two-rectangles.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:22:17.466000",
     "start_time": "2016-07-08T02:22:17.460000"
    },
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "import matplotlib\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:22:20.062000",
     "start_time": "2016-07-08T02:22:17.976000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((50000L, 8L, 8L), (50000L, 2L, 4L))"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Create images with random rectangles and bounding boxes. \n",
    "num_imgs = 50000\n",
    "\n",
    "img_size = 8\n",
    "min_rect_size = 1\n",
    "max_rect_size = 4\n",
    "num_objects = 2\n",
    "\n",
    "bboxes = np.zeros((num_imgs, num_objects, 4))\n",
    "imgs = np.zeros((num_imgs, img_size, img_size))\n",
    "\n",
    "for i_img in range(num_imgs):\n",
    "    for i_object in range(num_objects):\n",
    "        w, h = np.random.randint(min_rect_size, max_rect_size, size=2)\n",
    "        x = np.random.randint(0, img_size - w)\n",
    "        y = np.random.randint(0, img_size - h)\n",
    "        imgs[i_img, x:x+w, y:y+h] = 1.\n",
    "        bboxes[i_img, i_object] = [x, y, w, h]\n",
    "        \n",
    "imgs.shape, bboxes.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:22:23.851000",
     "start_time": "2016-07-08T02:22:23.599000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAPsAAAEACAYAAAB1UADIAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nrBcVAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x44efa20>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "i = 0\n",
    "plt.imshow(imgs[i].T, cmap='Greys', interpolation='none', origin='lower', extent=[0, img_size, 0, img_size])\n",
    "for bbox in bboxes[i]:\n",
    "    plt.gca().add_patch(matplotlib.patches.Rectangle((bbox[0], bbox[1]), bbox[2], bbox[3], ec='r', fc='none'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:23:09.791000",
     "start_time": "2016-07-08T02:23:09.657000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((50000L, 64L), 3.1477043194172438e-17, 1.0000000000000004)"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Reshape and normalize the data to mean 0 and std 1. \n",
    "X = (imgs.reshape(num_imgs, -1) - np.mean(imgs)) / np.std(imgs)\n",
    "X.shape, np.mean(X), np.std(X)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:23:21.661000",
     "start_time": "2016-07-08T02:23:21.646000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((50000L, 8L), 0.28096500000000002, 0.17486042655500986)"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Normalize x, y, w, h by img_size, so that all values are between 0 and 1.\n",
    "# Important: Do not shift to negative values (e.g. by setting to mean 0), because the IOU calculation needs positive w and h.\n",
    "y = bboxes.reshape(num_imgs, -1) / img_size\n",
    "y.shape, np.mean(y), np.std(y)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:23:45.972000",
     "start_time": "2016-07-08T02:23:45.966000"
    },
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# Split training and test.\n",
    "i = int(0.8 * num_imgs)\n",
    "train_X = X[:i]\n",
    "test_X = X[i:]\n",
    "train_y = y[:i]\n",
    "test_y = y[i:]\n",
    "test_imgs = imgs[i:]\n",
    "test_bboxes = bboxes[i:]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:23:51.640000",
     "start_time": "2016-07-08T02:23:46.869000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using Theano backend.\n"
     ]
    }
   ],
   "source": [
    "# Build the model.\n",
    "from keras.models import Sequential\n",
