C3W3_Assignment_handwriting_digit.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "view-in-github"
   },
   "source": [
    "<a href=\"https://colab.research.google.com/github/xerocopy/Advanced-Computer-Vision-with-TensorFlow/blob/main/C3W3_Assignment_handwriting_digit.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "k6x_sGPQocpw"
   },
   "source": [
    "# Week 3 Assignment: Image Segmentation of Handwritten Digits\n",
    "\n",
    "<img src='https://drive.google.com/uc?export=view&id=1-WBX7w_R9abwdGmAUBFWWEcJ0LkMppq2' alt='m2nist digits'>\n",
    "\n",
    "In this week's assignment, you will build a model that predicts the segmentation masks (pixel-wise label map) of handwritten digits. This model will be trained on the [M2NIST dataset](https://www.kaggle.com/farhanhubble/multimnistm2nist), a multi digit MNIST. If you've done the ungraded lab on the CamVid dataset, then many of the steps here will look familiar.\n",
    "\n",
    "You will build a Convolutional Neural Network (CNN) from scratch for the downsampling path and use a Fully Convolutional Network, FCN-8, to upsample and produce the pixel-wise label map. The model will be evaluated using the intersection over union (IOU) and Dice Score. Finally, you will download the model and upload it to the grader in Coursera to get your score for the assignment."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "cnr0kbM0g2LD"
   },
   "source": [
    "## Exercises\n",
    "\n",
    "We've given you some boilerplate code to work with and these are the 5 exercises you need to fill out before you can successfully get the segmentation masks.\n",
    "\n",
    "* [Exercise 1 - Define the Basic Convolution Block](#exercise-1)\n",
    "* [Exercise 2 - Define the Downsampling Path](#exercise-2)\n",
    "* [Exercise 3 - Define the FCN-8 decoder](#exercise-3)\n",
    "* [Exercise 4 - Compile the Model](#exercise-4)\n",
    "* [Exercise 5 - Model Training](#exercise-5)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "MZ3g9dJxSxmN"
   },
   "source": [
    "## Imports\n",
    "\n",
    "As usual, let's start by importing the packages you will use in this lab."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "#!pip install tensorflow_datasets"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "aifz2907kxYN",
    "outputId": "2be3ee9c-dfd0-4674-e783-dcc4cd8c422c"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tensorflow version 2.8.0\n"
     ]
    }
   ],
   "source": [
    "try:\n",
    "  # %tensorflow_version only exists in Colab.\n",
    "  %tensorflow_version 2.x\n",
    "except Exception:\n",
    "  pass\n",
    "\n",
    "import os\n",
    "import zipfile\n",
    "\n",
    "import PIL.Image, PIL.ImageFont, PIL.ImageDraw\n",
    "import numpy as np\n",
    "from matplotlib import pyplot as plt\n",
    "\n",
    "import tensorflow as tf\n",
    "import tensorflow_datasets as tfds\n",
    "from sklearn.model_selection import train_test_split\n",
    "\n",
    "print(\"Tensorflow version \" + tf.__version__)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "7RYh6cCzXE6R"
   },
   "source": [
    "## Download the dataset"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "QUGGF3wfqYni"
   },
   "source": [
    "[M2NIST](https://www.kaggle.com/farhanhubble/multimnistm2nist) is a **multi digit** [MNIST](http://yann.lecun.com/exdb/mnist/). \n",
    "Each image has up to 3 digits from MNIST digits and the corresponding labels file has the segmentation masks.\n",
    "\n",
    "The dataset is available on [Kaggle](https://www.kaggle.com) and you can find it [here](https://www.kaggle.com/farhanhubble/multimnistm2nist)\n",
    "\n",
    "To make it easier for you, we're hosting it on Google Cloud so you can download without Kaggle credentials.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "ROok0i9rMcu0",
    "outputId": "72b7356c-1731-42e5-bbce-5a064375f068"
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "SYSTEM_WGETRC = c:/progra~1/wget/etc/wgetrc\n",
      "syswgetrc = C:\\Program Files (x86)\\GnuWin32/etc/wgetrc\n",
      "--2022-03-28 20:19:44--  https://storage.googleapis.com/laurencemoroney-blog.appspot.com/m2nist.zip\n",
      "æ­£åœ¨è§£æž�ä¸»æœº storage.googleapis.com... 142.250.67.16, 142.250.71.80, 172.217.24.48, ...\n",
      "Connecting to storage.googleapis.com|142.250.67.16|:443... å·²è¿žæŽ¥ã€‚\n",
      "WARNING: cannot verify storage.googleapis.com's certificate, issued by `/C=US/O=Google Trust Services LLC/CN=GTS CA 1C3':\n",
      "  Unable to locally verify the issuer's authority.\n",
      "WARNING: certificate common name `*.storage.googleapis.com' doesn't match requested host name `storage.googleapis.com'.\n",
      "å·²å�‘å‡º HTTP è¯·æ±‚ï¼Œæ­£åœ¨ç­‰å¾…å›žåº”... 200 OK\n",
      "é•¿åº¦ï¼š17378168 (17M) [application/zip]\n",
      "Saving to: `/tmp/m2nist.zip'\n",
      "\n",
      "     0K .......... .......... .......... .......... ..........  0% 1.32M 13s\n",
      "    50K .......... .......... .......... .......... ..........  0% 2.09M 10s\n",
      "   100K .......... .......... .......... .......... ..........  0% 4.62M 8s\n",
      "   150K .......... .......... .......... .......... ..........  1% 3.45M 7s\n",
      "   200K .......... .......... .......... .......... ..........  1% 6.27M 6s\n",
      "   250K .......... .......... .......... .......... ..........  1% 4.16M 6s\n",
      "   300K .......... .......... .......... .......... ..........  2% 4.88M 5s\n",
      "   350K .......... .......... .......... .......... ..........  2% 4.23M 5s\n",
      "   400K .......... .......... .......... .......... ..........  2% 5.09M 5s\n",
      "   450K .......... .......... .......... .......... ..........  2% 6.91M 5s\n",
      "   500K .......... .......... .......... .......... ..........  3% 6.44M 4s\n",
      "   550K .......... .......... .......... .......... ..........  3% 6.44M 4s\n",
      "   600K .......... .......... .......... .......... ..........  3% 7.07M 4s\n",
      "   650K .......... .......... .......... .......... ..........  4% 6.04M 4s\n",
      "   700K .......... .......... .......... .......... ..........  4% 6.10M 4s\n",
      "   750K .......... .......... .......... .......... ..........  4% 6.45M 4s\n",
      "   800K .......... .......... .......... .......... ..........  5% 6.96M 4s\n",
      "   850K .......... .......... .......... .......... ..........  5% 5.86M 4s\n",
      "   900K .......... .......... .......... .......... ..........  5% 6.49M 4s\n",
      "   950K .......... .......... .......... .......... ..........  5% 6.75M 3s\n",
      "  1000K .......... .......... .......... .......... ..........  6% 5.33M 3s\n",
      "  1050K .......... .......... .......... .......... ..........  6% 6.83M 3s\n",
      "  1100K .......... .......... .......... .......... ..........  6% 6.85M 3s\n",
      "  1150K .......... .......... .......... .......... ..........  7% 5.14M 3s\n",
      "  1200K .......... .......... .......... .......... ..........  7% 7.12M 3s\n",
      "  1250K .......... .......... .......... .......... ..........  7% 6.53M 3s\n",
      "  1300K .......... .......... .......... .......... ..........  7% 7.07M 3s\n",
      "  1350K .......... .......... .......... .......... ..........  8% 6.66M 3s\n",
      "  1400K .......... .......... .......... .......... ..........  8% 6.80M 3s\n",
      "  1450K .......... .......... .......... .......... ..........  8% 6.65M 3s\n",
      "  1500K .......... .......... .......... .......... ..........  9% 6.99M 3s\n",
      "  1550K .......... .......... .......... .......... ..........  9% 5.74M 3s\n",
