assignment3/CSI5139_Assignment3.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# CSI 5139 F00: Assignment 3\n",
    "\n",
    "# content\n",
    "\n",
    "[1. Data Preparation](#Data_Preparation_Part)\n",
    "\n",
    "- [1.1 Provided](#Provided)\n",
    "\n",
    "- [1.2 Provided](#Provided)\n",
    "\n",
    "[2. Model Design (7 layers)](#model_design)\n",
    "\n",
    "- [2.1 Model structure](#model_structure)\n",
    "\n",
    "- [2.2 Train the model](#train_the_model)\n",
    "\n",
    "- [2.3 Evaluate the model](#evaluate_the_model)\n",
    "\n",
    "[2. VGG16 transfer learning](#vgg16_transfer_learning)\n",
    "\n",
    "- [2.1 Get all data sets](#get_all_data_sets)\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a id='Data_Preparation_Part'></a>\n",
    "\n",
    "# 1. Data Preparation\n",
    "\n",
    "<a id='Provided'></a>\n",
    "\n",
    "## 1.1 Provided by the assignment start\n",
    "\n",
    "_This notebook contains simple starter code to show loading and displaying a pair of test images and the associated flow._"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/azhang/.local/lib/python3.5/site-packages/matplotlib/__init__.py:1038: UserWarning: Bad val \"GTKAgg\" on line #1\n",
      "\t\"backend: GTKAgg\n",
      "\"\n",
      "\tin file \"/home/azhang/.config/matplotlib/matplotlibrc\"\n",
      "\tKey backend: Unrecognized backend string 'gtkagg': valid strings are ['Qt5Cairo', 'WX', 'Qt4Agg', 'Qt4Cairo', 'TkCairo', 'Qt5Agg', 'WXCairo', 'pdf', 'MacOSX', 'WebAgg', 'nbAgg', 'cairo', 'template', 'WXAgg', 'pgf', 'svg', 'GTK3Cairo', 'GTK3Agg', 'ps', 'agg', 'TkAgg']\n",
      "  (val, error_details, msg))\n"
     ]
    }
   ],
   "source": [
    "# Adapted helper routine from Aurelien Geron: Hand-on Machine Learning with Scikit-learn & Tensorflow.\n",
    "\n",
    "# Common imports\n",
    "import numpy as np\n",
    "import os\n",
    "import skimage \n",
    "\n",
    "# to make this notebook's output stable across runs\n",
    "np.random.seed(42)\n",
    "\n",
    "# To plot pretty figures\n",
    "%matplotlib inline\n",
    "import matplotlib\n",
    "import matplotlib.pyplot as plt\n",
    "plt.rcParams['axes.labelsize'] = 14\n",
    "plt.rcParams['xtick.labelsize'] = 12\n",
    "plt.rcParams['ytick.labelsize'] = 12\n",
    "\n",
    "# Where to save the figures\n",
    "PROJECT_ROOT_DIR = \"./\"\n",
    "NB_ID = \"assignment3\"\n",
    "\n",
    "# create the directory if it does not exist\n",
    "os.makedirs(os.path.join(PROJECT_ROOT_DIR, \"images\", NB_ID), exist_ok = True)\n",
    "        \n",
    "def save_fig(fig_id, tight_layout=True):\n",
    "    path = os.path.join(PROJECT_ROOT_DIR, \"images\", NB_ID, fig_id + \".png\")\n",
    "    print(\"Saving figure\", fig_id)\n",
    "    if tight_layout:\n",
    "        plt.tight_layout()\n",
    "    plt.savefig(path, format='png', dpi=300)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's load the image pairs and the corresponding flow "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_images = np.load(os.path.join(PROJECT_ROOT_DIR, NB_ID, 'rot_images_train.npy'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_flows = np.load(os.path.join(PROJECT_ROOT_DIR, NB_ID, 'rot_flows_train.npy'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Size of training images: 1860 x 6 x 64 x 64\n",
      "\n",
      "Size of flow results: 1860 x 4 x 64 x 64\n",
      "\n"
     ]
    }
   ],
   "source": [
    "print('Size of training images: {0} x {1} x {2} x {3}\\n'.format(*train_images.shape))\n",
    "print('Size of flow results: {0} x {1} x {2} x {3}\\n'.format(*train_flows.shape))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "For RGB images, we need the channels last, e.g., 64x64x3."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "def make_RGB(img_arr, img_no, seq_no ): \n",
    "    test = np.zeros(img_arr.shape[2:]+(3,))  \n",
    "    test[:,:,2] = img_arr[img_no,0+3*seq_no,:,:] \n",
    "    test[:,:,1] = img_arr[img_no,1+3*seq_no,:,:] \n",
    "    test[:,:,0] = img_arr[img_no,2+3*seq_no,:,:]\n",
    "    return test"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "def make_Greyscale(img_arr, img_no, seq_no ): \n",
    "    return img_arr[img_no,seq_no,:,:] "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Saving figure example_data\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 576x576 with 6 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "img_num = 100\n",
    "\n",
    "fig, axes = plt.subplots(2, 2,figsize=(8,8))\n",
    "axes[0, 0].imshow(make_RGB(train_images, img_num, 0 ))\n",
    "axes[0, 0].set_title('Start')    \n",
    "\n",
    "axes[0, 1].imshow(make_RGB(train_images, img_num, 1 ))\n",
    "axes[0, 1].set_title('Goal')    \n",
    "\n",
    "# Find the min and max of the flow.\n",
    "from matplotlib import colors\n",
    "fl_min = np.min(train_flows[img_num,0:2,:,:])\n",
    "fl_max = np.max(train_flows[img_num,0:2,:,:])\n",
    "fl_norm = colors.Normalize(vmin=fl_min, vmax=fl_max)\n",
    "\n",
    "im = axes[1, 0].imshow(make_Greyscale(train_flows, img_num, 0 ))\n",
    "im.set_norm(fl_norm)\n",
    "axes[1, 0].set_title('Flow in x')\n",
    "fig.colorbar(im,ax=axes[1, 0])\n",
    "\n",
    "im = axes[1, 1].imshow(make_Greyscale(train_flows, img_num, 1 ))\n",
    "im.set_norm(fl_norm)\n",
    "axes[1, 1].set_title('Flow in y')\n",
    "fig.colorbar(im,ax=axes[1, 1])\n",
    "\n",
    "save_fig(\"example_data\")\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Since all the train images and flows are already extracted out, we here get another $2$ necessary sets for the test."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "test_images = np.load(os.path.join(PROJECT_ROOT_DIR, NB_ID, 'rot_images_test.npy'))\n",
    "test_flows = np.load(os.path.join(PROJECT_ROOT_DIR, NB_ID, 'rot_flows_test.npy'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a id='tramsfer_the_format'></a>\n",