    "from keras.layers import Dense, Activation, Dropout\n",
    "from keras.optimizers import SGD\n",
    "model = Sequential([\n",
    "        Dense(256, input_dim=X.shape[-1]), \n",
    "        Activation('relu'), \n",
    "        Dropout(0.4), \n",
    "        Dense(y.shape[-1])\n",
    "    ])\n",
    "model.compile('adadelta', 'mse')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:33:22.619000",
     "start_time": "2016-07-08T02:27:43.964000"
    },
    "code_folding": [
     3,
     18
    ],
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 0\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0891 - val_loss: 0.0198\n",
      "Flipped 19332.0 training samples (48.33 %)\n",
      "Mean IOU: 0.17421432926\n",
      "Mean dist: 0.202194946413\n",
      "Mean mse: 0.00870554196174\n",
      "\n",
      "Epoch 1\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0188 - val_loss: 0.0188\n",
      "Flipped 7403.0 training samples (18.5075 %)\n",
      "Mean IOU: 0.201610919572\n",
      "Mean dist: 0.179289382942\n",
      "Mean mse: 0.00696734712723\n",
      "\n",
      "Epoch 2\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0150 - val_loss: 0.0204\n",
      "Flipped 4240.0 training samples (10.6 %)\n",
      "Mean IOU: 0.251106918999\n",
      "Mean dist: 0.153800096666\n",
      "Mean mse: 0.00559663701382\n",
      "\n",
      "Epoch 3\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "3s - loss: 0.0130 - val_loss: 0.0212\n",
      "Flipped 2280.0 training samples (5.7 %)\n",
      "Mean IOU: 0.277501980239\n",
      "Mean dist: 0.141010581968\n",
      "Mean mse: 0.00496034980357\n",
      "\n",
      "Epoch 4\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0119 - val_loss: 0.0217\n",
      "Flipped 1812.0 training samples (4.53 %)\n",
      "Mean IOU: 0.299665635151\n",
      "Mean dist: 0.131943238998\n",
      "Mean mse: 0.00453679827365\n",
      "\n",
      "Epoch 5\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0111 - val_loss: 0.0220\n",
      "Flipped 1437.0 training samples (3.5925 %)\n",
      "Mean IOU: 0.31418044106\n",
      "Mean dist: 0.125220445092\n",
      "Mean mse: 0.00419997932865\n",
      "\n",
      "Epoch 6\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0105 - val_loss: 0.0225\n",
      "Flipped 1526.0 training samples (3.815 %)\n",
      "Mean IOU: 0.331142524788\n",
      "Mean dist: 0.118991890303\n",
      "Mean mse: 0.00392323951083\n",
      "\n",
      "Epoch 7\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0100 - val_loss: 0.0228\n",
      "Flipped 1172.0 training samples (2.93 %)\n",
      "Mean IOU: 0.34543661596\n",
      "Mean dist: 0.113598305609\n",
      "Mean mse: 0.00367849045312\n",
      "\n",
      "Epoch 8\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0096 - val_loss: 0.0229\n",
      "Flipped 988.0 training samples (2.47 %)\n",
      "Mean IOU: 0.356540425907\n",
      "Mean dist: 0.109198101651\n",
      "Mean mse: 0.00348902813688\n",
      "\n",
      "Epoch 9\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0092 - val_loss: 0.0229\n",
      "Flipped 1041.0 training samples (2.6025 %)\n",
      "Mean IOU: 0.364558090872\n",
      "Mean dist: 0.106299435695\n",
      "Mean mse: 0.0033265020771\n",
      "\n",
      "Epoch 10\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0089 - val_loss: 0.0232\n",
      "Flipped 697.0 training samples (1.7425 %)\n",
      "Mean IOU: 0.374192726269\n",
      "Mean dist: 0.102362562652\n",
      "Mean mse: 0.00317120671055\n",
      "\n",
