Assignment_4.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Assignment 4\n",
    "### Abhi Ravikumar"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Here is a link to a sample Logistic Regression model for email spam/ham classification: http://jrmeyer.github.io/tutorial/2016/02/01/TensorFlow-Tutorial.html\n",
    "\n",
    "Follow the steps in the tutorial and write a Tensorflow program using your cloud ML instance.\n",
    "Questions:\n",
    "\n",
    "a)\n",
    "\n",
    "* What's the best accuracy you get on the test dataset?\n",
    "* What's the optimum value of the numEpochs and the learningRate.\n",
    "* Plot the error with Matplotlib\n",
    "\n",
    "b)\n",
    "\n",
    "* Add new neural network layers and training the model again. comapre the error function between these two models.\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/usr/lib/python3/dist-packages/matplotlib/__init__.py:1175: UserWarning:  This call to matplotlib.use() has no effect\n",
      "because the backend has already been chosen;\n",
      "matplotlib.use() must be called *before* pylab, matplotlib.pyplot,\n",
      "or matplotlib.backends is imported for the first time.\n",
      "\n",
      "  warnings.warn(_use_error_msg)\n"
     ]
    }
   ],
   "source": [
    "import matplotlib\n",
    "matplotlib.use('Agg')\n",
    "from __future__ import division\n",
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "import tarfile\n",
    "import os\n",
    "import matplotlib.pyplot as plt\n",
    "import time\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "loading training data\n",
      "loading test data\n"
     ]
    }
   ],
   "source": [
    "## Import Data\n",
    "\n",
    "def csv_to_numpy_array(filePath, delimiter):\n",
    "    return np.genfromtxt(filePath, delimiter=delimiter, dtype=None)\n",
    "\n",
    "def import_data():\n",
    "    if \"data\" not in os.listdir(os.getcwd()):\n",
    "        # Untar directory of data if we haven't already\n",
    "        tarObject = tarfile.open(\"data.tar.gz\")\n",
    "        tarObject.extractall()\n",
    "        tarObject.close()\n",
    "        print(\"Extracted tar to current directory\")\n",
    "    else:\n",
    "        # we've already extracted the files\n",
    "        pass\n",
    "\n",
    "    print(\"loading training data\")\n",
    "    trainX = csv_to_numpy_array(\"data/trainX.csv\", delimiter=\"\\t\")\n",
    "    trainY = csv_to_numpy_array(\"data/trainY.csv\", delimiter=\"\\t\")\n",
    "    print(\"loading test data\")\n",
    "    testX = csv_to_numpy_array(\"data/testX.csv\", delimiter=\"\\t\")\n",
    "    testY = csv_to_numpy_array(\"data/testY.csv\", delimiter=\"\\t\")\n",
    "    return trainX,trainY,testX,testY\n",
    "\n",
    "trainX,trainY,testX,testY = import_data()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# DATA SET PARAMETERS\n",
    "# Get our dimensions for our different variables and placeholders:\n",
    "# numFeatures = the number of words extracted from each email(great,dog,pill)\n",
    "numFeatures = trainX.shape[1]\n",
    "# numLabels = number of classes we are predicting (here just 2: Ham or Spam)\n",
    "numLabels = trainY.shape[1]\n",
    "\n",
    "# TRAINING SESSION PARAMETERS\n",
    "# number of times we iterate through training data\n",
    "# tensorboard shows that accuracy plateaus at ~25k epochs\n",
    "numEpochs = 1000\n",
    "# a smarter learning rate for gradientOptimizer\n",
    "learningRate = tf.train.exponential_decay(learning_rate=0.0008,\n",
    "                                          global_step= 1,\n",
    "                                          decay_steps=trainX.shape[0],\n",
    "                                          decay_rate= 0.95,\n",
    "                                          staircase=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# Define Hidden Layer\n",
    "hidden=4"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "### Placeholders\n",
    "\n",
