mnist.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Agora nosso objetivo é automatizar o reconhecimento de caracteres numéricos. <br><br> Para isso, usaremos uma técnica chamada Deep Learning (aprendizado profundo).<br><br>Atualmente, a biblioteca mais popular de Deep Learning é o Tensorflow, desenvolvido pelo Google\n",
    "Obs: Entender o funcionamento de uma rede neural é pré-requisito para entender esse algoritmo."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Começamos importando as bibliotecas necessárias:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\lbour\\Anaconda3\\lib\\site-packages\\h5py\\__init__.py:34: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n",
      "  from ._conv import register_converters as _register_converters\n"
     ]
    }
   ],
   "source": [
    "import numpy as np\n",
    "import tensorflow as tf\n",
    "import matplotlib.pyplot as plt\n",
    "from tensorflow.examples.tutorials.mnist import input_data    \n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Usaremos o famoso dataset MNIST que contém imagens digitalizadas de números manuscritos que serão usadas para treinar nosso modelo. <br>Um 'label' está associada  a cada imagem do dataset. Usaremos esse label ou etiqueta no formato 'one_hot' que poderia ser interpretada como<br> uma lista contendo a informação sobre qual número aquela \n",
    "imagem corresponde. <br>A lista contém 10 elementos, sendo 9 'zeros' e 1 'um'. O índice correspondente ao elemento 1\n",
    "é o dígito que está contido na imagem.<br>\n",
    "Exemplos:<br><center>\n",
    "    0   ==>   [1, 0, 0, 0, 0, 0, 0, 0, 0, 0]<br>\n",
    "    3   ==>   [0, 0, 0, 1, 0, 0, 0, 0, 0, 0]<br>\n",
    "    9   ==>   [0, 0, 0, 0, 0, 0, 0, 0, 0, 1]<br>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Extracting .\\train-images-idx3-ubyte.gz\n",
      "Extracting .\\train-labels-idx1-ubyte.gz\n",
      "Extracting .\\t10k-images-idx3-ubyte.gz\n",
      "Extracting .\\t10k-labels-idx1-ubyte.gz\n"
     ]
    }
   ],
   "source": [
    "mnist = input_data.read_data_sets(\".\", one_hot = True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Cada amostra de input é uma figura de 28x28 pixels com o desenho manuscrito de um dígito. Pegaremos um exemplo do dataset,  para fornecer uma amostra do input"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.text.Text at 0x19aeca8c4e0>"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAEICAYAAACQ6CLfAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAEHBJREFUeJzt3X+MVWV+x/H3Z9XNVtftioOEqMi6NQJpUhZvdJMVpNl2\nqyYNLqZG7a40tYt/WH8ktKkiRGM0kraLtWmzOlYr4o/tpvwyLWiUVMGabhktqyhSrZlBCchMXSvY\nJrvqt3/cgxlx7nOG++tc5vm8ksncOd9z5ny58OHce55z7qOIwMzy84WqGzCzajj8Zply+M0y5fCb\nZcrhN8uUw2+WKYc/Y5KelfRH3d7WeoPDPwFIGpT0W1X30YikyyXtkvSBpP2SVkn6StV95c7ht254\nAbggIr4CnAkcC9xRbUvm8E9gkk6S9E+ShiX9vHh82mGrfV3SvxdH5Q2SJo3a/puSXpD0vqSfSZrf\nTB8RsTsi9o1a9DHwa838Lmsfh39i+wLw98AZwDTg/4C/OWydq4A/BKYCHwF/DSDpVOCfqR+hJwF/\nAqyRNPnwnUiaVvwHMa1RI5LOl/Q/wAHgUuCvWvujWasc/gksIv47ItZExP9GxAHgTuCCw1ZbHRE7\nIuJDYDlwmaRjgO8BGyNiY0R8EhFPAwPAxWPsZ3dEfDUidid6eT4ifhU4DfgLYLAtf0hrmsM/gUk6\nXtJ9koYkfQBsAb5ahPuQt0c9HgKOA/qov1r4veKI/r6k94Hzqb9CaFpE7AGeBH7cyu+x1h1bdQPW\nUUuAs4HzImKfpNnAfwAatc7pox5PA34JjFD/T2F1RPygA30dC3y9A7/XjoCP/BPHcZK+NOrrWOBE\n6u/z3y9O5N06xnbfkzRL0vHA7cA/RsTHwCPA70r6HUnHFL9z/hgnDEtJ+v1D5wMknUH97cfmJv+c\n1iYO/8SxkXrQD33dRv2k2q9QP5L/G/WX24dbDTwE7AO+BFwPEBFvAwuApcAw9VcCf8oY/2aKE34H\nEyf8ZgEvSPoQ+FdgF9CJVxR2BOQP8zDLk4/8Zply+M0y5fCbZcrhN8tUV8f5+/r6Yvr06d3cpVlW\nBgcHGRkZUfmaLYZf0oXAPcAxwN9FxIrU+tOnT2dgYKCVXZpZQq1WG/e6Tb/sLy4R/VvgIurjuFdI\nmtXs7zOz7mrlPf+5wJsR8VZE/IL6tdoL2tOWmXVaK+E/lc/eFPJOsewzJC2WNCBpYHh4uIXdmVk7\ndfxsf0T0R0QtImqTJ3/uVnAzq0gr4d/DZ+8IO61YZmZHgVbCvw04S9LXJH0RuBx4oj1tmVmnNT3U\nFxEfSfpj4CnqQ30PRsSrbevMzDqqpXH+iNhI/VZSMzvK+PJes0w5/GaZcvjNMuXwm2XK4TfLlMNv\nlimH3yxTDr9Zphx+s0w5/GaZcvjNMuXwm2XK4TfLlMNvlimH3yxTDr9Zphx+s0w5/GaZcvjNMuXw\nm2XK4TfLlMNvlimH3yxTDr9Zphx+s0w5/GaZcvjNMuXwm2XK4TfLVEuz9NrRb2hoKFm///77k/U7\n77wzWZfUsBYRyW1nzpyZrN9xxx3J+sKFC5P13LUUfkmDwAHgY+CjiKi1oykz67x2HPl/MyJG2vB7\nzKyL/J7fLFOthj+AZyS9KGnxWCtIWixpQNLA8PBwi7szs3ZpNfznR8Rs4CLgWknzDl8hIvojohYR\ntcmTJ7e4OzNrl5bCHxF7iu/7gXXAue1oysw6r+nwSzpB0omHHgPfAXa0qzEz66xWzvZPAdYV47jH\nAo9FxJNt6cqOSOpcyl133ZXc9tFHH03WR0bSAzmpcfzx1FN27dqVrC9ZsiRZnzfvc+9CP9XX19dU\nTxNJ0+GPiLeA32hjL2bWRR7qM8uUw2+WKYffLFMOv1mmHH6zTPmW3qNA2a2ry5cvb1grG2oru622\nbPtp06Yl661c1Vk2zDg4OJisp4b6XnvttWZamlB85DfLlMNvlimH3yxTDr9Zphx+s0w5/GaZcvjN\nMuVx/qPAhg0bkvXUWHwrt9QCzJo1K1l/9tlnk/VWbp3dunVrsn7BBRck62W3BOfOR36zTDn8Zply\n+M0y5fCbZcrhN8uUw2+WKYffLFMe5+8BO3fuTNZff/31ZD11T33Z/fRl4/ArV65M1pctW5asL126\ntGGt7LMA5s6dm6yXfRZBSn9/f7K+ePGYs89NKD7ym2XK4TfLlMNvlimH3yxTDr9Zphx+s0w5/GaZ\n8jh/D5g5c2ayvm3btmQ9NVbf6lTUZePhrYyXl43zr127NllvZXrwhQsXJrfNQemRX9KDkvZL2jFq\n2SRJT0t6o/h+UmfbNLN2G8/L/oeACw9bdhOwOSLOAjYXP5vZUaQ0/BGxBXjvsMULgFXF41XAJW3u\ny8w6rNkTflMiYm/xeB8wpdGKkhZLGpA0MDw83OTuzKzdWj7bH/W7KxreYRER/RFRi4haK5M2mll7\nNRv+dyVNBSi+729fS2bWDc2G/wlgUfF4EZD+bGkz6zml4/ySHgfmA32S3gFuBVYAP5F0NTAEXNbJ\nJnM3Y8aMyvZddp3A2WefnayffPLJDWt33313ctsVK1Yk62X386feZrZ6/cNEUBr+iLiiQenbbe7F\nzLrIl/eaZcrhN8uUw2+WKYffLFMOv1mmfEvvBLBly5aGtbKP/S4b8iq73bhsGuzzzjuvYW3//vS1\nYWW37J5yyinJ+qZNm5L13PnIb5Yph98sUw6/WaYcfrNMOfxmmXL4zTLl8JtlyuP8E8Bjjz3WsFb2\n0dplt8WWjbWXbZ8ay2/lllyA6667LlmfM2dOsp47H/nNMuXwm2XK4TfLlMNvlimH3yxTDr9Zphx+\ns0x5nH+CKxunr3L7efPmJbdduXJlsu5x/Nb4yG+WKYffLFMOv1mmHH6zTDn8Zply+M0y5fCbZcrj\n/BPAlVde2bA2NDSU3HZkZCRZL/vc/4MHDybrKbfffnuy7nH8zio98kt6UNJ+STtGLbtN0h5J24uv\nizvbppm123he9j8EXDjG8rsjYnbxtbG9bZlZp5WGPyK2AO91oRcz66JWTvhdJ+nl4m3BSY1WkrRY\n0oCkgeHh4RZ2Z2bt1Gz4fwScCcwG9gI/bLRiRPRHRC0iamUfyGhm3dNU+CPi3Yj4OCI+Ae4Hzm1v\nW2bWaU2FX9LUUT9+F9