Transfer Learning & Bottlenecking.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "# Convolutional Deep Neural Networks in Keras with TensorFlow Backend\n",
    "---\n",
    "### Galen Ballew, February 2017"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "The purpose of this project is research and learning in *practical* implementation of CNN/DNN as one might expect in industry. For most problems that are considered understood as far as the network architecture needed to classify correctly, there already exists a network that does the job. Furthermore, these networks have been trained and their weights and biases are available. Importing networks and their parameters and modifying it to suit your current problem is called **transfer learning** because you are transferring the original networks ability to detect features, like edges for example, to a new network with a different set of output classes. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "In this notebook, you will find two things:  \n",
    "\n",
    "   1) An entire convolutional neural network written in Keras using TensorFlow backend. This network is a direct copy of the archtitecture used in my network for traffic sign classification. However, the original network was written directly in TensorFlow. You can view that repository [here.](https://github.com/galenballew/Traffic-Sign-CNN/blob/master/Traffic_Sign_Classifier.ipynb) The network itself is inside of the ** LeNet(x) ** function definition. Keras is a Theano and Tensorflow API that greatly simplfies writing of neural networks. Specifically, Keras increases ***readability*** and removes the need for explicity calculation of input/output dimensions between layers.\n",
    "    \n",
    "   2) Three case studies in transfer learning. All networks in this notebook (including my handwritten network) are trained on the [CIFAR10]() data set. Below, you will find three examples of using transfer learning to repurpose ResNet50, VGG16, and Inceptionv3 (AKA GoogLeNet) for the CIFAR10 data set. All of these networks were originally built and trained for the [ImageNet](http://image-net.org/) data set and competition. They are all winners in various years and tasks. By making a comparison of all four different networks, I hope to produce some understanding of how to use the knowledge, time, and effort that has already gone into deep learning in a practical, enterprise-facing manner."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "---\n",
    "## Rewriting in Keras\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 0 ns, sys: 0 ns, total: 0 ns\n",
      "Wall time: 20 µs\n",
      "Modules loaded.\n"
     ]
    }
   ],
   "source": [
    "import pickle\n",
    "import numpy as np\n",
    "import math\n",
    "import time\n",
    "%time\n",
    "\n",
    "# Fix error with TF and Keras\n",
    "import tensorflow as tf\n",
    "tf.python.control_flow_ops = tf\n",
    "\n",
    "print('Modules loaded.')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "[Documentation](https://keras.io/datasets/) for data sets included in Keras."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "from keras.datasets import cifar10\n",
    "\n",
    "(X_train, y_train), (X_test, y_test) = cifar10.load_data()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# Shuffle the data\n",
    "from sklearn.utils import shuffle\n",
    "X_train, y_train = shuffle(X_train, y_train) "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# Normalize the data features (rbg pixel values)\n",
    "def normalize_grayscale(image_data):\n",
    "    a = -0.5\n",
    "    b = 0.5\n",
    "    grayscale_min = 0\n",
    "    grayscale_max = 255\n",
    "    return a + ( ( (image_data - grayscale_min)*(b - a) )/( grayscale_max - grayscale_min ) )\n",
    "\n",
    "#X_normalized = normalize_grayscale(X_train)\n",
    "X_normalized = X_train / 255 - 0.5"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# One Hot encode the labels \n",
    "from sklearn.preprocessing import LabelBinarizer\n",
    "label_binarizer = LabelBinarizer()\n",
    "y_one_hot = label_binarizer.fit_transform(y_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "from keras.models import Sequential\n",
    "from keras.layers.core import Dense, Activation, Flatten, Dropout\n",
    "from keras.layers.convolutional import Convolution2D\n",
    "from keras.layers.pooling import MaxPooling2D\n",
    "from keras.regularizers import l2, activity_l2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "model = Sequential()\n",
    "\n",
    "#layer 1 - conv2d -> ReLU -> maxPool\n",
    "model.add(Convolution2D(32, 5, 5, input_shape=X_train.shape[1:]))\n",
    "model.add(Activation('relu'))\n",
    "model.add(MaxPooling2D((2, 2)))\n",
    "\n",
    "#layer 2 - conv2d -> ReLU -> maxPool\n",
    "model.add(Convolution2D(64, 5, 5))\n",
    "model.add(Activation('relu'))\n",
    "model.add(MaxPooling2D((2, 2)))\n",
    "\n",
    "model.add(Flatten())\n",
    "\n",
    "#layer 3 - fully connected -> relu -> dropout\n",
    "model.add(Dense(1024, activation='relu'))\n",
    "model.add(Dropout(0.5))\n",
    "\n",
    "#layer 4 - fully connected -> relu -> dropout\n",
    "model.add(Dense(512, activation='relu'))\n",
    "model.add(Dropout(0.5))\n",
    "\n",
    "#layer 5 - fully connexted -> softmax\n",
    "# TODO: L2 regularization applied to output correctly??\n",
    "model.add(Dense(10, activity_regularizer=activity_l2(1e-6)))\n",