    "# X = X-matrix / feature-matrix / data-matrix... It's a tensor to hold our email data. '\n",
    "# None' here means that we can hold any number of emails\n",
    "X = tf.placeholder(tf.float32, [None, numFeatures])\n",
    "# yGold = Y-matrix / label-matrix / labels... This will be our correct answers matrix. \n",
    "# Every row has either [1,0] for SPAM or [0,1] for HAM. 'None' here means that we can hold any number of emails\n",
    "yGold = tf.placeholder(tf.float32, [None, numLabels])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Values are randomly sampled from a Gaussian with a standard deviation of sqrt(6 / (numInputNodes + numOutputNodes + 1))\n",
    "\n",
    "hidden_weights = tf.Variable(tf.random_normal([numFeatures,hidden],\n",
    "                                       mean=0,\n",
    "                                       stddev=(np.sqrt(6/numFeatures+numLabels+1)),\n",
    "                                       name=\"hidden_weights\"))\n",
    "weights = tf.Variable(tf.random_normal([hidden,numLabels],\n",
    "                                       mean=0,\n",
    "                                       stddev=(np.sqrt(6/numFeatures+numLabels+1)),\n",
    "                                       name=\"weights\"))\n",
    "\n",
    "bias = tf.Variable(tf.random_normal([1,numLabels],\n",
    "                                    mean=0,\n",
    "                                    stddev=(np.sqrt(6/numFeatures+numLabels+1)),\n",
    "                                    name=\"bias\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "### Prediction Ops\n",
    "\n",
    "# INITIALIZE our weights and biases\n",
    "init_OP = tf.initialize_all_variables()\n",
    "\n",
    "# PREDICTION ALGORITHM i.e. FEEDFORWARD ALGORITHM\n",
    "apply_hidden_weights_OP = tf.matmul(X, hidden_weights, name=\"apply_hidden_weights\")\n",
    "apply_weights_OP = tf.matmul(apply_hidden_weights_OP, weights, name=\"apply_weights\")\n",
    "add_bias_OP = tf.add(apply_weights_OP, bias, name=\"add_bias\") \n",
    "activation_OP = tf.nn.sigmoid(add_bias_OP, name=\"activation\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# COST FUNCTION i.e. MEAN SQUARED ERROR\n",
    "cost_OP = tf.nn.l2_loss(activation_OP-yGold, name=\"squared_error_cost\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# OPTIMIZATION ALGORITHM i.e. GRADIENT DESCENT\n",
    "training_OP = tf.train.GradientDescentOptimizer(learningRate).minimize(cost_OP)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "### Visualization\n",
    "\n",
    "epoch_values=[]\n",
    "accuracy_values=[]\n",
    "cost_values=[]\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "step 0, training accuracy 0.4617\n",
      "step 0, cost 347.439\n",
      "step 0, change in cost 347.439\n",
      "step 10, training accuracy 0.463799\n",
      "step 10, cost 336.726\n",
      "step 10, change in cost 10.7136\n",
      "step 20, training accuracy 0.474292\n",
      "step 20, cost 323.819\n",
      "step 20, change in cost 12.9064\n",
      "step 30, training accuracy 0.474292\n",
      "step 30, cost 308.626\n",
      "step 30, change in cost 15.1938\n",
      "step 40, training accuracy 0.478489\n",
      "step 40, cost 292.422\n",
      "step 40, change in cost 16.2037\n",
      "step 50, training accuracy 0.50787\n",
      "step 50, cost 277.727\n",
      "step 50, change in cost 14.6953\n",
      "step 60, training accuracy 0.520462\n",
      "step 60, cost 266.572\n",
      "step 60, change in cost 11.1544\n",
      "step 70, training accuracy 0.533054\n",
      "step 70, cost 259.107\n",
      "step 70, change in cost 7.4649\n",
      "step 80, training accuracy 0.528856\n",
      "step 80, cost 254.257\n",
      "step 80, change in cost 4.85068\n",
      "step 90, training accuracy 0.521511\n",
      "step 90, cost 250.904\n",
      "step 90, change in cost 3.35303\n",
      "step 100, training accuracy 0.527807\n",
      "step 100, cost 248.324\n",