jRaF0z602l4/ySHgfmA32S3gFuBeZLmg0EMAhc08EerUTqvviye+bLlI3z\n33LLLcn6+vXrG9aWLFmS3HbTpk3Jel9fX7JuaaXhj4grxlj8QAd6MbMu8uW9Zply+M0y5fCbZcrh\nN8uUw2+WKd/SO06pS5Mn8pWLM2bMSNbXrFmTrF900UUNa08++WRy20ceeSRZv/HGG5N1S/OR3yxT\nDr9Zphx+s0w5/GaZcvjNMuXwm2XK4TfLlMf5C1u2bEnWU7eflo2Fr169uqmeJoKlS5c2rD311FPJ\nbXft2tXudmwUH/nNMuXwm2XK4TfLlMNvlimH3yxTDr9Zphx+s0xlM85fNlXYNdekP318ypQpDWs5\nj+N/+OGHyXrqeY2IdrdjR8BHfrNMOfxmmXL4zTLl8JtlyuE3y5TDb5Yph98sU+OZovt04GFgCvUp\nufsj4h5Jk4B/AKZTn6b7soj4eedabc26deuS9bJ7x+fPn9/Gbo4eO3fuTNYvvfTSZD31vEpKblv2\nOQnWmvEc+T8ClkTELOCbwLWSZgE3AZsj4ixgc/GzmR0lSsMfEXsj4qXi8QFgJ3AqsABYVay2Crik\nU02aWfsd0Xt+SdOBbwA/BaZExN6itI/62wIzO0qMO/ySvgysAW6MiA9G16J+kfaYF2pLWixpQNJA\n2fX1ZtY94wq/pOOoB//RiFhbLH5X0tSiPhXYP9a2EdEfEbWIqE3kCS3Njjal4Vf9lOwDwM6IWDmq\n9ASwqHi8CNjQ/vbMrFPGc0vvt4DvA69I2l4sWwqsAH4i6WpgCLisMy22x9y5c5P1sttLn3vuuYa1\nsqmkZ86cmayfc845yXqZoaGhhrWtW7cmt127dm2yvn79+mS97HlLDeeVTbF9ww03JOvWmtLwR8Tz\nQKO/wW+3tx0z6xZf4WeWKYffLFMOv1mmHH6zTDn8Zply+M0ylc1Hd5eNtS9cuDBZT413X3XVVclt\ny25dnTNnTrJeZvfu3Q1rIyMjyW1bGacfj2XLljWsXX/99S39bmuNj/xmmXL4zTLl8JtlyuE3y5TD\nb5Yph98sUw6/WaayGecvc++99ybrqbH0gYGBlvZdtn3ZWHtqrL5s2+OPPz5ZL7s+4uabb07Wy66f\nsOr4yG+WKYffLFMOv1mmHH6zTDn8Zply+M0y5fCbZcrj/IWy2YQ2bdrUsLZ8+fKW9n3fffcl62XT\nYPf19TW977LPxvc02ROXj/xmmXL4zTLl8JtlyuE3y5TDb5Yph98sUw6/WaY0js9tPx14GJgCBNAf\nEfdIug34ATBcrLo0IjamfletVotW7303s8ZqtRoDAwPjmmxhPBf5fAQsiYiXJJ0IvCjp6aJ2d0T8\nZbONmll1SsMfEXuBvcXjA5J2Aqd2ujEz66wjes8vaTrwDeCnxaLrJL0s6UFJJzXYZrGkAUkDw8PD\nY61iZhUYd/glfRlYA9wYER8APwLOBGZTf2Xww7G2i4j+iKhFRK3s+nkz655xhV/ScdSD/2hErAWI\niHcj4uOI+AS4Hzi3c22aWbuVhl/1j399ANgZEStHLZ86arXvAjva356Zdcp4zvZ/C/g+8Iqk7cWy\npcAVkmZTH/4bBK7pSIdm1hHjOdv/PDDWuGFyTN/Mepuv8DPLlMNvlimH3yxTDr9Zphx+s0w5/GaZ\ncvjNMuXwm2XK4TfLlMNvlimH3yxTDr9Zphx+s0w5/GaZKv3o7rbuTBoGhkYt6gNGutbAkenV3nq1\nL3BvzWpnb2dExLg+L6+r4f/czqWBiKhV1kBCr/bWq32Be2tWVb35Zb9Zphx+s0xVHf7+ivef0qu9\n9Wpf4N6aVUlvlb7nN7PqVH3kN7OKOPxmmaok/JIulLRL0puSbqqih0YkDUp6RdJ2SZXOJ17Mgbhf\n0o5RyyZJelrSG8X3MedIrKi32yTtKZ677ZIurqi30yX9i6TXJL0q6YZieaXPXaKvSp63rr/nl3QM\n8J/AbwPvANuAKyLita420oCkQaAWEZVfECJpHnAQeDgifr1Y9ufAexGxoviP86SI+LMe6e024GDV\n07YXs0lNHT2tPHAJ8AdU+Nwl+rqMCp63Ko785wJvRsRbEfEL4MfAggr66HkRsQV477DFC4BVxeNV\n1P/xdF2D3npCROyNiJeKxweAQ9PKV/rcJfqqRBXhPxV4e9TP71DhEzCGAJ6R9KKkxVU3M4YpEbG3\neLwPmFJlM2Monba9mw6bVr5nnrtmprtvN5/w+7zzI2I2cBFwbfHytidF/T1bL43Vjmva9m4ZY1r5\nT1X53DU73X27VRH+PcDpo34+rVjWEyJiT/F9P7CO3pt6/N1DMyQX3/dX3M+nemna9rGmlacHnrte\nmu6+ivBvA86S9DVJXwQuB56ooI/PkXRCcSIGSScA36H3ph5/AlhUPF4EbKiwl8/olWnbG00rT8XP\nXc9Ndx8RXf8CLqZ+xv+/gFuq6KFBX2cCPyu+Xq26N+Bx6i8Df0n93MjVwMnAZuAN4BlgUg/1thp4\nBXiZetCmVtTb+dRf0r8MbC++Lq76uUv0Vcnz5st7zTLlE35mmXL4zTLl8JtlyuE3y5TDb5Yph98s\nUw6/Wab+H3chFIrMeat8AAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x19aeaa197f0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "exemplo = mnist.train.images[1].reshape(28,28)\n",