    "model.add(Activation('softmax'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "Something to note here -- In my original model, regularization was applyed in the loss function, which was passed to the training operation (i.e. run this model and try to minimize this loss function). Keras introduces options for regularization that I am unfamiliar with. Specifically, you do not explicitly write a loss function. Instead, you can apply regularization on a layer by layer basis (*W_regularizer / b_regularizer*), or to the model's output (*activity_regularizer*). See more [here.](https://keras.io/regularizers/) \n",
    "\n",
    "Because I am training the network on a different data set, I can't be sure if I implemented L2 (Ridge) regularization in the same way as the original network. More investiation needs to be done in the source code of the API. I could not find any explanations via Google/StackOverflow/etc. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true,
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "# Compile and train the model\n",
    "# ~20-30 seconds per epoch on GPU\n",
    "model.compile('adam', 'categorical_crossentropy', ['accuracy'])\n",
    "%time history_10 = model.fit(X_normalized, y_one_hot, nb_epoch=10, validation_split=0.2, verbose=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "#save the weights\n",
    "model.save_weights('keras_10.h5')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "Check the validation accuracy for the last training epoch:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true,
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'acc': [0.4491, 0.59437499999999999, 0.64805000000000001, 0.68559999999999999, 0.71092500000000003, 0.73360000000000003, 0.75765000000000005, 0.77975000000000005, 0.793825, 0.81394999999999995], 'val_loss': [1.2008865358352661, 1.027778117275238, 0.96171257305145263, 0.91027058744430545, 0.89996965465545653, 0.86842247743606571, 0.86806706199646, 0.84969262714385985, 0.87266609010696417, 0.89330739226341249], 'loss': [1.512114985704422, 1.1559825409889222, 1.0104244817256927, 0.90804033093452452, 0.83008897619247435, 0.76539823830127718, 0.70053902325630191, 0.63937823104858393, 0.58790913615226748, 0.54148658394813542], 'val_acc': [0.56699999999999995, 0.64170000000000005, 0.66710000000000003, 0.68120000000000003, 0.69510000000000005, 0.69969999999999999, 0.70250000000000001, 0.71640000000000004, 0.71279999999999999, 0.71479999999999999]}\n"
     ]
    }
   ],
   "source": [
    "print(history_10.history)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# Preprocess data & one-hot encode the labels\n",
    "X_normalized_test = normalize_grayscale(X_test)\n",
    "y_one_hot_test = label_binarizer.fit_transform(y_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true,
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "10000/10000 [==============================] - 1s     \b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\n",
      "loss: 0.8927447565078735\n",
      "acc: 0.7118\n"
     ]
    }
   ],
   "source": [
    "# Evaluate model on test data\n",
    "metrics = model.evaluate(X_normalized_test, y_one_hot_test)\n",
    "for metric_i in range(len(model.metrics_names)):\n",
    "    metric_name = model.metrics_names[metric_i]\n",
    "    metric_value = metrics[metric_i]\n",
    "    print('{}: {}'.format(metric_name, metric_value))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "We can see here that the validation and test accuracy are essentially the same and have a very large error. This means the model is currently in a state of **high bias** and possibly requires more complexity. To resolve this, I will increase the number of training epochs to 30. Often, large networks with many output classes will require much more training than this, but it should give us an clue as to whether the network is flawed or simply needs to train more. We should expect a noticable increase in accuracy for both validation and test."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "%time history_30 = model.fit(X_normalized, y_one_hot, nb_epoch=30, validation_split=0.2, verbose=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "#save the weights\n",
    "model.save_weights('keras_30.h5')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "New validation accuracy with 30 training epochs:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.7152\n"
     ]
    }
   ],
   "source": [
    "print(history_30.history['val_acc'][-1])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "Evaluate the new model on test data:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      " 9984/10000 [============================>.] - ETA: 0s\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\bloss: 1.2196701302051545\n",
      "acc: 0.7076\n"
     ]
    }
   ],
   "source": [
    "metrics = model.evaluate(X_normalized_test, y_one_hot_test)\n",
    "for metric_i in range(len(model.metrics_names)):\n",
    "    metric_name = model.metrics_names[metric_i]\n",
    "    metric_value = metrics[metric_i]\n",
    "    print('{}: {}'.format(metric_name, metric_value))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "--- \n",
    "## Transfer Learning Protocol"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "#### The Four Possibilities:\n",
    "1) **New data set is small and similar to the previous:** Since the new data set is small, you run the risk of overfitting if you retrained everything. Instead, slice off the last fully connected layer and replace with with a new fully connected layer with the appropriate output size. This makes sense because the similarity of the obersations (i.e. pictures) means both the low-level (e.g. edges) and high-level features (e.g. shapes) will be similiar. Freeze the weights before the last layer and retrain!\n",