      "step 100, change in cost 2.58002\n",
      "step 110, training accuracy 0.524659\n",
      "step 110, cost 246.129\n",
      "step 110, change in cost 2.19507\n",
      "step 120, training accuracy 0.528856\n",
      "step 120, cost 244.127\n",
      "step 120, change in cost 2.00145\n",
      "step 130, training accuracy 0.533054\n",
      "step 130, cost 242.229\n",
      "step 130, change in cost 1.8985\n",
      "step 140, training accuracy 0.541448\n",
      "step 140, cost 240.39\n",
      "step 140, change in cost 1.8385\n",
      "step 150, training accuracy 0.546695\n",
      "step 150, cost 238.591\n",
      "step 150, change in cost 1.79904\n",
      "step 160, training accuracy 0.557188\n",
      "step 160, cost 236.821\n",
      "step 160, change in cost 1.77007\n",
      "step 170, training accuracy 0.564533\n",
      "step 170, cost 235.074\n",
      "step 170, change in cost 1.7468\n",
      "step 180, training accuracy 0.570829\n",
      "step 180, cost 233.347\n",
      "step 180, change in cost 1.72733\n",
      "step 190, training accuracy 0.582371\n",
      "step 190, cost 231.636\n",
      "step 190, change in cost 1.7104\n",
      "step 200, training accuracy 0.592865\n",
      "step 200, cost 229.94\n",
      "step 200, change in cost 1.69594\n",
      "step 210, training accuracy 0.603358\n",
      "step 210, cost 228.257\n",
      "step 210, change in cost 1.6835\n",
      "step 220, training accuracy 0.612802\n",
      "step 220, cost 226.584\n",
      "step 220, change in cost 1.67268\n",
      "step 230, training accuracy 0.620147\n",
      "step 230, cost 224.92\n",
      "step 230, change in cost 1.66385\n",
      "step 240, training accuracy 0.627492\n",
      "step 240, cost 223.264\n",
      "step 240, change in cost 1.65665\n",
      "step 250, training accuracy 0.63064\n",
      "step 250, cost 221.613\n",
      "step 250, change in cost 1.6507\n",
      "step 260, training accuracy 0.644281\n",
      "step 260, cost 219.967\n",
      "step 260, change in cost 1.64636\n",
      "step 270, training accuracy 0.652676\n",
      "step 270, cost 218.324\n",
      "step 270, change in cost 1.64325\n",
      "step 280, training accuracy 0.656873\n",
      "step 280, cost 216.682\n",
      "step 280, change in cost 1.64136\n",
      "step 290, training accuracy 0.663169\n",
      "step 290, cost 215.042\n",
      "step 290, change in cost 1.64053\n",
      "step 300, training accuracy 0.667366\n",
      "step 300, cost 213.401\n",
      "step 300, change in cost 1.64058\n",
      "step 310, training accuracy 0.671564\n",
      "step 310, cost 211.759\n",
      "step 310, change in cost 1.6418\n",
      "step 320, training accuracy 0.673662\n",
      "step 320, cost 210.116\n",
      "step 320, change in cost 1.64339\n",
      "step 330, training accuracy 0.684155\n",
      "step 330, cost 208.47\n",
      "step 330, change in cost 1.64569\n",
      "step 340, training accuracy 0.690451\n",
      "step 340, cost 206.821\n",
      "step 340, change in cost 1.64867\n",
      "step 350, training accuracy 0.699895\n",
      "step 350, cost 205.169\n",
      "step 350, change in cost 1.65207\n",
      "step 360, training accuracy 0.705142\n",
      "step 360, cost 203.514\n",
      "step 360, change in cost 1.65572\n",
      "step 370, training accuracy 0.710388\n",
      "step 370, cost 201.854\n",
      "step 370, change in cost 1.65959\n",
      "step 380, training accuracy 0.714586\n",
      "step 380, cost 200.19\n",
      "step 380, change in cost 1.6637\n",
      "step 390, training accuracy 0.721931\n",
      "step 390, cost 198.523\n",
      "step 390, change in cost 1.66777\n",
      "step 400, training accuracy 0.727177\n",
      "step 400, cost 196.851\n",
      "step 400, change in cost 1.67191\n",
      "step 410, training accuracy 0.729276\n",
      "step 410, cost 195.175\n",
      "step 410, change in cost 1.67577\n",
      "step 420, training accuracy 0.732424\n",
      "step 420, cost 193.495\n",
