Create(neuronsByLayers, activationFuncName, layersAutoConnectFunction=None, useBiasValue=1.0)`
LoadFromFile	`mlp = MicroMLP.LoadFromFile(filename)`

UsingMicroMLP speedly creation of a neural network :

frommicroMLPimportMicroMLPmlp=MicroMLP.Create([3,10,2],"Sigmoid",MicroMLP.LayersFullConnect)

UsingMicroMLP main class :

Name	Function
Constructor	`mlp = MicroMLP()`
GetLayer	`layer = mlp.GetLayer(layerIndex)`
GetLayerIndex	`idx = mlp.GetLayerIndex(layer)`
RemoveLayer	`mlp.RemoveLayer(layer)`
GetInputLayer	`inputLayer = mlp.GetInputLayer()`
GetOutputLayer	`outputLayer = mlp.GetOutputLayer()`
Learn	`ok = mlp.Learn(inputVectorNNValues, targetVectorNNValues)`
Test	`ok = mlp.Test(inputVectorNNValues, targetVectorNNValues)`
Predict	`outputVectorNNValues = mlp.Predict(inputVectorNNValues)`
QLearningLearnForChosenAction	`ok = mlp.QLearningLearnForChosenAction(stateVectorNNValues, rewardNNValue, pastStateVectorNNValues, chosenActionIndex, terminalState=True, discountFactorNNValue=None)`
QLearningPredictBestActionIndex	`bestActionIndex = mlp.QLearningPredictBestActionIndex(stateVectorNNValues)`
SaveToFile	`ok = mlp.SaveToFile(filename)`
AddExample	`ok = mlp.AddExample(inputVectorNNValues, targetVectorNNValues)`
ClearExamples	`mlp.ClearExamples()`
LearnExamples	`learnCount = mlp.LearnExamples(maxSeconds=30, maxCount=None, stopWhenLearned=True, printMAEAverage=True)`

Property	Example	Read/Write
Layers	`mlp.Layers`	get
LayersCount	`mlp.LayersCount`	get
IsNetworkComplete	`mlp.IsNetworkComplete`	get
MSE	`mlp.MSE`	get
MAE	`mlp.MAE`	get
MSEPercent	`mlp.MSEPercent`	get
MAEPercent	`mlp.MAEPercent`	get
ExamplesCount	`mlp.ExamplesCount`	get

UsingMicroMLP to learn the XOr problem (with hyperbolic tangent) :

frommicroMLPimportMicroMLPmlp=MicroMLP.Create(neuronsByLayers= [2,2,1],activationFuncName=MicroMLP.ACTFUNC_TANH,layersAutoConnectFunction=MicroMLP.LayersFullConnect )nnFalse=MicroMLP.NNValue.FromBool(False)nnTrue=MicroMLP.NNValue.FromBool(True)mlp.AddExample( [nnFalse,nnFalse], [nnFalse] )mlp.AddExample( [nnFalse,nnTrue ], [nnTrue ] )mlp.AddExample( [nnTrue ,nnTrue ], [nnFalse] )mlp.AddExample( [nnTrue ,nnFalse], [nnTrue ] )learnCount=mlp.LearnExamples()print("LEARNED :" )print("  - False xor False = %s"%mlp.Predict([nnFalse,nnFalse])[0].AsBool )print("  - False xor True  = %s"%mlp.Predict([nnFalse,nnTrue] )[0].AsBool )print("  - True  xor True  = %s"%mlp.Predict([nnTrue ,nnTrue] )[0].AsBool )print("  - True  xor False = %s"%mlp.Predict([nnTrue ,nnFalse])[0].AsBool )ifmlp.SaveToFile("mlp.json") :print("MicroMLP structure saved!" )

Variable	Description	Default
`mlp.Eta`	Weighting of the error correction	0.30
`mlp.Alpha`	Strength of connections plasticity	0.75
`mlp.Gain`	Network learning gain	0.99
`mlp.CorrectLearnedMAE`	Threshold of self-learning error	0.02