    "\n",
    "## 1.2 Transfer the format\n",
    "\n",
    "Here, all pairs of images (start and goal) are concatenated through the RGB channels into size $(64, 64, 6)$. And all pairs of ground truth labels (flow in x and y) are concatenated into size $(64, 64, 2)$.\n",
    "\n",
    "Note that, experimentally, concatenating data through last axis may be a good way for the model to learn."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "def GetInputSet(raw_images, concat_last_axis = True):\n",
    "    \"\"\"\n",
    "    Function: concatenate each pair images together as input.\n",
    "    \"\"\"\n",
    "    hm_images = len(raw_images)\n",
    "    dataset = []\n",
    "    for i in range(hm_images):\n",
    "        img_start = make_RGB(raw_images, i, 0)\n",
    "        img_goal = make_RGB(raw_images, i, 1)\n",
    "        # concatenate on the last dimension\n",
    "        if concat_last_axis:\n",
    "            imgs_concat = np.concatenate([img_start, img_goal], axis = -1)\n",
    "        else:\n",
    "            imgs_concat = np.concatenate([img_start, img_goal], axis = 0)\n",
    "        dataset.append(imgs_concat)\n",
    "    dataset = np.array(dataset)\n",
    "    return dataset\n",
    "\n",
    "def GetOutputset(raw_gts, concat_last_axis = True):\n",
    "    \"\"\"\n",
    "    Function: get the corresponding ground truth vectors.\n",
    "    \"\"\"\n",
    "    hm_gts = len(raw_gts)\n",
    "    dataset = []\n",
    "    for i in range(hm_gts):\n",
    "        # expand the last dimension for concatenating\n",
    "        gt_x = np.expand_dims(make_Greyscale(raw_gts, i, 0), axis=-1)\n",
    "        gt_y = np.expand_dims(make_Greyscale(raw_gts, i, 1), axis=-1)\n",
    "        # concatenate on the last dimension\n",
    "        if concat_last_axis:\n",
    "            gt_concat = np.concatenate([gt_x, gt_y], axis = -1)\n",
    "        else:\n",
    "            gt_concat = np.concatenate([gt_x, gt_y], axis = 0)\n",
    "        dataset.append(gt_concat)\n",
    "    dataset = np.array(dataset)\n",
    "    return dataset"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "And, of cause, before training, all data are shuffled."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Shape of the training images:\t (1860, 64, 64, 6)\n",
      "Shape of the training flows:\t (1860, 64, 64, 2)\n",
      "Shape of the test images:\t (444, 64, 64, 6)\n",
      "Shape of the test flows:\t (444, 64, 64, 2)\n"
     ]
    }
   ],
   "source": [
    "def ShuffleIndex(data):\n",
    "    \"\"\"\n",
    "    Function: return shuffled index\n",
    "    \"\"\"\n",
    "    shuffle_index = np.arange(len(data))\n",
    "    np.random.seed(100010)\n",
    "    np.random.shuffle(shuffle_index)\n",
    "    return shuffle_index\n",
    "\n",
    "def Shuffle(data_in, data_out):\n",
    "    \"\"\"\n",
    "    Function: shuffle dataset\n",
    "    \"\"\"\n",
    "    shuffle_index = ShuffleIndex(data_in)\n",
    "    data_in = data_in[shuffle_index]\n",
    "    data_out = data_out[shuffle_index]\n",
    "    return data_in, data_out\n",
    "\n",
    "train_in = GetInputSet(train_images)\n",
    "train_out = GetOutputset(train_flows)\n",
    "train_in, train_out = Shuffle(train_in, train_out)\n",
    "\n",
    "test_in = GetInputSet(test_images)\n",
    "test_out = GetOutputset(test_flows)\n",
    "\n",
    "print(\"Shape of the training images:\\t\", train_in.shape)\n",
    "print(\"Shape of the training flows:\\t\",train_out.shape)\n",
    "print(\"Shape of the test images:\\t\",test_in.shape)\n",
    "print(\"Shape of the test flows:\\t\",test_out.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a id='model_design'></a>\n",
    "\n",
    "# 2. Model design (7 layers)\n",
    "\n",
    "<a id='model_structure'></a>\n",
    "\n",
    "## 2.1 Model structure\n",
    "\n",
    "Four convolutional layers are built for learning the features of the image, then 2 deconvolutional layers are used for sizing back the feature maps to the flow image size.\n",
    "\n",
    "Loss function is designed based on mean square error, since the best way to learn the localization problems is MSE. \n",
    "\n",
    "Metrics function is defined by calculating the $l_1$ norm between prediction and ground truth. Then a threshold is set as $3e-3$ to filter the results. Specially, when the $l_1$ norm is less than the threshold, it will be set as the right prediction. Then, the number of the right predicition pixels are divided by the total number of pixels to get the accuracy.\n",
    "\n",
    "Details is printed out at the end."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/dtypes.py:523: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.\n",
      "  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\n",
      "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/dtypes.py:524: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.\n",
      "  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\n",
      "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/dtypes.py:525: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.\n",
      "  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\n",
      "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/dtypes.py:526: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.\n",
      "  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\n",
      "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/dtypes.py:527: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.\n",
      "  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\n",
      "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/dtypes.py:532: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.\n",
      "  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d (Conv2D)              (None, 64, 64, 64)        3520      \n",
      "_________________________________________________________________\n",