      "step 420, change in cost 1.67967\n",
      "step 430, training accuracy 0.737671\n",
      "step 430, cost 191.812\n",
      "step 430, change in cost 1.68317\n",
      "step 440, training accuracy 0.741868\n",
      "step 440, cost 190.126\n",
      "step 440, change in cost 1.68631\n",
      "step 450, training accuracy 0.748164\n",
      "step 450, cost 188.437\n",
      "step 450, change in cost 1.68922\n",
      "step 460, training accuracy 0.748164\n",
      "step 460, cost 186.745\n",
      "step 460, change in cost 1.69159\n",
      "step 470, training accuracy 0.751312\n",
      "step 470, cost 185.052\n",
      "step 470, change in cost 1.69347\n",
      "step 480, training accuracy 0.755509\n",
      "step 480, cost 183.357\n",
      "step 480, change in cost 1.69495\n",
      "step 490, training accuracy 0.759706\n",
      "step 490, cost 181.661\n",
      "step 490, change in cost 1.69571\n",
      "step 500, training accuracy 0.760755\n",
      "step 500, cost 179.965\n",
      "step 500, change in cost 1.69594\n",
      "step 510, training accuracy 0.762854\n",
      "step 510, cost 178.269\n",
      "step 510, change in cost 1.69559\n",
      "step 520, training accuracy 0.76915\n",
      "step 520, cost 176.575\n",
      "step 520, change in cost 1.69461\n",
      "step 530, training accuracy 0.773347\n",
      "step 530, cost 174.882\n",
      "step 530, change in cost 1.69289\n",
      "step 540, training accuracy 0.776495\n",
      "step 540, cost 173.191\n",
      "step 540, change in cost 1.69044\n",
      "step 550, training accuracy 0.781742\n",
      "step 550, cost 171.504\n",
      "step 550, change in cost 1.68745\n",
      "step 560, training accuracy 0.785939\n",
      "step 560, cost 169.82\n",
      "step 560, change in cost 1.68372\n",
      "step 570, training accuracy 0.790136\n",
      "step 570, cost 168.141\n",
      "step 570, change in cost 1.67924\n",
      "step 580, training accuracy 0.795383\n",
      "step 580, cost 166.467\n",
      "step 580, change in cost 1.67416\n",
      "step 590, training accuracy 0.798531\n",
      "step 590, cost 164.799\n",
      "step 590, change in cost 1.66824\n",
      "step 600, training accuracy 0.802728\n",
      "step 600, cost 163.137\n",
      "step 600, change in cost 1.6618\n",
      "step 610, training accuracy 0.809024\n",
      "step 610, cost 161.482\n",
      "step 610, change in cost 1.65451\n",
      "step 620, training accuracy 0.813221\n",
      "step 620, cost 159.836\n",
      "step 620, change in cost 1.64667\n",
      "step 630, training accuracy 0.81532\n",
      "step 630, cost 158.197\n",
      "step 630, change in cost 1.63823\n",
      "step 640, training accuracy 0.81532\n",
      "step 640, cost 156.568\n",
      "step 640, change in cost 1.62895\n",
      "step 650, training accuracy 0.816369\n",
      "step 650, cost 154.949\n",
      "step 650, change in cost 1.61919\n",
      "step 660, training accuracy 0.819517\n",
      "step 660, cost 153.34\n",
      "step 660, change in cost 1.60887\n",
      "step 670, training accuracy 0.822665\n",
      "step 670, cost 151.742\n",
      "step 670, change in cost 1.59792\n",
      "step 680, training accuracy 0.820567\n",
      "step 680, cost 150.156\n",
      "step 680, change in cost 1.58635\n",
      "step 690, training accuracy 0.825813\n",
      "step 690, cost 148.582\n",
      "step 690, change in cost 1.57437\n",
      "step 700, training accuracy 0.826863\n",
      "step 700, cost 147.02\n",
      "step 700, change in cost 1.56175\n",
      "step 710, training accuracy 0.83001\n",
      "step 710, cost 145.471\n",
      "step 710, change in cost 1.54883\n",
      "step 720, training accuracy 0.836306\n",
      "step 720, cost 143.936\n",
      "step 720, change in cost 1.53531\n",
      "step 730, training accuracy 0.841553\n",
      "step 730, cost 142.415\n",
      "step 730, change in cost 1.52135\n",
      "step 740, training accuracy 0.844701\n",