    "  \n",
    "2) **New data set is large and similar to the previous:**  Since there is more data, there is less risk of overfitting by retraining. Freeze the low-level feature weights and retrain the high-level features to get a better generalization. Don't forget to replace the last fully connected layer! *Optional: If your data set is large enough to handle it, you can initialize all the layers with their previous weights/biases and retrain the entire network.*\n",
    "  \n",
    "3) **New data set is small and different than the previous:** This is the most difficult situation to deal with. Intuitively, we know that the previous network is finely-tuned at each layer. However, we do not want any of the high-level features ***and*** we **cannot** afford to retrain them because we could overfit. Instead, remove all of the fully connected layers and all of the high-level convolutional layers. All that should remain are the first few low-level convolutional layers. Place a fully connected layer with the correct number of outputs, freeze the rest of the layers, and retrain.  \n",
    "  \n",
    "4) **New data set is large and different than the previous:** Retrain the entire network. It's usually a good idea to instantiate the previous model's weights/biases to speed up training (lot's of the low-level convolutions will have similiar weights/biases). Don't forget to replace the fully connected output layer. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "You can read more about the practicalities of transfer learning [here.](http://cs231n.github.io/transfer-learning/)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "In our case, the CIFAR10 and ImageNet data sets are very similiar and large. CIFAR10 has 50,000 training observations and 10,000 for validation. We will use approach #2 for all of our transfer learning examples in this project. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "---\n",
    "## Inception v3"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "The original Inception network was made by Google and aptly dubbed 'GoogLeNet' to pay homeage to Yann LeCunn and his pioneering of convolutional neural networks back in the 80's. GoogLeNet was built for the ImageNet Large Visual Recognition Challenge in 2012 and achieved an error rate of 6.67% (It was a really, *really* big deal). Now, Inception v3 achieves an error rate of 3.46% on the 2012 data set. \n",
    "\n",
    "Read more about the original Inception network [here](https://arxiv.org/abs/1409.4842) or about the new [Inception v3.](https://arxiv.org/abs/1512.00567)\n",
    "\n",
    "To understand what makes Inception so special watch this video about the [inception module](https://www.youtube.com/watch?v=VxhSouuSZDY). "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "from keras.applications.inception_v3 import InceptionV3\n",
    "from keras.models import Model\n",
    "from keras.layers import Dense, GlobalAveragePooling2D\n",
    "from keras import backend as K"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "#import inception with pre-trained weights. do not include fully connected layers\n",
    "inception_base = InceptionV3(weights='imagenet', include_top=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "Now that we have the InceptionV3 network, minus the fully connected \"classification\" layers, we will want to pipe the convolutional layer output into new fully connected layers with the correct number of classes for our dataset."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# add a global spatial average pooling layer\n",
    "x = inception_base.output\n",
    "x = GlobalAveragePooling2D()(x)\n",
    "# let's add a fully-connected layer\n",
    "x = Dense(512, activation='relu')(x)\n",
    "# and a fully connected output/classification layer. \n",
    "# CIFAR10 has 10 different classes (wowza)\n",
    "predictions = Dense(10, activation='softmax')(x)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "Read this [post](https://www.quora.com/What-is-global-average-pooling) about global average pooling. Essentially, this layer will take the average of all the feature maps from the last convolutional layer. Whereas those maps used to be fed into fully connected layers (we didn't import them), they will now feed into our new fully connected layer. What is important to know about global average pooling is that *it allows us the network to accept any Tensor/image size*, instead of expecting the size that it was originally trained on. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# create full model to train on \n",
    "inception_transfer = Model(input=inception_base.input, output=predictions)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# we want to train our new fully connected layers using the pre-trained weights\n",
    "# freeze all of the convolutional layers from Inceptionv3\n",
    "for layer in inception_base.layers:\n",
    "    layer.trainable = False"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "inception_transfer.compile(loss='categorical_crossentropy',\n",
    "              optimizer='rmsprop',\n",
    "              metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "---\n",
    "# Roadblock! (┛◉Д◉)┛彡┻━┻"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "source": [
    "```python\n",
    "history_inception10 = inception_transfer.fit(X_normalized, y_one_hot, nb_epoch=1, validation_split=0.2, verbose=0)\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "This code won't work! Let me tell you why:\n",
    "\n",
    "### Reason 1:\n",
    "The first reason is that it will not fit into memory. InceptionV3 is *much* bigger than the simple CNN model written above. The AWS GPU instance I am running on is apparently not PC Master Race and only has about 4GB of VRAM. Peasants. \n",
    "\n",
    "The work around for this is to not load the entire network into memory at once. In Keras, this is done using model.**fit_generator()** Check out the documentation [here.](https://keras.io/models/sequential/)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "Here is an example generator and fit_generator() call from the documentation:\n",
    "\n",
    "```python\n",
    "def generate_arrays_from_file(path):\n",
    "    while 1:\n",
    "    f = open(path)\n",
    "    for line in f:\n",
    "        # create Numpy arrays of input data\n",
    "        # and labels, from each line in the file\n",
    "        x, y = process_line(line)\n",
    "        yield (x, y)\n",
    "    f.close()\n",
    "\n",
    "model.fit_generator(generate_arrays_from_file('/my_file.txt'),\n",
    "        samples_per_epoch=10000, nb_epoch=10)\n",
    "```\n",
    "\n",
    "Essentially, this is batch processing. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "### Reason 2:\n",
    "#### Part 2a:\n",
    "This one is a bit trickier and has basically ZERO documentation. First, let me link to the little documentation there is because I will be referring to it. \n",
    "\n",
    "TensorFlow [tutorial](https://www.tensorflow.org/tutorials/image_retraining) on transfer learning with InceptionV3\n",
    "`retrain.py` from the tutorial's GitHub [repo](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/image_retraining/retrain.py)\n",
    "\n",
    "Looking at the `retrain.py` file, lines 100 and 247-257 are the culprits.\n",
    "\n",
    "```python\n",
    "# 100\n",
    "JPEG_DATA_TENSOR_NAME = 'DecodeJpeg/contents:0'\n",
    "\n",
    "#247\n",
    "with tf.Session() as sess:\n",
    "model_filename = os.path.join(\n",
    "    FLAGS.model_dir, 'classify_image_graph_def.pb')\n",
    "with gfile.FastGFile(model_filename, 'rb') as f:\n",
    "  graph_def = tf.GraphDef()\n",
    "  graph_def.ParseFromString(f.read())\n",
    "  bottleneck_tensor, jpeg_data_tensor, resized_input_tensor = (\n",
    "      tf.import_graph_def(graph_def, name='', return_elements=[\n",
    "          BOTTLENECK_TENSOR_NAME, JPEG_DATA_TENSOR_NAME,\n",
    "          RESIZED_INPUT_TENSOR_NAME]))\n",
    "return sess.graph, bottleneck_tensor, jpeg_data_tensor, resized_input_tensor\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "**DecodeJpeg** is a Tensor (fundamental object type in TensorFlow). You can check out its definition in the source code [here.](https://github.com/tensorflow/tensorflow/blob/27711108b5fce2e1692f9440631a183b3808fa01/tensorflow/core/kernels/decode_jpeg_op.cc)\n",
    "\n",
    "In a nutshell, the InceptionV3 network was written to parse images as `.jpeg` - ***not*** Numpy arrays. Obviously, using this Tensor is the workaround for training the model, however it requires writing explicit TensorFlow code which the project is trying to avoid when possible. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "#### Part 2b:\n",
    "The Keras implementation of InceptionV3 has a lower limit of the image dimensions it will accept for `input_size`. The lower limit is 139x139 pixels. CIFAR10 is 32x32. \n",
    "\n",
    "What this means is we have to use the TensorFlow library to resize CIFAR to an appropriate size (default is 299x299). [tf.images.resize_images()](https://www.tensorflow.org/api_docs/python/tf/image/resize_images) will do that, but at a cost. When you blow a 32x32 image up to 299x299 a lot of the granular features (i.e. edges) become muddled. The entire reason we want to transfer learn in the first place is to take advantage of the excellent low-level feature detection of the pre-trained InceptionV3 network. This is one possible reason to explain low accuracy scores. \n",
    "\n",
    "\n",
    "*Interesting discovery -* if you dig around in the Keras source code for its applications (i.e. pre-trained networks that are included with the source code), you'll find that almost all of them have a method named `preprocess_input()` that will normalize your pixel data for you. Take a look at the end of the [source code](https://github.com/fchollet/keras/blob/3a7cd05b488e327c1adccaaff10c78390d53b5a8/keras/applications/inception_v3.py). What's extra interesting is that this method isn't listed in the documentation! [Proof.](https://keras.io/applications/#inceptionv3)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "---\n",
    "# Solution:\n",
    "## Bottlenecking"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "From the TensorFlow [website](https://www.tensorflow.org/tutorials/image_retraining#bottlenecks): \n",
    "\n",
    "> 'Bottleneck' is an informal term we often use for the layer just before the final output layer that actually does the classification. This penultimate layer has been trained to output a set of values that's good enough for the classifier to use to distinguish between all the classes it's been asked to recognize. That means it has to be a meaningful and compact summary of the images, since it has to contain enough information for the classifier to make a good choice in a very small set of values. The reason our final layer retraining can work on new classes is that it turns out the kind of information needed to distinguish between all the 1,000 classes in ImageNet is often also useful to distinguish between new kinds of objects.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "### TL;DR\n",
    "1) Import pre-trained CNN  \n",
    "2) Remove fully connected layers  \n",
    "3) Run new data set through network as-is  \n",
    "4) Use output as input to train small fully connected network with correct number of classes  \n",
    "5) Attach the fully connected network to the CNN"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "### Wait... what?\n",
    "Don't worry. It's cool. Let me explain - \n",
    "\n",
    "The original approach was to take a pre-trained network, slice off the end (the part that classifies inputs), stick a new classification network on to replace it, *freeze* the convolutional layers, and then train the new part of the network. Easy to say, not quite as easy to do, but *definitely* much more **expensive** to do. \n",
    "\n",
    "By creating bottlenecks, we simply record the feature map outputs as they are before they would have been fed into the classification layers that were removed. In other words, bottlenecks require zero training. They are produced by the model predicting on your new data set. It is exponentially quicker to use a model to predict than it is to train it (think of all the computation that goes into stochastic/mini-batch gradient descent!) In fact, since bottlenecking does not require parallel computing, it can be done on a CPU if you have the time.  \n",
    "\n",
    "Long story short - Bottlenecking is genius. It will save you a lot of time. It's effective.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "# Inception v3.....v2"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## Create bottlenecks\n",
    "\n",
    "**WARNING**\n",
    "There is explicit TensorFlow code below. Not a lot, but still. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "from keras.applications.inception_v3 import InceptionV3\n",
    "from keras.layers import Input, AveragePooling2D\n",
    "from keras.applications.inception_v3 import preprocess_input\n",
    "from sklearn.model_selection import train_test_split\n",
    "from keras.models import Model\n",
    "from keras.datasets import cifar10\n",
    "import pickle\n",
    "import tensorflow as tf\n",
    "import keras.backend as K"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "batch_size = 16\n",
    "#inception default input size\n",
    "h, w, ch = 299, 299, 3\n",
    "\n",
    "#tensor. will receive cifar10 images as input, gets passed to resize_images\n",
    "img_placeholder = tf.placeholder(\"uint8\", (None, 32, 32, 3))\n",
    "\n",
    "#tensor. resized images. gets passed into Session() \n",
    "resize_op = tf.image.resize_images(img_placeholder, (h, w), method=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# create a generator for batch processing\n",
    "# this gen is written as if you could run through ALL of the data\n",
    "# AWS instance doesn't have enough memory to hold the entire training bottleneck in memory\n",
    "# so we will call for 10000 samples when we call it\n",
    "def gen(session, data, labels, batch_size):\n",
    "    def _f():\n",
    "        start = 0\n",
    "        end = start + batch_size\n",
    "        n = data.shape[0]\n",
    "        while True:\n",
    "            # run takes in a tensor/function and performs it.\n",
    "            # almost always, that function will take a Tensor as input\n",
    "            # when run is called, it takes a feed_dict param which translates\n",
    "            # Tensors into actual data/integers/floats/etc\n",
    "            # this is so you can write a network and only have to change the \n",
    "            # data being passed in one place instead of everywhere\n",
    "            \n",
    "            # X_batch is resized\n",
    "            X_batch = session.run(resize_op, {img_placeholder: data[start:end]})\n",
    "            # X_batch is normalized\n",
    "            X_batch = preprocess_input(X_batch)\n",
    "            y_batch = labels[start:end]\n",
    "            start += batch_size\n",
    "            end += batch_size\n",
    "            if start >= n:\n",
    "                start = 0\n",
    "                end = batch_size\n",
    "                print(\"Bottleneck predictions completed.\")\n",
    "\n",
    "            yield (X_batch, y_batch)\n",
    "\n",
    "    return _f"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "The generator had to be modeled to only use part of the data set. The AWS instance I'm on doesn't have enough memory to have 50,000 images in memory before writing them. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "def create_model():\n",
    "    input_tensor = Input(shape=(h, w, ch))\n",
    "    #inceptionv3 without the fully connected layers, input_size is == X_batch size\n",
    "    model = InceptionV3(input_tensor=input_tensor, include_top=False)\n",
    "    x = model.output\n",
    "    #average pooling flattens the dimensionality\n",
    "    x = AveragePooling2D((8, 8), strides=(8, 8))(x)\n",
    "    # make x the model output\n",
    "    model = Model(model.input, x)\n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# don't care about test data right now\n",
    "(X_train, y_train), (_, _) = cifar10.load_data()\n",
    "X_train, X_val, y_train, y_val = train_test_split(X_train, y_train, test_size=0.2, random_state=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Bottleneck training\n",
      "Bottleneck validation\n"
     ]
    }
   ],
   "source": [
    "with tf.Session() as sess:\n",
    "    # fetch session so Keras API can work \n",
    "    K.set_session(sess)\n",
    "    K.set_learning_phase(1)\n",
    "\n",
    "    model = create_model()\n",
    "\n",
    "    print('Bottleneck training')\n",
    "    train_gen = gen(sess, X_train, y_train, batch_size)\n",
    "    # store predictions, only 8000, 8000%batch_size == 0\n",
    "    bottleneck_features_train = model.predict_generator(train_gen(), 2000)\n",
    "    data = {'features': bottleneck_features_train, 'labels': y_train[:2000]}\n",
    "    # save bottleneck\n",
    "    pickle.dump(data, open('inception_train_bottleneck.p', 'wb'))\n",
    "\n",
    "    print('Bottleneck validation')\n",
    "    val_gen = gen(sess, X_val, y_val, batch_size)\n",
    "    bottleneck_features_validation = model.predict_generator(val_gen(), 2000)\n",
    "    data = {'features': bottleneck_features_validation, 'labels': y_val[:2000]}\n",
    "    pickle.dump(data, open('inception_validation_bottleneck.p', 'wb'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## Transfer Learning Inceptionv3"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "import pickle\n",
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "from keras.layers import Input, Flatten, Dense\n",
    "from keras.models import Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "def load_bottleneck_data(training_file, validation_file):\n",
    "    with open(training_file, 'rb') as f:\n",
    "        train_data = pickle.load(f)\n",
    "    with open(validation_file, 'rb') as f:\n",
    "        validation_data = pickle.load(f)\n",
    "\n",
    "    X_train = train_data['features']\n",
    "    y_train = train_data['labels']\n",
    "    X_val = validation_data['features']\n",
    "    y_val = validation_data['labels']\n",
    "\n",
    "    return X_train, y_train, X_val, y_val"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "03 - mcnulty.ipynb\n",
      "inception_bottleneck_100.h5\n",
      "inception_bottleneck.h5\n",
      "inception_bottleneck.ipynb\n",
      "inception_bottleneck_weights.h5\n",
      "inception_train_bottleneck.p\n",
      "inception_validation_bottleneck.p\n",
      "keras_10.h5\n",
      "keras_30.h5\n",
      "resnet_bottleneck_weights.h5\n",
      "resnet_train_bottleneck.p\n",
      "resnet_validate_bottleneck.p\n",
      "val_acc.png\n",
      "val_loss.png\n",
      "vgg_bottleneck_weights.h5\n",
      "vgg_train_bottleneck.p\n",
      "vgg_validation_bottleneck.p\n"
     ]
    }
   ],
   "source": [
    "%%bash\n",
    "ls"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# training bottleneck currently only has 100 predictions to speed up training\n",
    "# TODO: create and use full bottlenecks\n",
    "X_train, y_train, X_val, y_val = load_bottleneck_data('inception_train_bottleneck.p',\n",
    "                                                      'inception_validation_bottleneck.p')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(2000, 1, 1, 2048) (2000, 1)\n",
      "(2000, 1, 1, 2048) (2000, 1)\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "True"
      ]
     },
     "execution_count": 33,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "print(X_train.shape, y_train.shape)\n",
    "print(X_val.shape, y_val.shape)\n",
    "\n",
    "nb_classes = len(np.unique(y_train))\n",
    "nb_classes == 10"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# feature maps output 2048 dimensional data after averagepooling\n",
    "input_shape = X_train.shape[1:]\n",
    "# no documention on layers.Input() other than it is parameter for \n",
    "# applications.InceptionV3() and other pre-trained networks\n",
    "inp = Input(shape=input_shape)\n",
    "x = Flatten()(inp)\n",
    "x = Dense(nb_classes, activation='softmax')(x)\n",
    "model = Model(inp, x)\n",
    "model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "epochs = 10\n",
    "batch_size = 128"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "with tf.Session() as sess:\n",
    "    # fetch session so Keras API can work \n",
    "    K.set_session(sess)\n",
    "    K.set_learning_phase(1)\n",
    "    %time inception_history = model.fit(X_train, y_train, nb_epoch=epochs, \n",
    "                                  batch_size=batch_size,\n",
    "                                  validation_data=(X_val, y_val), shuffle=True, verbose=0)\n",
    "    model.save_weights('inception_bottleneck_weights.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'acc': [0.34250000047683715, 0.64550000143051145, 0.72099999999999997, 0.76400000000000001, 0.79000000381469726, 0.82850000190734863, 0.846500002861023, 0.86800000286102297, 0.88300000047683713, 0.90150000095367433], 'val_loss': [1.4375709362030029, 1.1400248184204103, 1.0458512754440308, 1.0018450679779052, 0.97828116703033452, 0.9620008239746094, 0.95265592861175541, 0.94571116352081297, 0.94115495872497557, 0.93911342859268188], 'loss': [1.9024671058654785, 1.164107654094696, 0.92374193811416627, 0.79141586589813229, 0.69847158288955691, 0.6280431690216064, 0.5676971120834351, 0.51723335957527161, 0.47384246182441714, 0.43497211360931398], 'val_acc': [0.52450000095367433, 0.62150000047683718, 0.63950000047683719, 0.65100000143051151, 0.66400000095367429, 0.67050000095367435, 0.67400000047683717, 0.67750000095367435, 0.6770000009536743, 0.67899999904632569]}\n"
     ]
    }
   ],
