{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<img width=200 src=\"ENSL.svg\"/> \n",
    "author: Aurélien Garivier\n",
    "\n",
    "# An experiment with the MNIST dataset\n",
    "\n",
    "\n",
    "The Mnist dataset provides a (somewhat old but) very classical example of Machine Learning. The goal is simple and useful: recognize hand-written digits. This was an important challenge for the postal companies in the 1980s.  \n",
    "\n",
    "It is available on the webpage of <a href=\"http://yann.lecun.com/exdb/mnist/\">Yann LeCun</a>, a very famous researcher in Machine Learning in general and neural networks in particular. \n",
    "\n",
    "</p>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Populating the interactive namespace from numpy and matplotlib\n"
     ]
    }
   ],
   "source": [
    "# loading a lot of facilities for numerical computations and graphs\n",
    "%pylab inline \n",
    "pylab.rcParams['figure.figsize'] = (20, 6) # to have larger plot\n",
    "seed(240979) #initialize random number generator\n",
    "\n",
    "# set working directory\n",
    "workingdir = './'"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1. Loading the data set\n",
    "\n",
    "### 1.1 Description of the data set\n",
    "The MNIST data contains $n\\_tot=70000$ images of size $dimX\\times dimY=28\\times28$, and their *labels* (a digit between $0$ and $9$). \n",
    "The images are stored as lines of the matrix *mnist.data*. \n",
    "Each image can be reshaped as a matrix $im \\in \\mathcal{M}_{dimX, dimY}(\\mathbb{R})$, where $im[i,j]$ is the intensity level of pixel (i,j): 0 means 'white', 255 means 'black'.\n",
    "\n",
    "The labels, stored in *mnist.target*, are numbers (here, we cast them to integers) between $0$ and $9$. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Loaded 70000 images of size 28x28\n",
      "Number of classes: 10\n",
      "Classes:\n",
      "[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]\n"
     ]
    }
   ],
   "source": [
    "# deprecated...\n",
    "#from sklearn.datasets import fetch_mldata # si besoin: sudo pip install -U scikit-learn ou (pour une install locale)  pip install --user --install-option=\"--prefix=\" -U scikit-learn\n",
    "#mnist = fetch_mldata('MNIST original', data_home='./')\n",
    "#mnist.target = mnist.target.astype(int) # by default the digits are floating numbers: convert to integers\n",
    "\n",
    "# new way to load MNIST dataset\n",
    "from sklearn.datasets import fetch_openml\n",
    "mnist = lambda:None\n",
    "mnist.data, mnist.target = fetch_openml(\"mnist_784\", version=1, return_X_y=True, as_frame=False)\n",
    "mnist.target = [int(d) for d in mnist.target]\n",
    "\n",
    "# alternative: loading a small dataset instead\n",
    "#from sklearn import datasets, svm, metrics\n",
    "#from sklearn.model_selection import train_test_split\n",
    "#digits = datasets.load_digits()\n",
    "#_, axes = plt.subplots(nrows=1, ncols=4, figsize=(10, 3))\n",
    "#for ax, image, label in zip(axes, digits.images, digits.target):\n",
    "#    ax.set_axis_off()\n",
    "#    ax.imshow(image, cmap=plt.cm.gray_r, interpolation=\"nearest\")\n",
    "#    ax.set_title(\"Training: %i\" % label)\n",
    "#n_samples = len(digits.images)\n",
    "#mnist = lambda:None\n",
    "#mnist.data = digits.images.reshape((n_samples, -1))\n",
    "#mnist.target = digits.target\n",
    "\n",
    "# defining general variables for use throuhout the notebook  \n",
    "n_tot = len(mnist.data)\n",
    "dimX = int(sqrt(len(mnist.data[0]))); dimY = dimX # nb of pixels in each dimension\n",
    "nc = len(unique(mnist.target)) # number of classes\n",
    "\n",
    "print(\"Loaded %d images of size %dx%d\"%(n_tot, dimX, dimY))\n",
    "print(\"Number of classes: %d\"%(nc))\n",
    "print(\"Classes:\")\n",
    "print(sorted(unique(mnist.target)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.2 What does the data look like?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 504x504 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1440x432 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Showing one image \n",
    "figure(figsize = (7,7))\n",
    "ii = random.randint(0, n_tot-1)\n",
    "plt.imshow(reshape(mnist.data[ii], [dimX, dimY], order='C'), cmap=\"Greys\", interpolation=\"none\") # or \"nearest\"\n",
    "ax = plt.gca()\n",
    "ax.set_xticks(arange(-0.5, dimX, 1))\n",
    "ax.set_yticks(arange(-0.5, dimY, 1))\n",
    "ax.set_xticklabels(arange(0, dimX+1, 1))\n",
    "ax.set_yticklabels(arange(0, dimY+1, 1))\n",
    "ax.grid(color='r', linestyle='--', linewidth=1)\n",
    "plt.show()\n",
    "\n",
    "# Showing a few images in the dataset with their labels\n",
    "nbRows = 5\n",
    "nbCols = 15\n",
    "I = [[0 for k in range(nbCols)] for j in range(nbRows)]\n",
    "I[0] = [random.randint(0, n_tot-1) for k in range(nbCols)]\n",
    "I[0][0] = ii\n",
    "M = np.concatenate([reshape(mnist.data[i], [dimX,dimY], order='C') for i in I[0]], axis=1)\n",
    "\n",
    "for j in range(1, nbRows):\n",
    "    I[j] = [random.randint(0, n_tot-1) for k in range(nbCols)]\n",
    "    M = concatenate([M, np.concatenate([np.reshape(mnist.data[i], [dimX, dimY], order='C') for i in I[j]], axis=1)], axis=0)\n",