      "Epoch 11\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0086 - val_loss: 0.0233\n",
      "Flipped 821.0 training samples (2.0525 %)\n",
      "Mean IOU: 0.381939290872\n",
      "Mean dist: 0.0995234924274\n",
      "Mean mse: 0.00304664076566\n",
      "\n",
      "Epoch 12\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0083 - val_loss: 0.0235\n",
      "Flipped 639.0 training samples (1.5975 %)\n",
      "Mean IOU: 0.389027211857\n",
      "Mean dist: 0.0974743195474\n",
      "Mean mse: 0.00293538438011\n",
      "\n",
      "Epoch 13\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0082 - val_loss: 0.0235\n",
      "Flipped 678.0 training samples (1.695 %)\n",
      "Mean IOU: 0.399173566656\n",
      "Mean dist: 0.0949032411027\n",
      "Mean mse: 0.00282684413067\n",
      "\n",
      "Epoch 14\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0080 - val_loss: 0.0236\n",
      "Flipped 618.0 training samples (1.545 %)\n",
      "Mean IOU: 0.403507223408\n",
      "Mean dist: 0.0929981797996\n",
      "Mean mse: 0.00273148774462\n",
      "\n",
      "Epoch 15\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0078 - val_loss: 0.0239\n",
      "Flipped 750.0 training samples (1.875 %)\n",
      "Mean IOU: 0.410588897039\n",
      "Mean dist: 0.0908761495889\n",
      "Mean mse: 0.0026544331145\n",
      "\n",
      "Epoch 16\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0076 - val_loss: 0.0241\n",
      "Flipped 686.0 training samples (1.715 %)\n",
      "Mean IOU: 0.418996945724\n",
      "Mean dist: 0.0888948065314\n",
      "Mean mse: 0.0025722494028\n",
      "\n",
      "Epoch 17\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0075 - val_loss: 0.0238\n",
      "Flipped 581.0 training samples (1.4525 %)\n",
      "Mean IOU: 0.424184888081\n",
      "Mean dist: 0.0873501393816\n",
      "Mean mse: 0.00251163484627\n",
      "\n",
      "Epoch 18\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0073 - val_loss: 0.0239\n",
      "Flipped 571.0 training samples (1.4275 %)\n",
      "Mean IOU: 0.430039911385\n",
      "Mean dist: 0.0856274159229\n",
      "Mean mse: 0.00243802334432\n",
      "\n",
      "Epoch 19\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0072 - val_loss: 0.0239\n",
      "Flipped 452.0 training samples (1.13 %)\n",
      "Mean IOU: 0.433272926262\n",
      "Mean dist: 0.0845405565164\n",
      "Mean mse: 0.00238937732887\n",
      "\n",
      "Epoch 20\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0071 - val_loss: 0.0243\n",
      "Flipped 403.0 training samples (1.0075 %)\n",
      "Mean IOU: 0.442589569635\n",
      "Mean dist: 0.081736119599\n",
      "Mean mse: 0.002300822613\n",
      "\n",
      "Epoch 21\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "3s - loss: 0.0070 - val_loss: 0.0243\n",
      "Flipped 426.0 training samples (1.065 %)\n",
      "Mean IOU: 0.437372472621\n",
      "Mean dist: 0.0825716755323\n",
      "Mean mse: 0.00229706401744\n",
      "\n",
      "Epoch 22\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0069 - val_loss: 0.0242\n",
      "Flipped 322.0 training samples (0.805 %)\n",
      "Mean IOU: 0.447041620476\n",
      "Mean dist: 0.0808203517427\n",
      "Mean mse: 0.00223702435172\n",
      "\n",
      "Epoch 23\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0069 - val_loss: 0.0241\n",
      "Flipped 323.0 training samples (0.8075 %)\n",
      "Mean IOU: 0.444742584252\n",
      "Mean dist: 0.0812139063405\n",
      "Mean mse: 0.00224665052732\n",
      "\n",
