Skip to content

Latest commit

 

History

History
254 lines (208 loc) · 9.96 KB

README.md

File metadata and controls

254 lines (208 loc) · 9.96 KB
Raw

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

HC²

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)

Using MicroMLP static functions :

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

Using MicroMLP speedly creation of a neural network :

from microMLP import MicroMLP
mlp = MicroMLP.Create([3, 10, 2], "Sigmoid", MicroMLP.LayersFullConnect)

Using MicroMLP 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

Using MicroMLP to learn the XOr problem (with hyperbolic tangent) :

from microMLP import MicroMLP

mlp = 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 )

if mlp.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
Graphe Activation function name Const Detail
HC² "Heaviside" MicroMLP.ACTFUNC_HEAVISIDE Heaviside binary step
HC² "Sigmoid" MicroMLP.ACTFUNC_SIGMOID Logistic (sigmoid or soft step)
HC² "TanH" MicroMLP.ACTFUNC_TANH Hyperbolic tangent
HC² "SoftPlus" MicroMLP.ACTFUNC_SOFTPLUS SoftPlus rectifier
HC² "ReLU" MicroMLP.ACTFUNC_RELU Rectified linear unit
HC² "Gaussian" MicroMLP.ACTFUNC_GAUSSIAN Gaussian function
Layers auto-connect function Detail
MicroMLP.LayersFullConnect Network fully connected

Using MicroMLP.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

Using MicroMLP.InputLayer(Layer) class :

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

Using MicroMLP.OutputLayer(Layer) class :

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

Using MicroMLP.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

Using MicroMLP.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

Using MicroMLP.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

Using MicroMLP.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()

Using MicroMLP.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 for HC² ;')

Keep it simple, stupid 👍