    "plt.imshow(M, cmap=\"Greys\", interpolation=\"none\")\n",
    "for j in range(nbRows):\n",
    "    for k in range(nbCols):\n",
    "        plt.text(dimX*k+2, dimX*j+5, mnist.target[I[j][k]], fontsize=14, color=\"red\")\n",
    "ax = plt.gca()\n",
    "ax.set_xticks(np.arange(-0.5, dimX*nbCols, dimX))\n",
    "ax.set_yticks(np.arange(-0.5, dimY*nbRows, dimY))\n",
    "ax.set_xticklabels([])\n",
    "ax.set_yticklabels([])\n",
    "ax.grid(color='k', linestyle='-', linewidth=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 2. What is the goal?\n",
    "\n",
    "\n",
    "### 2.1 Classifiers\n",
    "We must construct a *classifier* $\\mathcal{M}_{dimX, dimY}(\\mathbb{R})\\to \\{0,\\dots,9\\}$ which maps every possible data $X$ (an image) to a label $d$ (a digit).\n",
    "This is a **classification problem** with $n_c = 10$ classes.\n",
    "\n",
    "Classically, this classifiers is assumed to take as input:\n",
    "  * a matrix (table) $X\\in\\mathcal{M}_{n, p}(\\mathbb{R})$, where $X_{i,.}$ contains the relevant information on the $i^{th}$ example (here, image);\n",
    "  * a vector (list) $y\\in\\mathcal{Y}^n$ of *labels* (here, for each image, the corresponding digit)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Classifier: # abstract base class, just providing a template for the classifiers\n",
    "    def __init__(self):\n",
    "        pass\n",
    "    \n",
    "    def fit(self, X, y): \n",
    "        pass\n",
    "    \n",
    "    def predict(self, X): # returns a digit (0,1,...,9) for each line of X\n",
    "        return(apply_along_axis(self.predictOne, axis=1, arr=X))\n",
    "    \n",
    "    def predictOne(self, x):\n",
    "        #return(nc) # nc=10 means \"I don't know\"\n",
    "        return(random.choice(range(nc))) # random choice\n",
    "\n",
    "randomClassifier = Classifier()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2.2 Dividing the data into a training and a testing set\n",
    "\n",
    "\n",
    "We will separate the data into two sets: the *training set* and the *testing set.\n",
    "We will use the *training set* only to teach the machine recognize the digits.  \n",
    "The *testing set*, on the other hand, will be used to evaluate the performance of our classifier. It should *not* be used in the construction of the classifier! In usual challenges, it is hidden to the participants."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [
    {
     "ename": "TypeError",
     "evalue": "list indices must be integers or slices, not list",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mTypeError\u001b[0m                                 Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-40-0c9c4698feff>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[1;32m      1\u001b[0m \u001b[0mI\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m[\u001b[0m\u001b[0;36m1\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;36m2\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 2\u001b[0;31m \u001b[0mmnist\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mtarget\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mI\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
      "\u001b[0;31mTypeError\u001b[0m: list indices must be integers or slices, not list"
     ]
    }
   ],
   "source": [
    "I = [1,2]\n",
    "mnist.target[I]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "class DataSet:\n",
    "    def __init__(self, data, target): \n",
    "        self.data = data # input : images\n",
    "        self.X = data # default: features = data itself\n",
    "        self.target = target # output: labels\n",
    "        self.n = len(data)\n",
    "        \n",
    "# defining the dataSets\n",
    "# training set: the images we may use to construct the classifier\n",
    "\n",
    "n_train = 1000\n",
    "I = random.choice(range(n_tot), n_train, replace=False)\n",
    "trainingSet = DataSet(mnist.data[I, :], [mnist.target[i] for i in I]) \n",
    "\n",
    "# testing set: we may not use them, except at the end to assert the quality of the classifier\n",
    "mask = np.ones(n_tot, dtype=bool) #np.ones_like(a,dtype=bool)\n",
    "mask[I] = False\n",
    "testingSet = DataSet(np.array(mnist.data)[mask, :], np.array(mnist.target)[mask])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Training set of size 1000\n",
      "Testing set of size 69000\n",
      "Histogram of the labels in the training set:\n",
      "[ 93 107 100 109  97 101  85 107  91 110]\n",
      "Histogram of the labels in the testing set:\n",
      "[6810 7770 6890 7032 6727 6212 6791 7186 6734 6848]\n"
     ]
    }
   ],
   "source": [
    "print(\"Training set of size %d\"%(trainingSet.n))\n",
    "print(\"Testing set of size %d\"%(testingSet.n))\n",
    "print(\"Histogram of the labels in the training set:\")\n",
    "print(np.histogram(trainingSet.target, [-0.5+k for k in range(11)])[0])\n",
    "print(\"Histogram of the labels in the testing set:\")\n",
    "print(np.histogram(testingSet.target, [-0.5+k for k in range(11)])[0])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Evaluating a classifier\n",
    "The most natural performance measure is the *proportion of misclassified images*. \n",
    "One may compute the *training error* \n",
