Reseau neuronal basique fonctionnel

sobek/network.py

@@ -1,5 +1,4 @@
 import numpy as np
-import math
 
 class network:
 
@@ -7,71 +6,71 @@ class network:
         if type(inputLayerSize) != int:
             raise TypeError("The input layer size must be an int!")
 
-        self.__weights = []
+        self.weights = []
         self.__inputLayerSize = inputLayerSize
         oldLayerSize = inputLayerSize
         for layerSize in layerSizes:
-            self.__weights.append( np.random.random((layerSize, oldLayerSize)) )
+            self.weights.append( np.random.randn(layerSize, oldLayerSize) )
             oldLayerSize = layerSize
-        self.__biases = [[0]*layerSize for layerSize in layerSizes]
+        self.biases = [np.random.randn(layerSize) for layerSize in layerSizes]
-        self.__weights = np.array(self.__weights, dtype=object)
-        self.__biases = np.array(self.__biases, dtype=object)
 
     def __reLu(value, derivative=False):
         if (derivative):
-            return 0 if (value == 0) else 1
+            return 0 if (value < 0) else 1
         return max(0, value)
 
     def __sigmoid(value, derivative=False):
         if (derivative):
             return network.__sigmoid(value) * (1 - network.__sigmoid(value))
-        return 1/(1+math.exp(-value))
+        return 1.0/(1.0+np.exp(-value))
 
     def process(self, _input, __storeValues=False):
         if type(_input) != np.ndarray:
             raise TypeError("The input must be a vector!")
         if _input.size != self.__inputLayerSize:
             raise ValueError("The input vector has the wrong size!")
-        #if _input.dtype != np.float64:
+        if _input.dtype != np.float64:
-        #    raise TypeError("The input vector must contain floats!")
+            print(_input.dtype)
+            raise TypeError("The input vector must contain floats!")
 
         if (__storeValues):
             self.activations = []
             self.outputs = []
+            self.outputs.append(_input)
         
-        for layerWeights, bias in zip(self.__weights, self.__biases):
+        for layerWeights, layerBias in zip(self.weights, self.biases):
             
-            _input = np.matmul(layerWeights, _input)
+            _input = np.dot(layerWeights, _input)
-            _input = np.add(_input, bias)
+            _input = np.add(_input, layerBias)
 
             if (__storeValues): 
-                self.activations.append(_input.copy())
+                self.activations.append(_input)
 
             #activation function application
-            for neuron in range(len(_input)):
+            #for i in range(len(_input)):
-                _input[neuron] = network.__sigmoid(_input[neuron])
+            #    _input[i] = network.__sigmoid(_input)
+            _input = network.__sigmoid(_input)
 
             #On peut comparer la performance si on recalcul plus tard
             if (__storeValues): 
-                self.outputs.append(_input.copy())
+                self.outputs.append(_input)
-
-        self.activations = np.array(self.activations, dtype=object)
-        self.outputs = np.array(self.outputs, dtype=object)
 
         return _input
 
 
 
     def train(self, inputs, desiredOutputs, learningRate):
-        errorSumsWeights = [[[0]*len(neuron) for neuron in layer] for layer in self.__weights]
+        if (len(inputs) != len(desiredOutputs)):
-        errorSumsBiases = [[0]*len(layer) for layer in self.__biases]
+            raise ValueError("The inputs and desired outputs vectors must have the same amount of data !")
-        self.__errors = [[0]*len(layer) for layer in self.__weights]
 
         for _input, desiredOutput in zip(inputs, desiredOutputs):
 
-            #rempli self.activations et self.outputs
+            errorSumsWeights = [np.zeros(layer.shape) for layer in self.weights]
-            self.__output = self.process(_input, True)
+            errorSumsBiases = [np.zeros(layer.shape) for layer in self.biases]
+            self.__errors = [np.zeros(len(layer)) for layer in self.weights]
 
+            #rempli self.activations et self.outputs
+            self.process(_input, True)
             self.__desiredOutput = desiredOutput
 
