AllShape_MultiBox.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "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": 96,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((1000, 16, 16, 3), (1000, 3, 4), (1000, 3), (1000, 3))"
      ]
     },
     "execution_count": 96,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import cairo\n",
    "num_imgs = 1000\n",
    "\n",
    "img_size = 16\n",
    "min_object_size = 2\n",
    "max_object_size = 8\n",
    "num_objects = 3\n",
    "\n",
    "bboxes = np.zeros((num_imgs, num_objects, 4))\n",
    "imgs = np.zeros((num_imgs, img_size, img_size, 4), dtype=np.uint8)  # format: BGRA\n",
    "shapes = np.zeros((num_imgs, num_objects), dtype=int)\n",
    "num_shapes = 3\n",
    "shape_labels = ['rectangle', 'circle', 'triangle']\n",
    "colors = np.zeros((num_imgs, num_objects), dtype=int)\n",
    "num_colors = 3\n",
    "color_labels = ['r', 'g', 'b']\n",
    "\n",
    "for i_img in range(num_imgs):\n",
    "    surface = cairo.ImageSurface.create_for_data(imgs[i_img], cairo.FORMAT_ARGB32, img_size, img_size)\n",
    "    cr = cairo.Context(surface)\n",
    "\n",
    "    # Fill background white.\n",
    "    cr.set_source_rgb(1, 1, 1)\n",
    "    cr.paint()\n",
    "    \n",
    "    # TODO: Try no overlap here.\n",
    "    # Draw random shapes.\n",
    "    for i_object in range(num_objects):\n",
    "        shape = np.random.randint(num_shapes)\n",
    "        shapes[i_img, i_object] = shape\n",
    "        if shape == 0:  # rectangle\n",
    "            w, h = np.random.randint(min_object_size, max_object_size, size=2)\n",
    "            x = np.random.randint(0, img_size - w)\n",
    "            y = np.random.randint(0, img_size - h)\n",
    "            bboxes[i_img, i_object] = [x, y, w, h]\n",
    "            cr.rectangle(x, y, w, h)            \n",
    "        elif shape == 1:  # circle   \n",
    "            r = 0.5 * np.random.randint(min_object_size, max_object_size)\n",
    "            x = np.random.randint(r, img_size - r)\n",
    "            y = np.random.randint(r, img_size - r)\n",
    "            bboxes[i_img, i_object] = [x - r, y - r, 2 * r, 2 * r]\n",
    "            cr.arc(x, y, r, 0, 2*np.pi)\n",
    "        elif shape == 2:  # triangle\n",
    "            w, h = np.random.randint(min_object_size, max_object_size, size=2)\n",
    "            x = np.random.randint(0, img_size - w)\n",
    "            y = np.random.randint(0, img_size - h)\n",
    "            bboxes[i_img, i_object] = [x, y, w, h]\n",
    "            cr.move_to(x, y)\n",
    "            cr.line_to(x+w, y)\n",
    "            cr.line_to(x+w, y+h)\n",
    "            cr.line_to(x, y)\n",
    "            cr.close_path()\n",
    "        \n",
    "        # Put some variation to the colors by adding a small random offset to the rgb values.\n",
    "        color = np.random.randint(num_colors)\n",
    "        colors[i_img, i_object] = color\n",
    "        max_offset = 0.3\n",
    "        r_offset, g_offset, b_offset = max_offset * 2. * (np.random.rand(3) - 0.5)\n",
    "        if color == 0:\n",
    "            cr.set_source_rgb(1-max_offset+r_offset, 0+g_offset, 0+b_offset)\n",
    "        elif color == 1:\n",
    "            cr.set_source_rgb(0+r_offset, 1-max_offset+g_offset, 0+b_offset)\n",
    "        elif color == 2:\n",
    "            cr.set_source_rgb(0+r_offset, 0-max_offset+g_offset, 1+b_offset)\n",
    "        cr.fill()\n",
    "        \n",
    "imgs = imgs[..., 2::-1]  # is BGRA, convert to RGB\n",
    "\n",
    "imgs.shape, bboxes.shape, shapes.shape, colors.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 97,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f0a4e9e0450>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "i = 2\n",
    "plt.imshow(imgs[i], interpolation='none', origin='lower', extent=[0, img_size, 0, img_size])\n",
    "for bbox, shape, color in zip(bboxes[i], shapes[i], colors[i]):\n",
    "    plt.gca().add_patch(matplotlib.patches.Rectangle((bbox[0], bbox[1]), bbox[2], bbox[3], ec='k', fc='none'))\n",
    "    plt.annotate(shape_labels[shape], (bbox[0], bbox[1] + bbox[3] + 0.7), color=color_labels[color], clip_on=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 98,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(1000, 3, 16, 16)\n"
     ]
    }
   ],
   "source": [
    "imgs = np.transpose(imgs,(0,3,1,2))\n",
    "print imgs.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 99,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((1000, 3, 16, 16), 0.37750297691993473, 0.29189804501498112)"
      ]
     },
     "execution_count": 99,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "X = (imgs - 128.) / 255.\n",
    "X.shape, np.mean(X), np.std(X)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 100,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((1000, 30), True)"
      ]
     },
     "execution_count": 100,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "colors_onehot = np.zeros((num_imgs, num_objects, num_colors))\n",
    "for i_img in range(num_imgs):\n",
    "    for i_object in range(num_objects):\n",
    "        colors_onehot[i_img, i_object, colors[i_img, i_object]] = 1\n",
    "\n",
    "shapes_onehot = np.zeros((num_imgs, num_objects, num_shapes))\n",
    "for i_img in range(num_imgs):\n",
    "    for i_object in range(num_objects):\n",
    "        shapes_onehot[i_img, i_object, shapes[i_img, i_object]] = 1\n",
    "        \n",
    "y = np.concatenate([bboxes / img_size, shapes_onehot, colors_onehot], axis=-1).reshape(num_imgs, -1)\n",
    "y.shape, np.all(np.argmax(colors_onehot, axis=-1) == colors)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 101,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(800, 3, 16, 16) (800, 30)\n"
     ]
    }
   ],
   "source": [
    "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:]\n",
    "print train_X.shape,train_y.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 102,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from torchsummary import summary\n",
    "import torch \n",
    "import torch.nn as nn\n",
    "from torch.autograd import Variable\n",
    "import torch.nn.functional as F"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 103,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "class CNNModel(nn.Module):\n",
    "    def __init__(self):\n",
    "        super(CNNModel,self).__init__()\n",
    "\n",
    "        #conv 1\n",
    "        self.cnn1=nn.Conv2d(in_channels=3, out_channels=32, kernel_size=3, stride=1, padding=2)\n",
    "        self.relu1=nn.ReLU()\n",
    "\n",
    "        #Maxpool 1\n",
    "        self.maxpool1=nn.MaxPool2d(kernel_size=2)\n",
    "\n",
    "        #conv 2\n",
    "        self.cnn2=nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, stride=1, padding=2)\n",
    "        self.relu2=nn.ReLU()\n",
    "\n",
    "        #Maxpool 2\n",
    "        self.maxpool2=nn.MaxPool2d(kernel_size=2)\n",
    "        \n",
    "        #conv 3\n",
    "        self.cnn3=nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=2)\n",
    "        self.relu3=nn.ReLU()\n",
    "\n",
    "        #Maxpool 3\n",
    "        self.maxpool3=nn.MaxPool2d(kernel_size=2)\n",
    "        \n",
    "        #FC 1\n",
    "        self.fc1=nn.Linear(1152,256)\n",
    "        self.relu4=nn.ReLU()\n",
    "            \n",
    "        #Drop out \n",
    "        self.dropout1 = nn.Dropout(0.2)\n",
    "        \n",
    "        #FC 2\n",
    "        self.fc2=nn.Linear(256,30)\n",
    "        \n",
    "\n",
    "    def forward(self,x):\n",
    "        '''\n",
    "        ori=x[0][0].data.numpy()\n",
    "        print ('original : ',ori.shape)\n",
    "        plt.imshow(ori, cmap=plt.get_cmap('gray'))\n",
    "        plt.show(),\n",
    "        '''\n",
    "        #conv1\n",
    "        out=self.cnn1(x)\n",
    "        out=self.relu1(out)\n",
    "        '''\n",
    "        #print type(out)\n",
    "        intm=out[0][0].data.numpy()\n",
    "        print ('conv 1: ',intm.shape)\n",
    "        #imgplot = plt.imshow(intm)\n",
    "        plt.imshow(intm, cmap=plt.get_cmap('gray'))\n",
    "        plt.show(),\n",
    "        '''\n",
    "        \n",
    "        #max poo1 1\n",
    "        out=self.maxpool1(out)\n",
    "        '''\n",
    "        ap=out[0][0].data.numpy()\n",
    "        print ('maxpool 1 : ',ap.shape)\n",
    "        plt.imshow(ap, cmap=plt.get_cmap('gray'))\n",
    "        plt.show(),\n",
    "        '''\n",
    "        \n",
    "        \n",
    "        #conv2\n",
    "        out=self.cnn2(out)\n",
    "        out=self.relu2(out)\n",
    "        '''\n",
    "        intm=out[0][0].data.numpy()\n",
    "        print ('conv 2: ',intm.shape)\n",
    "        imgplot = plt.imshow(intm)\n",
    "        plt.imshow(intm, cmap=plt.get_cmap('gray'))\n",
    "        plt.show(),\n",
    "        '''\n",
    "        \n",
    "        #max poo1 2\n",
    "        out=self.maxpool2(out)\n",
    "        '''\n",
    "        ap=out[0][0].data.numpy()\n",
    "        print ('maxpool 2 : ',ap.shape)\n",
    "        plt.imshow(ap, cmap=plt.get_cmap('gray'))\n",
    "        plt.show()\n",
    "        '''\n",
    "        \n",
    "        #conv3\n",
    "        out=self.cnn3(out)\n",
    "        out=self.relu3(out)\n",
    "        '''\n",
    "        intm=out[0][0].data.numpy()\n",
    "        print ('conv 3: ',intm.shape)\n",
    "        imgplot = plt.imshow(intm)\n",
    "        plt.imshow(intm, cmap=plt.get_cmap('gray'))\n",
    "        plt.show(),\n",
    "        '''\n",
    "        \n",
    "        #max poo1 3\n",
    "        out=self.maxpool3(out)\n",
    "        \"\"\"ap=out[0][0].data.numpy()\n",
    "        print ('maxpool 3 : ',ap.shape)\n",
    "        plt.imshow(ap, cmap=plt.get_cmap('gray'))\n",
    "        plt.show()\n",
    "        \"\"\"\n",
    "        \n",
    "        out=out.view(out.size(0),-1)        \n",
    "        #fc1\n",
    "        out=self.fc1(out)\n",
    "        out=self.relu4(out) \n",
    "        \n",
    "        \n",
    "        #dropout\n",
    "        out=self.dropout1(out)\n",
    "        \n",
    "        #fc2\n",
    "        out=self.fc2(out)\n",
    "        \n",
    "        return out"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 104,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "----------------------------------------------------------------\n",
      "        Layer (type)               Output Shape         Param #\n",
      "================================================================\n",
      "            Conv2d-1        [-1, 32L, 18L, 18L]             896\n",
      "              ReLU-2        [-1, 32L, 18L, 18L]               0\n",
      "         MaxPool2d-3          [-1, 32L, 9L, 9L]               0\n",
      "            Conv2d-4        [-1, 64L, 11L, 11L]           18496\n",
      "              ReLU-5        [-1, 64L, 11L, 11L]               0\n",
      "         MaxPool2d-6          [-1, 64L, 5L, 5L]               0\n",
      "            Conv2d-7         [-1, 128L, 7L, 7L]           73856\n",
      "              ReLU-8         [-1, 128L, 7L, 7L]               0\n",
      "         MaxPool2d-9         [-1, 128L, 3L, 3L]               0\n",
      "           Linear-10                 [-1, 256L]          295168\n",
      "             ReLU-11                 [-1, 256L]               0\n",
      "          Dropout-12                 [-1, 256L]               0\n",
      "           Linear-13                  [-1, 30L]            7710\n",
      "================================================================\n",
      "Total params: 396126\n",
      "Trainable params: 396126\n",
      "Non-trainable params: 0\n",
      "----------------------------------------------------------------\n",
      "None\n"
     ]
    }
   ],
   "source": [
    "model = CNNModel()\n",
    "\n",
    "print summary(model, (3, 16,16))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 105,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "model=CNNModel()\n",
    "criterion=nn.MSELoss()\n",
    "learning_rate=0.01\n",
    "optimizer=torch.optim.Adam(model.parameters(),lr=learning_rate)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 106,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "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]  # TODO: Check if its more performant if tensor elements are accessed directly below.\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 dist(bbox1, bbox2):\n",
    "    return np.sqrt(np.sum(np.square(bbox1[:2] - bbox2[:2])))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 107,
   "metadata": {},
   "outputs": [],
   "source": [
    "num_epochs_flipping = 100\n",
    "\n",
    "flipped_train_y = np.array(train_y)\n",
    "flipped = np.zeros((len(train_y), num_epochs_flipping ))\n",
    "ious_epoch = np.zeros((len(train_y), num_epochs_flipping ))\n",
    "dists_epoch = np.zeros((len(train_y), num_epochs_flipping ))\n",
    "mses_epoch = np.zeros((len(train_y), num_epochs_flipping ))\n",
    "acc_shapes_epoch = np.zeros((len(train_y), num_epochs_flipping ))\n",
    "acc_colors_epoch = np.zeros((len(train_y), num_epochs_flipping ))\n",
    "#print flipped,ious_epoch,flipped.shape,ious_epoch.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 108,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "epoch 0, loss 0.243518233299\n",
      "Mean IOU: 0.0\n",
      "Mean dist: 10.2385147444\n",
      "Mean mse: 48.2319692879\n",
      "Accuracy shapes: 0.33875\n",
      "Accuracy colors: 0.3325\n",
      "----------------------------------------------------------------\n",
      "epoch 1, loss 48.1049575806\n",
      "Mean IOU: 0.00271587529485\n",
      "Mean dist: 0.433211106271\n",
      "Mean mse: 0.0891320936352\n",
      "Accuracy shapes: 0.345833333333\n",
      "Accuracy colors: 0.331666666667\n",
      "----------------------------------------------------------------\n",
      "epoch 2, loss 0.214309871197\n",
      "Mean IOU: 0.0\n",
      "Mean dist: 0.520148350472\n",
      "Mean mse: 0.125893356001\n",
      "Accuracy shapes: 0.320833333333\n",
      "Accuracy colors: 0.335833333333\n",
      "----------------------------------------------------------------\n",
      "epoch 3, loss 0.250291883945\n",
      "Mean IOU: 0.0\n",
      "Mean dist: 0.525583667173\n",
      "Mean mse: 0.128562310383\n",
      "Accuracy shapes: 0.327916666667\n",
      "Accuracy colors: 0.335833333333\n",
      "----------------------------------------------------------------\n",
      "epoch 4, loss 0.253443270922\n",
      "Mean IOU: 0.0\n",
      "Mean dist: 0.525359451079\n",
      "Mean mse: 0.128635441969\n",
      "Accuracy shapes: 0.32875\n",
      "Accuracy colors: 0.335833333333\n",
      "----------------------------------------------------------------\n",
      "epoch 5, loss 0.25322291255\n",
      "Mean IOU: 0.0\n",
      "Mean dist: 0.519963904815\n",
      "Mean mse: 0.126305194354\n",
      "Accuracy shapes: 0.33125\n",
      "Accuracy colors: 0.335833333333\n",
      "----------------------------------------------------------------\n",
      "epoch 6, loss 0.250451922417\n",
      "Mean IOU: 5.77047031664e-08\n",
      "Mean dist: 0.512399267335\n",
      "Mean mse: 0.122856400562\n",
      "Accuracy shapes: 0.325\n",
      "Accuracy colors: 0.335416666667\n",
      "----------------------------------------------------------------\n",
      "epoch 7, loss 0.246468320489\n",
      "Mean IOU: 1.0329900811e-05\n",
      "Mean dist: 0.501439060945\n",
      "Mean mse: 0.117813689057\n",
      "Accuracy shapes: 0.326666666667\n",
      "Accuracy colors: 0.334583333333\n",
      "----------------------------------------------------------------\n",
      "epoch 8, loss 0.241068214178\n",
      "Mean IOU: 0.000142446379904\n",
      "Mean dist: 0.488211650497\n",
      "Mean mse: 0.11170718332\n",
      "Accuracy shapes: 0.329166666667\n",
      "Accuracy colors: 0.332083333333\n",
      "----------------------------------------------------------------\n",
      "epoch 9, loss 0.234581977129\n",
      "Mean IOU: 0.000499223835972\n",
      "Mean dist: 0.469997932587\n",
      "Mean mse: 0.103385747578\n",
      "Accuracy shapes: 0.33\n",
      "Accuracy colors: 0.329583333333\n",
      "----------------------------------------------------------------\n",
      "epoch 10, loss 0.226271241903\n",
      "Mean IOU: 0.00184835420171\n",
      "Mean dist: 0.448412373105\n",
      "Mean mse: 0.0938521578224\n",
      "Accuracy shapes: 0.340416666667\n",
      "Accuracy colors: 0.33125\n",
      "----------------------------------------------------------------\n",
      "epoch 11, loss 0.216226384044\n",
      "Mean IOU: 0.00652892631689\n",
      "Mean dist: 0.421678115553\n",
      "Mean mse: 0.0827595635836\n",
      "Accuracy shapes: 0.332916666667\n",
      "Accuracy colors: 0.331666666667\n",
      "----------------------------------------------------------------\n",
      "epoch 12, loss 0.204065516591\n",
      "Mean IOU: 0.0202717687126\n",
      "Mean dist: 0.387487607249\n",
      "Mean mse: 0.0689442994803\n",
      "Accuracy shapes: 0.337916666667\n",
      "Accuracy colors: 0.335\n",
      "----------------------------------------------------------------\n",
      "epoch 13, loss 0.1890963763\n",
      "Mean IOU: 0.0463790830758\n",
      "Mean dist: 0.347621404795\n",
      "Mean mse: 0.0537181894975\n",
      "Accuracy shapes: 0.327916666667\n",
      "Accuracy colors: 0.33375\n",
      "----------------------------------------------------------------\n",
      "epoch 14, loss 0.173093602061\n",
      "Mean IOU: 0.088481510362\n",
      "Mean dist: 0.305226199788\n",
      "Mean mse: 0.0392197100122\n",
      "Accuracy shapes: 0.327916666667\n",
      "Accuracy colors: 0.30875\n",
      "----------------------------------------------------------------\n",
      "epoch 15, loss 0.157894730568\n",
      "Mean IOU: 0.126137785668\n",
      "Mean dist: 0.281115486454\n",
      "Mean mse: 0.0318612796229\n",
      "Accuracy shapes: 0.352916666667\n",
      "Accuracy colors: 0.324166666667\n",
      "----------------------------------------------------------------\n",
      "epoch 16, loss 0.1490996629\n",
      "Mean IOU: 0.128507212618\n",
      "Mean dist: 0.289333520219\n",
      "Mean mse: 0.0386121815693\n",
      "Accuracy shapes: 0.329583333333\n",
      "Accuracy colors: 0.33625\n",
      "----------------------------------------------------------------\n",
      "epoch 17, loss 0.153965622187\n",
      "Mean IOU: 0.108720015881\n",
      "Mean dist: 0.316921944856\n",
      "Mean mse: 0.0503278431522\n",
      "Accuracy shapes: 0.336666666667\n",
      "Accuracy colors: 0.324166666667\n",
      "----------------------------------------------------------------\n",
      "epoch 18, loss 0.165214315057\n",
      "Mean IOU: 0.10246814531\n",
      "Mean dist: 0.328126540429\n",
      "Mean mse: 0.0531447706015\n",
      "Accuracy shapes: 0.329166666667\n",
