

In [75]:

    
import numpy as np
import math

Neural Net Class

Weights is a three dimensional vector with the number of layers needed, and then rows and columns of the vector. Bias is also a three dimensional vector, but it only has one column. The values are then randomized.



In [249]:

    
class NeuralNet(object):
    def __init__(self, layers, rows, col):
        self.rows=rows
        self.col=col  
        self.weights=np.empty((layers, rows, col))
        self.bias=np.empty((layers, rows, 1))
        for i in range(layers):
            self.weights[i]=np.random.rand(rows, col)
            self.bias[i]=np.random.rand(rows, 1)
    def feedforward(self, train, layer, training_point):
        return (np.dot(self.weights[layer], train.input[training_point])+self.bias)[0]
    def sigmoid(self, train, layer, training_point):
        output=self.feedforward(train, layer, training_point)
        for i in range(train.dim):
            output[i]=1/(1+math.exp(-output[i]))
        return output
    def error(self, train, layer, training_point):
        return self.sigmoid(train, layer, training_point)-train.target[training_point]
    def grad(self, train, layer, training_point):
        for row in range(self.rows):
            err=self.error( train, layer, training_point)[row]
            sig=self.sigmoid( train, layer, training_point)[row]
            for col in range(self.col):
                self.weights[layer][row][col] -= .1*err*(sig)*(1-sig)*(train.input[layer][col])



In [247]:

    
class Trainer:
    def __init__(self, max_points, dim):
        self.dim=dim
        self.training_points=0
        self.input=np.empty((max_points, dim, 1))
        self.target=np.empty((max_points, dim, 1))
    def add_point(self, inp, tar):
        self.input[self.training_points]=inp
        self.target[self.training_points]=tar
        self.training_points+=1



In [251]:

    
y=Trainer(1,2)
y.add_point(np.array(([5], [1])), np.array(([.99], [.01])))
x=NeuralNet(1,2,2)
for i in range(10000):
    x.grad(y,0, 0)
print x.error(y, 0, 0)









    



[[ 0.00156434]
 [ 0.00031271]]



In [ ]:



In [ ]:



In [ ]: