stuffForFriday_Tobias.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Contents of this part:\n",
    "* loading and visualizing the MNIST dataset (training and test subsets)\n",
    "* build and evaluate a simple CNN model\n",
    "    * Experiments:\n",
    "        * changing network structure/adding more layers: conv, pool, dropout, or fully connected\n",
    "        * changing meta parameters\n",
    "        * reduce the size of the training set to see how many samples are required for a good model\n",
    "* Beyond accuracy: some additional metrics of predictive performance \n",
    "* visualize network architecture (does not work yet; requires graphviz and Pydot libraries)\n",
    "* Inspecting a learned network: visualizing learned network weights, visualizing network activations, ...\n",
    "* What are the important features in a given input image: systematic occlusion experiment"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now, let's train an extremely simple CNN:\n",
    "First, we have to load the keras libraries and prepare the MNIST dataset (in this version it is already split into a training and a test set)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "from keras.models import Sequential\n",
    "from keras.layers import Dense\n",
    "from keras.datasets import mnist\n",
    "from keras.layers import Dropout\n",
    "from keras.layers import Flatten\n",
    "from keras.layers.convolutional import Convolution2D\n",
    "from keras.layers.convolutional import MaxPooling2D\n",
    "from keras.utils import np_utils\n",
    "from keras import backend as K\n",
    "K.set_image_dim_ordering('th')\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline \n",
    "# todo: replace link to MNIST dataset:\n",
    "(X_train, y_train), (X_test, y_test) = mnist.load_data(\"mnist.pkl.gz\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "Let's look at some of the MNIST samples:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAT8AAAD8CAYAAAABraMFAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n5kdETmLzIyIn/R+KxB3xQTzcVwAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x1269b8d50>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.subplot(221)\n",
    "plt.imshow(X_train[0], cmap=plt.get_cmap('gray'))\n",
    "plt.subplot(222)\n",
    "plt.imshow(X_train[1], cmap=plt.get_cmap('gray'))\n",
    "plt.subplot(223)\n",
    "plt.imshow(X_train[2], cmap=plt.get_cmap('gray'))\n",
    "plt.subplot(224)\n",
    "plt.imshow(X_train[3], cmap=plt.get_cmap('gray'))\n",
    "# show the plot\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# take one channel only:\n",
    "X_train = X_train.reshape(X_train.shape[0], 1, 28, 28).astype('float32')\n",
    "X_test = X_test.reshape(X_test.shape[0], 1, 28, 28).astype('float32')\n",
    "# normalize inputs from 0-255 to 0-1\n",
    "X_train = X_train / 255\n",
    "X_test = X_test / 255\n",
    "# one hot encode outputs\n",
    "y_train = np_utils.to_categorical(y_train)\n",
    "y_test = np_utils.to_categorical(y_test)\n",
    "num_classes = y_test.shape[1]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Time to define and train our first (not-so)-deep neural network!\n",
    "### Suggestions for experimenting:\n",
    "You can experiment a bit with the number of epochs, batch size, or even with the *number of training samples* (can we still train a good model with, say 10000 instead of 60000 training samples?)\n",
    "Try what happens with predictive performance if you add a dropout layer, additional Conv/Pool layers, and/or another fully connected layer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def model1():\n",
    "    model = Sequential()\n",
    "    model.add(Convolution2D(32, 5, 5, border_mode='valid', input_shape=(1, 28, 28), activation='relu'))\n",
    "    model.add(MaxPooling2D(pool_size=(2, 2)))\n",
    "    # model.add(Dropout(0.2))\n",
    "    model.add(Flatten())\n",
    "    model.add(Dense(128, activation='relu'))\n",
    "    model.add(Dense(num_classes, activation='softmax'))\n",
    "    # Compile model\n",
    "    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n",
    "    return model   "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/anaconda/lib/python2.7/site-packages/ipykernel_launcher.py:3: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(32, (5, 5), padding=\"valid\", activation=\"relu\", input_shape=(1, 28, 28...)`\n",
      "  This is separate from the ipykernel package so we can avoid doing imports until\n",
      "/anaconda/lib/python2.7/site-packages/keras/models.py:848: UserWarning: The `nb_epoch` argument in `fit` has been renamed `epochs`.\n",
      "  warnings.warn('The `nb_epoch` argument in `fit` '\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 60000 samples, validate on 10000 samples\n",
      "Epoch 1/10\n",
      "93s - loss: 0.2247 - acc: 0.9359 - val_loss: 0.0891 - val_acc: 0.9721\n",
      "Epoch 2/10\n",
      "95s - loss: 0.0649 - acc: 0.9810 - val_loss: 0.0482 - val_acc: 0.9839\n",
      "Epoch 3/10\n",
      "96s - loss: 0.0453 - acc: 0.9861 - val_loss: 0.0462 - val_acc: 0.9844\n",
      "Epoch 4/10\n",
      "95s - loss: 0.0338 - acc: 0.9894 - val_loss: 0.0354 - val_acc: 0.9884\n",
      "Epoch 5/10\n",
      "95s - loss: 0.0266 - acc: 0.9919 - val_loss: 0.0431 - val_acc: 0.9860\n",
      "Epoch 6/10\n",
      "95s - loss: 0.0220 - acc: 0.9931 - val_loss: 0.0412 - val_acc: 0.9864\n",
      "Epoch 7/10\n",
      "96s - loss: 0.0165 - acc: 0.9947 - val_loss: 0.0367 - val_acc: 0.9882\n",
      "Epoch 8/10\n",
      "96s - loss: 0.0141 - acc: 0.9954 - val_loss: 0.0414 - val_acc: 0.9872\n",
      "Epoch 9/10\n",
      "96s - loss: 0.0113 - acc: 0.9964 - val_loss: 0.0333 - val_acc: 0.9894\n",
      "Epoch 10/10\n",
      "95s - loss: 0.0082 - acc: 0.9977 - val_loss: 0.0385 - val_acc: 0.9885\n",
      "Baseline Error: 1.15%\n"
     ]
    }
   ],
   "source": [
    "# build the model\n",
    "model = model1()\n",
    "# Fit the model\n",
    "model.fit(X_train, y_train, validation_data=(X_test, y_test), nb_epoch=10, batch_size=200, verbose=2)\n",
    "# Final evaluation of the model\n",
    "scores = model.evaluate(X_test, y_test, verbose=0)\n",
    "print(\"Baseline Error: %.2f%%\" % (100-scores[1]*100))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "It's always good to inspect some additional metrics of predictive performance. If you're not familiar with them you could check their definition:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "             precision    recall  f1-score   support\n",
      "\n",
      "          0       0.99      0.99      0.99       980\n",
      "          1       0.99      1.00      0.99      1135\n",
      "          2       0.98      1.00      0.99      1032\n",
      "          3       0.97      1.00      0.99      1010\n",
      "          4       0.99      0.99      0.99       982\n",
      "          5       0.99      0.98      0.99       892\n",
      "          6       1.00      0.99      0.99       958\n",
      "          7       0.99      0.98      0.99      1028\n",
      "          8       0.99      0.98      0.99       974\n",
      "          9       0.99      0.98      0.98      1009\n",
      "\n",
      "avg / total       0.99      0.99      0.99     10000\n",
      "\n"
     ]
    }
   ],
   "source": [
    "Y_pred = model.predict(X_test)\n",
    "# Convert one-hot to index\n",
    "y_pred = np.argmax(Y_pred, axis=1)\n",
    "yt = np.argmax(y_test, axis=1)\n",
    "from sklearn.metrics import classification_report\n",
    "print(classification_report(yt, y_pred))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We can visualize the architecture of the neural network, but currently the graphviz back-end and the pydot wrapper is not available in our environment. Anyways, this is how it would work: "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "ename": "ImportError",
     "evalue": "No module named visualize_util",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mImportError\u001b[0m                               Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-10-be8c52e570b4>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m()\u001b[0m\n\u001b[0;32m----> 1\u001b[0;31m \u001b[0;32mfrom\u001b[0m \u001b[0mkeras\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mutils\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mvisualize_util\u001b[0m \u001b[0;32mimport\u001b[0m \u001b[0mplot\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m      2\u001b[0m \u001b[0mplot\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmodel\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;31m# , to_file='model.png')\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;31mImportError\u001b[0m: No module named visualize_util"
     ]
    }
   ],
   "source": [
    "from keras.utils.visualize_util import plot\n",
    "plot(model) # , to_file='model.png')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Visualize weights and convolutions:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import numpy.ma as ma\n",
    "from mpl_toolkits.axes_grid1 import make_axes_locatable\n",
    "\n",
    "def nice_imshow(ax, data, vmin=None, vmax=None, cmap=None):\n",
    "    \"\"\"Wrapper around pl.imshow\"\"\"\n",
    "    if cmap is None:\n",
    "        cmap = cm.jet\n",
    "    if vmin is None:\n",
    "        vmin = data.min()\n",
    "    if vmax is None:\n",
    "        vmax = data.max()\n",
    "    divider = make_axes_locatable(ax)\n",
    "    cax = divider.append_axes(\"right\", size=\"5%\", pad=0.05)\n",
    "    im = ax.imshow(data, vmin=vmin, vmax=vmax, interpolation='nearest', cmap=cmap)\n",
    "    plt.colorbar(im, cax=cax)\n",
    "def make_mosaic(imgs, nrows, ncols, border=1):\n",
    "    \"\"\"\n",
    "    Given a set of images with all the same shape, makes a\n",
    "    mosaic with nrows and ncols\n",
    "    \"\"\"\n",
    "    nimgs = imgs.shape[0]\n",
    "    imshape = imgs.shape[1:]\n",
    "    \n",
    "    mosaic = ma.masked_all((nrows * imshape[0] + (nrows - 1) * border,\n",
    "                            ncols * imshape[1] + (ncols - 1) * border),\n",
    "                            dtype=np.float32)\n",
    "    \n",
    "    paddedh = imshape[0] + border\n",
    "    paddedw = imshape[1] + border\n",
    "    for i in xrange(nimgs):\n",
    "        row = int(np.floor(i / ncols))\n",
    "        col = i % ncols\n",
    "        \n",
    "        mosaic[row * paddedh:row * paddedh + imshape[0],\n",
    "               col * paddedw:col * paddedw + imshape[1]] = imgs[i]\n",
    "    return mosaic\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "('W shape : ', (32, 5, 5))\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAS8AAAD+CAYAAABry4PmAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n4+U4TiJx4+U4TiL5/yKf79J4YCSVAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x2aab2812e190>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Visualize weights\n",
    "W = model.layers[0].W.get_value(borrow=True)\n",
    "W = np.squeeze(W)\n",
    "print(\"W shape : \", W.shape)\n",
    "\n",
    "nice_imshow(plt.gca(), make_mosaic(W, 6, 6), cmap=plt.cm.binary)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "It's very useful to be able to save or load module weights, for later model inspection, for continuing to train, or for transfer learning (transferring a learned model to a different dataset).\n",
    "Question: What is the difference between model.save and model.save_weights?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from keras.models import load_model\n",
    "\n",
    "model.save('my_model.h5')  # creates a HDF5 file 'my_model.h5'\n",
    "del model  # deletes the existing model\n",
    "\n",
    "# returns a compiled model\n",
    "# identical to the previous one\n",
    "model = load_model('my_model.h5')\n",
    "model.save_weights('my_model_weights.h5')\n",
    "model.load_weights('my_model_weights.h5')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "It may also be instructive to look at the activations in a specific network layer, given a specific input sample. If you are interested in how this is done you can google \"visualizing network activations keras\"...."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## What is the classifier \"looking at\" in a given image? Systematic occlusion experiment: "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "So what are the key regions in the following digit for the classifier?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x129130ad0>"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAADXVJREFUeJzt3WuIXfW5x/HfT08KYgNeQsPEBqeKHGgSSMIgQmOM9lg8\noRBLQOqLJgelqVjLKfbF8UI4wbyR2otFpJDQ2KjVVtKWRKzxMhyUQi2JMVUntjGnpDRhTFKs1Hgh\nNXnOi1npmcbZ/72zb2tPnu8Hhuy9nnV5XOaXtfZea9bfESEA+ZxVdwMA6kH4gaQIP5AU4QeSIvxA\nUoQfSIrwA0kRfiApwg8k9S/93JhtbicEeiwi3Mp8HR35bV9n+w+299m+o5N1Aegvt3tvv+2zJe2V\ndK2kA5J2SLoxIvYUluHID/RYP478l0vaFxF/jIhjkn4qaUUH6wPQR52E/yJJf570/kA17Z/YXmN7\np+2dHWwLQJf1/Au/iNggaYPEaT8wSDo58h+UNHfS+09X0wBMA52Ef4eky2x/xvYnJH1Z0rbutAWg\n19o+7Y+Ij2zfJukZSWdL2hQRY13rDEBPtX2pr62N8Zkf6Lm+3OQDYPoi/EBShB9IivADSRF+ICnC\nDyRF+IGkCD+QFOEHkiL8QFKEH0iK8ANJEX4gKcIPJEX4gaQIP5AU4QeSIvxAUoQfSIrwA0kRfiAp\nwg8kRfiBpAg/kBThB5Ii/EBShB9IivADSRF+IKm2h+iWJNv7Jb0r6bikjyJipBtNAei9jsJfuToi\n/tKF9QDoI077gaQ6DX9Iet72y7bXdKMhAP3R6Wn/kog4aPtTkp6z/fuIeHHyDNU/CvzDAAwYR0R3\nVmSvk3Q0Ir5TmKc7GwPQUES4lfnaPu23fa7tmSdfS/qCpNfbXR+A/urktH+2pF/aPrmexyJie1e6\nAtBzXTvtb2ljnPYDPdfz034A0xvhB5Ii/EBShB9IivADSRF+IKlu/FYfBtjChQuL9fXr1xfry5cv\nL9bPOqt8/Dhx4kTD2pYtW4rL3n333cX6+Ph4sX711Vc3rI2OjhaX/eCDD4r1MwFHfiApwg8kRfiB\npAg/kBThB5Ii/EBShB9Iiuv808CMGTOK9auuuqph7aGHHiouOzQ0VKw3+5Xv0nX8ZsuvXLmyuGyz\na+1z584t1pctW9awtnr16uKyjz76aLF+JuDIDyRF+IGkCD+QFOEHkiL8QFKEH0iK8ANJcZ1/Gli8\neHGxvn17+8MlNPud+Ntuu61Yf//999ve9sUXX1ysv/fee8X6Aw88UKwfO3asYa3Zf3cGHPmBpAg/\nkBThB5Ii/EBShB9IivADSRF+IKmm1/ltb5L0RUmHI2J+Ne0CST+TNCxpv6QbIuKvvWvzzDZv3rxi\nfdu2bW2vu9nz6e+8885ifdeuXW1vu5k5c+YU61u3bi3WzzvvvGL9vvvua1hrtl8yaOXI/2NJ150y\n7Q5JoxFxmaTR6j2AaaRp+CPiRUlvnzJ5haTN1evNkq7vcl8Aeqzdz/yzI+Lk/ZFvSZrdpX4A9EnH\n9/ZHRNhu+KA222skrel0OwC6q90j/yHbQ5JU/Xm40YwRsSEiRiJipM1tAeiBdsO/TdLJx5+ullT+\nWhbAwGkaftuPS/qNpH+1fcD2zZLulXSt7Tcl/Vv1HsA00vQzf0Tc2KD0+S73ktbatWuL9VmzZhXr\nTz31VMPa7bffXlx23759xXovzZ8/v1hftGhRR+vv5DkHGXCHH5AU4QeSIvxAUoQfSIrwA0kRfiAp\nNxuCuasbK9wGfCbbuHFjsX7TTTcV680eYX3FFVc0rO3Zs6e4bK+Vhhd/9tlni8suXbq0WH/hhReK\n9WuuuaZYP1NFhFuZjyM/kBThB5Ii/EBShB9IivADSRF+ICnCDyTFEN19MDJSfohRs3stjh49WqzX\neS2/dB1fktavX9+wduWVVxaXbbZf7rnnnmIdZRz5gaQIP5AU4QeSIvxAUoQfSIrwA0kRfiAprvOj\naHh4uFi/9dZbi/Vmjw4vGR8fL9Z3797d9rrBkR9Ii/ADSRF+ICnCDyRF+IGkCD+QFOEHkmp6nd/2\nJklflHQ4IuZX09ZJ+qqkI9Vsd0XEr3rV5HTX7PftFyxYUKxfeOGFxforr7xy2j21qtnw4HPmzCnW\nOxkXYnR0tFh/55132l43Wjvy/1jSdVNM/35ELKx+CD4wzTQNf0S8KOntPvQCoI86+cz/Dduv2t5k\n+/yudQSgL9oN/w8lXSJpoaRxSd9tNKPtNbZ32t7Z5rYA9EBb4Y+IQxFxPCJOSNoo6fLCvBsiYiQi\nyk+xBNBXbYXf9tCkt1+S9Hp32gHQL61c6ntc0jJJs2wfkPTfkpbZXigpJO2X9LUe9gigB9zJddjT\n3pjdv40NkHPOOadYf+KJJ4r15cuXF+v9/H94qhUrVhTrq1atalhbuXJlcdklS5YU6y+99FKxnlVE\nuJX5uMMPSIrwA0kRfiApwg8kRfiBpAg/kBSX+qaBZcuWFevNhgAvGRsbK9affvrpYv3BBx8s1m+5\n5ZaGtb179xaXXbp0abF+5MiRYj0rLvUBKCL8QFKEH0iK8ANJEX4gKcIPJEX4gaS4zo+OHD9+vFgv\n/f167LHHisuWfh0YjXGdH0AR4QeSIvxAUoQfSIrwA0kRfiApwg8k1fS5/chteHi4o+WPHj3asHb/\n/fd3tG50hiM/kBThB5Ii/EBShB9IivADSRF+ICnCDyTV9Dq/7bmSHpY0W1JI2hARP7B9gaSfSRqW\ntF/SDRHx1961ijqsXbu2o+WffPLJhrVdu3Z1tG50ppUj/0eSvhURn5V0haSv2/6spDskjUbEZZJG\nq/cApomm4Y+I8YjYVb1+V9Ibki6StELS5mq2zZKu71WTALrvtD7z2x6WtEjSbyXNjojxqvSWJj4W\nAJgmWr633/YnJf1c0jcj4m/2/z8mLCKi0fP5bK+RtKbTRgF0V0tHftszNBH8n0TEL6rJh2wPVfUh\nSYenWjYiNkTESES0P5okgK5rGn5PHOJ/JOmNiPjepNI2Saur16slbe1+ewB6pZXT/s9J+oqk12zv\nrqbdJeleSU/YvlnSnyTd0JsW0Uvz5s0r1leuXNnR+p955pmOlkfvNA1/RPxaUqPngH++u+0A6Bfu\n8AOSIvxAUoQfSIrwA0kRfiApwg8kxaO7k1u8eHGxPnPmzGK92RDvH3744Wn3hP7gyA8kRfiBpAg/\nkBThB5Ii/EBShB9IivADSXGdP7lZs2YV682u44+NjRXrW7ZsOe2e0B8c+YGkCD+QFOEHkiL8QFKE\nH0iK8ANJEX4gKa7zJ7dq1aqOln/kkUe61An6jSM/kBThB5Ii/EBShB9IivADSRF+ICnCDyTV9Dq/\n7bmSHpY0W1JI2hARP7C9TtJXJR2pZr0rIn7Vq0bRG3v27CnWFyxY0KdO0G+t3OTzkaRvRcQu2zMl\nvWz7uar2/Yj4Tu/aA9ArTcMfEeOSxqvX79p+Q9JFvW4MQG+d1md+28OSFkn6bTXpG7Zftb3J9vkN\nlllje6ftnR11CqCrWg6/7U9K+rmkb0bE3yT9UNIlkhZq4szgu1MtFxEbImIkIka60C+ALmkp/LZn\naCL4P4mIX0hSRByKiOMRcULSRkmX965NAN3WNPy2LelHkt6IiO9Nmj40abYvSXq9++0B6JVWvu3/\nnKSvSHrN9u5q2l2SbrS9UBOX//ZL+lpPOkRPbd++vVi/9NJLi/UdO3Z0sx30USvf9v9akqcocU0f\nmMa4ww9IivADSRF+ICnCDyRF+IGkCD+QlJsNwdzVjdn92xiQVERMdWn+YzjyA0kRfiApwg8kRfiB\npAg/kBThB5Ii/EBS/R6i+y+S/jTp/axq2iAa1N4GtS+J3trVzd4ubnXGvt7k87GN2zsH9dl+g9rb\noPYl0Vu76uqN034gKcIPJFV3+DfUvP2SQe1tUPuS6K1dtfRW62d+APWp+8gPoCa1hN/2dbb/YHuf\n7Tvq6KER2/ttv2Z7d91DjFXDoB22/fqkaRfYfs72m9WfUw6TVlNv62wfrPbdbtvLa+ptru3/sb3H\n9pjt/6ym17rvCn3Vst/6ftpv+2xJeyVdK+mApB2SboyI8ljRfWJ7v6SRiKj9mrDtpZKOSno4IuZX\n074t6e2IuLf6h/P8iPivAeltnaSjdY/cXA0oMzR5ZGlJ10v6D9W47wp93aAa9lsdR/7LJe2LiD9G\nxDFJP5W0ooY+Bl5EvCjp7VMmr5C0uXq9WRN/efquQW8DISLGI2JX9fpdSSdHlq513xX6qkUd4b9I\n0p8nvT+gwRryOyQ9b/tl22vqbmYKs6th0yXpLUmz62xmCk1Hbu6nU0aWHph9186I193GF34ftyQi\nFkr6d0lfr05vB1JMfGYbpMs1LY3c3C9TjCz9D3Xuu3ZHvO62OsJ/UNLcSe8/XU0bCBFxsPrzsKRf\navBGHz50cpDU6s/DNffzD4M0cvNUI0trAPbdII14XUf4d0i6zPZnbH9C0pclbauhj4+xfW71RYxs\nnyvpCxq80Ye3SVpdvV4taWuNvfyTQRm5udHI0qp53w3ciNcR0fcfScs18Y3//0q6u44eGvR1iaTf\nVT9jdfcm6XFNnAb+XRPfjdws6UJJo5LelPS8pAsGqLdHJL0m6VVNBG2opt6WaOKU/lVJu6uf5XXv\nu0Jftew37vADkuILPyApwg8kRfiBpAg/kBThB5Ii/EBShB9IivADSf0fdMlH+mrHFx0AAAAASUVO\nRK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x1289649d0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.imshow(X_test[12,0,:,:], cmap=plt.get_cmap('gray'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[  1.09467305e-11,   1.67104715e-11,   4.28287716e-09,\n",
       "          6.12939175e-06,   3.14096997e-06,   2.57675445e-08,\n",
       "          7.44461404e-13,   1.01103907e-07,   4.69509899e-07,\n",
       "          9.99990106e-01]], dtype=float32)"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.predict(X_test[12:13,:,:,:])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's take a part away from that digit...."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import copy\n",
    "xx = copy.copy(X_test[12:13,:,:,:])\n",
    "xx[0,0,5:17,5:16] = 0"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x129232cd0>"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nKcIPJEX4gaQIP5DU/wBuJfyfMK+4JQAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x1291556d0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.imshow(xx[0,0,:,:], cmap=plt.get_cmap('gray'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now the classifier thinks it's most likely a 7, as we can see when predicting again (question: what do you think would happen if we had \"European\" 1s in the dataset, too?!):"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[  2.21284972e-06,   1.91220941e-04,   4.62521275e-05,\n",
       "          1.39423264e-06,   3.27059533e-04,   3.48505793e-08,\n",
       "          2.39841436e-09,   9.98668909e-01,   2.51685597e-05,\n",
       "          7.37696479e-04]], dtype=float32)"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.predict(xx[0:1,:,:,:])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's try to do the occlusion systematically, sliding a black square across the image:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "def occlude(X, size=10):\n",
    "    Y = np.zeros(((X.shape[1]-size)*(X.shape[0]-size), 1, 28, 28))\n",
    "    for i in xrange(0, X.shape[0]-size):\n",
    "        for j in xrange(0, X.shape[1]-size):\n",
    "            XX = X.copy()\n",
    "            XX[i:(i+size), j:(j+size)] = 0\n",
    "            Y[j+i*(X.shape[1]-size),0,:,:] = XX\n",
    "    return Y"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "xx = copy.copy(X_test[12:13,:,:,:])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "Y = occlude(xx[0,0,:,:])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We now have 324 partially occluded images generated from our original '9' digit. Have a look at some of them if you like:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(324, 1, 28, 28)\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x127d82850>"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x12926e690>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "print(Y.shape)\n",
    "plt.imshow(Y[158,0,:,:], cmap=plt.get_cmap('gray'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now, let's make the predictions for all of them using our classifier:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "pp = model.predict(Y)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[  1.09467305e-11,   1.67104715e-11,   4.28287716e-09,\n",
       "          6.12939175e-06,   3.14096997e-06,   2.57675445e-08,\n",
       "          7.44461404e-13,   1.01103907e-07,   4.69509899e-07,\n",
       "          9.99990106e-01]], dtype=float32)"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pred = model.predict(X_test[12:13,:,:,:])\n",
    "pred"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [],
   "source": [
    "myDiff = ([pred[0,9]]*len(pp[:,9])) -pp[:,9]\n",
    "diffMatrix = np.reshape(myDiff, (18,18), order='C')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x127e6fc10>"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x127d9a790>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "DD = np.zeros((28,28))\n",
    "DD[5:23,5:23] = diffMatrix\n",
    "plt.imshow(DD)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "....so in this case, the importance map is not too surprising or enlightening. If you want you can experiment with the size of the sliding black square. We used 10x10 which is quite big for the 28x28 thumbnails.\n",
    "\n",
    "You could also experiment with some other digits from the MNIST test dataset...."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 2",
   "language": "python",
   "name": "python2"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.13"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}