      "step 740, cost 140.907\n",
      "step 740, change in cost 1.50705\n",
      "step 750, training accuracy 0.8468\n",
      "step 750, cost 139.415\n",
      "step 750, change in cost 1.4924\n",
      "step 760, training accuracy 0.850997\n",
      "step 760, cost 137.938\n",
      "step 760, change in cost 1.47729\n",
      "step 770, training accuracy 0.850997\n",
      "step 770, cost 136.476\n",
      "step 770, change in cost 1.46204\n",
      "step 780, training accuracy 0.853095\n",
      "step 780, cost 135.029\n",
      "step 780, change in cost 1.44632\n",
      "step 790, training accuracy 0.853095\n",
      "step 790, cost 133.599\n",
      "step 790, change in cost 1.43037\n",
      "step 800, training accuracy 0.856243\n",
      "step 800, cost 132.185\n",
      "step 800, change in cost 1.41428\n",
      "step 810, training accuracy 0.858342\n",
      "step 810, cost 130.787\n",
      "step 810, change in cost 1.39792\n",
      "step 820, training accuracy 0.859391\n",
      "step 820, cost 129.405\n",
      "step 820, change in cost 1.38136\n",
      "step 830, training accuracy 0.86149\n",
      "step 830, cost 128.041\n",
      "step 830, change in cost 1.36461\n",
      "step 840, training accuracy 0.863589\n",
      "step 840, cost 126.693\n",
      "step 840, change in cost 1.34776\n",
      "step 850, training accuracy 0.863589\n",
      "step 850, cost 125.362\n",
      "step 850, change in cost 1.33081\n",
      "step 860, training accuracy 0.864638\n",
      "step 860, cost 124.049\n",
      "step 860, change in cost 1.31372\n",
      "step 870, training accuracy 0.866737\n",
      "step 870, cost 122.752\n",
      "step 870, change in cost 1.29664\n",
      "step 880, training accuracy 0.870934\n",
      "step 880, cost 121.473\n",
      "step 880, change in cost 1.27941\n",
      "step 890, training accuracy 0.874082\n",
      "step 890, cost 120.21\n",
      "step 890, change in cost 1.2622\n",
      "step 900, training accuracy 0.875131\n",
      "step 900, cost 118.965\n",
      "step 900, change in cost 1.24501\n",
      "step 910, training accuracy 0.87618\n",
      "step 910, cost 117.738\n",
      "step 910, change in cost 1.22773\n",
      "step 920, training accuracy 0.87723\n",
      "step 920, cost 116.527\n",
      "step 920, change in cost 1.21063\n",
      "step 930, training accuracy 0.87723\n",
      "step 930, cost 115.333\n",
      "step 930, change in cost 1.19348\n",
      "step 940, training accuracy 0.879328\n",
      "step 940, cost 114.157\n",
      "step 940, change in cost 1.1764\n",
      "step 950, training accuracy 0.881427\n",
      "step 950, cost 112.998\n",
      "step 950, change in cost 1.15937\n",
      "step 960, training accuracy 0.882476\n",
      "step 960, cost 111.855\n",
      "step 960, change in cost 1.14247\n",
      "step 970, training accuracy 0.884575\n",
      "step 970, cost 110.73\n",
      "step 970, change in cost 1.12569\n",
      "step 980, training accuracy 0.886674\n",
      "step 980, cost 109.621\n",
      "step 980, change in cost 1.10899\n",
      "step 990, training accuracy 0.886674\n",
      "step 990, cost 108.528\n",
      "step 990, change in cost 1.09241\n",
      "final accuracy on test set: 0.933333\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "[<matplotlib.lines.Line2D at 0x7fa7c6c18a90>]"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAhMAAAFvCAYAAADwq0kDAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nRERE5APo/wHK1OXHMbcP+gAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fa7c6b9b080>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Turn on interactive plotting\n",
    "plt.ion()\n",
    "# Create the main, super plot\n",
    "fig = plt.figure()\n",
    "# Create two subplots on their own axes and give titles\n",
    "ax1 = plt.subplot(\"211\")\n",
    "ax1.set_title(\"TRAINING ACCURACY\", fontsize=10)\n",
    "ax2 = plt.subplot(\"212\")\n",
    "ax2.set_title(\"TRAINING COST\", fontsize=10)\n",
