neuralNetwork.py

gistfile1.txt

# rewrite by aaaddress1@chroot.org
# refer: github.com/makeyourownneuralnetwork/makeyourownneuralnetwork/blob/master/part2_neural_network_mnist_data.ipynb
import matplotlib, numpy, os, pickle
def sigmoid(x):
    return 1 / (1 + numpy.exp(-x))


def saveModel():
    global wih # weight of (input -> hidden)  layer ... ( 100, 28^2 )
    global who # weight of (hidden -> output) layer ... (  10,  100 )   
    pickle.dump(wih, open('wih.bin', 'wb'))
    pickle.dump(who, open('who.bin', 'wb'))

def loadModel():
    global wih # weight of (input -> hidden)  layer ... ( 100, 28^2 )
    global who # weight of (hidden -> output) layer ... (  10,  100 )   
    wih = pickle.load(open('wih.bin', 'rb'))
    who = pickle.load(open('who.bin', 'rb'))
    
def testWritingPics():
    import matplotlib.pyplot as plt
    import matplotlib.image as mpimg

    with open('mnist_test.csv') as fp:
        testSet = fp.read().splitlines()

    print(f'[+] len(TestData) = {len(testSet)}')
    index = int(input('[!] choose a handwriting pics? '))
    inputWeight = numpy.asfarray( testSet[index].split(',')[1:] ) / 255. * 0.99 + 0.01
    
    plt.imshow(inputWeight.reshape((28, 28)))
    plt.show()
    
    ans = numpy.argmax(query(inputWeight))
    print(f'[+] neural network thinks it\'s {ans} ?')

def query(inputWeight):
    global wih # weight of (input -> hidden)  layer ... ( 100, 28^2 )
    global who # weight of (hidden -> output) layer ... (  10,  100 )   
    inputWeight = numpy.array(inputWeight, ndmin=2).T
    hiddenOut = sigmoid(numpy.dot(wih, inputWeight))
    return sigmoid(numpy.dot(who, hiddenOut))

def train(inputWeight, targetWeight):
    global learningrate
    global wih # weight of (input -> hidden)  layer ... ( 100, 28^2 )
    global who # weight of (hidden -> output) layer ... (  10,  100 )   

    # issue ref: github.com/makeyourownneuralnetwork/makeyourownneuralnetwork/issues/27
    inputWeight = numpy.array(inputWeight, ndmin=2).T   # transform into (28^2, 1)
    targetWeight = numpy.array(targetWeight, ndmin=2).T # transform into (10,   1) 
    
    # hidden layer input & output.
    hiddenOut = sigmoid(numpy.dot(wih, inputWeight))

    # output layer input & output.
    outlayerOut = sigmoid(numpy.dot(who, hiddenOut))

    # update weight for output layer.
    errOutPredict = targetWeight - outlayerOut
    who += learningrate * numpy.dot( 
        (errOutPredict * outlayerOut * (1.0 - outlayerOut)), 
        numpy.transpose(hiddenOut)
    )
    
    # update weight for hidden layer.
    errHiddenPredict = numpy.dot(who.T, errOutPredict) 
    wih += learningrate * numpy.dot(
        (errHiddenPredict * hiddenOut * (1.0 - hiddenOut)), 
        numpy.transpose(inputWeight)
    )
    
def main():
    global learningrate
    global trainSet
    global wih # weight of (input -> hidden)  layer ... ( 100, 28x28 )
    global who # weight of (hidden -> output) layer ... (  10,   100 )   

    # parse training data.  
    with open('mnist_train.csv') as fp:
        trainSet = fp.read().splitlines()
    assert len( trainSet[0].split(',')[1:] ) == 28**2, '[!] no train, no gain.'

    # init for neural network.
    learningrate = 0.15
    inputNodes = 28 **2
    hiddenNodes = 100
    outputNodes = 10
    wih = numpy.random.normal(0.0, pow(inputNodes, -0.5), (hiddenNodes, inputNodes))
    who = numpy.random.normal(0.0, pow(hiddenNodes, -0.5), (outputNodes, hiddenNodes))

    epochos = 1 # training times.
    for _ in range(epochos):
        for x, record in enumerate(trainSet):
            print(f'[+] ({x}/{ len(trainSet) }) : training handwriting -> { record[0] }')
            inputWeight = (numpy.asfarray( record.split(',')[1: ] ) / 255.0 * 0.99) + 0.01
            targetWeight = numpy.zeros(outputNodes) + 0.01
            targetWeight[int( record[0] )] = 0.99
            train( inputWeight, targetWeight )

    saveModel()
    print('[+] done, we cool?')
    