      "conv2d_1 (Conv2D)            (None, 32, 32, 128)       73856     \n",
      "_________________________________________________________________\n",
      "conv2d_2 (Conv2D)            (None, 32, 32, 256)       295168    \n",
      "_________________________________________________________________\n",
      "conv2d_3 (Conv2D)            (None, 16, 16, 512)       1180160   \n",
      "_________________________________________________________________\n",
      "conv2d_transpose (Conv2DTran (None, 32, 32, 256)       1179904   \n",
      "_________________________________________________________________\n",
      "conv2d_transpose_1 (Conv2DTr (None, 64, 64, 128)       295040    \n",
      "_________________________________________________________________\n",
      "conv2d_4 (Conv2D)            (None, 64, 64, 2)         2306      \n",
      "=================================================================\n",
      "Total params: 3,029,954\n",
      "Trainable params: 3,029,954\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "import os\n",
    "os.environ[\"CUDA_DEVICE_ORDER\"]=\"PCI_BUS_ID\"\n",
    "os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"1\"\n",
    "# import tf.keras\n",
    "import tensorflow.keras.backend as K\n",
    "K.clear_session()\n",
    "import tensorflow as tf\n",
    "from tensorflow.keras.models import Sequential\n",
    "import tensorflow.keras.layers as L\n",
    "from tensorflow.keras.callbacks import TensorBoard, ReduceLROnPlateau\n",
    "import tensorflow.keras.regularizers as R\n",
    "from tensorflow.keras.optimizers import Adam\n",
    "\n",
    "def Loss(y_true, y_pred):\n",
    "    \"\"\"\n",
    "    Loss Function: use the MSE principle.\n",
    "    \"\"\"\n",
    "    weight_vector = tf.where(y_true==0., 0.2*K.ones_like(y_true), 0.8*K.ones_like(y_true))\n",
    "    # return K.mean(K.square(y_true - y_pred)*weight_vector)\n",
    "    return K.mean(K.square(y_pred - y_true))\n",
    "\n",
    "def Metrics(y_true, y_pred):\n",
    "    \"\"\"\n",
    "    Metrics Function: set if the difference between prediction and \n",
    "    ground truth is less than 3e-3, set it as right prediction. \n",
    "    Then calculate the total number of the right predictions out of \n",
    "    the whole flow image.\n",
    "    \"\"\"\n",
    "    # weight_vector = tf.where(y_true==0., 0.2*K.ones_like(y_true), 0.8*K.ones_like(y_true))\n",
    "    value = K.abs(y_true - y_pred)\n",
    "    false_pred = K.cast(K.greater(value, 3.e-3), 'float32') # * weight_vector\n",
    "    whole_ones = K.ones_like(value) # * weight_vector\n",
    "    return K.sum(1. - false_pred) / K.sum(whole_ones)\n",
    "\n",
    "K.clear_session()\n",
    "\n",
    "model = Sequential()\n",
    "model.add(L.Conv2D(64, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                   activation='relu', input_shape=(64, 64, 6)))\n",
    "model.add(L.Conv2D(128, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu'))\n",
    "model.add(L.Conv2D(256, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                   activation='relu'))\n",
    "model.add(L.Conv2D(512, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu'))\n",
    "model.add(L.Conv2DTranspose(256, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu'))\n",
    "model.add(L.Conv2DTranspose(128, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu'))\n",
    "model.add(L.Conv2D(2, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                   activation='linear'))\n",
    "\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a id='train_the_model'></a>\n",
    "\n",
    "## 2.2 Train the model\n",
    "\n",
    "Some parameters are set as:\n",
    "1. learning rate = $0.0008$;\n",
    "2. learning rate decay rate = $0.8$;\n",
    "3. batch size = $10$;\n",
    "4. total epochs = $15$."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/10\n",
      "1860/1860 [==============================] - 11s 6ms/step - loss: 0.4929 - Metrics: 0.1870\n",
      "Epoch 2/10\n",
      "1860/1860 [==============================] - 8s 4ms/step - loss: 0.2323 - Metrics: 0.2187\n",
      "Epoch 3/10\n",
      "1860/1860 [==============================] - 9s 5ms/step - loss: 0.1625 - Metrics: 0.2640\n",
      "Epoch 4/10\n",
      "1860/1860 [==============================] - 9s 5ms/step - loss: 0.1183 - Metrics: 0.2783\n",
      "Epoch 5/10\n",
      "1860/1860 [==============================] - 9s 5ms/step - loss: 0.0915 - Metrics: 0.2361\n",
      "Epoch 6/10\n",
      "1860/1860 [==============================] - 9s 5ms/step - loss: 0.0758 - Metrics: 0.2298\n",
      "Epoch 7/10\n",
      "1860/1860 [==============================] - 9s 5ms/step - loss: 0.0619 - Metrics: 0.2470\n",
      "Epoch 8/10\n",
      "1860/1860 [==============================] - 9s 5ms/step - loss: 0.0550 - Metrics: 0.2450\n",
      "Epoch 9/10\n",
      "1860/1860 [==============================] - 8s 4ms/step - loss: 0.0467 - Metrics: 0.2937\n",
      "Epoch 10/10\n",
      "1860/1860 [==============================] - 9s 5ms/step - loss: 0.0406 - Metrics: 0.2933\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<tensorflow.python.keras.callbacks.History at 0x7f0728452f98>"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# GPU memory management\n",
    "config = tf.ConfigProto()\n",
    "config.gpu_options.allow_growth = True\n",
    "K.set_session(tf.Session(config=config))\n",
    "\n",
    "reduce_lr = ReduceLROnPlateau(monitor='loss', factor=0.8, patience=3, min_lr=1e-5)\n",
    "model.compile(loss=Loss, optimizer=Adam(lr=0.0008), metrics=[Metrics])\n",
    "\n",
    "model.fit(train_in, train_out, batch_size=10, epochs=10,\n",
    "#             validation_data=(test_in, test_out),\n",
    "            callbacks=[reduce_lr])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a id='evaluate_the_model'></a>\n",
    "\n",
    "## 2.3 Evaluate the model\n",