      "  1600K .......... .......... .......... .......... ..........  9% 5.96M 3s\n",
      "  1650K .......... .......... .......... .......... .......... 10% 6.46M 3s\n",
      "  1700K .......... .......... .......... .......... .......... 10% 7.37M 3s\n",
      "  1750K .......... .......... .......... .......... .......... 10% 6.87M 3s\n",
      "  1800K .......... .......... .......... .......... .......... 10% 6.78M 3s\n",
      "  1850K .......... .......... .......... .......... .......... 11% 6.58M 3s\n",
      "  1900K .......... .......... .......... .......... .......... 11% 6.92M 3s\n",
      "  1950K .......... .......... .......... .......... .......... 11% 5.14M 3s\n",
      "  2000K .......... .......... .......... .......... .......... 12% 6.78M 3s\n",
      "  2050K .......... .......... .......... .......... .......... 12% 6.82M 3s\n",
      "  2100K .......... .......... .......... .......... .......... 12% 6.86M 3s\n",
      "  2150K .......... .......... .......... .......... .......... 12% 6.95M 3s\n",
      "  2200K .......... .......... .......... .......... .......... 13% 6.77M 3s\n",
      "  2250K .......... .......... .......... .......... .......... 13% 6.89M 3s\n",
      "  2300K .......... .......... .......... .......... .......... 13% 6.85M 3s\n",
      "  2350K .......... .......... .......... .......... .......... 14% 5.07M 3s\n",
      "  2400K .......... .......... .......... .......... .......... 14% 6.69M 3s\n",
      "  2450K .......... .......... .......... .......... .......... 14% 6.90M 3s\n",
      "  2500K .......... .......... .......... .......... .......... 15% 6.63M 3s\n",
      "  2550K .......... .......... .......... .......... .......... 15% 7.11M 3s\n",
      "  2600K .......... .......... .......... .......... .......... 15% 6.98M 3s\n",
      "  2650K .......... .......... .......... .......... .......... 15% 6.54M 3s\n",
      "  2700K .......... .......... .......... .......... .......... 16% 6.83M 2s\n",
      "  2750K .......... .......... .......... .......... .......... 16% 4.38M 2s\n",
      "  2800K .......... .......... .......... .......... .......... 16% 8.70M 2s\n",
      "  2850K .......... .......... .......... .......... .......... 17% 6.87M 2s\n",
      "  2900K .......... .......... .......... .......... .......... 17% 7.53M 2s\n",
      "  2950K .......... .......... .......... .......... .......... 17% 6.16M 2s\n",
      "  3000K .......... .......... .......... .......... .......... 17% 7.03M 2s\n",
      "  3050K .......... .......... .......... .......... .......... 18% 7.67M 2s\n",
      "  3100K .......... .......... .......... .......... .......... 18% 6.03M 2s\n",
      "  3150K .......... .......... .......... .......... .......... 18% 5.20M 2s\n",
      "  3200K .......... .......... .......... .......... .......... 19% 6.59M 2s\n",
      "  3250K .......... .......... .......... .......... .......... 19% 5.53M 2s\n",
      "  3300K .......... .......... .......... .......... .......... 19% 6.61M 2s\n",
      "  3350K .......... .......... .......... .......... .......... 20% 7.49M 2s\n",
      "  3400K .......... .......... .......... .......... .......... 20% 6.11M 2s\n",
      "  3450K .......... .......... .......... .......... .......... 20% 8.47M 2s\n",
      "  3500K .......... .......... .......... .......... .......... 20% 6.07M 2s\n",
      "  3550K .......... .......... .......... .......... .......... 21% 5.20M 2s\n",
      "  3600K .......... .......... .......... .......... .......... 21% 6.90M 2s\n",
      "  3650K .......... .......... .......... .......... .......... 21% 6.21M 2s\n",
      "  3700K .......... .......... .......... .......... .......... 22% 6.51M 2s\n",
      "  3750K .......... .......... .......... .......... .......... 22% 7.56M 2s\n",
      "  3800K .......... .......... .......... .......... .......... 22% 6.90M 2s\n",
      "  3850K .......... .......... .......... .......... .......... 22% 6.99M 2s\n",
      "  3900K .......... .......... .......... .......... .......... 23% 6.16M 2s\n",
      "  3950K .......... .......... .......... .......... .......... 23% 5.16M 2s\n",
      "  4000K .......... .......... .......... .......... .......... 23% 7.15M 2s\n",
      "  4050K .......... .......... .......... .......... .......... 24% 7.08M 2s\n",
      "  4100K .......... .......... .......... .......... .......... 24% 7.30M 2s\n",
      "  4150K .......... .......... .......... .......... .......... 24% 6.42M 2s\n",
      "  4200K .......... .......... .......... .......... .......... 25% 6.59M 2s\n",
      "  4250K .......... .......... .......... .......... .......... 25% 6.88M 2s\n",
      "  4300K .......... .......... .......... .......... .......... 25% 6.79M 2s\n",
      "  4350K .......... .......... .......... .......... .......... 25% 5.30M 2s\n",
      "  4400K .......... .......... .......... .......... .......... 26% 5.89M 2s\n",
      "  4450K .......... .......... .......... .......... .......... 26% 7.65M 2s\n",
      "  4500K .......... .......... .......... .......... .......... 26% 6.76M 2s\n",
      "  4550K .......... .......... .......... .......... .......... 27% 7.01M 2s\n",
      "  4600K .......... .......... .......... .......... .......... 27% 6.75M 2s\n",
      "  4650K .......... .......... .......... .......... .......... 27% 7.29M 2s\n",
      "  4700K .......... .......... .......... .......... .......... 27% 6.40M 2s\n",
      "  4750K .......... .......... .......... .......... .......... 28% 5.10M 2s\n",
      "  4800K .......... .......... .......... .......... .......... 28% 6.88M 2s\n",
      "  4850K .......... .......... .......... .......... .......... 28% 6.76M 2s\n",
      "  4900K .......... .......... .......... .......... .......... 29% 6.75M 2s\n",
      "  4950K .......... .......... .......... .......... .......... 29% 7.83M 2s\n",
      "  5000K .......... .......... .......... .......... .......... 29% 6.27M 2s\n",
      "  5050K .......... .......... .......... .......... .......... 30% 6.51M 2s\n",
      "  5100K .......... .......... .......... .......... .......... 30% 7.84M 2s\n",
      "  5150K .......... .......... .......... .......... .......... 30% 4.61M 2s\n",
      "  5200K .......... .......... .......... .......... .......... 30% 7.08M 2s\n",
      "  5250K .......... .......... .......... .......... .......... 31% 7.02M 2s\n",
      "  5300K .......... .......... .......... .......... .......... 31% 7.27M 2s\n",
      "  5350K .......... .......... .......... .......... .......... 31% 6.23M 2s\n",
      "  5400K .......... .......... .......... .......... .......... 32% 6.88M 2s\n",
      "  5450K .......... .......... .......... .......... .......... 32% 6.68M 2s\n",
      "  5500K .......... .......... .......... .......... .......... 32% 6.81M 2s\n",
      "  5550K .......... .......... .......... .......... .......... 32% 5.27M 2s\n",
      "  5600K .......... .......... .......... .......... .......... 33% 6.35M 2s\n",
      "  5650K .......... .......... .......... .......... .......... 33% 6.96M 2s\n",
      "  5700K .......... .......... .......... .......... .......... 33% 6.98M 2s\n",
      "  5750K .......... .......... .......... .......... .......... 34% 6.61M 2s\n",
      "  5800K .......... .......... .......... .......... .......... 34% 6.99M 2s\n",
      "  5850K .......... .......... .......... .......... .......... 34% 6.20M 2s\n",
      "  5900K .......... .......... .......... .......... .......... 35% 7.72M 2s\n",
      "  5950K .......... .......... .......... .......... .......... 35% 4.94M 2s\n",