Activation function name	Const	Detail
`"Heaviside"`	MicroMLP.ACTFUNC_HEAVISIDE	Heaviside binary step
`"Sigmoid"`	MicroMLP.ACTFUNC_SIGMOID	Logistic (sigmoid or soft step)
`"TanH"`	MicroMLP.ACTFUNC_TANH	Hyperbolic tangent
`"SoftPlus"`	MicroMLP.ACTFUNC_SOFTPLUS	SoftPlus rectifier
`"ReLU"`	MicroMLP.ACTFUNC_RELU	Rectified linear unit
`"Gaussian"`	MicroMLP.ACTFUNC_GAUSSIAN	Gaussian function

Layers auto-connect function	Detail
`MicroMLP.LayersFullConnect`	Network fully connected

UsingMicroMLP.Layer class :

Name	Function
Constructor	`layer = MicroMLP.Layer(parentMicroMLP, activationFuncName=None, neuronsCount=0)`
GetLayerIndex	`idx = layer.GetLayerIndex()`
GetNeuron	`neuron = layer.GetNeuron(neuronIndex)`
GetNeuronIndex	`idx = layer.GetNeuronIndex(neuron)`
AddNeuron	`layer.AddNeuron(neuron)`
RemoveNeuron	`layer.RemoveNeuron(neuron)`
GetMeanSquareError	`mse = layer.GetMeanSquareError()`
GetMeanAbsoluteError	`mae = layer.GetMeanAbsoluteError()`
GetMeanSquareErrorAsPercent	`mseP = layer.GetMeanSquareErrorAsPercent()`
GetMeanAbsoluteErrorAsPercent	`maeP = layer.GetMeanAbsoluteErrorAsPercent()`
Remove	`layer.Remove()`

Property	Example	Read/Write
ParentMicroMLP	`layer.ParentMicroMLP`	get
ActivationFuncName	`layer.ActivationFuncName`	get
Neurons	`layer.Neurons`	get
NeuronsCount	`layer.NeuronsCount`	get

UsingMicroMLP.InputLayer(Layer) class :

Name	Function
Constructor	`inputLayer = MicroMLP.InputLayer(parentMicroMLP, neuronsCount=0)`
SetInputVectorNNValues	`ok = inputLayer.SetInputVectorNNValues(inputVectorNNValues)`

UsingMicroMLP.OutputLayer(Layer) class :

Name	Function
Constructor	`outputLayer = MicroMLP.OutputLayer(parentMicroMLP, activationFuncName, neuronsCount=0)`
GetOutputVectorNNValues	`outputVectorNNValues = outputLayer.GetOutputVectorNNValues()`
ComputeTargetLayerError	`ok = outputLayer.ComputeTargetLayerError(targetVectorNNValues)`

UsingMicroMLP.Neuron class :

Name	Function
Constructor	`neuron = MicroMLP.Neuron(parentLayer)`
GetNeuronIndex	`idx = neuron.GetNeuronIndex()`
GetInputConnections	`connections = neuron.GetInputConnections()`
GetOutputConnections	`connections = neuron.GetOutputConnections()`
AddInputConnection	`neuron.AddInputConnection(connection)`
AddOutputConnection	`neuron.AddOutputConnection(connection)`
RemoveInputConnection	`neuron.RemoveInputConnection(connection)`
RemoveOutputConnection	`neuron.RemoveOutputConnection(connection)`
SetBias	`neuron.SetBias(bias)`
GetBias	`neuron.GetBias()`
SetOutputNNValue	`neuron.SetOutputNNValue(nnvalue)`
ComputeValue	`neuron.ComputeValue()`
ComputeError	`neuron.ComputeError(targetNNValue=None)`
Remove	`neuron.Remove()`