      "Epoch 24\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0068 - val_loss: 0.0243\n",
      "Flipped 403.0 training samples (1.0075 %)\n",
      "Mean IOU: 0.453926726631\n",
      "Mean dist: 0.078654691172\n",
      "Mean mse: 0.00216548576696\n",
      "\n",
      "Epoch 25\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0067 - val_loss: 0.0241\n",
      "Flipped 318.0 training samples (0.795 %)\n",
      "Mean IOU: 0.454372876255\n",
      "Mean dist: 0.0789630339262\n",
      "Mean mse: 0.00215756388899\n",
      "\n",
      "Epoch 26\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0066 - val_loss: 0.0242\n",
      "Flipped 384.0 training samples (0.96 %)\n",
      "Mean IOU: 0.455093751177\n",
      "Mean dist: 0.0784194052206\n",
      "Mean mse: 0.00213678858996\n",
      "\n",
      "Epoch 27\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0066 - val_loss: 0.0245\n",
      "Flipped 319.0 training samples (0.7975 %)\n",
      "Mean IOU: 0.457610800536\n",
      "Mean dist: 0.0772177815774\n",
      "Mean mse: 0.0021002412861\n",
      "\n",
      "Epoch 28\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0065 - val_loss: 0.0245\n",
      "Flipped 363.0 training samples (0.9075 %)\n",
      "Mean IOU: 0.463026173604\n",
      "Mean dist: 0.0760770275642\n",
      "Mean mse: 0.00206011307056\n",
      "\n",
      "Epoch 29\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0065 - val_loss: 0.0242\n",
      "Flipped 335.0 training samples (0.8375 %)\n",
      "Mean IOU: 0.460029429252\n",
      "Mean dist: 0.0773609536022\n",
      "Mean mse: 0.00208323971074\n",
      "\n",
      "Epoch 30\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0064 - val_loss: 0.0245\n",
      "Flipped 234.0 training samples (0.585 %)\n",
      "Mean IOU: 0.46961803655\n",
      "Mean dist: 0.0749676349377\n",
      "Mean mse: 0.00203222502343\n",
      "\n",
      "Epoch 31\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0064 - val_loss: 0.0247\n",
      "Flipped 387.0 training samples (0.9675 %)\n",
      "Mean IOU: 0.470608679921\n",
      "Mean dist: 0.0740146431432\n",
      "Mean mse: 0.00199135766526\n",
      "\n",
      "Epoch 32\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0063 - val_loss: 0.0245\n",
      "Flipped 282.0 training samples (0.705 %)\n",
      "Mean IOU: 0.470118171211\n",
      "Mean dist: 0.0738592915286\n",
      "Mean mse: 0.00197842837846\n",
      "\n",
      "Epoch 33\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0063 - val_loss: 0.0245\n",
      "Flipped 293.0 training samples (0.7325 %)\n",
      "Mean IOU: 0.474404012902\n",
      "Mean dist: 0.0735462556776\n",
      "Mean mse: 0.001959014029\n",
      "\n",
      "Epoch 34\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0063 - val_loss: 0.0247\n",
      "Flipped 273.0 training samples (0.6825 %)\n",
      "Mean IOU: 0.47574437323\n",
      "Mean dist: 0.0728054629133\n",
      "Mean mse: 0.00194009840461\n",
      "\n",
      "Epoch 35\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "3s - loss: 0.0062 - val_loss: 0.0245\n",
      "Flipped 285.0 training samples (0.7125 %)\n",
      "Mean IOU: 0.477151418694\n",
      "Mean dist: 0.0726782049133\n",
      "Mean mse: 0.00192376103654\n",
      "\n",
      "Epoch 36\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0061 - val_loss: 0.0247\n",
      "Flipped 280.0 training samples (0.7 %)\n",
      "Mean IOU: 0.478953860581\n",
      "Mean dist: 0.0720496281562\n",
      "Mean mse: 0.00189495957262\n",
      "\n",