   "source": [
    "print(inception_history.history)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "---\n",
    "# VGG16"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "Made in 2012 by the Oxford **V**isual **G**eometry **G**group. It achieves 7.5% top-5 error on ILSVRC-2012-val, 7.4% top-5 error on ILSVRC-2012-test.  \n",
    "Available [here.](https://gist.github.com/baraldilorenzo/07d7802847aaad0a35d3)  \n",
    "Read the paper [here.](https://arxiv.org/abs/1409.1556)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## Bottlenecks VGG16"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "from keras.applications.vgg16 import VGG16, preprocess_input\n",
    "from keras.layers import Input, AveragePooling2D\n",
    "from sklearn.model_selection import train_test_split\n",
    "from keras.models import Model\n",
    "from keras.datasets import cifar10\n",
    "import pickle\n",
    "import tensorflow as tf\n",
    "import keras.backend as K"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "batch_size = 16\n",
    "h, w, ch = 224, 224, 3\n",
    "img_placeholder = tf.placeholder(\"uint8\", (None, 32, 32, 3))\n",
    "resize_op = tf.image.resize_images(img_placeholder, (h, w), method=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "def create_model_vgg():\n",
    "    input_tensor = Input(shape=(h, w, ch))\n",
    "    model = VGG16(input_tensor=input_tensor, include_top=False)\n",
    "    x = model.output\n",
    "    x = AveragePooling2D((7, 7))(x)\n",
    "    model = Model(model.input, x)\n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "True"
      ]
     },
     "execution_count": 41,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "(X_train, y_train), (_, _) = cifar10.load_data()\n",
    "X_train, X_val, y_train, y_val = train_test_split(X_train, y_train, test_size=0.2, random_state=0)\n",
    "nb_classes = len(np.unique(y_train))\n",
    "nb_classes == 10"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "with tf.Session() as sess:\n",
    "    K.set_session(sess)\n",
    "    K.set_learning_phase(1)\n",
    "\n",
    "    model = create_model_vgg()\n",
    "\n",
    "    train_gen = gen(sess, X_train, y_train, batch_size)\n",
    "    bottleneck_features_train = model.predict_generator(train_gen(), 2000)\n",
    "    data = {'features': bottleneck_features_train, 'labels': y_train[:2000]}\n",
    "    pickle.dump(data, open('vgg_train_bottleneck.p', 'wb'))\n",
    "\n",
    "    val_gen = gen(sess, X_val, y_val, batch_size)\n",
    "    bottleneck_features_validation = model.predict_generator(val_gen(), 2000)\n",
    "    data = {'features': bottleneck_features_validation, 'labels': y_val[:2000]}\n",
    "    pickle.dump(data, open('vgg_validation_bottleneck.p', 'wb'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## Transfer Learning VGG16"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "X_train, y_train, X_val, y_val = load_bottleneck_data('vgg_train_bottleneck.p',\n",
    "                                                      'vgg_validation_bottleneck.p')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "input_shape = X_train.shape[1:]\n",
    "inp = Input(shape=input_shape)\n",
    "x = Flatten()(inp)\n",
    "x = Dense(nb_classes, activation='softmax')(x)\n",
    "model = Model(inp, x)\n",
    "model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "with tf.Session() as sess:\n",
    "    # fetch session so Keras API can work \n",
    "    K.set_session(sess)\n",
    "    K.set_learning_phase(1)\n",
    "    %time vgg_history =model.fit(X_train, y_train, nb_epoch=epochs, batch_size=batch_size,\n",
    "                       validation_data=(X_val, y_val), shuffle=True, verbose=0)\n",
    "    model.save_weights('vgg_bottleneck_weights.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'acc': [0.39550000000000002, 0.66600000000000004, 0.75600000000000001, 0.80549999999999999, 0.84499999999999997, 0.87050000000000005, 0.89049999999999996, 0.90700000000000003, 0.92000000000000004, 0.9375], 'val_loss': [1.3788541226387023, 1.0295256323814392, 0.98009510564804081, 0.83149981689453123, 0.79790244710445402, 0.8015463349819183, 0.77138957583904266, 0.73167234730720521, 0.76594497299194331, 0.72016408586502079], 'loss': [2.2644250402450563, 1.0324710443019867, 0.7252017570734024, 0.57768816328048711, 0.46883150166273119, 0.40063909685611726, 0.34498759192228318, 0.3094297588169575, 0.26612326717376711, 0.23840149456262588], 'val_acc': [0.5605, 0.65900000000000003, 0.6855, 0.72250000000000003, 0.73450000000000004, 0.73450000000000004, 0.73599999999999999, 0.75700000000000001, 0.74950000000000006, 0.76200000000000001]}\n"
     ]
    }
   ],
   "source": [
    "print(vgg_history.history)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "---\n",
    "## ResNet50"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "The Microsoft Research team that won ILSRVC 2015 competition submitted several versions of ResNet. ResNet152 (152 layers) actually surpassed human performance on ImageNet dataset for the first time in history. ResNet50 (50 layers) is one of the version provided in experiments. Read their paper [here.](https://arxiv.org/abs/1512.03385)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "from keras.applications.resnet50 import ResNet50, preprocess_input\n",
    "from keras.layers import Input, AveragePooling2D\n",
    "from sklearn.model_selection import train_test_split\n",
    "from keras.models import Model\n",
    "from keras.datasets import cifar10\n",
    "import pickle\n",
    "import tensorflow as tf\n",