    "$$ L_{mis}^{train}(classifier) = \\frac{1}{n} \\sum_{i=1}^n \\mathbb{1}\\{classifier\\big(train\\_images[i]\\big) \\neq train\\_labels[i]\\big)\\}$$ \n",
    "and the *testing error*\n",
    "$$ L_{mis}^{test}(classifier) = \\frac{1}{n_{test}} \\sum_{i=1}^{n_{test}} \\mathbb{1}\\{classifier\\big(test\\_images[i]\\big) \\neq test\\_labels[i]\\big)\\}\\;.$$ \n",
    "The testing error is the only correct estimator of the performance of the classifier on future data, because the training sample is used to train the classifier (after all, it may just store them and learn the answers by heart).\n",
    "\n",
    "Of course, for the purely random classifier, the results are predictible! (guess the results before running the next cell)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Random classifier: misclassification on the training set=0.907\n",
      "Random classifier: misclassification on the testing set=0.900667\n"
     ]
    }
   ],
   "source": [
    "def evaluate(classifier, X, y): # (X,y) is a testing set\n",
    "    return(mean(classifier.predict(X) != y))\n",
    "\n",
    "print(\"Random classifier: misclassification on the training set=%g\"%(evaluate(randomClassifier, trainingSet.X, trainingSet.target)))\n",
    "print(\"Random classifier: misclassification on the testing set=%g\"%(evaluate(randomClassifier, testingSet.X, testingSet.target)))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 3. Without Machine Learning: designing a classifier by hand\n",
    "\n",
    "\n",
    "### 3.1 An expert system\n",
    "The first, most natural approach, is to try to recognize the features from a few characteristics. For example:\n",
    "  - the contour of a '0' is black, while its center is white,\n",
    "  - a '1' has a more or less vertical bar in the middle, and nothing else,\n",
    "  - the top, bottom, and second diagonal of a '2' are black,\n",
    "  - ...\n",
    "  \n",
    "Thus, it seems natural to build simple \"detectors\" that sense the presence or absence of writing in the top, bottom, left, right, center, outer part, and diagonals of the image. Then, we can cook a classifier based on these characteristics of the image. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAABG0AAACdCAYAAAAOs4XNAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjMuNCwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8QVMy6AAAACXBIWXMAAAsTAAALEwEAmpwYAAAMBUlEQVR4nO3db4xdZZ0H8O/VpmLSbaOhAaPW0LCOG6vpBhCp2U1HXQgShXdqgikJQkyJNhJjlEQh8d/6RkIIVWRfYNxF2F0jLwQ0JLZRQXe3SBV0jRp0ibEsqcQ/JSsQe/bFKct05nZ7LvPn+d2Zzyc5Ge7JQH+deebew3ee872jrusCAAAAQC0vaD0AAAAAAAsJbQAAAAAKEtoAAAAAFCS0AQAAAChIaAMAAABQkNAGAAAAoKB1J/uE0Wh0ZZIrk+SUnHLWlmxZ9qEAAAAA1oqf5WeHu67bPP/8qOu6wf+RmdFMd3NuXtLBAAAAANay2cw+0HXd2fPPuz0KAAAAoCChDQAAAEBBQhsAAACAgk5aRHwys5ldijmYYvuyb9Dn7ex2Lu8gTIX9o/2DPq/Sc8sk3V8snaFrxXNLO6PRqPUI/8drEUNN4+sQ7Qx9bqm+XlzLLD/XLbVUukYZZ+hzS2KnDQAAAEBJQhsAAACAgoQ2AAAAAAUtutMGAACAusb1e+i5YbWo3l+zWHbaAAAAABQktAEAAAAoSGgDAAAAUJDQBgAAAKAgRcQAAABrjHJiptFqLx0ex04bAAAAgIKENgAAAAAFCW0AAAAAChLaAAAAABSkiBgAAADlxJSyFkuHx7HTBgAAAKAgoQ0AAABAQUIbAAAAgIKENgAAAAAFKSIGAABgLOXErASlwydmpw0AAABAQUIbAAAAgIKENgAAAAAFCW0AAAAAClJEDAAAwGDKiVkMpcOTsdMGAAAAoCChDQAAAEBBQhsAAACAgoQ2AAAAAAUpIgYAAGBRlBMzjtLhxbPTBgAAAKAgoQ0AAABAQUIbAAAAgIKENgAAAAAFLbqIWLkU+0f7W48AALBqrbXrbcWlq4dy4rXFz+7ysNMGAAAAoCChDQAAAEBBQhsAAACAgoQ2AAAAAAUJbQAAABah67oFB+ONRqMFB9PH93G4xT4/CG0AAAAAChLaAAAAABQktAEAAAAoSGgDAAAAUNC61gMAAHBi1csdFa7CeON+Nqr/PLcy7uviuaUO63a45Vi3dtoAAAAAFCS0AQAAAChIaAMAAABQkNAGAAAAoCBFxAAAACtAOfFwyonbsB6HW6n1aKcNAAAAQEFCGwAAAICChDYAAAAABQltAAAAAApSRAwAANCIcuLhlBMvLetsuJbrzE4bAAAAgIKENgAAAAAFCW0AAAAAChLaAAAAABSkiBgAAKAQ5cTDKScexvoZrtr6sdMGAAAAoCChDQAAAEBBQptJ7d2bnHFGcsopyVlnJd/5TuuJqOjb307e8Y7k5S9PRqPk1ltbT0RVn/lMcs45ycaNyebNydvfnjz8cOupqOqmm5LXv75fLxs3Juedl9x1V+upmAIfTdIlubH1INR03XX99crc4/TTW09FUacnuTXJ40n+J8mPk/xty4Eo65fpX3vmH19vOdQUEtpM4o47kj17kmuuSR58MNmxI7nwwuTRR1tPRjVHjiTbtiU33JC8+MWtp6Gy/fuT3buT++9PvvWtZN265K1vTZ54ovVkVPSKVySf/Wzygx8kBw4kb35zcsklyY9+1HoyCjs3yRVJfth6EGqbmUkOHXrueOih1hNR0KYk9yUZJbkoyV8leX/6AAfmOyd9yPfs8ddJjib555ZDTSFFxJP43OeSyy5Lrriif3zjjck3vpF8/vP9b8vhWW97W38k/ZqBE/nmN49//OUvJ5s2Jffd1++6gbkuvvj4x5/6VP8a9L3v9TtwYJ6NSf4pyeVJPt54Fopbt87umuIqlBN/OMmhJLvmnPvVik4wzFovJ65SOnx43uPLk/whyb80mOVEpmFd2Gkz1NNPJw88kJx//vHnzz+//w05wFL44x+To0eTl7yk9SRU9+c/J7ff3u/s27Gj9TQU9cUk/5pkX+tBqO+RR/rbus84I3nXu/rHMM8lSf4tye1J/jvJg0muajkQU+XyJP+Y/rY6hhPaDHX4cH+BfNppx58/7bTkscfazASsPnv2JNu3910lMM5DDyUbNiQvelHyvvclX/ta8rrXtZ6Kgt6b5MwkH2s9CPWde27fv3fPPcktt/TXtjt2JL/9bevJKGZrkt1JHklyQZIbkvx9BDec3N+lXz//0HqQKeT2qEnN32rWdQvPATwfV1+dfPe7/fHCF7aehqpmZpKDB5Pf/S756leTXbv6bqRt2xoPRiWvTvLpJH+T5JnGszAFLrzw+MdvfGOydWvypS/1r01wzAuSHEhyzbHHB5P8ZfrQ5qZGMzEdrkjy79Gv9nzYaTPUqaf2/xM1f1fN448v3H0DMKkPfjD5ylf6MuKtW1tPQ2Xr1ydnnpmcfXbfp7Z9e3L99a2nopjzkmxO8nD60OaZJDvT/4b8mSTrm03GVNiwIXnta5Of/7z1JBRzKMlP5p37zyRbGszC9Nic5OIkt7QeZEoJbYZav75/i+977z3+/L336hIAFmfPnuS22/rA5jWvaT0N0+bo0eSpp1pPQTF3JtmWZPuc4z/S91BsT/J0k6mYGn/6U/LTnyYve1nrSTiJrusWHMvpviQz8869Osl/LeufujRGo9GCYzWYhr/XZUmeSv8a1NJK/7wsFbdHTeLqq5P3vCd5wxuSN70p+cIXkt/8pu8UgLmOHEl+8Yv+n48e7d8W/uDB5KUvTbb4XQRzXHVV/45Rd97Zlw8/u5tvw4b+gLk+8pHkoouSV76yL62+7bb+1qi77mo9GcX8/tgx15NJnkjy45Ufh+o+9KH+HQu3bOl3kX/iE8mTT/a3X8Ic1ye5P/3tUXekfwvnD+S526VgnPemD2yOtB5kSgltJvHOd/aFbJ/8ZHLoUN8fcPfdyate1XoyqjlwIJmdfe7xtdf2x65dfdEfPGvv3v7jW