      "  6000K .......... .......... .......... .......... .......... 35% 5.87M 2s\n",
      "  6050K .......... .......... .......... .......... .......... 35% 8.72M 2s\n",
      "  6100K .......... .......... .......... .......... .......... 36% 6.65M 2s\n",
      "  6150K .......... .......... .......... .......... .......... 36% 7.40M 2s\n",
      "  6200K .......... .......... .......... .......... .......... 36% 6.85M 2s\n",
      "  6250K .......... .......... .......... .......... .......... 37% 7.04M 2s\n",
      "  6300K .......... .......... .......... .......... .......... 37% 6.11M 2s\n",
      "  6350K .......... .......... .......... .......... .......... 37% 5.11M 2s\n",
      "  6400K .......... .......... .......... .......... .......... 38% 7.32M 2s\n",
      "  6450K .......... .......... .......... .......... .......... 38% 6.34M 2s\n",
      "  6500K .......... .......... .......... .......... .......... 38% 6.94M 2s\n",
      "  6550K .......... .......... .......... .......... .......... 38% 6.52M 2s\n",
      "  6600K .......... .......... .......... .......... .......... 39% 6.98M 2s\n",
      "  6650K .......... .......... .......... .......... .......... 39% 6.80M 2s\n",
      "  6700K .......... .......... .......... .......... .......... 39% 6.86M 2s\n",
      "  6750K .......... .......... .......... .......... .......... 40% 5.23M 2s\n",
      "  6800K .......... .......... .......... .......... .......... 40% 6.71M 2s\n",
      "  6850K .......... .......... .......... .......... .......... 40% 6.34M 2s\n",
      "  6900K .......... .......... .......... .......... .......... 40% 7.47M 2s\n",
      "  6950K .......... .......... .......... .......... .......... 41% 6.59M 2s\n",
      "  7000K .......... .......... .......... .......... .......... 41% 7.12M 2s\n",
      "  7050K .......... .......... .......... .......... .......... 41% 6.84M 2s\n",
      "  7100K .......... .......... .......... .......... .......... 42% 6.81M 2s\n",
      "  7150K .......... .......... .......... .......... .......... 42% 4.79M 2s\n",
      "  7200K .......... .......... .......... .......... .......... 42% 7.01M 2s\n",
      "  7250K .......... .......... .......... .......... .......... 43% 6.80M 2s\n",
      "  7300K .......... .......... .......... .......... .......... 43% 7.12M 2s\n",
      "  7350K .......... .......... .......... .......... .......... 43% 6.78M 2s\n",
      "  7400K .......... .......... .......... .......... .......... 43% 6.85M 2s\n",
      "  7450K .......... .......... .......... .......... .......... 44% 7.15M 2s\n",
      "  7500K .......... .......... .......... .......... .......... 44% 6.35M 1s\n",
      "  7550K .......... .......... .......... .......... .......... 44% 5.11M 1s\n",
      "  7600K .......... .......... .......... .......... .......... 45% 6.73M 1s\n",
      "  7650K .......... .......... .......... .......... .......... 45% 8.06M 1s\n",
      "  7700K .......... .......... .......... .......... .......... 45% 6.80M 1s\n",
      "  7750K .......... .......... .......... .......... .......... 45% 6.06M 1s\n",
      "  7800K .......... .......... .......... .......... .......... 46% 6.62M 1s\n",
      "  7850K .......... .......... .......... .......... .......... 46% 6.73M 1s\n",
      "  7900K .......... .......... .......... .......... .......... 46% 7.50M 1s\n",
      "  7950K .......... .......... .......... .......... .......... 47% 4.93M 1s\n",
      "  8000K .......... .......... .......... .......... .......... 47% 6.88M 1s\n",
      "  8050K .......... .......... .......... .......... .......... 47% 6.95M 1s\n",
      "  8100K .......... .......... .......... .......... .......... 48% 6.90M 1s\n",
