Proj2_Data_Feats_Gen.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "Data Generation Standard (v1)\n",
    "    x_data ==> X_train\n",
    "Data Generation Features of Pre-trained models (v2, v3)\n",
    "    X_train ==> X_train_feats_v2\n",
    "            ==> X_train_feats_v3"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "keras 2.0.5\n",
      "tensorflow 1.2.0\n"
     ]
    }
   ],
   "source": [
    "import sys\n",
    "## To get helper functions from another folder\n",
    "# sys.path.insert(0, '../') # if _helper_basics_ is in previous folder\n",
    "now_i_am_at = 'home' # home dso test\n",
    "if now_i_am_at=='home': sys.path.insert(0, 'E:/Leonard HDD/Dropbox/DSO/Tasks/')\n",
    "elif now_i_am_at=='dso': sys.path.insert(0, 'D:/Dropbox/DSO/Tasks')\n",
    "\n",
    "from _helper_basics_ import *\n",
    "\n",
    "%matplotlib inline\n",
    "plt.rcParams['savefig.dpi'] = 100\n",
    "\n",
    "%load_ext autoreload\n",
    "%autoreload 2\n",
    "\n",
    "print('keras', keras.__version__)\n",
    "print('tensorflow', tf.__version__)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def show_object(obj):\n",
    "    # Reshape 784 array into 28x28 image\n",
    "    image = obj.reshape([28,28])\n",
    "    fig, axes = plt.subplots(1, )\n",
    "    fig.subplots_adjust(hspace=0.5, wspace=0.5)\n",
    "    plt.imshow(image, cmap='gray_r')\n",
    "    plt.show()\n",
    "#shuffling function\n",
    "def unison_shuffled_copies(a, b):\n",
    "    assert len(a) == len(b)\n",
    "    p = np.random.permutation(len(a))\n",
    "    return a[p], b[p]\n",
    "\n",
    "num_examples_per_class = 2000\n",
    "classes = ['airplane','alarm clock','ambulance','angel','ant','anvil','apple','axe', 'banana','bandage','barn','baseball bat','baseball',\n",
    "           'basket', 'basketball','bathtub','beach','bear','beard','bed','bee','belt', 'bicycle','binoculars','birthday cake','blueberry', \n",
    "           'book','boomerang','bottlecap','bowtie','bracelet','brain', 'bread','broom','bulldozer','bus','bush','butterfly','cactus','cake']\n",
    "num_classes = len(classes)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Data Generation: Standard\n",
    "x_data ==> X_train"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "x_data (80000, 784)\n",
      "y_data (80000,)\n"
     ]
    }
   ],
   "source": [
    "## x_data\n",
    "x_data = np.load(\"./x_data_40_classes_2k.npy\")\n",
    "\n",
    "## y_data\n",
    "labels = [np.full((num_examples_per_class,), classes.index(qdraw)) for qdraw in classes]\n",
    "y_data = np.concatenate(labels,axis=0)\n",
    "\n",
    "## Shuffling\n",
    "x_data,y_data = unison_shuffled_copies(x_data,y_data)\n",
    "print('x_data',x_data.shape)\n",
    "print('y_data',y_data.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "np.save('x_data.npy',x_data)\n",
    "np.save('y_data.npy',y_data)\n",
    "\n",
    "np.save('X_train.npy',X_train)\n",
    "np.save('Y_train.npy',Y_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "X_train (80000, 28, 28, 1)\n",
      "Y_train (80000, 40)\n"
     ]
    }
   ],
   "source": [
    "X_train = x_data.reshape(80000,28,28,1)\n",
    "Y_train = to_categorical(y_data, num_classes=num_classes)\n",
    "print('X_train',X_train.shape)\n",
    "print('Y_train',Y_train.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "## Normalised and splitting into train/test set\n",
    "\n",
    "X_train = X_train/255.\n",
    "X_train = X_train.astype(np.float16)\n",
    "\n",
    "indices = np.arange(len(Y_train))\n",
    "np.random.shuffle(indices)\n",
    "\n",
    "X_train = X_train[:int(.8*len(Y_train))]\n",
    "Y_train = Y_train[:int(.8*len(Y_train))]\n",
    "X_test = X_train[int(.8*len(Y_train)):]\n",
    "Y_test = Y_train[int(.8*len(Y_train)):]\n",
    "\n",
    "np.save('X_train_normalised_spliced.npy',X_train)\n",
    "np.save('Y_train_normalised_spliced.npy',Y_train)\n",
    "np.save('X_test_normalised_spliced.npy',X_test)\n",
    "np.save('Y_test_normalised_spliced.npy',Y_test)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Visualisation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n435yAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x2bd4ae4d978>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "classes[y_data[idx]] bed\n",