      "Epoch 37\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0061 - val_loss: 0.0249\n",
      "Flipped 238.0 training samples (0.595 %)\n",
      "Mean IOU: 0.485612802479\n",
      "Mean dist: 0.0705254064983\n",
      "Mean mse: 0.00185026440083\n",
      "\n",
      "Epoch 38\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0061 - val_loss: 0.0245\n",
      "Flipped 220.0 training samples (0.55 %)\n",
      "Mean IOU: 0.477454585999\n",
      "Mean dist: 0.0719046474651\n",
      "Mean mse: 0.00188276242328\n",
      "\n",
      "Epoch 39\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0061 - val_loss: 0.0247\n",
      "Flipped 235.0 training samples (0.5875 %)\n",
      "Mean IOU: 0.481772569811\n",
      "Mean dist: 0.0713024149384\n",
      "Mean mse: 0.00187054984367\n",
      "\n",
      "Epoch 40\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0061 - val_loss: 0.0248\n",
      "Flipped 264.0 training samples (0.66 %)\n",
      "Mean IOU: 0.483492848349\n",
      "Mean dist: 0.0707738993554\n",
      "Mean mse: 0.00184634708223\n",
      "\n",
      "Epoch 41\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0061 - val_loss: 0.0249\n",
      "Flipped 253.0 training samples (0.6325 %)\n",
      "Mean IOU: 0.481306290104\n",
      "Mean dist: 0.0707721782897\n",
      "Mean mse: 0.00184881686819\n",
      "\n",
      "Epoch 42\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0061 - val_loss: 0.0247\n",
      "Flipped 259.0 training samples (0.6475 %)\n",
      "Mean IOU: 0.484973269763\n",
      "Mean dist: 0.0702191452585\n",
      "Mean mse: 0.00181671099039\n",
      "\n",
      "Epoch 43\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "3s - loss: 0.0060 - val_loss: 0.0247\n",
      "Flipped 226.0 training samples (0.565 %)\n",
      "Mean IOU: 0.485699224635\n",
      "Mean dist: 0.0705432710211\n",
      "Mean mse: 0.00182664435223\n",
      "\n",
      "Epoch 44\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0060 - val_loss: 0.0245\n",
      "Flipped 242.0 training samples (0.605 %)\n",
      "Mean IOU: 0.486526667704\n",
      "Mean dist: 0.0702527203698\n",
      "Mean mse: 0.00181987211074\n",
      "\n",
      "Epoch 45\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0060 - val_loss: 0.0250\n",
      "Flipped 218.0 training samples (0.545 %)\n",
      "Mean IOU: 0.490067178897\n",
      "Mean dist: 0.0689831817917\n",
      "Mean mse: 0.00179033876822\n",
      "\n",
      "Epoch 46\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0059 - val_loss: 0.0248\n",
      "Flipped 245.0 training samples (0.6125 %)\n",
      "Mean IOU: 0.491345980976\n",
      "Mean dist: 0.0687047519983\n",
      "Mean mse: 0.00177435903349\n",
      "\n",
      "Epoch 47\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0059 - val_loss: 0.0247\n",
      "Flipped 239.0 training samples (0.5975 %)\n",
      "Mean IOU: 0.489264187585\n",
      "Mean dist: 0.0694518844395\n",
      "Mean mse: 0.0017843305106\n",
      "\n",
      "Epoch 48\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0059 - val_loss: 0.0248\n",
      "Flipped 266.0 training samples (0.665 %)\n",
      "Mean IOU: 0.488129088089\n",
      "Mean dist: 0.0695082726052\n",
      "Mean mse: 0.00178704787256\n",
      "\n",
      "Epoch 49\n",
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/1\n",
      "2s - loss: 0.0059 - val_loss: 0.0248\n",
      "Flipped 197.0 training samples (0.4925 %)\n",
      "Mean IOU: 0.487095655783\n",
      "Mean dist: 0.0694985856093\n",
      "Mean mse: 0.00177344207584\n",