      "  8150K .......... .......... .......... .......... .......... 48% 6.81M 1s\n",
      "  8200K .......... .......... .......... .......... .......... 48% 6.51M 1s\n",
      "  8250K .......... .......... .......... .......... .......... 48% 6.77M 1s\n",
      "  8300K .......... .......... .......... .......... .......... 49% 6.47M 1s\n",
      "  8350K .......... .......... .......... .......... .......... 49% 5.48M 1s\n",
      "  8400K .......... .......... .......... .......... .......... 49% 6.69M 1s\n",
      "  8450K .......... .......... .......... .......... .......... 50% 6.86M 1s\n",
      "  8500K .......... .......... .......... .......... .......... 50% 6.89M 1s\n",
      "  8550K .......... .......... .......... .......... .......... 50% 6.85M 1s\n",
      "  8600K .......... .......... .......... .......... .......... 50% 6.76M 1s\n",
      "  8650K .......... .......... .......... .......... .......... 51% 6.75M 1s\n",
      "  8700K .......... .......... .......... .......... .......... 51% 6.47M 1s\n",
      "  8750K .......... .......... .......... .......... .......... 51% 5.33M 1s\n",
      "  8800K .......... .......... .......... .......... .......... 52% 7.07M 1s\n",
      "  8850K .......... .......... .......... .......... .......... 52% 6.80M 1s\n",
      "  8900K .......... .......... .......... .......... .......... 52% 6.67M 1s\n",
      "  8950K .......... .......... .......... .......... .......... 53% 6.89M 1s\n",
      "  9000K .......... .......... .......... .......... .......... 53% 6.61M 1s\n",
      "  9050K .......... .......... .......... .......... .......... 53% 7.08M 1s\n",
      "  9100K .......... .......... .......... .......... .......... 53% 6.85M 1s\n",
      "  9150K .......... .......... .......... .......... .......... 54% 4.89M 1s\n",
      "  9200K .......... .......... .......... .......... .......... 54% 7.54M 1s\n",
      "  9250K .......... .......... .......... .......... .......... 54% 6.60M 1s\n",
      "  9300K .......... .......... .......... .......... .......... 55% 6.53M 1s\n",
      "  9350K .......... .......... .......... .......... .......... 55% 8.10M 1s\n",
      "  9400K .......... .......... .......... .......... .......... 55% 6.30M 1s\n",
      "  9450K .......... .......... .......... .......... .......... 55% 6.38M 1s\n",
      "  9500K .......... .......... .......... .......... .......... 56% 6.58M 1s\n",
      "  9550K .......... .......... .......... .......... .......... 56% 5.45M 1s\n",
      "  9600K .......... .......... .......... .......... .......... 56% 6.25M 1s\n",
      "  9650K .......... .......... .......... .......... .......... 57% 7.07M 1s\n",
      "  9700K .......... .......... .......... .......... .......... 57% 7.06M 1s\n",
      "  9750K .......... .......... .......... .......... .......... 57% 6.33M 1s\n",
      "  9800K .......... .......... .......... .......... .......... 58% 6.76M 1s\n",
      "  9850K .......... .......... .......... .......... .......... 58% 7.20M 1s\n",
      "  9900K .......... .......... .......... .......... .......... 58% 7.02M 1s\n",
      "  9950K .......... .......... .......... .......... .......... 58% 5.50M 1s\n",
      " 10000K .......... .......... .......... .......... .......... 59% 6.38M 1s\n",
      " 10050K .......... .......... .......... .......... .......... 59% 6.53M 1s\n",
      " 10100K .......... .......... .......... .......... .......... 59% 6.74M 1s\n",
      " 10150K .......... .......... .......... .......... .......... 60% 6.59M 1s\n",
      " 10200K .......... .......... .......... .......... .......... 60% 7.09M 1s\n",
      " 10250K .......... .......... .......... .......... .......... 60% 7.29M 1s\n",
      " 10300K .......... .......... .......... .......... .......... 60% 6.55M 1s\n",