      "y_data[idx] 19\n",
      "np.argmax(Y_train[idx]) 19\n"
     ]
    }
   ],
   "source": [
    "idx = 5000\n",
    "\n",
    "## visual of x,y_data\n",
    "show_object(x_data[idx])\n",
    "print('classes[y_data[idx]]', classes[y_data[idx]])\n",
    "print('y_data[idx]',y_data[idx])\n",
    "\n",
    "## To verify correctnessof X,Y_train\n",
    "print('np.argmax(Y_train[idx])',np.argmax(Y_train[idx]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Data Generation: Features from Pre-trained Models\n",
    "    X_train ==> X_train_feats_v2\n",
    "    X_train ==> X_train_feats_v3"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Simple Concatenate to form pseudo 3 channel image"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "X_train_3 (80000, 28, 28, 3)\n"
     ]
    }
   ],
   "source": [
    "## Generate 3 channels\n",
    "X_train_3 = np.append(np.append(X_train,X_train,axis=3), X_train, axis=3)\n",
    "print('X_train_3',X_train_3.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### V2: Xception"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "X_train_feat_v2 (80000, 1, 1, 2048)\n",
      "Y_train (80000, 40)\n",
      "time taken to generate feats : 472.32990431785583\n",
      "time taken to save feats : 0.0\n"
     ]
    }
   ],
   "source": [
    "Archi_dir = 'v2/'\n",
    "\n",
    "## Loading the Pre-trained Xception Model\n",
    "Inp = Input(shape=(28,28,3))\n",
    "Xception_model = Xception(include_top=False, input_tensor=Inp)\n",
    "for layer in Xception_model.layers:    layer.trainable = False\n",
    "Outp = Xception_model(Inp)\n",
    "Get_Feats_v2 = Model(Inp,Outp)\n",
    "Get_Feats_v2.summary()\n",
    "\n",
    "## Generating output of Xception as features\n",
    "start_time = time.time()\n",
    "X_train_feat_v2 = Get_Feats_v2.predict(X_train_3)\n",
    "print('X_train_feat_v2',X_train_feat_v2.shape)\n",
    "print('Y_train',Y_train.shape)\n",
    "print('time taken to generate feats : {}'.format(time.time()-start_time))\n",
    "\n",
    "start_time = time.time()\n",
    "np.save(Archi_dir+'X_train_feat_v2.npy',X_train_feat_v2)\n",
    "print('time taken to save feats : {}'.format(time.time()-start_time))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nSUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x2bd6eb034a8>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "classes[y_data[idx]] beard\n",
      "y_data[idx] 18\n",
      "np.argmax(Y_train[idx]) 18\n"
     ]
    }
   ],
   "source": [
    "## Sanity check\n",
    "\n",
    "idx = 0; channel = 0\n",
    "show_object(X_train_3[idx,:,:,channel])\n",
    "print('classes[y_data[idx]]', classes[y_data[idx]])\n",
    "print('y_data[idx]',y_data[idx])\n",
    "print('np.argmax(Y_train[idx])',np.argmax(Y_train[idx]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "X_train_feat_v2 (80000, 1, 1, 2048)\n",
      "Y_train (80000, 40)\n"
     ]
    }
   ],
   "source": [
    "## Load\n",
    "X_train_feat_v2 = np.load(Archi_dir+'X_train_feat_v2.npy')\n",
    "Y_train = np.load('Y_train.npy')\n",
    "\n",
    "print('X_train_feat_v2',X_train_feat_v2.shape)\n",
    "print('Y_train',Y_train.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### V3:MobileNet"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "Archi_dir = 'v3/'\n",
    "\n",
    "Inp = Input(shape=(28,28,3))\n",
    "Mobilenet_model = Mobilenet(include_top=False, input_tensor=Inp)\n",
    "for layer in Mobilenet_model.layers:    layer.trainable = False\n",
    "Outp = Mobilenet_model(Inp)\n",
    "Get_Feats_v3 = Model(Inp,Outp)\n",
    "Get_Feats_v3.summary()\n",
    "\n",
    "start_time = time.time()\n",
    "X_train_feat_v3 = Get_Feats_v3.predict(X_train_3)\n",
    "print('X_train_feat_v3',X_train_feat_v3.shape)\n",
    "print('Y_train',Y_train.shape)\n",
    "print('time taken to generate feats : {}'.format(time.time()-start_time))\n",
    "\n",
    "start_time = time.time()\n",
    "# np.save(Archi_dir+'X_train_feat_v3.npy',X_train_feat_v3)\n",
    "print('time taken to save feats : {}'.format(time.time()-start_time))"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.5.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}