95y/Plrr02uu27Fx6G4xx5LLr20/7hpU/823/fck1xwQevJgGn2618n7353/8Ybmzf3nTbf/75rXBY4kP4dpD6dvuT80WMf9zacidp2pt+NdWnjOaaZ0GZSu3f3B/x/du7sS6rhZKwTJiH0ZRFmT/4prFW3t75pgWly97EDhtifpN4NW9NFpw0AAABAQXbaAAAArBLjylUrltNWMO7rUrmc1vdxuMrfx0nZaQMAAABQkNAGAAAAoCChDQAAAEBBQhsAAACAghQRAwAArGLKiYerUk7s+zPcaiodHsdOGwAAAICChDYAAAAABQltAAAAAAoS2gAAAAAUpIgYAKCw1V6wCLShnHi45S4n9nUfbi2+JtppAwAAAFCQ0AYAAACgIKENAAAAQEFCGwAAAICChDYAAACk67oFB+ONRqPjjuf77ykhPjHrsSe0AQAAAChIaAMAAABQkNAGAAAAoCChDQAAAEBB61oPwNqhZIsk2Zd9rUcAgKniGoqWxpW/WpMstbVaMjyEnTYAAAAABQltAAAAAAoS2gAAAAAUJLQBAAAAKEgRMQAAAIMpJ2YxlA5Pxk4bAAAAgIKENgAAAAAFCW0AAAAAChLaAAAAABSkiBgAAIBFUU7MOEqHF89OGwAAAICChDYAAAAABQltAAAAAAoS2gAAAAAUpIgYAACAJaeceG1ROrw87LQBAAAAKEhoAwAAAFCQ0AYAAACgIKENAAAAQEGKiAEAAFgRyolXB6XDK8dOGwAAAICChDYAAAAABQltAAAAAAoS2gAAAAAUpIgYAACAZpQT16Z0uC07bQAAAAAKEtoAAAAAFCS0AQAAAChIaAMAAABQkCJiAAAASlFO3IbS4XrstAEAAAAoSGgDAAAAUJDQBgAAAKAgoQ0AAABAQaNJioZmRjPdzbl5GccBAAAAWFtmM/tA13Vnzz9vpw0AAABAQUIbAAAAgIKENgAAAAAFCW0AAAAACjppEfFoNLoyyZXHHm5L8vByD8WqcGqSw62HYGpYLwxlrTAJ64WhrBUmYb0wlLXCJGa6rvuL+Scneveo0Wh0YFybMcxnrTAJ64WhrBUmYb0wlLXCJKwXhrJWmMSJ1ovbowAAAAAKEtoAAAAAFDRpaPPFZZmC1chaYRLWC0NZK0zCemEoa4VJWC8MZa0wibHrZaJOGwAAAABWhtujAAAAAAoS2gAAAAAUJLQBAAAAKEhoAwAAAFCQ0AYAAACgoP8FBlSfGDc0p8IAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<Figure size 1440x432 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "from scipy.linalg import toeplitz # for trigonal matrices\n",
    "\n",
    "# hand-made masks to detect some characteristics of the images\n",
    "\n",
    "top = 0; Mtop = np.concatenate((np.ones(28*8), np.zeros(28*20)))\n",
    "bottom = 1 ; Mbottom = np.concatenate((np.zeros(28*20), np.ones(28*8)))\n",
    "left = 2; Mleft = np.tile(np.concatenate((np.ones(8), np.zeros(20))), 28)\n",
    "right = 3; Mright = np.tile(np.concatenate((np.zeros(20), np.ones(8))), 28)\n",
    "center = 4; Mcenter = np.concatenate((np.zeros(28*10), np.tile(np.concatenate((np.zeros(10), np.ones(8), np.zeros(10))), 8), np.zeros(28*10)))\n",
    "outside = 5; Moutside = np.concatenate((np.ones(28*8), np.tile(np.concatenate((np.ones(8), np.zeros(12), np.ones(8))), 12), np.ones(28*8)))\n",
    "diag1 = 6; Mdiag1 = np.reshape(toeplitz(np.concatenate((np.ones(8), np.zeros(20)))), [28*28])\n",
    "diag2 = 7; Mdiag2 = np.reshape(np.flipud(toeplitz(np.concatenate((np.ones(8), np.zeros(20))))), [28*28])\n",
    "\n",
    "masks = [Mtop, Mbottom, Mleft, Mright, Mcenter, Moutside, Mdiag1, Mdiag2]\n",
    "p = len(masks)\n",
    "\n",
    "# Show the masks\n",
    "\n",
    "M = np.concatenate([np.reshape(masks[i], [dimX,dimY], order='C') for i in range(p)], axis=1)\n",
    "plt.imshow(M, cmap=\"Greys\", interpolation=\"none\")\n",
    "for j in range(p):\n",
    "    plt.text(dimX*j+12, 15, j, fontsize=14, color=\"red\")\n",
    "ax = plt.gca()\n",
    "ax.set_xticks(np.arange(-0.5, dimX*p, dimX))\n",
    "ax.set_yticks(np.arange(-0.5, dimY, dimY))\n",
    "ax.set_xticklabels([])\n",
    "ax.set_yticklabels([])\n",
    "ax.grid(color='m', linestyle='-', linewidth=10)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The masks will be used to sum the \"total amount of writing\" in each of the chosen areas of the images. \n",
    "Each image will be represented by a tuple of values called *features*.\n",
    "The first job is thus to build, for each image, its feature vector. \n",
    "Here, the feature vector is made of $p=8$ boolean features, which indicate whether each zone is \"written\" or not.\n",
    "\n",
    "We must choose a *threshold* for each mask to decide whether the zone is \"written\" or not. \n",
    "We could hand-tune these thresholds; for simplicity, here, we chose that a zone is \"written\" if less than $\\alpha = 45\\%$ of the images in the training set are \"more written\" in that zone.\n",
    "\n",
    "Doing so, we introduce an important intermediate step: for every image we compute a *feature vector* $x$=featureVector(image), and then the goal is to construction a mapping $\\psi:x\\mapsto \\{0,\\dots,d\\}$. The classifier is of the form: $image \\mapsto \\psi\\big(featureVector(image)\\big)$."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [],
   "source": [
    "class FeatureMaker: # abstract class, serves as a template for all feature making classes\n",
    "    def __init__(self):\n",
    "        self.p = dimX*dimY # number of features\n",
    "    \n",
    "    def computeFeatures(self, image):\n",
    "        return(np.array(image)/255) # by default, the feature is the input itself\n",
    "    \n",
    "    def computeAllFeatures(self, dataSet): # stores the features in vector X of the dataSet\n",
    "        dataSet.X = np.array([self.computeFeatures(im) for im in dataSet.data]) \n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Features: \n",