    "plt.imshow(exemplo, cmap='gray_r')\n",
    "plt.title('Label: {}'.format(np.argmax(mnist.train.labels[1])))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Começamos a nos concentrar na elaboração do modelo. O primeiro passo é definir os parâmetros:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "taxa_de_aprendizagem = .01   # velocidade com que o nosso algoritmo vai aprender, normalmente são usados valores baixos já que\n",
    "                             # faremos várias iterações no modelo. Valores altos frequentemente levam a piores resultados\n",
    "\n",
    "epocas = 200                 # quantidade de iterações\n",
    "lote = 100                   # dividimos nosso dataset em 'x' lotes para que nosso computador consiga dar conta do processamento\n",
    "exibicao = 10                # exibir informações a cada 'x' épocas\n",
    "\n",
    "n_input = 784                # tamanho do input: 28 * 28\n",
    "n_classes = 10               # número de classes possíveis: de 0 a 9\n",
    "\n",
    "n_camada_oculta = 256        # quantidade de neurônios da camada oculta"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Iniciamos pesos e viéses do modelo com valores aleatórios entre -1 e 1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "pesos_camada_oculta = tf.Variable(tf.random_normal([n_input, n_camada_oculta]))\n",
    "pesos_output = tf.Variable(tf.random_normal([n_camada_oculta, n_classes]))\n",
    "\n",
    "vies_camada_oculta = tf.Variable(tf.random_normal([n_camada_oculta]))\n",
    "vies_output = tf.Variable(tf.random_normal([n_classes]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Placeholder é uma variável que iremos atribuir dado em algum momento posterior. Isso permite que possamos criar nossa operações e construir nosso grafo sem precisar dos dados a priori"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "x = tf.placeholder(tf.float32, [None, 784])\n",
    "y = tf.placeholder(tf.float32, [None, n_classes])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Realizamos as operações necessárias e utilizamos a função de ativação Relu\n",
    "\n",
    "Relu é uma função que retorna 0 para valores negativos e o próprio valor para números não negativos. Exemplo:\n",
    "<code><center>\n",
    "def relu(x):\n",
    "    if x < 0:\n",
    "        return 0\n",
    "else:\n",
    "        return x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "camada_oculta = tf.add(tf.matmul(x, pesos_camada_oculta), vies_camada_oculta)\n",
    "camada_oculta = tf.nn.relu(camada_oculta)\n",
    "\n",
    "logits = tf.add(tf.matmul(camada_oculta, pesos_output), vies_output)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Definir a perda (ou custo) e o otimizador (Gradiente Descendente)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "perda = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=y))\n",
