NN Implementation

NN.json

{"cells": [{"metadata": {"trusted": true, "collapsed": false}, "cell_type": "code", "source": "import numpy as np\n\n#Utility Functions\ndef loadDataSet():\n    #Last column Target\n    dataset = np.array([[1,1,1,1],\n                        [1,1,0,1],\n                        [1,0,1,1],\n                        [1,0,0,0]])\n    return dataset\n\ndef sigmoid(z,derivate=False):\n    if derivate:\n        #Derivate Sigmoid Function\n        return z * (1-z)\n    #Simoidal Function\n    return 1.0 / (1.0 + np.e**(-z))\n\ndef classify(Test_X):\n    a_1 = sigmoid(np.dot(Test_X,weights_1.T))\n    a_1 = np.hstack((np.ones((1,1)),a_1))\n    a_2 = sigmoid(np.dot(a_1,weights_2.T))\n    print(\"Prob Of Pos : %f\"%a_2)\n    if a_2 >= 0.5:\n        print(\"1\")\n    else:\n        print(\"0\")\n\n#Arquitecture\n'''\n    3 nodes input layer - include bias\n    3 nodes hidden layer - include bias\n    1 node output layer\n'''\n        \n#Calc - HardCode for the arquitecture above\ndataset = loadDataSet()\nnumber_observations = dataset.shape[0]\nweights_1 = np.random.random((2,3))\nweights_2 = np.random.random((1,3))\n\nmaxIter = 100000\nactivateDebug = False\nfor iterNumber in range(maxIter):\n    data = np.random.permutation(dataset)\n    X = data[:,:3]\n    y = data[:,-1:]\n\n    #FeedForward\n    a_1 = sigmoid(np.dot(X,weights_1.T))\n    a_1 = np.hstack((np.ones((number_observations,1)),a_1))\n    a_2 = sigmoid(np.dot(a_1,weights_2.T))\n   \n    #BackPropagation\n    delta_2 = (y - a_2) * sigmoid(a_2,True)\n    delta_1 = delta_2.dot(weights_2) * sigmoid(a_1,True)\n    \n    if iterNumber % 100 == 0 and activateDebug:\n        print(\"Error : %f - %d\"%(np.sum(delta_2),iterNumber))\n   \n    #DELTA1:\n    '''\n        1)Product delta_1 * activation - in this case X o a_0\n        2)Sum by columns to get acumulate\n        3)Throw away first column\n        4)Make a reshape\n        5)Tile the vector, to be equal dimension of weights\n    '''\n    DELTA_1 = np.tile((np.sum(delta_1 * X,axis=0)[1:]).reshape(2,1),3)\n    \n    #DELTA2:\n    DELTA_2 = np.sum(delta_2 * a_1,axis=0).reshape(1,3)\n    \n    weights_1 += DELTA_1\n    weights_2 += DELTA_2\n\n#Classify\nclassify(np.array([[1,1,1]]))\nclassify(np.array([[1,1,0]]))\nclassify(np.array([[1,0,1]]))\nclassify(np.array([[1,0,0]]))\n", "outputs": [], "execution_count": null}, {"metadata": {"trusted": true, "collapsed": true}, "cell_type": "code", "source": "", "outputs": [], "execution_count": null}], "nbformat": 4, "metadata": {"language_info": {"file_extension": ".py", "name": "python", "nbconvert_exporter": "python", "mimetype": "text/x-python", "version": "3.4.3", "pygments_lexer": "ipython3", "codemirror_mode": {"version": 3, "name": "ipython"}}, "kernelspec": {"language": "python", "display_name": "Python 3", "name": "python3"}}, "nbformat_minor": 0}