    "\n",
    "The model is then evaluated on the test set."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "444/444 [==============================] - 1s 2ms/step\n",
      "The accuracy on test set is: \t 0.1628412468744828\n"
     ]
    }
   ],
   "source": [
    "results = model.evaluate(test_in, test_out)\n",
    "print(\"The accuracy on test set is: \\t\", results[1])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "In order the make the results more intuitive, a random pair of images are selected out and the prediction of the pair with the ground truth labels are plotted at the end of the cell."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Prediction and ground truth of the example images in the test set.\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 576x576 with 8 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "random_test_number = 101\n",
    "\n",
    "test_example_images = np.expand_dims(test_in[random_test_number], axis=0)\n",
    "test_example_gt = test_out[random_test_number]\n",
    "test_example_pred = model.predict(test_example_images)\n",
    "test_example_pred = np.squeeze(test_example_pred)\n",
    "test_example_gt = np.squeeze(test_example_gt)\n",
    "\n",
    "fl_min = np.min(test_example_gt)\n",
    "fl_max = np.max(test_example_gt)\n",
    "fl_norm = colors.Normalize(vmin=fl_min, vmax=fl_max)\n",
    "\n",
    "fig, axes = plt.subplots(2, 2,figsize=(8,8))\n",
    "\n",
    "print(\"Prediction and ground truth of the example images in the test set.\")\n",
    "im = axes[0, 0].imshow(test_example_pred[..., 0])\n",
    "# im = axes[0, 0].imshow(test_example_pred[:len(test_example_pred)//2, :])\n",
    "im.set_norm(fl_norm)\n",
    "axes[0, 0].set_title('Prediction: Flow in x')\n",
    "fig.colorbar(im,ax=axes[0, 0])\n",
    "\n",
    "im = axes[0, 1].imshow(test_example_pred[..., 1])\n",
    "# im = axes[0, 1].imshow(test_example_pred[len(test_example_pred)//2:, :])\n",
    "im.set_norm(fl_norm)\n",
    "axes[0, 1].set_title('Prediction: Flow in y')\n",
    "fig.colorbar(im,ax=axes[0, 1])\n",
    "\n",
    "im = axes[1, 0].imshow(test_example_gt[..., 0])\n",
    "# im = axes[1, 0].imshow(test_example_gt[:len(test_example_gt)//2, ...])\n",
    "im.set_norm(fl_norm)\n",
    "axes[1, 0].set_title('Ground Truth: Flow in x')\n",
    "fig.colorbar(im,ax=axes[1, 0])\n",
    "\n",
    "im = axes[1, 1].imshow(test_example_gt[..., 1])\n",
    "# im = axes[1, 1].imshow(test_example_gt[len(test_example_gt)//2:, ...])\n",
    "im.set_norm(fl_norm)\n",
    "axes[1, 1].set_title('Ground Truth: Flow in y')\n",
    "fig.colorbar(im,ax=axes[1, 1])\n",
    "\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a id='vgg16_transfer_learning'></a>\n",
    "\n",
    "# 3. VGG16 transfer learning\n",
    "\n",
    "<a id='get_all_data_sets'></a>\n",
    "\n",
    "## 3.1 Get all data sets\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Shape of the rotation training images:\t (1860, 64, 64, 6)\n",
      "Shape of the rotation training flows:\t (1860, 64, 64, 2)\n",
      "Shape of the rotation test images:\t (444, 64, 64, 6)\n",
      "Shape of the rotation test flows:\t (444, 64, 64, 2)\n",
      "Shape of the mixed training images:\t (1860, 64, 64, 6)\n",
      "Shape of the mixed training flows:\t (1860, 64, 64, 2)\n",
      "Shape of the mixed test images:\t (444, 64, 64, 6)\n",
      "Shape of the mixed test flows:\t (444, 64, 64, 2)\n"
     ]
    }
   ],
   "source": [
    "mix_train_images = np.load(os.path.join(PROJECT_ROOT_DIR, NB_ID, 'mix_images_train.npy'))\n",
    "mix_train_flows = np.load(os.path.join(PROJECT_ROOT_DIR, NB_ID, 'mix_flows_train.npy'))\n",
    "mix_test_images = np.load(os.path.join(PROJECT_ROOT_DIR, NB_ID, 'mix_images_test.npy'))\n",
    "mix_test_flows = np.load(os.path.join(PROJECT_ROOT_DIR, NB_ID, 'mix_flows_test.npy'))\n",
    "\n",
    "mix_train_in = GetInputSet(mix_train_images, concat_last_axis=True)\n",
    "mix_train_out = GetOutputset(mix_train_flows, concat_last_axis=True)\n",
    "mix_train_in, mix_train_out = Shuffle(mix_train_in, mix_train_out)\n",
    "\n",
    "mix_test_in = GetInputSet(mix_test_images, concat_last_axis=True)\n",
    "mix_test_out = GetOutputset(mix_test_flows, concat_last_axis=True)\n",
    "\n",
    "rot_train_in = GetInputSet(train_images, concat_last_axis=True)\n",
    "rot_train_out = GetOutputset(train_flows, concat_last_axis=True)\n",
    "rot_train_in, rot_train_out = Shuffle(rot_train_in, rot_train_out)\n",
    "\n",
    "rot_test_in = GetInputSet(test_images, concat_last_axis=True)\n",
    "rot_test_out = GetOutputset(test_flows, concat_last_axis=True)\n",
    "\n",
    "print(\"Shape of the rotation training images:\\t\", rot_train_in.shape)\n",
    "print(\"Shape of the rotation training flows:\\t\",rot_train_out.shape)\n",
    "print(\"Shape of the rotation test images:\\t\",rot_test_in.shape)\n",
    "print(\"Shape of the rotation test flows:\\t\",rot_test_out.shape)\n",
    "\n",
    "print(\"Shape of the mixed training images:\\t\", mix_train_in.shape)\n",
    "print(\"Shape of the mixed training flows:\\t\",mix_train_out.shape)\n",
    "print(\"Shape of the mixed test images:\\t\",mix_test_in.shape)\n",
    "print(\"Shape of the mixed test flows:\\t\",mix_test_out.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "block1_conv1 (Conv2D)        (None, 64, 64, 64)        1792      \n",
      "_________________________________________________________________\n",
      "block1_conv2 (Conv2D)        (None, 64, 64, 64)        36928     \n",
      "_________________________________________________________________\n",
      "block1_pool (MaxPooling2D)   (None, 32, 32, 64)        0         \n",
      "_________________________________________________________________\n",
      "block2_conv1 (Conv2D)        (None, 32, 32, 128)       73856     \n",
      "_________________________________________________________________\n",