      "\n"
     ]
    }
   ],
   "source": [
    "# Flip bboxes during training.\n",
    "# Note: The validation loss is always quite big here because we don't flip the bounding boxes for the validation data. \n",
    "def IOU(bbox1, bbox2):\n",
    "    '''Calculate overlap between two bounding boxes [x, y, w, h] as the area of intersection over the area of unity'''\n",
    "    x1, y1, w1, h1 = bbox1[0], bbox1[1], bbox1[2], bbox1[3]\n",
    "    x2, y2, w2, h2 = bbox2[0], bbox2[1], bbox2[2], bbox2[3]\n",
    "\n",
    "    w_I = min(x1 + w1, x2 + w2) - max(x1, x2)\n",
    "    h_I = min(y1 + h1, y2 + h2) - max(y1, y2)\n",
    "    if w_I <= 0 or h_I <= 0:  # no overlap\n",
    "        return 0\n",
    "    I = w_I * h_I\n",
    "\n",
    "    U = w1 * h1 + w2 * h2 - I\n",
    "\n",
    "    return I / U\n",
    "\n",
    "def distance(bbox1, bbox2):\n",
    "    return np.sqrt(np.sum(np.square(bbox1[:2] - bbox2[:2])))\n",
    "\n",
    "num_epochs = 50\n",
    "flipped_train_y = np.array(train_y)\n",
    "flipped = np.zeros((len(flipped_train_y), num_epochs))\n",
    "ious_epoch = np.zeros((len(flipped_train_y), num_epochs))\n",
    "dists_epoch = np.zeros((len(flipped_train_y), num_epochs))\n",
    "mses_epoch = np.zeros((len(flipped_train_y), num_epochs))\n",
    "\n",
    "for epoch in range(num_epochs):\n",
    "    print 'Epoch', epoch\n",
    "    model.fit(train_X, flipped_train_y, nb_epoch=1, validation_data=(test_X, test_y), verbose=2)\n",
    "    pred_y = model.predict(train_X)\n",
    "\n",
    "    for i, (pred_bboxes, exp_bboxes) in enumerate(zip(pred_y, flipped_train_y)):\n",
    "        \n",
    "        flipped_exp_bboxes = np.concatenate([exp_bboxes[4:], exp_bboxes[:4]])\n",
    "        \n",
    "        mse = np.mean(np.square(pred_bboxes - exp_bboxes))\n",
    "        mse_flipped = np.mean(np.square(pred_bboxes - flipped_exp_bboxes))\n",
    "        \n",
    "        iou = IOU(pred_bboxes[:4], exp_bboxes[:4]) + IOU(pred_bboxes[4:], exp_bboxes[4:])\n",
    "        iou_flipped = IOU(pred_bboxes[:4], flipped_exp_bboxes[:4]) + IOU(pred_bboxes[4:], flipped_exp_bboxes[4:])\n",
    "        \n",
    "        dist = distance(pred_bboxes[:4], exp_bboxes[:4]) + distance(pred_bboxes[4:], exp_bboxes[4:])\n",
    "        dist_flipped = distance(pred_bboxes[:4], flipped_exp_bboxes[:4]) + distance(pred_bboxes[4:], flipped_exp_bboxes[4:])\n",
    "        \n",
    "        if mse_flipped < mse:  # you can also use iou or dist here\n",
    "            flipped_train_y[i] = flipped_exp_bboxes\n",
    "            flipped[i, epoch] = 1\n",
    "            mses_epoch[i, epoch] = mse_flipped / 2.\n",
    "            ious_epoch[i, epoch] = iou_flipped / 2.\n",
    "            dists_epoch[i, epoch] = dist_flipped / 2.\n",
    "        else:\n",
    "            mses_epoch[i, epoch] = mse / 2.\n",
    "            ious_epoch[i, epoch] = iou / 2.\n",
    "            dists_epoch[i, epoch] = dist / 2.\n",
    "            \n",
    "    print 'Flipped {} training samples ({} %)'.format(np.sum(flipped[:, epoch]), np.mean(flipped[:, epoch]) * 100.)\n",
    "    print 'Mean IOU: {}'.format(np.mean(ious_epoch[:, epoch]))\n",
    "    print 'Mean dist: {}'.format(np.mean(dists_epoch[:, epoch]))\n",
    "    print 'Mean mse: {}'.format(np.mean(mses_epoch[:, epoch]))\n",
    "    print"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:38:08.340000",