      "Accuracy colors: 0.3275\n",
      "----------------------------------------------------------------\n",
      "epoch 19, loss 0.169459164143\n",
      "Mean IOU: 0.110402197903\n",
      "Mean dist: 0.314100057445\n",
      "Mean mse: 0.0454202037829\n",
      "Accuracy shapes: 0.322916666667\n",
      "Accuracy colors: 0.337083333333\n",
      "----------------------------------------------------------------\n",
      "epoch 20, loss 0.16132889688\n",
      "Mean IOU: 0.125064535276\n",
      "Mean dist: 0.290763424873\n",
      "Mean mse: 0.0360215657728\n",
      "Accuracy shapes: 0.349166666667\n",
      "Accuracy colors: 0.332916666667\n",
      "----------------------------------------------------------------\n",
      "epoch 21, loss 0.152993127704\n",
      "Mean IOU: 0.119817779323\n",
      "Mean dist: 0.280127759708\n",
      "Mean mse: 0.0310024148969\n",
      "Accuracy shapes: 0.33875\n",
      "Accuracy colors: 0.342916666667\n",
      "----------------------------------------------------------------\n",
      "epoch 22, loss 0.148418784142\n",
      "Mean IOU: 0.107727439574\n",
      "Mean dist: 0.279597578677\n",
      "Mean mse: 0.0313227833968\n",
      "Accuracy shapes: 0.333333333333\n",
      "Accuracy colors: 0.340416666667\n",
      "----------------------------------------------------------------\n",
      "epoch 23, loss 0.148261517286\n",
      "Mean IOU: 0.0896089820847\n",
      "Mean dist: 0.288527390062\n",
      "Mean mse: 0.0348842152668\n",
      "Accuracy shapes: 0.317916666667\n",
      "Accuracy colors: 0.349166666667\n",
      "----------------------------------------------------------------\n",
      "epoch 24, loss 0.150935664773\n",
      "Mean IOU: 0.0781153503913\n",
      "Mean dist: 0.291636752859\n",
      "Mean mse: 0.0369708370892\n",
      "Accuracy shapes: 0.34625\n",
      "Accuracy colors: 0.349583333333\n",
      "----------------------------------------------------------------\n",
      "epoch 25, loss 0.153300762177\n",
      "Mean IOU: 0.0773842696625\n",
      "Mean dist: 0.292943864413\n",
      "Mean mse: 0.0379857050473\n",
      "Accuracy shapes: 0.339166666667\n",
      "Accuracy colors: 0.35375\n",
      "----------------------------------------------------------------\n",
      "epoch 26, loss 0.154279008508\n",
      "Mean IOU: 0.0751158001191\n",
      "Mean dist: 0.293176356724\n",
      "Mean mse: 0.0382416484575\n",
      "Accuracy shapes: 0.34375\n",
      "Accuracy colors: 0.360416666667\n",
      "----------------------------------------------------------------\n",
      "epoch 27, loss 0.154400244355\n",
      "Mean IOU: 0.0793582361607\n",
      "Mean dist: 0.288962380022\n",
      "Mean mse: 0.0368730310401\n",
      "Accuracy shapes: 0.345416666667\n",
      "Accuracy colors: 0.3475\n",
      "----------------------------------------------------------------\n",
      "epoch 28, loss 0.153370454907\n",
      "Mean IOU: 0.0844633627212\n",
      "Mean dist: 0.285843789251\n",
      "Mean mse: 0.0357333425582\n",
      "Accuracy shapes: 0.33875\n",
      "Accuracy colors: 0.36125\n",
      "----------------------------------------------------------------\n",
      "epoch 29, loss 0.151578813791\n",
      "Mean IOU: 0.0946329600526\n",
      "Mean dist: 0.280997227402\n",
      "Mean mse: 0.0340647752657\n",
      "Accuracy shapes: 0.336666666667\n",
      "Accuracy colors: 0.344166666667\n",
      "----------------------------------------------------------------\n",
      "epoch 30, loss 0.150616630912\n",
      "Mean IOU: 0.108874921525\n",
      "Mean dist: 0.276245871771\n",
      "Mean mse: 0.0324126806193\n",
      "Accuracy shapes: 0.335416666667\n",
      "Accuracy colors: 0.350416666667\n",
      "----------------------------------------------------------------\n",
      "epoch 31, loss 0.1486787498\n",
      "Mean IOU: 0.117540690708\n",
      "Mean dist: 0.275760717925\n",
      "Mean mse: 0.0311280248731\n",
      "Accuracy shapes: 0.330833333333\n",
      "Accuracy colors: 0.34875\n",
      "----------------------------------------------------------------\n",
      "epoch 32, loss 0.148213252425\n",
      "Mean IOU: 0.127687933789\n",
      "Mean dist: 0.27553591991\n",
      "Mean mse: 0.0311541732359\n",
      "Accuracy shapes: 0.338333333333\n",
      "Accuracy colors: 0.3475\n",
      "----------------------------------------------------------------\n",
      "epoch 33, loss 0.148414433002\n",
      "Mean IOU: 0.132176718326\n",
      "Mean dist: 0.275224676694\n",
      "Mean mse: 0.0309502537573\n",
      "Accuracy shapes: 0.329166666667\n",
      "Accuracy colors: 0.347916666667\n",
      "----------------------------------------------------------------\n",
      "epoch 34, loss 0.148909017444\n",
      "Mean IOU: 0.137317636819\n",
      "Mean dist: 0.277510200875\n",
      "Mean mse: 0.0315628241274\n",
      "Accuracy shapes: 0.328333333333\n",
      "Accuracy colors: 0.33625\n",
      "----------------------------------------------------------------\n",
      "epoch 35, loss 0.150297865272\n",
      "Mean IOU: 0.137007239122\n",
      "Mean dist: 0.278225031279\n",
      "Mean mse: 0.0321568061091\n",
      "Accuracy shapes: 0.31625\n",
      "Accuracy colors: 0.35\n",
      "----------------------------------------------------------------\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "epoch 36, loss 0.150190189481\n",
      "Mean IOU: 0.140198574647\n",
      "Mean dist: 0.276904196246\n",
      "Mean mse: 0.03163590149\n",
      "Accuracy shapes: 0.328333333333\n",
      "Accuracy colors: 0.348333333333\n",
      "----------------------------------------------------------------\n",
      "epoch 37, loss 0.149491488934\n",
      "Mean IOU: 0.135782391211\n",
      "Mean dist: 0.276307946901\n",
      "Mean mse: 0.0319212725971\n",
      "Accuracy shapes: 0.344166666667\n",
      "Accuracy colors: 0.341666666667\n",
      "----------------------------------------------------------------\n",
      "epoch 38, loss 0.148268818855\n",
      "Mean IOU: 0.139723241607\n",
      "Mean dist: 0.272340940573\n",
      "Mean mse: 0.0307430154904\n",
      "Accuracy shapes: 0.338333333333\n",
      "Accuracy colors: 0.35375\n",
      "----------------------------------------------------------------\n",
      "epoch 39, loss 0.148283794522\n",