      "block2_conv2 (Conv2D)        (None, 32, 32, 128)       147584    \n",
      "_________________________________________________________________\n",
      "block2_pool (MaxPooling2D)   (None, 16, 16, 128)       0         \n",
      "_________________________________________________________________\n",
      "block3_conv1 (Conv2D)        (None, 16, 16, 256)       295168    \n",
      "_________________________________________________________________\n",
      "block3_conv2 (Conv2D)        (None, 16, 16, 256)       590080    \n",
      "_________________________________________________________________\n",
      "block3_conv3 (Conv2D)        (None, 16, 16, 256)       590080    \n",
      "_________________________________________________________________\n",
      "block3_pool (MaxPooling2D)   (None, 8, 8, 256)         0         \n",
      "_________________________________________________________________\n",
      "block4_conv1 (Conv2D)        (None, 8, 8, 512)         1180160   \n",
      "_________________________________________________________________\n",
      "block4_conv2 (Conv2D)        (None, 8, 8, 512)         2359808   \n",
      "_________________________________________________________________\n",
      "block4_conv3 (Conv2D)        (None, 8, 8, 512)         2359808   \n",
      "=================================================================\n",
      "Total params: 7,635,264\n",
      "Trainable params: 0\n",
      "Non-trainable params: 7,635,264\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "import tensorflow.keras.applications as App\n",
    "K.clear_session()\n",
    "\n",
    "base_model = App.vgg16.VGG16(input_shape=(64, 64, 3), include_top=False, weights='imagenet')\n",
    "vgg16_layers = Sequential()\n",
    "for layer in base_model.layers[:14]:\n",
    "    vgg16_layers.add(layer)\n",
    "\n",
    "base_model.trainable = False\n",
    "vgg16_layers.trainable = False\n",
    "vgg16_layers.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d (Conv2D)              (None, 64, 64, 64)        3520      \n",
      "_________________________________________________________________\n",
      "conv2d_1 (Conv2D)            (None, 64, 64, 3)         1731      \n",
      "_________________________________________________________________\n",
      "sequential (Sequential)      (None, 8, 8, 512)         7635264   \n",
      "_________________________________________________________________\n",
      "conv2d_transpose_1 (Conv2DTr (None, 16, 16, 256)       1179904   \n",
      "_________________________________________________________________\n",
      "conv2d_transpose_2 (Conv2DTr (None, 32, 32, 128)       295040    \n",
      "_________________________________________________________________\n",
      "conv2d_transpose_3 (Conv2DTr (None, 64, 64, 64)        73792     \n",
      "_________________________________________________________________\n",
      "conv2d_transpose_5 (Conv2DTr (None, 64, 64, 2)         1154      \n",
      "=================================================================\n",
      "Total params: 9,190,405\n",
      "Trainable params: 1,555,141\n",
      "Non-trainable params: 7,635,264\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "header = L.Conv2D(64, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                   activation='relu', input_shape=(64, 64, 6))\n",
    "header2 = L.Conv2D(3, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                   activation='relu')\n",
    "\n",
    "regression_layer1 = L.Conv2DTranspose(512, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu')\n",
    "regression_layer2 = L.Conv2DTranspose(256, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu')\n",
    "regression_layer3 = L.Conv2DTranspose(128 , (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu')\n",
    "regression_layer4 = L.Conv2DTranspose(64, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu')\n",
    "regression_layer5 = L.Conv2DTranspose(32, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu')\n",
    "regression_layer6 = L.Conv2D(1, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                   activation='linear')\n",
    "regression_layer7 = L.Conv2DTranspose(2, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                   activation='linear')\n",
    "\n",
    "vgg16_based_model = Sequential([header, header2, vgg16_layers, regression_layer2,\n",
    "                                regression_layer3, regression_layer4, regression_layer7])\n",
    "\n",
    "vgg16_based_model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/10\n",
      "1860/1860 [==============================] - 8s 4ms/step - loss: 0.4548 - Metrics: 0.0953\n",
      "Epoch 2/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.2611 - Metrics: 0.2223\n",
      "Epoch 3/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.2305 - Metrics: 0.2496\n",
      "Epoch 4/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.2059 - Metrics: 0.2831\n",
      "Epoch 5/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.1838 - Metrics: 0.3114\n",
      "Epoch 6/10\n",
      "1860/1860 [==============================] - 6s 3ms/step - loss: 0.1609 - Metrics: 0.3250\n",
      "Epoch 7/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.1454 - Metrics: 0.3286\n",
      "Epoch 8/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.1307 - Metrics: 0.3494\n",
      "Epoch 9/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.1209 - Metrics: 0.3584\n",
      "Epoch 10/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.1133 - Metrics: 0.3686\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<tensorflow.python.keras.callbacks.History at 0x7f068075fa20>"
      ]
     },
     "execution_count": 18,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "reduce_lr = ReduceLROnPlateau(monitor='loss', factor=0.8, patience=1, min_lr=1e-5)\n",
    "\n",