if __name__ == "__main__":
    main()

neuralNetwork.py

# rewrite by aaaddress1@chroot.org
# refer: github.com/makeyourownneuralnetwork/makeyourownneuralnetwork/blob/master/part2_neural_network_mnist_data.ipynb
import matplotlib, numpy, os, pickle
def sigmoid(x):
    return 1 / (1 + numpy.exp(-x))


def saveModel():
    global wih # weight of (input -> hidden)  layer ... ( 100, 28^2 )
    global who # weight of (hidden -> output) layer ... (  10,  100 )   
    pickle.dump(wih, open('wih.bin', 'wb'))
    pickle.dump(who, open('who.bin', 'wb'))

def loadModel():
    global wih # weight of (input -> hidden)  layer ... ( 100, 28^2 )
    global who # weight of (hidden -> output) layer ... (  10,  100 )   
    wih = pickle.load(open('wih.bin', 'rb'))
    who = pickle.load(open('who.bin', 'rb'))
    
def testWritingPics():
    import matplotlib.pyplot as plt
    import matplotlib.image as mpimg

    with open('mnist_test.csv') as fp:
        testSet = fp.read().splitlines()

    print(f'[+] len(TestData) = {len(testSet)}')
    index = int(input('[!] choose a handwriting pics? '))
    inputWeight = numpy.asfarray( testSet[index].split(',')[1:] ) / 255. * 0.99 + 0.01
    
    plt.imshow(inputWeight.reshape((28, 28)))
    plt.show()
    
    ans = numpy.argmax(query(inputWeight))
    print(f'[+] neural network thinks it\'s {ans} ?')

def query(inputWeight):
    global wih # weight of (input -> hidden)  layer ... ( 100, 28^2 )
    global who # weight of (hidden -> output) layer ... (  10,  100 )   
    inputWeight = numpy.array(inputWeight, ndmin=2).T
    hiddenOut = sigmoid(numpy.dot(wih, inputWeight))
    return sigmoid(numpy.dot(who, hiddenOut))

def train(inputWeight, targetWeight):
    global learningrate
    global wih # weight of (input -> hidden)  layer ... ( 100, 28^2 )
    global who # weight of (hidden -> output) layer ... (  10,  100 )   

    # issue ref: github.com/makeyourownneuralnetwork/makeyourownneuralnetwork/issues/27
    inputWeight = numpy.array(inputWeight, ndmin=2).T   # transform into (28^2, 1)
    targetWeight = numpy.array(targetWeight, ndmin=2).T # transform into (10,   1) 
    
    # hidden layer input & output.
    hiddenOut = sigmoid(numpy.dot(wih, inputWeight))

    # output layer input & output.
    outlayerOut = sigmoid(numpy.dot(who, hiddenOut))

    # update weight for output layer.
    errOutPredict = targetWeight - outlayerOut
    who += learningrate * numpy.dot( 
        (errOutPredict * outlayerOut * (1.0 - outlayerOut)), 
        numpy.transpose(hiddenOut)
    )
    
    # update weight for hidden layer.
    errHiddenPredict = numpy.dot(who.T, errOutPredict) 
    wih += learningrate * numpy.dot(
        (errHiddenPredict * hiddenOut * (1.0 - hiddenOut)), 
        numpy.transpose(inputWeight)
    )
    
def main():
    global learningrate
    global trainSet
    global wih # weight of (input -> hidden)  layer ... ( 100, 28x28 )
    global who # weight of (hidden -> output) layer ... (  10,   100 )   

    # parse training data.  
    with open('mnist_train.csv') as fp:
        trainSet = fp.read().splitlines()
    assert len( trainSet[0].split(',')[1:] ) == 28**2, '[!] no train, no gain.'

    # init for neural network.
    learningrate = 0.15
    inputNodes = 28 **2
    hiddenNodes = 100
    outputNodes = 10
    wih = numpy.random.normal(0.0, pow(inputNodes, -0.5), (hiddenNodes, inputNodes))
    who = numpy.random.normal(0.0, pow(hiddenNodes, -0.5), (outputNodes, hiddenNodes))

    epochos = 1 # training times.
    for _ in range(epochos):
        for x, record in enumerate(trainSet):
            print(f'[+] ({x}/{ len(trainSet) }) : training handwriting -> { record[0] }')
            inputWeight = (numpy.asfarray( record.split(',')[1: ] ) / 255.0 * 0.99) + 0.01
            targetWeight = numpy.zeros(outputNodes) + 0.01
            targetWeight[int( record[0] )] = 0.99
            train( inputWeight, targetWeight )

    saveModel()
    print('[+] done, we cool?')
    

if __name__ == "__main__":
    main()