      "Mean IOU: 0.137163038245\n",
      "Mean dist: 0.270436670718\n",
      "Mean mse: 0.0299974031593\n",
      "Accuracy shapes: 0.339583333333\n",
      "Accuracy colors: 0.33875\n",
      "----------------------------------------------------------------\n",
      "epoch 40, loss 0.147818669677\n",
      "Mean IOU: 0.133115877483\n",
      "Mean dist: 0.270029101156\n",
      "Mean mse: 0.030221434354\n",
      "Accuracy shapes: 0.339583333333\n",
      "Accuracy colors: 0.34\n",
      "----------------------------------------------------------------\n",
      "epoch 41, loss 0.147420048714\n",
      "Mean IOU: 0.133043313658\n",
      "Mean dist: 0.270566239546\n",
      "Mean mse: 0.0303996879387\n",
      "Accuracy shapes: 0.3425\n",
      "Accuracy colors: 0.339583333333\n",
      "----------------------------------------------------------------\n",
      "epoch 42, loss 0.147689729929\n",
      "Mean IOU: 0.132240453478\n",
      "Mean dist: 0.270560124549\n",
      "Mean mse: 0.0307291523167\n",
      "Accuracy shapes: 0.327916666667\n",
      "Accuracy colors: 0.338333333333\n",
      "----------------------------------------------------------------\n",
      "epoch 43, loss 0.147824794054\n",
      "Mean IOU: 0.129094186458\n",
      "Mean dist: 0.273552055376\n",
      "Mean mse: 0.0312795198741\n",
      "Accuracy shapes: 0.349583333333\n",
      "Accuracy colors: 0.347083333333\n",
      "----------------------------------------------------------------\n",
      "epoch 44, loss 0.148255601525\n",
      "Mean IOU: 0.130451364635\n",
      "Mean dist: 0.267161905847\n",
      "Mean mse: 0.0301833447618\n",
      "Accuracy shapes: 0.34\n",
      "Accuracy colors: 0.33625\n",
      "----------------------------------------------------------------\n",
      "epoch 45, loss 0.148455619812\n",
      "Mean IOU: 0.126990809219\n",
      "Mean dist: 0.271933042457\n",
      "Mean mse: 0.0309516700061\n",
      "Accuracy shapes: 0.340833333333\n",
      "Accuracy colors: 0.342083333333\n",
      "----------------------------------------------------------------\n",
      "epoch 46, loss 0.148421704769\n",
      "Mean IOU: 0.133684932423\n",
      "Mean dist: 0.268800657307\n",
      "Mean mse: 0.0306468624904\n",
      "Accuracy shapes: 0.345416666667\n",
      "Accuracy colors: 0.341666666667\n",
      "----------------------------------------------------------------\n",
      "epoch 47, loss 0.14776904881\n",
      "Mean IOU: 0.135683719006\n",
      "Mean dist: 0.268010149945\n",
      "Mean mse: 0.030158445632\n",
      "Accuracy shapes: 0.334583333333\n",
      "Accuracy colors: 0.343333333333\n",
      "----------------------------------------------------------------\n",
      "epoch 48, loss 0.147641658783\n",
      "Mean IOU: 0.137646292927\n",
      "Mean dist: 0.26780944052\n",
      "Mean mse: 0.0299958563695\n",
      "Accuracy shapes: 0.34375\n",
      "Accuracy colors: 0.34\n",
      "----------------------------------------------------------------\n",
      "epoch 49, loss 0.147490739822\n",
      "Mean IOU: 0.139328525507\n",
      "Mean dist: 0.268765325835\n",
      "Mean mse: 0.0304408041343\n",
      "Accuracy shapes: 0.338333333333\n",
      "Accuracy colors: 0.345416666667\n",
      "----------------------------------------------------------------\n",
      "epoch 50, loss 0.147541731596\n",
      "Mean IOU: 0.14355965413\n",
      "Mean dist: 0.269750643682\n",
      "Mean mse: 0.0307287097375\n",
      "Accuracy shapes: 0.348333333333\n",
      "Accuracy colors: 0.35375\n",
      "----------------------------------------------------------------\n",
      "epoch 51, loss 0.147484764457\n",
      "Mean IOU: 0.145370771089\n",
      "Mean dist: 0.26703550184\n",
      "Mean mse: 0.0303648278171\n",
      "Accuracy shapes: 0.343333333333\n",
      "Accuracy colors: 0.334583333333\n",
      "----------------------------------------------------------------\n",
      "epoch 52, loss 0.147584244609\n",
      "Mean IOU: 0.150250961456\n",
      "Mean dist: 0.26723606725\n",
      "Mean mse: 0.0303213387435\n",
      "Accuracy shapes: 0.3475\n",
      "Accuracy colors: 0.348333333333\n",
      "----------------------------------------------------------------\n",
      "epoch 53, loss 0.147904232144\n",
      "Mean IOU: 0.151740314187\n",
      "Mean dist: 0.26627086128\n",
      "Mean mse: 0.030487305284\n",
      "Accuracy shapes: 0.339166666667\n",
      "Accuracy colors: 0.353333333333\n",
      "----------------------------------------------------------------\n",
      "epoch 54, loss 0.147395178676\n",
      "Mean IOU: 0.148239611503\n",
      "Mean dist: 0.267509763426\n",
      "Mean mse: 0.0305691031505\n",
      "Accuracy shapes: 0.329583333333\n",
      "Accuracy colors: 0.34625\n",
      "----------------------------------------------------------------\n",
      "epoch 55, loss 0.147719144821\n",
      "Mean IOU: 0.148122964292\n",
      "Mean dist: 0.26627870975\n",
      "Mean mse: 0.0300311248681\n",
      "Accuracy shapes: 0.3425\n",
      "Accuracy colors: 0.3575\n",
      "----------------------------------------------------------------\n",
      "epoch 56, loss 0.147556588054\n",
      "Mean IOU: 0.149944915411\n",
      "Mean dist: 0.26552918603\n",
      "Mean mse: 0.0301978665164\n",
      "Accuracy shapes: 0.3475\n",
      "Accuracy colors: 0.354166666667\n",
      "----------------------------------------------------------------\n",
      "epoch 57, loss 0.147107675672\n",
      "Mean IOU: 0.143834818063\n",
      "Mean dist: 0.265289748393\n",
      "Mean mse: 0.02977000375\n",
      "Accuracy shapes: 0.345833333333\n",
      "Accuracy colors: 0.354583333333\n",
      "----------------------------------------------------------------\n",
      "epoch 58, loss 0.147237077355\n",
      "Mean IOU: 0.141964069565\n",
      "Mean dist: 0.263682732192\n",
      "Mean mse: 0.0295182845764\n",
      "Accuracy shapes: 0.325833333333\n",
      "Accuracy colors: 0.339166666667\n",
      "----------------------------------------------------------------\n",
      "epoch 59, loss 0.147136643529\n",
      "Mean IOU: 0.140822343167\n",
      "Mean dist: 0.26326638008\n",
      "Mean mse: 0.0295256680661\n",
      "Accuracy shapes: 0.336666666667\n",
      "Accuracy colors: 0.3525\n",
      "----------------------------------------------------------------\n",