    "vgg16_based_model.compile(loss=Loss, optimizer=Adam(lr=0.0003), metrics=[Metrics])\n",
    "vgg16_based_model.fit(rot_train_in, rot_train_out, batch_size=10, epochs=10,\n",
    "#             validation_data=(test_in, test_out),\n",
    "            callbacks=[reduce_lr])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "444/444 [==============================] - 1s 3ms/step\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "[0.1407922286171097, 0.3877041098233816]"
      ]
     },
     "execution_count": 19,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "random_test_number = 101\n",
    "\n",
    "rot_test_example_images = np.expand_dims(rot_test_in[random_test_number], axis=0)\n",
    "rot_test_example_gt = rot_test_out[random_test_number]\n",
    "rot_test_example_pred = vgg16_based_model.predict(rot_test_example_images)\n",
    "rot_test_example_pred = np.squeeze(rot_test_example_pred)\n",
    "rot_test_example_gt = np.squeeze(rot_test_example_gt)\n",
    "\n",
    "rot_fl_min = np.min(rot_test_example_gt)\n",
    "rot_fl_max = np.max(rot_test_example_gt)\n",
    "rot_fl_norm = colors.Normalize(vmin=fl_min, vmax=fl_max)\n",
    "\n",
    "vgg16_based_model.evaluate(rot_test_in, rot_test_out)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d (Conv2D)              (None, 64, 64, 64)        3520      \n",
      "_________________________________________________________________\n",
      "conv2d_1 (Conv2D)            (None, 64, 64, 3)         1731      \n",
      "_________________________________________________________________\n",
      "sequential (Sequential)      multiple                  7635264   \n",
      "_________________________________________________________________\n",
      "conv2d_transpose_1 (Conv2DTr (None, 16, 16, 256)       1179904   \n",
      "_________________________________________________________________\n",
      "conv2d_transpose_2 (Conv2DTr (None, 32, 32, 128)       295040    \n",
      "_________________________________________________________________\n",
      "conv2d_transpose_3 (Conv2DTr (None, 64, 64, 64)        73792     \n",
      "_________________________________________________________________\n",
      "conv2d_transpose_5 (Conv2DTr (None, 64, 64, 2)         1154      \n",
      "=================================================================\n",
      "Total params: 9,190,405\n",
      "Trainable params: 1,555,141\n",
      "Non-trainable params: 7,635,264\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "K.clear_session()\n",
    "\n",
    "base_model = App.vgg16.VGG16(input_shape=(128, 64, 3), include_top=False, weights='imagenet')\n",
    "vgg16_layers = Sequential()\n",
    "for layer in base_model.layers[:14]:\n",
    "    vgg16_layers.add(layer)\n",
    "\n",
    "base_model.trainable = False\n",
    "vgg16_layers.trainable = False\n",
    "\n",
    "header = L.Conv2D(64, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                   activation='relu', input_shape=(64, 64, 6))\n",
    "header2 = L.Conv2D(3, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                   activation='relu')\n",
    "\n",
    "regression_layer1 = L.Conv2DTranspose(512, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu')\n",
    "regression_layer2 = L.Conv2DTranspose(256, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu')\n",
    "regression_layer3 = L.Conv2DTranspose(128 , (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu')\n",
    "regression_layer4 = L.Conv2DTranspose(64, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu')\n",
    "regression_layer5 = L.Conv2DTranspose(32, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                   activation='relu')\n",
    "regression_layer6 = L.Conv2D(1, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                   activation='linear')\n",
    "regression_layer7 = L.Conv2DTranspose(2, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                   activation='linear')\n",
    "\n",
    "vgg16_based_model = Sequential([header, header2, vgg16_layers, regression_layer2,\n",
    "                                regression_layer3, regression_layer4, regression_layer7])\n",
    "\n",
    "vgg16_based_model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.6220 - Metrics: 0.2028\n",
      "Epoch 2/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.6217 - Metrics: 0.2302\n",
      "Epoch 3/10\n",
      "1860/1860 [==============================] - 6s 3ms/step - loss: 0.6217 - Metrics: 0.2492\n",
      "Epoch 4/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.6217 - Metrics: 0.2475\n",
      "Epoch 5/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.6216 - Metrics: 0.2454\n",
      "Epoch 6/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.5509 - Metrics: 0.1487\n",
      "Epoch 7/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.3140 - Metrics: 0.2641\n",
      "Epoch 8/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.2549 - Metrics: 0.3601\n",
      "Epoch 9/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.2232 - Metrics: 0.6325\n",
      "Epoch 10/10\n",
      "1860/1860 [==============================] - 7s 4ms/step - loss: 0.2039 - Metrics: 0.8587\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<tensorflow.python.keras.callbacks.History at 0x7f074cd2dc88>"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "reduce_lr = ReduceLROnPlateau(monitor='loss', factor=0.8, patience=1, min_lr=1e-5)\n",
    "\n",
    "vgg16_based_model.compile(loss=Loss, optimizer=Adam(lr=0.0003), metrics=[Metrics])\n",
    "vgg16_based_model.fit(mix_train_in, mix_train_out, batch_size=10, epochs=10,\n",
    "#             validation_data=(test_in, test_out),\n",
    "            callbacks=[reduce_lr])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "444/444 [==============================] - 1s 1ms/step\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "[0.22147919157066862, 0.8751388423077695]"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "mix_test_example_images = np.expand_dims(mix_test_in[random_test_number], axis=0)\n",