      "[False, False, False, False, False, False, False, False]\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAWgAAAFlCAYAAADGe3ILAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjMuNCwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8QVMy6AAAACXBIWXMAAAsTAAALEwEAmpwYAAAPy0lEQVR4nO3dW6idd5nH8d/T2nphUqlmW0qtE0fKeChMHDZloB4raj1A9aKDvZAOFOqFBQWFERHUiwEZPIwXg1Cx2AEPKOrYizK1VLEjTkujRE0n0wOS0diQpFZpvWlt+5+LrMBuTZq1stfKerLX5wNh7/3u/37X83a13755s9abGmMEgH7OWvYAAByfQAM0JdAATQk0QFMCDdCUQAM09bzT+WA7duwYO3fuPJ0PCdDa/v378/DDD9fxvndaA71z587s3r37dD4kQGvr6+sn/N6mLnFU1ZVVdV9VPVhVH9vMvgB4plMOdFWdneTfkrwjyauTXFNVr57XYACrbjNn0JcleXCM8esxxhNJvpnkqvmMBcBmAn1Rkt9u+PrAZNszVNX1VbW7qnYfOXJkEw8HsFo2E+jj/anjX9x5aYxx4xhjfYyxvra2tomHA1gtmwn0gSQXb/j6pUke2tw4AByzmUDfk+SSqnp5VZ2b5H1JbpnPWACc8uugxxhPVtUNSW5LcnaSm8YY985tMoAVt6k3qowxbk1y65xmAWAD9+IAaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaOp5m/nhqtqf5LEkTyV5coyxPo+hANhkoCfePMZ4eA77AWADlzgAmtpsoEeSH1TVz6rq+nkMBMBRm73EcfkY46GqekmS26vqf8cYd25cMAn39Unyspe9bJMPB7A6NnUGPcZ4aPLxcJLvJbnsOGtuHGOsjzHW19bWNvNwACvllANdVS+oqu3HPk/ytiR75zUYwKrbzCWOC5J8r6qO7efrY4z/nMtUAJx6oMcYv07yt3OcBYANvMwOoCmBBmhKoAGaEmiApgQaoCmBBmhKoAGaEmiApgQaoCmBBmhKoAGaEmiApubxdxKyhTz++OMzrd+1a9fUa++7776Z9j3GmGn95M6KLbzrXe+aeu0NN9ww077f/va3zzoOZyhn0ABNCTRAUwIN0JRAAzQl0ABNCTRAUwIN0JRAAzQl0ABNCTRAU97qzTP8/ve/n2n9/fffP/XaV73qVTPt++qrr55p/SOPPDL12j179sy071ndddddU689cODATPu+4oorZlp/zjnnzLSePpxBAzQl0ABNCTRAUwIN0JRAAzQl0ABNCTRAUwIN0JRAAzQl0ABNCTRAU+7FwTM88cQTM63fvn371Gt//OMfz7TvHTt2zLS+k5/+9KdTr339618/077vvvvumda/7nWvm2k9fTiDBmhKoAGaEmiApgQaoCmBBmhKoAGaEmiApgQaoCmBBmhKoAGaEmiAptyLg2f49re/PdP68847b+q1Z/K9NWY1yz+XWd11110zrXcvjjOXM2iApgQaoKmTBrqqbqqqw1W1d8O2F1XV7VX1wOTj+YsdE2D1THMG/dUkVz5r28eS3DHGuCTJHZOvAZijkwZ6jHFnkkeetfmqJDdPPr85yXvmOxYAp3oN+oIxxsEkmXx8yYkWVtX1VbW7qnYfOXLkFB8OYPUs/A8Jxxg3jjHWxxjra2tri344gC3jVAN9qKouTJLJx8PzGwmA5NQDfUuSayefX5vk+/MZB4BjpnmZ3TeS/HeSv6mqA1V1XZLPJHlrVT2Q5K2TrwGYo5O+1XuMcc0JvvWWOc9CA1/84hdnWn/WWd7rdDwvfOELp1577rnnzrTvH/7whzOt/+hHPzrTevrwXxdAUwIN0JRAAzQl0ABNCTRAUwIN0JRAAzQl0ABNCTRAUwIN0JRAAzR10ntxsFpmvbfGn//856nXPv744zPt+/nPf/5M6zu5+OKLp147633S9+7de/JFbAnOoAGaEmiApgQaoCmBBmhKoAGaEmiApgQaoCmBBmhKoAGaEmiApgQaoCn34mBTDh8+PPXaBx98cKZ9v+Y1r5l1nDb++Mc/Tr32sccem2nf27dvn3EazlTOoAGaEmiApgQaoCmBBmhKoAGaEmiApgQaoCmBBmhKoAGaEmiAprzVGxbgwIEDU6999NFHZ9r35ZdfPus4nKGcQQM0JdAATQk0QFMCDdCUQAM0JdAATQk0QFMCDdCUQAM0JdAATQk0QFPuxcEzXHfddTOt//SnP72gSVbHGGOm9W984xsXNAndOIMGaEqgAZo6aaCr6qaqOlxVezds+1RV/a6q9kx+vXOxYwKsnmnOoL+a5MrjbP/CGGPX5Net8x0LgJMGeoxxZ5JHTsMsAGywmWvQN1TVLyeXQM4/0aKqur6qdlfV7iNHjmzi4QBWy6kG+ktJXpFkV5KDST53ooVjjBvHGOtjjPW1tbVTfDiA1XNKgR5jHBpjPDXGeDrJl5NcNt+xADilQFfVhRu+fG+SvSdaC8CpOek7CavqG0nelGRHVR1I8skkb6qqXUlGkv1JPrC4EQFW00kDPca45jibv7KAWQDYwL04eIZPfOITM61/85vfPPXaV77ylbOOsxKqaqb1L37xixc0Cd14qzdAUwIN0JRAAzQl0ABNCTRAUwIN0JRAAzQl0ABNCTRAUwIN0JRAAzTlXhw8w9lnnz3T+je84Q0LmoQTufLK4/0VoWxFzqABmhJogKYEGqApgQZoSqABmhJogKYEGqApgQZoSqABmhJogKa81RsW4Lbbbpt67Y4dO2ba97Zt22YdhzOUM2iApgQaoCmBBmhKoAGaEmiApgQaoCmBBmhKoAGaEmiApgQaoCmBBmjKvThgAQ4cODD12ksuuWSmfZ933nmzjsMZyhk0QFMCDdCUQAM0JdAATQk0QFMCDdCUQAM0JdAATQk0QFMCDdCUQAM05V4csABVNfXas85ynsTx+TcDoKmTBrqqLq6qH1XVvqq6t6o+NNn+oqq6vaoemHw8f/HjAqyOac6gn0zykTHGq5L8fZIPVtWrk3wsyR1jjEuS3DH5GoA5OWmgxxgHxxg/n3z+WJJ9SS5KclWSmyfLbk7yngXNCLCSZroGXVU7k7w2yd1JLhhjHEyORjzJS07wM9dX1e6q2n3kyJFNjguwOqYOdFVtS/KdJB8eYzw67c+NMW4cY6yPMdbX1tZOZUaAlTRVoKvqnByN89fGGN+dbD5UVRdOvn9hksOLGRFgNU3zKo5K8pUk+8YYn9/wrVuSXDv5/Nok35//eACra5o3qlye5P1JflVVeybbPp7kM0m+VVXXJflNkqsXMiHAijppoMcYP0lyordFvWW+4wBwjLd6w5Lt3bt3pvV/+MMfZlp//vneQ3am8lZvgKYEGqApgQZoSqABmhJogKYEGqApgQZoSqABmhJogKYEGqApgQZoyr04YMkuvfTSmda7t8bqcAYN0JRAAzQl0ABNCTRAUwIN0JRAAzQl0ABNCTRAUwIN0JRAAzQl0ABNuRcHLMAYY+q1Tz/99AIn4UzmDBqgKYEGaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaMpbvWEBqmrqtffcc89M+z506NBM6y+44IKZ1tOHM2iApgQaoCmBBmhKoAGaEmiApgQaoCmBBmhKoAGaEmiApgQaoCmBBmjKvThgyZ588smZ1j/11FMLmoRunEEDNHXSQFfVxVX1o6raV1X3VtWHJts/VVW/q6o9k1/vXPy4AKtjmkscTyb5yBjj51W1PcnPqur2yfe+MMb47OLGA1hdJw30GONgkoOTzx+rqn1JLlr0YACrbqZr0FW1M8lrk9w92XRDVf2yqm6qqvPnPRzAKps60FW1Lcl3knx4jPFoki8leUWSXTl6hv25E/zc9VW1u6p2HzlyZPMTA6yIqQJdVefkaJy/Nsb4bpKMMQ6NMZ4aYzyd5MtJLjvez44xbhxjrI8x1tfW1uY1N8CWN82rOCrJV5LsG2N8fsP2Czcse2+SvfMfD2B1TfMqjsuTvD/Jr6pqz2Tbx5NcU1W7kowk+5N8YAHzAaysaV7F8ZMkx/srim+d/zgAHOOdhABNuRcHLMC73/3uqdf+4he/mGnf27Ztm3UczlDOoAGaEmiApgQaoCmBBmhKoAGaEmiApgQaoCmBBmhKoAGaEmiAprzVGxbgiiuuWMhaVoszaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgKYEGaKrGGKfvwaqOJPm/43xrR5KHT9sgy+M4t55VOVbHuTh/NcZYO943TmugT6Sqdo8x1pc9x6I5zq1nVY7VcS6HSxwATQk0QFNdAn3jsgc4TRzn1rMqx+o4l6DFNWgA/lKXM2gAnmWpga6qK6vqvqp6sKo+tsxZFq2q9lfVr6pqT1XtXvY881JVN1XV4arau2Hbi6rq9qp6YPLx/GXOOA8nOM5PVdXvJs/pnqp65zJnnIequriqflRV+6rq3qr60GT7lnpOn+M4Wz2nS7vEUVVnJ7k/yVuTHEhyT5Jrxhj/s5SBFqyq9idZH2NsqdeSVtUbkvwpyb+PMS6dbPuXJI+MMT4z+R/v+WOMf1rmnJt1guP8VJI/jTE+u8zZ5qmqLkxy4Rjj51W1PcnPkrwnyT9mCz2nz3Gc/5BGz+kyz6AvS/LgGOPXY4wnknwzyVVLnIdTMMa4M8kjz9p8VZKbJ5/fnKP/4p/RTnCcW84Y4+AY4+eTzx9Lsi/JRdliz+lzHGcrywz0RUl+u+HrA2n4D2iORpIfVNXPqur6ZQ+zYBeMMQ4mR/9DSPKSJc+zSDdU1S8nl0DO6N/2P1tV7Uzy2iR3Zws/p886zqTRc7rMQNdxtm3ll5RcPsb4uyTvSPLByW+ZObN9KckrkuxKcjDJ55Y6zRxV1bYk30ny4THGo8ueZ1GOc5ytntNlBvpAkos3fP3SJA8taZaFG2M8NPl4OMn3cvQSz1Z1aHKN79i1vsNLnmchxhiHxhhPjTGeTvLlbJHntKrOydFofW2M8d3J5i33nB7vOLs9p8sM9D1JLqmql1fVuUnel+SWJc6zMFX1gskfRKSqXpDkbUn2PvdPndFuSXLt5PNrk3x/ibMszLFgTbw3W+A5rapK8pUk+8YYn9/wrS31nJ7oOLs9p0t9o8rkJSz/muTsJDeNMf55acMsUFX9dY6eNSfJ85J8fasca1V9I8mbcvQuYIeSfDLJfyT5VpKXJflNkqvHGGf0H7Cd4DjflKO/FR5J9if5wLHrtGeqqnpdkv9K8qskT082fzxHr89umef0OY7zmjR6Tr2TEKAp7yQEaEqgAZoSaICmBBqgKYEGaEqgAZoSaICmBBqgqf8HTJF9JukCpy8AAAAASUVORK5CYII=\n",
      "text/plain": [
       "<Figure size 1440x432 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "class LogicalFeatureMaker(FeatureMaker):    \n",
    "    def __init__(self, masks, thresholds=[]):\n",
    "        self.masks = masks\n",