     "start_time": "2016-07-08T02:38:07.482000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.text.Text at 0x122c06a0>"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYkAAAEPCAYAAAC3NDh4AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nJCRJI5kkJEkjmSQkSSOZJCRJI/1/xhWdaamtd0wAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x1650c940>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Show flippings for a few training samples.\n",
    "plt.pcolor(flipped[:300], cmap='Greys')\n",
    "plt.xlabel('Epoch')\n",
    "plt.ylabel('Training sample')\n",
    "\n",
    "# plt.savefig('plots/bw-two-rectangles_flips.png', dpi=300)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:38:09.494000",
     "start_time": "2016-07-08T02:38:09.091000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(0, 1)"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXcAAAEACAYAAABI5zaHAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nRUSCkIK7iEgQUnAXEQlCCu4iIkHo/wHiNq2PBRMbaQAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x164fef60>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Plot metrics on the training data. \n",
    "mean_ious_epoch = np.mean(ious_epoch, axis=0)\n",
    "mean_dists_epoch = np.mean(dists_epoch, axis=0)\n",
    "mean_mses_epoch = np.mean(mses_epoch, axis=0)\n",
    "plt.plot(mean_ious_epoch, label='Mean IOU')  # between predicted and assigned true bboxes\n",
    "plt.plot(mean_dists_epoch, label='Mean distance')  # relative to image size\n",
    "plt.plot(mean_mses_epoch, label='Mean mse')\n",
    "plt.annotate(np.round(np.max(mean_ious_epoch), 3), (len(mean_ious_epoch)-1, mean_ious_epoch[-1]+0.03), horizontalalignment='right', color='b')\n",
    "plt.annotate(np.round(np.min(mean_dists_epoch), 3), (len(mean_dists_epoch)-1, mean_dists_epoch[-1]+0.03), horizontalalignment='right', color='g')\n",
    "plt.annotate(np.round(np.min(mean_mses_epoch), 3), (len(mean_mses_epoch)-1, mean_mses_epoch[-1]+0.03), horizontalalignment='right', color='r')\n",
    "\n",
    "plt.legend()\n",
    "plt.ylim(0, 1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:38:19.683000",
     "start_time": "2016-07-08T02:38:19.452000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(10000L, 2L, 4L)"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Predict bounding boxes on the test images.\n",
    "pred_y = model.predict(test_X)\n",
    "pred_bboxes = pred_y * img_size\n",
    "pred_bboxes = pred_bboxes.reshape(len(pred_bboxes), num_objects, -1)\n",
    "pred_bboxes.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:38:27.705000",
     "start_time": "2016-07-08T02:38:26.975000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAArkAAACyCAYAAABGBtdeAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nNLkAAACoDk0uAAAAqvP/PLurt16sBz4AAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x18a2ef98>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Show a few images and predicted bounding boxes from the test dataset. \n",
    "plt.figure(figsize=(12, 3))\n",
    "for i_subplot in range(1, 5):\n",
    "    plt.subplot(1, 4, i_subplot)\n",
    "    i = np.random.randint(len(test_X))\n",
    "    plt.imshow(test_imgs[i].T, cmap='Greys', interpolation='none', origin='lower', extent=[0, img_size, 0, img_size])\n",
    "    for pred_bbox, exp_bbox in zip(pred_bboxes[i], test_bboxes[i]):\n",
    "        plt.gca().add_patch(matplotlib.patches.Rectangle((pred_bbox[0], pred_bbox[1]), pred_bbox[2], pred_bbox[3], ec='r', fc='none'))\n",
    "\n",
    "# plt.savefig('plots/bw-two-rectangles_with-flipping.png', dpi=300)"
   ]
  },
  {
   "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": "2.7.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}