             #Somme de matrice ?
@@ -84,17 +83,23 @@ class network:
 
         total = 0
         
+        
         errorSumsWeights = np.multiply(errorSumsWeights, -(learningRate/len(inputs)))
-        self.__weights = np.add(self.__weights, errorSumsWeights)
+        self.weights = np.add(self.weights, errorSumsWeights)
 
         errorSumsBiases = np.multiply(errorSumsBiases, -(learningRate/len(inputs)))
-        self.__biases = np.add(self.__biases, errorSumsBiases)
+        self.biases = np.add(self.biases, errorSumsBiases)
 
-        print(self.__biases)
+        #print(self.__biases)
         
         """
         for layerNumber in range(len(errorSumsWeights)):
                 for neuronNumber in range(len(errorSumsWeights[layerNumber])):
+
+                    errorSumsBiases[layerNumber][neuronNumber] = errorSumsBiases[layerNumber][neuronNumber] / len(inputs)
+                    total += errorSumsBiases[layerNumber][neuronNumber]
+                    self.biases[layerNumber][neuronNumber] -= learningRate * errorSumsBiases[layerNumber][neuronNumber]
+                    
                     for weightNumber in range(len(errorSumsWeights[layerNumber][neuronNumber])):
 
                         #Probablement faisable avec une multiplication de matrices
@@ -103,24 +108,24 @@ class network:
                         total += errorSumsWeights[layerNumber][neuronNumber][weightNumber]
 
                         #Probablement faisable avec une somme de matrices
-                        self.__weights[layerNumber][neuronNumber][weightNumber] -= learningRate * errorSumsWeights[layerNumber][neuronNumber][weightNumber]
+                        self.weights[layerNumber][neuronNumber][weightNumber] -= learningRate * errorSumsWeights[layerNumber][neuronNumber][weightNumber]
 
-        print("Error : " + str(total))"""
+        #print("Error : " + str(total))"""
 
     def __Error(self, layer, neuron):
         if (self.__errors[layer][neuron] == 0 ):
-            self.__errors[layer][neuron] = self.__ErrorFinalLayer(neuron) if (layer == len(self.__weights)-1) else self.__ErrorHiddenLayer(layer, neuron)
+            self.__errors[layer][neuron] = self.__ErrorFinalLayer(neuron) if (layer == len(self.weights)-1) else self.__ErrorHiddenLayer(layer, neuron)
         return self.__errors[layer][neuron]
         
     def __ErrorFinalLayer(self, neuron):
-        return network.__sigmoid(self.activations[len(self.activations)-1][neuron], True) * (self.__output[neuron] - self.__desiredOutput[neuron])
+        return network.__sigmoid(self.activations[-1][neuron], derivative=True) * (self.outputs[-1][neuron] - self.__desiredOutput[neuron])
 
     def __ErrorHiddenLayer(self, layer, neuron):
         upperLayerLinksSum = 0
         #Probablement faisable avec une multiplication de matrices
-        for upperLayerNeuron in range(len(self.__weights[layer+1])):
+        for upperLayerNeuron in range(len(self.weights[layer+1])):
-            upperLayerLinksSum += self.__weights[layer+1][upperLayerNeuron][neuron] * self.__errors[layer+1][upperLayerNeuron]
+            upperLayerLinksSum += self.weights[layer+1][upperLayerNeuron][neuron] * self.__errors[layer+1][upperLayerNeuron]
-        return network.__sigmoid(self.activations[layer][neuron], True) * upperLayerLinksSum
+        return network.__sigmoid(self.activations[layer][neuron], derivative=True) * upperLayerLinksSum
 
     def __PartialDerivative(self, layer, neuron, weight):
-        return self.__Error(layer, neuron) * self.outputs[layer-1][weight]
+        return self.__Error(layer, neuron) * self.outputs[layer][weight]

testLearning.py

@@ -4,42 +4,36 @@ from sobek.network import network
 
 random.seed()
 
-myNetwork = network(10, 10, 10)
+myNetwork = network(10, 10)
 
-learningRate = 1
+learningRate = 3
 
-for j in range(100):
+for j in range(10000):
+    rand = []
     inputs = []
-    inputs2 = []
     desiredOutputs = []
     
     if (j%50 == 0):
         print(j)
 
-    for i in range(1000):
+    for i in range(10):
-        inputs.append([(random.randrange(10)/10)])
+        rand.append( random.randrange(10)/10)
-    inputs = np.array(inputs, dtype=object)
 