    "mix_test_example_gt = mix_test_out[random_test_number]\n",
    "mix_test_example_pred = vgg16_based_model.predict(mix_test_example_images)\n",
    "mix_test_example_pred = np.squeeze(mix_test_example_pred)\n",
    "mix_test_example_gt = np.squeeze(mix_test_example_gt)\n",
    "\n",
    "mix_fl_min = np.min(mix_test_example_gt)\n",
    "mix_fl_max = np.max(mix_test_example_gt)\n",
    "mix_fl_norm = colors.Normalize(vmin=fl_min, vmax=fl_max)\n",
    "\n",
    "vgg16_based_model.evaluate(mix_test_in, mix_test_out)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Prediction and ground truth of the example images in the test set.\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x504 with 8 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "fig, axes = plt.subplots(2, 4, figsize=(12,7))\n",
    "\n",
    "print(\"Prediction and ground truth of the example images in the test set.\")\n",
    "im = axes[0, 0].imshow(rot_test_example_pred[..., 0])\n",
    "im.set_norm(rot_fl_norm)\n",
    "axes[0, 0].set_title('Rotation Prediction in x')\n",
    "# fig.colorbar(im,ax=axes[0, 0])\n",
    "\n",
    "im = axes[0, 1].imshow(rot_test_example_pred[..., 1])\n",
    "im.set_norm(rot_fl_norm)\n",
    "axes[0, 1].set_title('Rotation Prediction in y')\n",
    "# fig.colorbar(im,ax=axes[0, 1])\n",
    "\n",
    "im = axes[1, 0].imshow(rot_test_example_gt[..., 0])\n",
    "im.set_norm(rot_fl_norm)\n",
    "axes[1, 0].set_title('Rotation Ground Truth in x')\n",
    "# fig.colorbar(im,ax=axes[1, 0])\n",
    "\n",
    "im = axes[1, 1].imshow(rot_test_example_gt[..., 1])\n",
    "im.set_norm(rot_fl_norm)\n",
    "axes[1, 1].set_title('Rotation Ground Truth in y')\n",
    "# fig.colorbar(im,ax=axes[1, 1])\n",
    "\n",
    "#############################################################\n",
    "im = axes[0, 2].imshow(mix_test_example_pred[..., 0])\n",
    "im.set_norm(mix_fl_norm)\n",
    "axes[0, 2].set_title('Mixed Prediction in x')\n",
    "# fig.colorbar(im,ax=axes[0, 2])\n",
    "\n",
    "im = axes[0, 3].imshow(mix_test_example_pred[..., 1])\n",
    "im.set_norm(mix_fl_norm)\n",
    "axes[0, 3].set_title('Mixed Prediction in y')\n",
    "# fig.colorbar(im,ax=axes[0, 3])\n",
    "\n",
    "im = axes[1, 2].imshow(mix_test_example_gt[..., 0])\n",
    "im.set_norm(mix_fl_norm)\n",
    "axes[1, 2].set_title('Mixed Ground Truth in x')\n",
    "# fig.colorbar(im,ax=axes[1, 2])\n",
    "\n",
    "im = axes[1, 3].imshow(mix_test_example_gt[..., 1])\n",
    "im.set_norm(mix_fl_norm)\n",
    "axes[1, 3].set_title('Mixed Ground Truth in y')\n",
    "# fig.colorbar(im,ax=axes[1, 3])\n",
    "\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "__________________________________________________________________________________________________\n",
      "Layer (type)                    Output Shape         Param #     Connected to                     \n",
      "==================================================================================================\n",
      "input_1 (InputLayer)            (None, 64, 64, 3)    0                                            \n",
      "__________________________________________________________________________________________________\n",
      "input_2 (InputLayer)            (None, 64, 64, 3)    0                                            \n",
      "__________________________________________________________________________________________________\n",
      "conv2d (Conv2D)                 (None, 32, 32, 64)   1792        input_1[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_2 (Conv2D)               (None, 32, 32, 64)   1792        input_2[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_1 (Conv2D)               (None, 16, 16, 128)  73856       conv2d[0][0]                     \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_3 (Conv2D)               (None, 16, 16, 128)  73856       conv2d_2[0][0]                   \n",
      "__________________________________________________________________________________________________\n",
      "concatenate (Concatenate)       (None, 16, 16, 256)  0           conv2d_1[0][0]                   \n",
      "                                                                 conv2d_3[0][0]                   \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_4 (Conv2D)               (None, 16, 16, 256)  590080      concatenate[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_5 (Conv2D)               (None, 16, 16, 512)  1180160     conv2d_4[0][0]                   \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_transpose (Conv2DTranspo (None, 32, 32, 256)  1179904     conv2d_5[0][0]                   \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_transpose_1 (Conv2DTrans (None, 64, 64, 128)  295040      conv2d_transpose[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_6 (Conv2D)               (None, 64, 64, 2)    2306        conv2d_transpose_1[0][0]         \n",
      "==================================================================================================\n",
      "Total params: 3,398,786\n",
      "Trainable params: 3,398,786\n",
      "Non-trainable params: 0\n",
      "__________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "from tensorflow.keras.models import Model\n",
    "\n",
    "K.clear_session()\n",
    "\n",
    "def FirstTwoLayers(x):\n",
    "    x = L.Conv2D(64, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                               activation='relu')(x)\n",
    "    x = L.Conv2D(128, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                               activation='relu')(x)\n",
    "    return x\n",
    "\n",
    "def NonSeqModel(input_start, input_goal):\n",
    "    feature_start = FirstTwoLayers(input_start)\n",