    "import keras.backend as K"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "def create_model_resnet():\n",
    "    input_tensor = Input(shape=(h, w, ch))\n",
    "    model = ResNet50(input_tensor=input_tensor, include_top=False)\n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "(X_train, y_train), (_, _) = cifar10.load_data()\n",
    "X_train, X_val, y_train, y_val = train_test_split(X_train, y_train, test_size=0.2, random_state=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "with tf.Session() as sess:\n",
    "    K.set_session(sess)\n",
    "    K.set_learning_phase(1)\n",
    "\n",
    "    model = create_model_resnet()\n",
    "\n",
    "    train_gen = gen(sess, X_train, y_train, batch_size)\n",
    "    bottleneck_features_train = model.predict_generator(train_gen(), 2000)\n",
    "    data = {'features': bottleneck_features_train, 'labels': y_train[:2000]}\n",
    "    pickle.dump(data, open('resnet_train_bottleneck.p', 'wb'))\n",
    "\n",
    "    val_gen = gen(sess, X_val, y_val, batch_size)\n",
    "    bottleneck_features_validation = model.predict_generator(val_gen(), 2000)\n",
    "    data = {'features': bottleneck_features_validation, 'labels': y_val[:2000]}\n",
    "    pickle.dump(data, open('resnet_validate_bottleneck.p', 'wb'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## Transfer Learning Resnet50"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "X_train, y_train, X_val, y_val = load_bottleneck_data('resnet_train_bottleneck.p', \n",
    "                                                      'resnet_train_bottleneck.p')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "nb_classes=10"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "input_shape = X_train.shape[1:]\n",
    "inp = Input(shape=input_shape)\n",
    "x = Flatten()(inp)\n",
    "x = Dense(nb_classes, activation='softmax')(x)\n",
    "model = Model(inp, x)\n",
    "model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "with tf.Session() as sess:\n",
    "    # fetch session so Keras API can work \n",
    "    K.set_session(sess)\n",
    "    K.set_learning_phase(1)\n",
    "    %time resnet_history =model.fit(X_train, y_train, nb_epoch=epochs, batch_size=batch_size,\n",
    "                       validation_data=(X_val, y_val), shuffle=True, verbose=0)\n",
    "    model.save_weights('resnet_bottleneck_weights.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'acc': [0.52500000000000002, 0.78100000000000003, 0.84399999999999997, 0.89049999999999996, 0.92449999999999999, 0.94399999999999995, 0.96499999999999997, 0.98150000000000004, 0.98650000000000004, 0.99250000000000005], 'val_loss': [0.71555527520179751, 0.49753920614719394, 0.37569377863407133, 0.27851561969518662, 0.22171158787608147, 0.18626745122671126, 0.14767943611741066, 0.12359493595361709, 0.12178987002372742, 0.090790085643529889], 'loss': [1.3432180433273315, 0.6748421652317047, 0.5018155208826065, 0.38168321341276168, 0.30449971109628676, 0.24512132728099822, 0.19791801410913468, 0.16246009841561318, 0.13528480681777, 0.11536293536424637], 'val_acc': [0.78349999999999997, 0.84999999999999998, 0.90249999999999997, 0.9425, 0.95650000000000002, 0.97099999999999997, 0.98799999999999999, 0.99450000000000005, 0.98950000000000005, 0.999]}\n"
     ]
    }
   ],
   "source": [
    "print(resnet_history.history)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "---\n",
    "# Results"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "plt.style.use('fivethirtyeight')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "x_ax = range(10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 58,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAABCgAAAJgCAYAAABIs1LHAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f5652fa7080>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "fig = plt.figure(figsize=(16,8))\n",
    "plt.plot(x_ax, resnet_history.history['val_acc'], label='ResNet50')\n",
    "plt.plot(x_ax, vgg_history.history['val_acc'], label='VGG16')\n",
    "plt.plot(x_ax, inception_history.history['val_acc'], label='InceptionV3')\n",
    "plt.suptitle('Validation Accuracy')\n",
    "plt.xlabel('Epoch', rotation=45)\n",
    "plt.ylabel('Accuracy')\n",
    "fig.savefig('val_acc.png')\n",
    "plt.legend()\n",
    "plt.show();"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 59,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAABCgAAAJgCAYAAABIs1LHAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAAAAbEdAAQAAAAAAbPf/A8lliwKWjOnlAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f5652ee8a58>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "fig = plt.figure(figsize=(16,8))\n",
    "plt.plot(x_ax, resnet_history.history['val_loss'], label='ResNet50')\n",
    "plt.plot(x_ax, vgg_history.history['val_loss'], label='VGG16')\n",
    "plt.plot(x_ax, inception_history.history['val_loss'], label='InceptionV3')\n",
    "plt.suptitle('Validation Loss')\n",
    "plt.xlabel('Epoch', rotation=45)\n",
    "plt.ylabel('Loss')\n",
    "fig.savefig('val_loss.png')\n",
    "plt.legend()\n",
    "plt.show();"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "source": [
    "---\n",
    "## Next Steps\n",
    "\n",
    "I need to combine the networks and have it predict on new images from start to finish. "
   ]
  }
 ],
 "metadata": {
  "anaconda-cloud": {},
  "kernelspec": {
   "display_name": "Python [conda root]",
   "language": "python",
   "name": "conda-root-py"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.5.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}