-    for i in range(1000):
+    for i in range(10):
-        desiredOutputs.append([0]*10)
+        desiredOutputs.append(np.zeros(10))
-        desiredOutputs[i][9 - int(inputs[i][0]*10)] = 1.0
+        desiredOutputs[i][9 - int(rand[i]*10)] = 1.0
-    desiredOutputs = np.array(desiredOutputs, dtype=object)
 
-    #for i in range(1000):
+    for i in range(10):
-    #    inputs2.append([0]*10)
+        inputs.append(np.zeros(10))
-    #    inputs2[i][int(inputs[i][0]*10)] = 1.0
+        inputs[i][int(rand[i]*10)] = 1.0
-    inputs2 = np.array(inputs2, dtype=object)
     
-    if (j%10000 == 0):
+    myNetwork.train(inputs, desiredOutputs, learningRate)
-        learningRate*= 0.1
-        
-    myNetwork.train(desiredOutputs, desiredOutputs, learningRate)
 
 test = []
-test.append([0]*10)
+test.append(np.zeros(10))
-test.append([0]*10)
+test.append(np.zeros(10))
 test[0][1] = 1.0
-test[1][8] = 1.0
+test[1][5] = 1.0
-test = np.array(test, dtype=object)
+print(test[0])
 print(myNetwork.process(test[0]))
 print(myNetwork.process(test[1]))

testLearningNAND.py

@@ -0,0 +1,63 @@
+import numpy as np
+import random
+from sobek.network import network
+
+random.seed()
+
+myNetwork = network(2, 1)
+
+learningRate = 3
+
+test = []
+result = []
+test.append(np.zeros(2))
+test.append(np.zeros(2))
+test.append(np.zeros(2))
+test.append(np.zeros(2))
+test[1][1] = 1.0
+test[2][0] = 1.0
+test[3][0] = 1.0
+test[3][1] = 1.0
+result.append(np.ones(1))
+result.append(np.ones(1))
+result.append(np.ones(1))
+result.append(np.zeros(1))
+
+for j in range(10000):
+    inputs = []
+    desiredOutputs = []
+    
+    if (j%1000 == 0):
+        print(j)
+
+    random.shuffle(test)
+
+    for i in range(4):
+        if (test[i][0] == 1.0) and (test[i][1] == 1.0):
+            result[i][0] = 0.0
+        else:
+            result[i][0] = 1.0
+    
+    myNetwork.train(test, result, learningRate)
+
+test = []
+result = []
+test.append(np.zeros(2))
+test.append(np.zeros(2))
+test.append(np.zeros(2))
+test.append(np.zeros(2))
+test[1][1] = 1.0
+test[2][0] = 1.0
+test[3][0] = 1.0
+test[3][1] = 1.0
+result.append(np.ones(1))
+result.append(np.ones(1))
+result.append(np.ones(1))
+result.append(np.zeros(1))
+
+print(myNetwork.weights)
+print(myNetwork.biases)
+print("0 0 : " + str(myNetwork.process(test[0])) + " == 1 ?")
+print("0 1 : " + str(myNetwork.process(test[1])) + " == 1 ?")
+print("1 0 : " + str(myNetwork.process(test[2])) + " == 1 ?")
+print("1 1 : " + str(myNetwork.process(test[3])) + " == 0 ?")

testNAND.py

@@ -0,0 +1,25 @@
+import numpy as np
+import random
+from sobek.network import network
+
+myNetwork = network(2, 1)
+
+test = []
+test.append(np.zeros(2))
+test.append(np.zeros(2))
+test.append(np.zeros(2))
+test.append(np.zeros(2))
+test[1][1] = 1.0
+test[2][0] = 1.0
+test[3][0] = 1.0
+test[3][1] = 1.0
+
+myNetwork.weights = [np.array([[-10.0, -10.0]])]
+myNetwork.biases = [np.array([15.0])]
+print(myNetwork.weights)
+print(myNetwork.biases)
+
+print("0 0 : " + str(myNetwork.process(test[0])) + " == 1 ?")
+print("0 1 : " + str(myNetwork.process(test[1])) + " == 1 ?")
+print("1 0 : " + str(myNetwork.process(test[2])) + " == 1 ?")
+print("1 1 : " + str(myNetwork.process(test[3])) + " == 0 ?")