    "        self.p = len(masks)\n",
    "        self.thresholds = thresholds\n",
    "\n",
    "    def computeFeatures(self, image):\n",
    "        return([self.isWritten(image, self.masks[k], self.thresholds[k]) for k in range(self.p)])\n",
    "                \n",
    "    def isWritten(self, image, mask, threshold): \n",
    "        return(sum(image*mask) > threshold)\n",
    "\n",
    "    def computeThresholds(self, dataSet):\n",
    "        self.trainingSet = dataSet\n",
    "        self.thresholds = [self.computeThreshold(mask) for mask in masks]\n",
    "          \n",
    "    def computeThreshold(self, mask):\n",
    "        alpha = 0.45 # value obtained after some hand tuning...\n",
    "        return(sorted([sum(im*mask) for im in trainingSet.data])[int(alpha*len(trainingSet.data))])\n",
    "\n",
    "logicalFeatureMaker = LogicalFeatureMaker(masks, 8000 * np.array([0.5, 0.7, 0.1, 0.2, 0.7, 1.3, 2, 2]))\n",
    "logicalFeatureMaker.computeThresholds(trainingSet) # overwrite arbitrary features \n",
    "\n",
    "plt.imshow(np.reshape(mnist.data[ii], [dimX, dimY], order='C'), cmap=\"Greys\", interpolation=\"none\")\n",
    "print(\"Features: \")\n",
    "print(logicalFeatureMaker.computeFeatures(mnist.data[ii]))\n",
    "\n",
    "logicalFeatureMaker.computeAllFeatures(trainingSet) # updates trainingSet.X\n",
    "logicalFeatureMaker.computeAllFeatures(testingSet) # updates testingSet.X"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {},
   "outputs": [],
   "source": [
    "class HandMadeClassifier(Classifier):\n",
    "    def __init__(self):\n",
    "        self.featureMaker = logicalFeatureMaker\n",
    "        self.p = logicalFeatureMaker.p\n",
    "        \n",
    "    def fit(self, X, y): # no data fitting!\n",
    "        pass\n",
    "    \n",
    "    def predict(self, X):\n",
    "        return([self.predictOne(x) for x in X])\n",
    "\n",
    "    def predictOne(self, x):\n",
    "        if  x[outside] and not x[center]:\n",
    "            return(0)\n",
    "        elif not x[left] and not x[right] and x[center] and not x[outside]:\n",
    "            return(1)\n",
    "        elif x[top] and x[diag2] and x[bottom] and not x[center]:\n",
    "            return(2)\n",
    "        elif x[top] and x[bottom] and x[center] and x[right]:\n",
    "            return(3)\n",
    "        elif x[diag2] and x[bottom] and not x[top] and not x[outside]:\n",
    "            return(4)\n",
    "        elif x[top] and x[bottom] and x[diag1] and not x[diag2]:\n",
    "            return(5)\n",
    "        elif x[top] and x[center] and x[bottom] and x[diag1] and not x[right]:\n",
    "            return(6)\n",
    "        elif x[top] and x[diag2] and not x[bottom] and not x[left]:\n",
    "            return(7)\n",
    "        elif x[top] and x[center] and x[bottom] and x[left]:\n",
    "            return(8)\n",
    "        elif x[top] and x[center] and x[bottom] and x[right] and not x[left]:\n",
    "            return(9)\n",
    "        else:\n",
    "            # return(10) # 10 means: I don't know\n",
    "            return(random.randint(0, 10)) # random guessing"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "training error: 1\n",
      "testing error: 0.806348\n"
     ]
    }
   ],
   "source": [
    "handMadeClassifier = HandMadeClassifier()\n",
    "\n",
    "trainingError = evaluate(handMadeClassifier, trainingSet.X, trainingSet.target)\n",
    "print(\"training error: %g\" % (trainingError))\n",
    "\n",
    "testingError = evaluate(handMadeClassifier, testingSet.X, testingSet.target)\n",
    "print(\"testing error: %g\" % (testingError))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "* -> Training and testing error are slightly better than random classification (90%), but not by far! They are really not satisfactory. It would require a huge effort to obtain decent performance. \n",
    "\n",
    "* One can observe here, as expected, that training and testing errors coincide statistically.\n",
    "\n",
    "### 3.2 Questions\n",
    "* Try improving the decision rule coded in the function HandMadeClassifier.predictOne\n",
    "* Try adding new, relevant features. What is the best score you can reach?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Nearest-neighbor classifier\n",
    "Distance-based methods consider some notion of *distance* in the feature space, and assume that if to feature vectors are close, then the corresponding targets are likely to be equal.\n",
    "\n",
    "### 7.1 What distance?\n",
    "The first problem, of course, is to provide the feature space with a relevant distance. There are a few default choices, and in particular the Euclidian distance (o $L^2$ distance) in $\\mathbb{R}^p$ :\n",
    "$$d(x,y) = \\sqrt{\\sum_{i=1}^p (x_i-y_i)^2}$$\n",
    "\n",
    "Is this distance really relevant on images? \n",
    "A better alternative would be the *earthmoving distance*, but it is much more costly in computation time and we will see that the $L^2$ distance already performs well.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "If you code it by yourself, the result will be good butthe computation time ridiculously high...\n",
    "\n",