    "\n",
    "otimizador = tf.train.GradientDescentOptimizer(learning_rate=taxa_de_aprendizagem).minimize(perda)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### INICIAR A SESSÃO E CALCULAR A ACURÁCIA DAS NOSSA PREVIÇÕES\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Época: 0001 custo= 5.352\n",
      "Época: 0011 custo= 2.775\n",
      "Época: 0021 custo= 2.158\n",
      "Época: 0031 custo= 1.287\n",
      "Época: 0041 custo= 0.218\n",
      "Época: 0051 custo= 0.591\n",
      "Época: 0061 custo= 0.311\n",
      "Época: 0071 custo= 0.142\n",
      "Época: 0081 custo= 0.518\n",
      "Época: 0091 custo= 0.006\n",
      "Época: 0101 custo= 0.009\n",
      "Época: 0111 custo= 0.011\n",
      "Época: 0121 custo= 0.658\n",
      "Época: 0131 custo= 0.006\n",
      "Época: 0141 custo= 0.075\n",
      "Época: 0151 custo= 0.109\n",
      "Época: 0161 custo= 0.011\n",
      "Época: 0171 custo= 0.003\n",
      "Época: 0181 custo= 0.009\n",
      "Época: 0191 custo= 0.023\n",
      "\n",
      " ACURÁCIA ==>   0.9375\n"
     ]
    }
   ],
   "source": [
    "init = tf.global_variables_initializer() # iniciailiza as variáveis globais\n",
    "\n",
    "\n",
    "with tf.Session() as sess:\n",
    "    sess.run(init)\n",
    "    \n",
    "    ### CICLOS DE TREINAMENTO\n",
    "    for epoca in range(epocas):\n",
    "        lote_total = int(mnist.train.num_examples/lote)\n",
    "        \n",
    "        for i in range(lote_total):\n",
    "            lote_x, lote_y = mnist.train.next_batch(lote)\n",
    "            \n",
    "            # Executar otimização e calcular perda\n",
    "            sess.run(otimizador, feed_dict={x: lote_x, y:lote_y})\n",
    "      \n",
    "        if epoca % exibicao == 0:\n",
    "            c = sess.run(perda, feed_dict={x: lote_x, y: lote_y})\n",
    "            print(\"Época:\", '%04d' % (epoca+1), \"custo=\", \\\n",
    "                \"{:.3f}\".format(c))\n",
    "            \n",
    "    ### TESTAR O MODELO\n",
    "    previsoes_corretas = tf.equal(tf.argmax(logits, 1), tf.argmax(y, 1))\n",
    "\n",
    "    ### CALCULAR ACURÁCIA\n",
    "    acuracia = tf.reduce_mean(tf.cast(previsoes_corretas, \"float\"))\n",
    "\n",
    "    test_size = 256\n",
    "    print(\"\\n ACURÁCIA ==>  \", acuracia.eval({x: mnist.test.images[:test_size], y: mnist.test.labels[:test_size]}))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<b>FONTES: <br>\n",
    "    - https://www.tensorflow.org/get_started/mnist/beginners)\n",
    "    - https://www.oreilly.com/learning/not-another-mnist-tutorial-with-tensorflow\n"
   ]
  }
 ],
 "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": "3.6.1"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}