    "    feature_goal = FirstTwoLayers(input_goal)\n",
    "\n",
    "    x = L.Concatenate()([feature_start, feature_goal])\n",
    "    x = L.Conv2D(256, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                       activation='relu')(x)\n",
    "    x = L.Conv2D(512, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                       activation='relu')(x)\n",
    "    x = L.Conv2DTranspose(256, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                       activation='relu')(x)\n",
    "    x = L.Conv2DTranspose(128, (3, 3), strides=(2, 2), padding='same', \\\n",
    "                       activation='relu')(x)\n",
    "    x = L.Conv2D(2, (3, 3), strides=(1, 1), padding='same', \\\n",
    "                       activation='linear')(x)\n",
    "    return Model(inputs=[input_start, input_goal], outputs=x)\n",
    "\n",
    "\n",
    "model_input_start = L.Input(shape=(64,64,3))\n",
    "model_input_goal = L.Input(shape=(64,64,3))\n",
    "\n",
    "nonseqentialModel = NonSeqModel(model_input_start, model_input_goal)\n",
    "nonseqentialModel.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "def TransferInputData(input_data):\n",
    "    \"\"\"\n",
    "    Function: transfer training data to the format we want.\n",
    "    \"\"\"\n",
    "    return [input_data[..., :3], input_data[..., 3:]]\n",
    "\n",
    "rot_train_in_sep = TransferInputData(rot_train_in)\n",
    "rot_test_in_sep = TransferInputData(rot_test_in)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/10\n",
      "1860/1860 [==============================] - 9s 5ms/step - loss: 0.3802 - Metrics: 0.2758\n",
      "Epoch 2/10\n",
      "1860/1860 [==============================] - 8s 5ms/step - loss: 0.1566 - Metrics: 0.2051\n",
      "Epoch 3/10\n",
      "1860/1860 [==============================] - 8s 4ms/step - loss: 0.1020 - Metrics: 0.1951\n",
      "Epoch 4/10\n",
      "1860/1860 [==============================] - 8s 4ms/step - loss: 0.0686 - Metrics: 0.1527\n",
      "Epoch 5/10\n",
      "1860/1860 [==============================] - 8s 4ms/step - loss: 0.0519 - Metrics: 0.1939\n",
      "Epoch 6/10\n",
      "1860/1860 [==============================] - 8s 4ms/step - loss: 0.0427 - Metrics: 0.2407\n",
      "Epoch 7/10\n",
      "1860/1860 [==============================] - 9s 5ms/step - loss: 0.0336 - Metrics: 0.3030\n",
      "Epoch 8/10\n",
      "1860/1860 [==============================] - 9s 5ms/step - loss: 0.0294 - Metrics: 0.3108\n",
      "Epoch 9/10\n",
      "1860/1860 [==============================] - 9s 5ms/step - loss: 0.0306 - Metrics: 0.3425\n",
      "Epoch 10/10\n",
      "1860/1860 [==============================] - 9s 5ms/step - loss: 0.0197 - Metrics: 0.4033\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<tensorflow.python.keras.callbacks.History at 0x7f05f4675b00>"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# GPU memory management\n",
    "config = tf.ConfigProto()\n",
    "config.gpu_options.allow_growth = True\n",
    "K.set_session(tf.Session(config=config))\n",
    "\n",
    "reduce_lr = ReduceLROnPlateau(monitor='loss', factor=0.8, patience=3, min_lr=1e-5)\n",
    "\n",
    "nonseqentialModel.compile(loss=Loss, optimizer=Adam(lr=0.0008), metrics=[Metrics])\n",
    "\n",
    "nonseqentialModel.fit(rot_train_in_sep, rot_train_out, batch_size=10, epochs=10,\n",
    "#             validation_data=(test_in, test_out),\n",
    "            callbacks=[reduce_lr])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "444/444 [==============================] - 1s 2ms/step\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "[0.022587707232650335, 0.535259076603898]"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "nonseqentialModel.evaluate(rot_test_in_sep, rot_test_out)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Prediction and ground truth of the example images in the test set.\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 576x576 with 8 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "random_test_number = 101\n",
    "\n",
    "test_example_images_1 = np.expand_dims(rot_test_in_sep[0][random_test_number], axis=0)\n",
    "test_example_images_2 = np.expand_dims(rot_test_in_sep[1][random_test_number], axis=0)\n",
    "test_example_images = [test_example_images_1, test_example_images_2]\n",
    "test_example_gt = rot_test_out[random_test_number]\n",
    "test_example_pred = nonseqentialModel.predict(test_example_images)\n",
    "test_example_pred = np.squeeze(test_example_pred)\n",
    "test_example_gt = np.squeeze(test_example_gt)\n",
    "\n",
    "fl_min = np.min(test_example_gt)\n",
    "fl_max = np.max(test_example_gt)\n",
    "fl_norm = colors.Normalize(vmin=fl_min, vmax=fl_max)\n",
    "\n",
    "fig, axes = plt.subplots(2, 2,figsize=(8,8))\n",
    "\n",
    "print(\"Prediction and ground truth of the example images in the test set.\")\n",
    "im = axes[0, 0].imshow(test_example_pred[..., 0])\n",
    "im.set_norm(fl_norm)\n",
    "axes[0, 0].set_title('Prediction: Flow in x')\n",
    "fig.colorbar(im,ax=axes[0, 0])\n",
    "\n",
    "im = axes[0, 1].imshow(test_example_pred[..., 1])\n",
    "im.set_norm(fl_norm)\n",
    "axes[0, 1].set_title('Prediction: Flow in y')\n",
    "fig.colorbar(im,ax=axes[0, 1])\n",
    "\n",
    "im = axes[1, 0].imshow(test_example_gt[..., 0])\n",
    "im.set_norm(fl_norm)\n",
    "axes[1, 0].set_title('Ground Truth: Flow in x')\n",
    "fig.colorbar(im,ax=axes[1, 0])\n",
    "\n",
    "im = axes[1, 1].imshow(test_example_gt[..., 1])\n",
    "# im = axes[1, 1].imshow(test_example_gt[len(test_example_gt)//2:, ...])\n",
    "im.set_norm(fl_norm)\n",
    "axes[1, 1].set_title('Ground Truth: Flow in y')\n",
    "fig.colorbar(im,ax=axes[1, 1])\n",
    "\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "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": "3.5.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}