mnist_10_99.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Précision > 99,3%"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Modification du réseau précédent pour augementer le nombre de canaux de sortie de la première couche. Le but est de prendre en compte plus de traits caractérisques de nos images d'entrée."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "from __future__ import print_function\n",
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "from tensorflow.keras.utils import to_categorical\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()\n",
    "x_train = x_train/255.\n",
    "x_test  = x_test/255.\n",
    "y_train = to_categorical(y_train, dtype='int8')\n",
    "y_test  = to_categorical(y_test, dtype='int8')\n",
    "\n",
    "train_size = x_train.shape[0] # Nombre de ligne de la matrice x_train\n",
    "test_size  = x_test.shape[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "#Nombre d'époques pour l'apprentissage:\n",
    "n_epoch = 15\n",
    "\n",
    "# Taille du batch d'entraînement:\n",
    "batch_size = 100\n",
    "\n",
    "# Incrément pour l'apprentissage\n",
    "starter_learning_rate = 0.003\n",
    "# taux de décroissance du pas d'apprentissage\n",
    "decay_rate = 0.7\n",
    "# Nombre d'époques à faire avant de changer de pas\n",
    "decay_epoch = 3"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "X = tf.placeholder(tf.float32, shape=(None,28,28))\n",
    "Xd = tf.reshape(X, [-1,28,28,1])\n",
    "\n",
    "\n",
    "W1 = tf.Variable(tf.truncated_normal([6,6,1,12], stddev=0.1))\n",
    "b1 = tf.Variable(tf.ones([12])/10)\n",
    "stride1 = 1\n",
    "Y1cnv = tf.nn.conv2d(Xd, W1, strides=[1, stride1, stride1, 1], padding='SAME')\n",
    "Y1 = tf.nn.relu(Y1cnv + b1)\n",
    "\n",
    "W2 = tf.Variable(tf.truncated_normal([5,5,12,12], stddev=0.1))\n",
    "b2 = tf.Variable(tf.ones([12])/10)\n",
    "stride2 = 2\n",
    "Y2cnv = tf.nn.conv2d(Y1, W2, strides=[1,stride2,stride2,1], padding='SAME')\n",
    "Y2 = tf.nn.relu(Y2cnv + b2)\n",
    "\n",
    "W3 = tf.Variable(tf.truncated_normal([4,4,12,24], stddev=0.1))\n",
    "b3 = tf.Variable(tf.ones([24])/10)\n",
    "stride3 = 2\n",
    "Y3cnv = tf.nn.conv2d(Y2, W3, strides=[1,stride3,stride3,1], padding='SAME')\n",
    "Y3 = tf.nn.relu(Y3cnv + b3)\n",
    "\n",
    "Y3d = tf.reshape(Y3, [-1, 7*7*24])\n",
    "\n",
    "# dropout_rate: probabilité qu'un neurone soit supprimé (doit être = 0 lors de la phase de test)\n",
    "dropout_rate = tf.placeholder(tf.float32)\n",
    "W4 = tf.Variable(tf.truncated_normal([7*7*24,200], stddev=0.1))\n",
    "b4 = tf.Variable(tf.ones([200])/10)\n",
    "Y4 = tf.nn.relu(tf.matmul(Y3d,W4) + b4)\n",
    "Y4d = tf.nn.dropout(Y4, rate=dropout_rate)\n",
    "\n",
    "W5 = tf.Variable(tf.truncated_normal([200, 10], stddev=0.1))\n",
    "b5 = tf.Variable(tf.zeros([10]))\n",
    "Ylogits = tf.matmul(Y4d,W5) + b5\n",
    "Y = tf.nn.softmax(Ylogits)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "Y_ = tf.placeholder(tf.float32, shape=(None,10))\n",
    "cross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=Ylogits, labels=Y_)\n",
    "cross_entropy = tf.reduce_sum(cross_entropy)\n",
    "\n",
    "is_correct = tf.equal(tf.argmax(Y,1), tf.argmax(Y_,1))\n",
    "accuracy = tf.reduce_mean(tf.cast(is_correct, tf.float32))\n",
    "\n",
    "global_step = tf.Variable(0, trainable=False)\n",
    "learning_rate = tf.train.exponential_decay(starter_learning_rate, global_step, decay_epoch*train_size//batch_size, decay_rate, staircase=True)\n",