      "epoch 60, loss 0.147499993443\n",
      "Mean IOU: 0.137011877451\n",
      "Mean dist: 0.26510634141\n",
      "Mean mse: 0.0299578231037\n",
      "Accuracy shapes: 0.320833333333\n",
      "Accuracy colors: 0.353333333333\n",
      "----------------------------------------------------------------\n",
      "epoch 61, loss 0.147381097078\n",
      "Mean IOU: 0.136758880346\n",
      "Mean dist: 0.266270376855\n",
      "Mean mse: 0.0300393046554\n",
      "Accuracy shapes: 0.31625\n",
      "Accuracy colors: 0.3675\n",
      "----------------------------------------------------------------\n",
      "epoch 62, loss 0.147157952189\n",
      "Mean IOU: 0.137844567352\n",
      "Mean dist: 0.264015757502\n",
      "Mean mse: 0.029754787466\n",
      "Accuracy shapes: 0.335416666667\n",
      "Accuracy colors: 0.359166666667\n",
      "----------------------------------------------------------------\n",
      "epoch 63, loss 0.147379919887\n",
      "Mean IOU: 0.139203433402\n",
      "Mean dist: 0.263663080639\n",
      "Mean mse: 0.029722110638\n",
      "Accuracy shapes: 0.35125\n",
      "Accuracy colors: 0.331666666667\n",
      "----------------------------------------------------------------\n",
      "epoch 64, loss 0.147663235664\n",
      "Mean IOU: 0.141776527219\n",
      "Mean dist: 0.262116920388\n",
      "Mean mse: 0.0293093592859\n",
      "Accuracy shapes: 0.3225\n",
      "Accuracy colors: 0.340833333333\n",
      "----------------------------------------------------------------\n",
      "epoch 65, loss 0.147089719772\n",
      "Mean IOU: 0.140731725534\n",
      "Mean dist: 0.2662599506\n",
      "Mean mse: 0.0299690405716\n",
      "Accuracy shapes: 0.31375\n",
      "Accuracy colors: 0.359166666667\n",
      "----------------------------------------------------------------\n",
      "epoch 66, loss 0.147171422839\n",
      "Mean IOU: 0.139750331076\n",
      "Mean dist: 0.263199710544\n",
      "Mean mse: 0.0293887373993\n",
      "Accuracy shapes: 0.332916666667\n",
      "Accuracy colors: 0.350416666667\n",
      "----------------------------------------------------------------\n",
      "epoch 67, loss 0.146905064583\n",
      "Mean IOU: 0.145157647712\n",
      "Mean dist: 0.264456104567\n",
      "Mean mse: 0.0297201778025\n",
      "Accuracy shapes: 0.337916666667\n",
      "Accuracy colors: 0.360833333333\n",
      "----------------------------------------------------------------\n",
      "epoch 68, loss 0.147062435746\n",
      "Mean IOU: 0.144298334805\n",
      "Mean dist: 0.264112484558\n",
      "Mean mse: 0.0296180161957\n",
      "Accuracy shapes: 0.336666666667\n",
      "Accuracy colors: 0.356666666667\n",
      "----------------------------------------------------------------\n",
      "epoch 69, loss 0.147209823132\n",
      "Mean IOU: 0.149538284336\n",
      "Mean dist: 0.261985687897\n",
      "Mean mse: 0.0291057976482\n",
      "Accuracy shapes: 0.332083333333\n",
      "Accuracy colors: 0.345416666667\n",
      "----------------------------------------------------------------\n",
      "epoch 70, loss 0.147025942802\n",
      "Mean IOU: 0.149668153101\n",
      "Mean dist: 0.264416809923\n",
      "Mean mse: 0.0296434541635\n",
      "Accuracy shapes: 0.335416666667\n",
      "Accuracy colors: 0.3525\n",
      "----------------------------------------------------------------\n",
      "epoch 71, loss 0.147151172161\n",
      "Mean IOU: 0.14439365286\n",
      "Mean dist: 0.266819599161\n",
      "Mean mse: 0.0302544965707\n",
      "Accuracy shapes: 0.325416666667\n",
      "Accuracy colors: 0.347916666667\n",
      "----------------------------------------------------------------\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "epoch 72, loss 0.147246822715\n",
      "Mean IOU: 0.14900103235\n",
      "Mean dist: 0.264306536533\n",
      "Mean mse: 0.0295666618577\n",
      "Accuracy shapes: 0.338333333333\n",
      "Accuracy colors: 0.352916666667\n",
      "----------------------------------------------------------------\n",
      "epoch 73, loss 0.147046878934\n",
      "Mean IOU: 0.144359381134\n",
      "Mean dist: 0.265903457042\n",
      "Mean mse: 0.029920403282\n",
      "Accuracy shapes: 0.337916666667\n",
      "Accuracy colors: 0.350833333333\n",
      "----------------------------------------------------------------\n",
      "epoch 74, loss 0.146823659539\n",
      "Mean IOU: 0.142046365631\n",
      "Mean dist: 0.265965135004\n",
      "Mean mse: 0.0301631691552\n",
      "Accuracy shapes: 0.346666666667\n",
      "Accuracy colors: 0.360416666667\n",
      "----------------------------------------------------------------\n",
      "epoch 75, loss 0.147230789065\n",
      "Mean IOU: 0.142833072298\n",
      "Mean dist: 0.264205296347\n",
      "Mean mse: 0.02970280001\n",
      "Accuracy shapes: 0.335\n",
      "Accuracy colors: 0.354583333333\n",
      "----------------------------------------------------------------\n",
      "epoch 76, loss 0.146893754601\n",
      "Mean IOU: 0.14309091518\n",
      "Mean dist: 0.264750672953\n",
      "Mean mse: 0.0296376107012\n",
      "Accuracy shapes: 0.3475\n",
      "Accuracy colors: 0.36375\n",
      "----------------------------------------------------------------\n",
      "epoch 77, loss 0.147187262774\n",
      "Mean IOU: 0.142559712917\n",
      "Mean dist: 0.266464256399\n",
      "Mean mse: 0.029851236947\n",
      "Accuracy shapes: 0.335416666667\n",
      "Accuracy colors: 0.34625\n",
      "----------------------------------------------------------------\n",
      "epoch 78, loss 0.147045254707\n",
      "Mean IOU: 0.139654419838\n",
      "Mean dist: 0.265024749164\n",
      "Mean mse: 0.029745310682\n",
      "Accuracy shapes: 0.345\n",
      "Accuracy colors: 0.349166666667\n",
      "----------------------------------------------------------------\n",
      "epoch 79, loss 0.146891474724\n",
      "Mean IOU: 0.137633403923\n",
      "Mean dist: 0.266483119033\n",
      "Mean mse: 0.0300375216606\n",
      "Accuracy shapes: 0.333333333333\n",
      "Accuracy colors: 0.349583333333\n",
      "----------------------------------------------------------------\n",
      "epoch 80, loss 0.146845564246\n",
      "Mean IOU: 0.138482338522\n",
      "Mean dist: 0.265698573917\n",
      "Mean mse: 0.029972007229\n",