    "The classifiers of scikitlearn are usually well coded, and much more efficient than a few \"quick and dirty\" lines.\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Proportion of misclassified images in testing set: 0.035\n",
      "elapsed: 1.32675 seconds\n"
     ]
    }
   ],
   "source": [
    "import time\n",
    "\n",
    "n_train = 60000\n",
    "I = random.choice(range(n_tot), n_train, replace=False)\n",
    "trainingSet = DataSet(np.array(mnist.data)[I, :], np.array(mnist.target)[I]) \n",
    "\n",
    "# testing set: we may not use them, except at the end to assert the quality of the classifier\n",
    "mask = np.ones(n_tot, dtype=bool) #np.ones_like(a,dtype=bool)\n",
    "mask[I] = False\n",
    "testingSet = DataSet(np.array(mnist.data)[mask, :], np.array(mnist.target)[mask])\n",
    "\n",
    "from sklearn.neighbors import KNeighborsClassifier\n",
    "k = 1\n",
    "nearestNeighborsClassifier2 = KNeighborsClassifier(n_neighbors=k) # L2 distance is the default\n",
    "nearestNeighborsClassifier2.fit(trainingSet.X, trainingSet.target)\n",
    "\n",
    "J = random.choice(range(len(testingSet.target)), 1000, replace=False)\n",
    "\n",
    "start_time = time.time()\n",
    "print(\"Proportion of misclassified images in testing set: %g\" %(evaluate(nearestNeighborsClassifier2, testingSet.X[J, :], testingSet.target[J])))\n",
    "print(\"elapsed: %g seconds\"%(time.time()-start_time))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "\n",
    "### 7.2 k-nearest neighbors\n",
    "The idea of the k-nearest neighbors classifier is very simple: \n",
    "* choose an integer $k$;\n",
    "* if you want to guess which target corresponds to some $x$ in the feature space, then simply\n",
    "  - look at the $k$ points of the training set whose feature vectors are closest to $x$, \n",
    "  - and take the most common label among the targets of these points \n",
    "\n",
    "Nearest neighbor classification with 1, 3 or 5 neighbors seems to reach around 3% errors, but still quite slow! (for example more than 6 minutes for a 10000 images testing set).\n",
    "Note that the fitting seems to get faster when more and more points are given (check!)\n",
    "\n",
    "### 7.3 Exercices:\n",
    "* Try another number of neighbors. Find the best such number\n",
    "* How does the performance improve with the size of the training set?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 7.4 Can we do better? \n",
    "The most obvious possibility is to improve the measure of distance between images: after all, a small translation can lead to large L2 distance but should be considered as harmless in terms of classification.\n",
    "\n",
    "The best possible distance would be an editing distance with no cost on translation nor small rotations. \n",
    "A popular choice recently is the *earthmoving distance*, which measures how much \"black\" much be move to go from one image to the other."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Distribution of the distance to k-th neighbor"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 79,
   "metadata": {},
   "outputs": [],
   "source": [
    "# normalization: all dimensions of x between 0 and 1\n",
    "n = trainingSet.X.shape[0]\n",
    "\n",
    "trainingSet.X = trainingSet.X / 255\n",
    "testingSet.X = testingSet.X / 255\n",
    "\n",
    "M = 100\n",
    "J = random.choice(range(len(testingSet.target)), M, replace=False)\n",
    "k = 1\n",
    "kdist = zeros(M)\n",
    "for i in range(M):\n",
    "    j = J[i]\n",
    "    alldists = apply_along_axis( lambda x: norm(testingSet.X[j,:]-x), axis=1, arr=trainingSet.X)\n",
    "    kdist[i] = sorted(alldists)[k]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 63,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([ 4.,  2.,  4.,  1., 13., 16., 24., 14., 12., 10.]),\n",
       " array([2.04600970e-05, 2.86253409e-05, 3.67905849e-05, 4.49558289e-05,\n",
       "        5.31210729e-05, 6.12863169e-05, 6.94515609e-05, 7.76168048e-05,\n",
       "        8.57820488e-05, 9.39472928e-05, 1.02112537e-04]),\n",
       " <a list of 10 Patch objects>)"
      ]
     },
     "execution_count": 63,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.hist(kdist)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 64,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "estimated dimension of the manifold supporting the data:  [1.27929335]\n"
     ]
    }
   ],
   "source": [
    "# estimated dimension of the manifold supporting the data\n",
    "Dmed = median(kdist)\n",
    "\n",
    "from scipy.optimize import fsolve\n",
    "\n",
    "f = lambda d: d*log(Dmed*sqrt(pi)) - special.loggamma(d)-log(log(2))+log(n)\n",
    "est_dim = fsolve(f, 28*28)\n",
    "\n",
    "# est_dim = log(n) / log(mean(kdist))\n",
    "\n",
    "print(\"estimated dimension of the manifold supporting the data: \", est_dim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 65,
   "metadata": {},
   "outputs": [],
   "source": [
    "fakeDim = 28*28\n",
    "fakeX = rand(trainingSet.X.shape[0], fakeDim)\n",
    "M = 100\n",
    "J = random.choice(range(len(testingSet.target)), M, replace=False)\n",
    "k = 1\n",
    "fakekdist = zeros(M)\n",
    "for i in range(M):\n",
    "    newX = rand(1, fakeDim)\n",
    "    alldists = apply_along_axis( lambda x: norm(newX-x), axis=1, arr=fakeX)\n",