    "# plt.tight_layout()\n",
    "\n",
    "# Create a tensorflow session\n",
    "sess = tf.Session()\n",
    "\n",
    "# Initialize all tensorflow variables\n",
    "sess.run(init_OP)\n",
    "\n",
    "## Ops for vizualization\n",
    "# argmax(activation_OP, 1) gives the label our model thought was most likely\n",
    "# argmax(yGold, 1) is the correct label\n",
    "correct_predictions_OP = tf.equal(tf.argmax(activation_OP,1),tf.argmax(yGold,1))\n",
    "# False is 0 and True is 1, what was our average?\n",
    "accuracy_OP = tf.reduce_mean(tf.cast(correct_predictions_OP, \"float\"))\n",
    "# Summary op for regression output\n",
    "activation_summary_OP = tf.histogram_summary(\"output\", activation_OP)\n",
    "# Summary op for accuracy\n",
    "accuracy_summary_OP = tf.scalar_summary(\"accuracy\", accuracy_OP)\n",
    "# Summary op for cost\n",
    "cost_summary_OP = tf.scalar_summary(\"cost\", cost_OP)\n",
    "# Summary ops to check how variables (W, b) are updating after each iteration\n",
    "weightSummary = tf.histogram_summary(\"weights\", weights.eval(session=sess))\n",
    "biasSummary = tf.histogram_summary(\"biases\", bias.eval(session=sess))\n",
    "# Merge all summaries\n",
    "#####all_summary_OPS = tf.merge_all_summaries()\n",
    "# Summary writer\n",
    "#writer = tf.train.SummaryWriter(\"summary_logs\", sess.graph_def)\n",
    "\n",
    "# Initialize reporting variables\n",
    "cost = 0\n",
    "diff = 1\n",
    "\n",
    "# Training epochs\n",
    "for i in range(numEpochs):\n",
    "    if i > 1 and diff < .0001:\n",
    "        print(\"change in cost %g; convergence.\"%diff)\n",
    "        break\n",
    "    else:\n",
    "        # Run training step\n",
    "        step = sess.run(training_OP, feed_dict={X: trainX, yGold: trainY})\n",
    "        # Report occasional stats\n",
    "        if i % 10 == 0:\n",
    "            # Add epoch to epoch_values\n",
    "            epoch_values.append(i)\n",
    "            # Generate accuracy stats on test data\n",
    "            #summary_results = sess.run(all_summary_OPS, feed_dict={X: trainX, yGold: trainY})\n",
    "            train_accuracy = sess.run(accuracy_OP, feed_dict={X: trainX, yGold: trainY})\n",
    "            newCost = sess.run(cost_OP, feed_dict={X: trainX, yGold: trainY})\n",
    "            # Add accuracy to live graphing variable\n",
    "            accuracy_values.append(train_accuracy)\n",
    "            # Add cost to live graphing variable\n",
    "            cost_values.append(newCost)\n",
    "            # Write summary stats to writer\n",
    "            #writer.add_summary(summary_results, i)\n",
    "            # Re-assign values for variables\n",
    "            diff = abs(newCost - cost)\n",
    "            cost = newCost\n",
    "            \n",
    "            #generate print statements\n",
    "            print(\"step %d, training accuracy %g\"%(i, train_accuracy))\n",
    "            print(\"step %d, cost %g\"%(i, newCost))\n",
    "            print(\"step %d, change in cost %g\"%(i, diff))\n",
    "\n",
    "            # Plot progress to our two subplots\n",
    "            accuracyLine, = ax1.plot(epoch_values, accuracy_values)\n",
    "            costLine, = ax2.plot(epoch_values, cost_values)\n",
    "            fig.canvas.draw()\n",
    "            time.sleep(1)\n",
    "\n",
    "\n",
    "# How well do we perform on held-out test data?\n",
    "print(\"final accuracy on test set: %s\" %str(sess.run(accuracy_OP, \n",
    "                                                     feed_dict={X: testX, \n",
    "                                                                yGold: testY})))\n",
    "plt.plot(epoch_values,accuracy_values)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "Optimum value of the numEpochs=1000 and the learningRate=0.8. Final accuracy on test set=0.93333"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.4.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}