    "\n",
    "# Optimisation\n",
    "optimizer = tf.train.AdamOptimizer(learning_rate)\n",
    "train_step = optimizer.minimize(cross_entropy, global_step=global_step)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "accuracy_train = []\n",
    "accuracy_test = []\n",
    "loss_train = []\n",
    "loss_test = []\n",
    "i_best = 0\n",
    "accuracy_best = 0"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoque: \n",
      "1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  "
     ]
    }
   ],
   "source": [
    "sess = tf.Session()\n",
    "\n",
    "init = tf.global_variables_initializer()\n",
    "sess.run(init)\n",
    "print('Epoque: ')\n",
    "\n",
    "for i in range(n_epoch*train_size//100):\n",
    "    epoch = (i*100)//train_size + 1\n",
    "    \n",
    "    # Chargement d'un batch d'images et de réponses correctes correspondantes\n",
    "    batch_train_X  = x_train[(i*100)%train_size:((i+1)*100)%train_size]\n",
    "    batch_train_Y_ = y_train[(i*100)%train_size:((i+1)*100)%train_size]\n",
    "    # Dictionnaire de données en entrée\n",
    "    train_data = {X: batch_train_X, Y_: batch_train_Y_, dropout_rate:0.25}\n",
    "\n",
    "    # train\n",
    "    sess.run(train_step, feed_dict=train_data)\n",
    "    \n",
    "    # Visualisation de la progression\n",
    "    if not (i*100)%(train_size):\n",
    "        #Affichage de l'évolution en ligne\n",
    "        print(epoch, ' ', end='')\n",
    "    if not i%50:\n",
    "        # Pour le batch d'entraînement\n",
    "        acc, loss  = sess.run([accuracy, cross_entropy], feed_dict=train_data)\n",
    "        accuracy_train.append(acc)\n",
    "        loss_train.append(loss/batch_size)\n",
    "        \n",
    "        # Pour le test: évaluation sur tout l'ensemble des données\n",
    "        test_data = {X: x_test, Y_: y_test, dropout_rate:0}\n",
    "        acc, loss = sess.run([accuracy, cross_entropy], feed_dict=test_data)\n",
    "        accuracy_test.append(acc)\n",
    "        loss_test.append(loss/test_size)\n",
    "        if acc>accuracy_best:\n",
    "            accuracy_best = acc\n",
    "            i_best = i"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Best score for iteration 6550 \t| Accuracy = 0.9933\n"
     ]
    }
   ],
   "source": [
    "# Représentation des résultats\n",
    "# Précision\n",
    "plt.plot(accuracy_test, c='coral', label='test')\n",
    "plt.plot(accuracy_train, c='steelblue', label='train')\n",
    "plt.xlabel('i//50')\n",
    "plt.title('Précision')\n",
    "plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)\n",
    "plt.show()\n",
    "# Fonction de coût\n",
    "plt.plot(loss_test, c='coral', label='test')\n",
    "plt.plot(loss_train, c='steelblue', label='train')\n",
    "plt.title('Entropie croisée moyenne')\n",
    "plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)\n",
    "plt.show()\n",
    "# Meilleur résultat sur un batch de test\n",
    "print(\"Best score for iteration {} \\t| Accuracy = {:.4f}\".format(i_best, accuracy_best ))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "sess.close()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Best score for iteration 6550 \t| Accuracy = 0.9933"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "99.33 % de taux de reconnaissance sur le jeu de test. Mission réussie :)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "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": "3.7.1"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}