    "    fakekdist[i] = sorted(alldists)[k]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 66,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([ 4.,  6.,  9., 20., 16., 18., 15.,  7.,  3.,  2.]),\n",
       " array([10.28909205, 10.33572082, 10.3823496 , 10.42897837, 10.47560715,\n",
       "        10.52223592, 10.5688647 , 10.61549347, 10.66212225, 10.70875102,\n",
       "        10.7553798 ]),\n",
       " <a list of 10 Patch objects>)"
      ]
     },
     "execution_count": 66,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.hist(fakekdist)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 93,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "estimated dimension if the true dimension is really  50  :  [78.55065441]\n",
      "vb1:  1.730219245836112e-13 u:  1.9218606737057904\n",
      "proba to be larger than median: 0.0\n"
     ]
    }
   ],
   "source": [
    "fakeDmed = median(fakekdist)\n",
    "\n",
    "f = lambda d: d*log(fakeDmed*sqrt(pi)) - special.loggamma(d/2+1)-log(log(2))+log(n)\n",
    "est_dim = fsolve(f, 28*28)\n",
    "\n",
    "# est_dim = log(n) / log(mean(kdist))\n",
    "\n",
    "print(\"estimated dimension if the true dimension is really \", fakeDim, \" : \", est_dim)\n",
    "\n",
    "vb1 = exp((fakeDim/2*log(pi)-special.loggamma(fakeDim/2+1)))\n",
    "u = fakeDmed\n",
    "print(\"vb1: \", vb1, \"u: \", u)\n",
    "print(\"proba to be larger than median:\", exp(-n * (u**fakeDim) * vb1) )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 72,
   "metadata": {},
   "outputs": [],
   "source": [
    "fakeDim = 50\n",
    "fakeX = rand(trainingSet.X.shape[0], fakeDim)\n",
    "M = 100\n",
    "J = random.choice(range(len(testingSet.target)), M, replace=False)\n",
    "k = 1\n",
    "fakekdist = zeros(M)\n",
    "for i in range(M):\n",
    "    newX = rand(1, fakeDim)\n",
    "    alldists = apply_along_axis( lambda x: norm(newX-x), axis=1, arr=fakeX)\n",
    "    fakekdist[i] = sorted(alldists)[k]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([ 1.,  1.,  2.,  3., 20., 22., 25., 20.,  3.,  3.]),\n",
       " array([1.58441277, 1.6408968 , 1.69738083, 1.75386486, 1.81034889,\n",
       "        1.86683292, 1.92331695, 1.97980097, 2.036285  , 2.09276903,\n",
       "        2.14925306]),\n",
       " <a list of 10 Patch objects>)"
      ]
     },
     "execution_count": 37,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.hist(fakekdist)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "5.674426527792628"
      ]
     },
     "execution_count": 42,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sqrt(fakeDim)/trainingSet.X.shape[0]**(1/fakeDim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "5779614360472.331"
      ]
     },
     "execution_count": 56,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from scipy import special\n",
    "1/exp((fakeDim/2)*log(pi)-special.loggamma(fakeDim/2+1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "1.5511210043330984e+25"
      ]
     },
     "execution_count": 52,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "special.gamma(fakeDim/2+1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 60,
   "metadata": {},
   "outputs": [
    {
     "ename": "NameError",
     "evalue": "name 'func' is not defined",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mNameError\u001b[0m                                 Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-60-f4f44d8c2abc>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m()\u001b[0m\n\u001b[1;32m      2\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      3\u001b[0m \u001b[0mf\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;32mlambda\u001b[0m \u001b[0md\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0md\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0mlog\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mDmed\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0msqrt\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mpi\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;34m-\u001b[0m \u001b[0mspecial\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mloggamma\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0md\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m-\u001b[0m\u001b[0mlog\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mlog\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m2\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m+\u001b[0m\u001b[0mlog\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mn\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 4\u001b[0;31m \u001b[0mfsolve\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mfunc\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;36m28\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0;36m28\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
      "\u001b[0;31mNameError\u001b[0m: name 'func' is not defined"
     ]
    }
   ],
   "source": [
    "from scipy.optimize import fsolve\n",
    "\n",
    "f = lambda d: d*log(Dmed*sqrt(pi)) - special.loggamma(d)-log(log(2))+log(n)\n",
    "fsolve(f, 28*28)"
   ]
  },
  {
   "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": "3.8.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}