Property	Example	Read/Write
ParentLayer	`neuron.ParentLayer`	get
ComputedOutput	`neuron.ComputedOutput`	get
ComputedDeltaError	`neuron.ComputedDeltaError`	get
ComputedSignalError	`neuron.ComputedSignalError`	get

UsingMicroMLP.Connection class :

Name	Function
Constructor	`connection = MicroMLP.Connection(neuronSrc, neuronDst, weight=None)`
UpdateWeight	`connection.UpdateWeight(eta, alpha)`
Remove	`connection.Remove()`

Property	Example	Read/Write
NeuronSrc	`connection.NeuronSrc`	get
NeuronDst	`connection.NeuronDst`	get
Weight	`connection.Weight`	get

UsingMicroMLP.Bias class :

Name	Function
Constructor	`bias = MicroMLP.Bias(neuronDst, value=1.0, weight=None)`
UpdateWeight	`bias.UpdateWeight(eta, alpha)`
Remove	`bias.Remove()`

Property	Example	Read/Write
NeuronDst	`bias.NeuronDst`	get
Value	`bias.Value`	get
Weight	`bias.Weight`	get

UsingMicroMLP.NNValue static functions :

Name	Function
FromPercent	`nnvalue = MicroMLP.NNValue.FromPercent(value)`
NewPercent	`nnvalue = MicroMLP.NNValue.NewPercent()`
FromByte	`nnvalue = MicroMLP.NNValue.FromByte(value)`
NewByte	`nnvalue = MicroMLP.NNValue.NewByte()`
FromBool	`nnvalue = MicroMLP.NNValue.FromBool(value)`
NewBool	`nnvalue = MicroMLP.NNValue.NewBool()`
FromAnalogSignal	`nnvalue = MicroMLP.NNValue.FromAnalogSignal(value)`
NewAnalogSignal	`nnvalue = MicroMLP.NNValue.NewAnalogSignal()`

UsingMicroMLP.NNValue class :

Name	Function
Constructor	`nnvalue = MicroMLP.NNValue(minValue, maxValue, value)`

Property	Example	Read/Write
AsFloat	`nnvalue.AsFloat = 639.513`	get / set
AsInt	`nnvalue.AsInt = 12345`	get / set
AsPercent	`nnvalue.AsPercent = 65`	get / set
AsByte	`nnvalue.AsByte = b'\x75'`	get / set
AsBool	`nnvalue.AsBool = True`	get / set
AsAnalogSignal	`nnvalue.AsAnalogSignal = 0.39472`	get / set

By JC`zic forHC² ;')

Keep it simple, stupid 👍

About

A micro neural network multilayer perceptron for MicroPython (used on ESP32 and Pycom modules)

micromlp.hc2.fr

Releases

No releases published

Packages

No packages published

Contributors2

Languages

Python100.0%

Movatterモバイル変換

License

jczic/MicroMLP

Folders and files

Latest commit

History

Repository files navigation

MicroMLP is a micro artificial neural network multilayer perceptron (principally used on ESP32 andPycom modules)

Very easy to integrate and very light with one file only :

MicroMLP features :

Use deep learning for :

UsingMicroMLP static functions :

UsingMicroMLP speedly creation of a neural network :

UsingMicroMLP main class :

UsingMicroMLP to learn the XOr problem (with hyperbolic tangent) :

UsingMicroMLP.Layer class :

UsingMicroMLP.InputLayer(Layer) class :

UsingMicroMLP.OutputLayer(Layer) class :

UsingMicroMLP.Neuron class :

UsingMicroMLP.Connection class :

UsingMicroMLP.Bias class :

UsingMicroMLP.NNValue static functions :

UsingMicroMLP.NNValue class :

By JC`zic forHC² ;')

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages0

Uh oh!

Contributors2

Uh oh!

Languages

MicroMLP is a micro artificial neural network multilayer perceptron (principally used on ESP32 and Pycom modules)

Packages