      "Accuracy shapes: 0.34125\n",
      "Accuracy colors: 0.335416666667\n",
      "----------------------------------------------------------------\n",
      "epoch 81, loss 0.147066906095\n",
      "Mean IOU: 0.138038036218\n",
      "Mean dist: 0.268435292548\n",
      "Mean mse: 0.0304336820775\n",
      "Accuracy shapes: 0.348333333333\n",
      "Accuracy colors: 0.344583333333\n",
      "----------------------------------------------------------------\n",
      "epoch 82, loss 0.146698012948\n",
      "Mean IOU: 0.142599179785\n",
      "Mean dist: 0.265076753006\n",
      "Mean mse: 0.0295929724805\n",
      "Accuracy shapes: 0.347916666667\n",
      "Accuracy colors: 0.337083333333\n",
      "----------------------------------------------------------------\n",
      "epoch 83, loss 0.146707266569\n",
      "Mean IOU: 0.141882418363\n",
      "Mean dist: 0.266653295674\n",
      "Mean mse: 0.029947612748\n",
      "Accuracy shapes: 0.339166666667\n",
      "Accuracy colors: 0.358333333333\n",
      "----------------------------------------------------------------\n",
      "epoch 84, loss 0.146935939789\n",
      "Mean IOU: 0.150575016173\n",
      "Mean dist: 0.261333318227\n",
      "Mean mse: 0.0289796721079\n",
      "Accuracy shapes: 0.333333333333\n",
      "Accuracy colors: 0.35625\n",
      "----------------------------------------------------------------\n",
      "epoch 85, loss 0.146890372038\n",
      "Mean IOU: 0.145095794085\n",
      "Mean dist: 0.264459592555\n",
      "Mean mse: 0.0292610265514\n",
      "Accuracy shapes: 0.3275\n",
      "Accuracy colors: 0.369166666667\n",
      "----------------------------------------------------------------\n",
      "epoch 86, loss 0.147048056126\n",
      "Mean IOU: 0.143860360583\n",
      "Mean dist: 0.26642397168\n",
      "Mean mse: 0.0298800938217\n",
      "Accuracy shapes: 0.33875\n",
      "Accuracy colors: 0.359166666667\n",
      "----------------------------------------------------------------\n",
      "epoch 87, loss 0.14673410356\n",
      "Mean IOU: 0.143836611419\n",
      "Mean dist: 0.266232183518\n",
      "Mean mse: 0.0299595068496\n",
      "Accuracy shapes: 0.333333333333\n",
      "Accuracy colors: 0.345833333333\n",
      "----------------------------------------------------------------\n",
      "epoch 88, loss 0.147112548351\n",
      "Mean IOU: 0.144753380573\n",
      "Mean dist: 0.266302544891\n",
      "Mean mse: 0.0296441104118\n",
      "Accuracy shapes: 0.3425\n",
      "Accuracy colors: 0.35\n",
      "----------------------------------------------------------------\n",
      "epoch 89, loss 0.147193074226\n",
      "Mean IOU: 0.144005546728\n",
      "Mean dist: 0.264594800684\n",
      "Mean mse: 0.0293319278184\n",
      "Accuracy shapes: 0.340416666667\n",
      "Accuracy colors: 0.3575\n",
      "----------------------------------------------------------------\n",
      "epoch 90, loss 0.147010341287\n",
      "Mean IOU: 0.142524097454\n",
      "Mean dist: 0.264937451108\n",
      "Mean mse: 0.0296373943836\n",
      "Accuracy shapes: 0.343333333333\n",
      "Accuracy colors: 0.352083333333\n",
      "----------------------------------------------------------------\n",
      "epoch 91, loss 0.147210240364\n",
      "Mean IOU: 0.144404426293\n",
      "Mean dist: 0.265823848861\n",
      "Mean mse: 0.029642232392\n",
      "Accuracy shapes: 0.340833333333\n",
      "Accuracy colors: 0.351666666667\n",
      "----------------------------------------------------------------\n",
      "epoch 92, loss 0.146844655275\n",
      "Mean IOU: 0.140579258205\n",
      "Mean dist: 0.264541179034\n",
      "Mean mse: 0.029655717734\n",
      "Accuracy shapes: 0.340416666667\n",
      "Accuracy colors: 0.3525\n",
      "----------------------------------------------------------------\n",
      "epoch 93, loss 0.147137731314\n",
      "Mean IOU: 0.145016978539\n",
      "Mean dist: 0.261761634771\n",
      "Mean mse: 0.0290545110883\n",
      "Accuracy shapes: 0.335833333333\n",
      "Accuracy colors: 0.350833333333\n",
      "----------------------------------------------------------------\n",
      "epoch 94, loss 0.146929427981\n",
      "Mean IOU: 0.145768236651\n",
      "Mean dist: 0.264718337232\n",
      "Mean mse: 0.0296954203801\n",
      "Accuracy shapes: 0.329583333333\n",
      "Accuracy colors: 0.347916666667\n",
      "----------------------------------------------------------------\n",
      "epoch 95, loss 0.147042810917\n",
      "Mean IOU: 0.138399806205\n",
      "Mean dist: 0.267448960204\n",
      "Mean mse: 0.0299679738358\n",
      "Accuracy shapes: 0.340833333333\n",
      "Accuracy colors: 0.341666666667\n",
      "----------------------------------------------------------------\n",
      "epoch 96, loss 0.146807432175\n",
      "Mean IOU: 0.141807006828\n",
      "Mean dist: 0.264308500753\n",
      "Mean mse: 0.0294924854883\n",
      "Accuracy shapes: 0.332083333333\n",
      "Accuracy colors: 0.348333333333\n",
      "----------------------------------------------------------------\n",
      "epoch 97, loss 0.146532386541\n",
      "Mean IOU: 0.148479553664\n",
      "Mean dist: 0.264586882786\n",
      "Mean mse: 0.0296642301455\n",
      "Accuracy shapes: 0.345416666667\n",
      "Accuracy colors: 0.34375\n",
      "----------------------------------------------------------------\n",
      "epoch 98, loss 0.147038400173\n",
      "Mean IOU: 0.138770580193\n",
      "Mean dist: 0.265384521915\n",
      "Mean mse: 0.0297368635986\n",
      "Accuracy shapes: 0.33875\n",
      "Accuracy colors: 0.345416666667\n",
      "----------------------------------------------------------------\n",
      "epoch 99, loss 0.147038444877\n",
      "Mean IOU: 0.146671417254\n",
      "Mean dist: 0.264613970782\n",
      "Mean mse: 0.0296216244706\n",
      "Accuracy shapes: 0.353333333333\n",
      "Accuracy colors: 0.349166666667\n",
      "----------------------------------------------------------------\n"
     ]
    }
   ],
   "source": [
    "loss_list=[]\n",
    "for epoch in range(num_epochs_flipping):\n",
    "    #print 'Epoch', epoch\n",
    "    inputs=Variable(torch.FloatTensor(train_X))\n",
    "    #print 'input size {}'.format(inputs.shape)\n",
    "    labels=Variable(torch.FloatTensor(flipped_train_y))\n",
    "    optimizer.zero_grad()\n",
    "    outputs = model(inputs)\n",
    "    loss=criterion(outputs,labels)\n",
    "    loss.backward()\n",
    "    optimizer.step()\n",
    "    #if epoch%100==0:\n",
    "    print ('epoch {}, loss {}'.format(epoch,loss.data[0])) \n",
    "    loss_list.append(loss.data[0])\n",
    "\n",
    "    pred_y = model(Variable(torch.FloatTensor(train_X))).data.numpy()\n",
    "    #print 'pred {}'.format(pred_y)\n",
    "    for sample, (pred, exp) in enumerate(zip(pred_y, flipped_train_y)):\n",
    "        \n",
    "        # TODO: Make this simpler.\n",
    "        \n",
    "        pred = pred.reshape(num_objects, -1)\n",
    "        exp = exp.reshape(num_objects, -1)\n",
    "        #print 'Pred {}'.format(pred.shape)\n",
    "        #print 'exp {}'.format(exp.shape)\n",
    "        pred_bboxes = pred[:, :4]\n",
    "        exp_bboxes = exp[:, :4]\n",
    "        \n",
    "        ious = np.zeros((num_objects, num_objects))\n",
    "        dists = np.zeros((num_objects, num_objects))\n",
    "        mses = np.zeros((num_objects, num_objects))\n",
    "        #print 'sizes of ious dists mses {}'.format(ious.shape)\n",
    "        for i, exp_bbox in enumerate(exp_bboxes):\n",
    "            for j, pred_bbox in enumerate(pred_bboxes):\n",
    "                ious[i, j] = IOU(exp_bbox, pred_bbox)\n",
    "                dists[i, j] = dist(exp_bbox, pred_bbox)\n",
    "                mses[i, j] = np.mean(np.square(exp_bbox - pred_bbox))\n",
    "        #print ious\n",
    "                \n",
    "        new_order = np.zeros(num_objects, dtype=int)\n",
    "        \n",
    "        #print 'new_order {} '.format(new_order.shape)\n",
    "        for i in range(num_objects):\n",
    "            # Find pred and exp bbox with maximum iou and assign them to each other (i.e. switch the positions of the exp bboxes in y).\n",
    "            #print 'ious.argmax() {}'.format(ious.argmax())\n",
    "            ind_exp_bbox, ind_pred_bbox = np.unravel_index(ious.argmax(), ious.shape)\n",
    "            #print 'ind_exp_bbox {}, ind_pred_bbox {}'.format(ind_exp_bbox, ind_pred_bbox)\n",
    "            ious_epoch[sample, epoch] += ious[ind_exp_bbox, ind_pred_bbox]\n",
    "            dists_epoch[sample, epoch] += dists[ind_exp_bbox, ind_pred_bbox]\n",
    "            mses_epoch[sample, epoch] += mses[ind_exp_bbox, ind_pred_bbox]\n",
    "            #print 'ious_epoch {}'.format(ious_epoch)\n",
    "            ious[ind_exp_bbox] = -1  # set iou of assigned bboxes to -1, so they don't get assigned again\n",
    "            ious[:, ind_pred_bbox] = -1\n",
    "            #print 'ious_epoch {}'.format(ious_epoch)\n",
    "            new_order[ind_pred_bbox] = ind_exp_bbox\n",
    "            #print 'new_order {} '.format(new_order)\n",
    "        \n",
    "        flipped_train_y[sample] = exp[new_order].flatten()\n",
    "        \n",
    "        flipped[sample, epoch] = 1. - np.mean(new_order == np.arange(num_objects, dtype=int))#np.array_equal(new_order, np.arange(num_objects, dtype=int))  # TODO: Change this to reflect the number of flips.\n",
    "        ious_epoch[sample, epoch] /= num_objects\n",
    "        dists_epoch[sample, epoch] /= num_objects\n",
    "        mses_epoch[sample, epoch] /= num_objects\n",
    "        \n",
    "        acc_shapes_epoch[sample, epoch] = np.mean(np.argmax(pred[:, 4:4+num_shapes], axis=-1) == np.argmax(exp[:, 4:4+num_shapes], axis=-1))\n",
    "        acc_colors_epoch[sample, epoch] = np.mean(np.argmax(pred[:, 4+num_shapes:4+num_shapes+num_colors], axis=-1) == np.argmax(exp[:, 4+num_shapes:4+num_shapes+num_colors], axis=-1))\n",
    "        #print 'Flipped {} % of all elements'.format(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 'Accuracy shapes: {}'.format(np.mean(acc_shapes_epoch[:, epoch]))\n",
    "    print 'Accuracy colors: {}'.format(np.mean(acc_colors_epoch[:, epoch]))\n",
    "    print '----------------------------------------------------------------'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 109,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((200, 3, 4), (200, 3), (200, 3))"
      ]
     },
     "execution_count": 109,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pred_y1  = model(Variable(torch.FloatTensor(test_X))).data.numpy()\n",
    "pred_y1 = pred_y1.reshape(len(pred_y1), num_objects, -1)\n",
    "pred_bboxes = pred_y1[..., :4] * img_size\n",
    "pred_shapes = np.argmax(pred_y1[..., 4:4+num_shapes], axis=-1).astype(int)  # take max from probabilities\n",
    "# print pred_y[..., 4+num_shapes:4+num_shapes+num_colors].shape\n",
    "# print np.argmax(pred_y[..., 5:8], axis=-1).shape\n",
    "pred_colors = np.argmax(pred_y1[..., 4+num_shapes:4+num_shapes+num_colors], axis=-1).astype(int)\n",
    "pred_bboxes.shape, pred_shapes.shape, pred_colors.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 110,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAA6IAAAHSCAYAAAD2RXZvAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAElFTkSuQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f0a4ea0ff10>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.figure(figsize=(16, 8))\n",
    "for i_subplot in range(1, 9):\n",
    "    plt.subplot(2, 4, i_subplot)\n",
    "    i = np.random.randint(len(test_X))\n",
    "    #print test_imgs[i].transpose(1,2,0).shape\n",
    "    plt.imshow(test_imgs[i].transpose(1,2,0), interpolation='none', origin='lower', extent=[0, 16, 0, 16])\n",
    "    for bbox, shape, color in zip(pred_bboxes[i], pred_shapes[i], pred_colors[i]):\n",
    "        plt.gca().add_patch(matplotlib.patches.Rectangle((bbox[0], bbox[1]), bbox[2], bbox[3], ec='k', fc='none'))\n",
    "        plt.annotate(shape_labels[shape], (bbox[0], bbox[1] + bbox[3] + 0.7), color=color_labels[color], clip_on=False, bbox={'fc': 'w', 'ec': 'none', 'pad': 1, 'alpha': 0.6})\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "torch.save(model.state_dict(), 'multibox_notflipped_chk.pt')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## The bounding boxes are centered to the objects in the data\n",
    "## We need to handle the case of flipping \n",
    "reference https://towardsdatascience.com/object-detection-with-neural-networks-a4e2c46b4491"
   ]
  }
 ],
 "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": "2.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}