      " 10350K .......... .......... .......... .......... .......... 61% 5.11M 1s\n",
      " 10400K .......... .......... .......... .......... .......... 61% 6.84M 1s\n",
      " 10450K .......... .......... .......... .......... .......... 61% 6.61M 1s\n",
      " 10500K .......... .......... .......... .......... .......... 62% 6.88M 1s\n",
      " 10550K .......... .......... .......... .......... .......... 62% 7.09M 1s\n",
      " 10600K .......... .......... .......... .......... .......... 62% 6.71M 1s\n",
      " 10650K .......... .......... .......... .......... .......... 63% 6.53M 1s\n",
      " 10700K .......... .......... .......... .......... .......... 63% 7.17M 1s\n",
      " 10750K .......... .......... .......... .......... .......... 63% 4.82M 1s\n",
      " 10800K .......... .......... .......... .......... .......... 63% 7.14M 1s\n",
      " 10850K .......... .......... .......... .......... .......... 64% 6.85M 1s\n",
      " 10900K .......... .......... .......... .......... .......... 64% 7.23M 1s\n",
      " 10950K .......... .......... .......... .......... .......... 64% 6.68M 1s\n",
      " 11000K .......... .......... .......... .......... .......... 65% 6.91M 1s\n",
      " 11050K .......... .......... .......... .......... .......... 65% 6.80M 1s\n",
      " 11100K .......... .......... .......... .......... .......... 65% 7.04M 1s\n",
      " 11150K .......... .......... .......... .......... .......... 65% 4.90M 1s\n",
      " 11200K .......... .......... .......... .......... .......... 66% 6.78M 1s\n",
      " 11250K .......... .......... .......... .......... .......... 66% 7.17M 1s\n",
      " 11300K .......... .......... .......... .......... .......... 66% 7.49M 1s\n",
      " 11350K .......... .......... .......... .......... .......... 67% 6.84M 1s\n",
      " 11400K .......... .......... .......... .......... .......... 67% 5.86M 1s\n",
      " 11450K .......... .......... .......... .......... .......... 67% 6.87M 1s\n",
      " 11500K .......... .......... .......... .......... .......... 68% 7.40M 1s\n",
      " 11550K .......... .......... .......... .......... .......... 68% 4.88M 1s\n",
      " 11600K .......... .......... .......... .......... .......... 68% 6.91M 1s\n",
      " 11650K .......... .......... .......... .......... .......... 68% 6.84M 1s\n",
      " 11700K .......... .......... .......... .......... .......... 69% 6.95M 1s\n",
      " 11750K .......... .......... .......... .......... .......... 69% 6.70M 1s\n",
      " 11800K .......... .......... .......... .......... .......... 69% 6.76M 1s\n",
      " 11850K .......... .......... .......... .......... .......... 70% 6.76M 1s\n",
      " 11900K .......... .......... .......... .......... .......... 70% 7.74M 1s\n",
      " 11950K .......... .......... .......... .......... .......... 70% 4.85M 1s\n",
      " 12000K .......... .......... .......... .......... .......... 71% 6.50M 1s\n",
      " 12050K .......... .......... .......... .......... .......... 71% 6.54M 1s\n",
      " 12100K .......... .......... .......... .......... .......... 71% 8.16M 1s\n",
      " 12150K .......... .......... .......... .......... .......... 71% 5.95M 1s\n",
      " 12200K .......... .......... .......... .......... .......... 72% 7.16M 1s\n",
      " 12250K .......... .......... .......... .......... .......... 72% 6.76M 1s\n",
      " 12300K .......... .......... .......... .......... .......... 72% 7.34M 1s\n",
      " 12350K .......... .......... .......... .......... .......... 73% 5.08M 1s\n",
      " 12400K .......... .......... .......... .......... .......... 73% 6.40M 1s\n",
      " 12450K .......... .......... .......... .......... .......... 73% 7.32M 1s\n",
      " 12500K .......... .......... .......... .......... .......... 73% 6.33M 1s\n",
      " 12550K .......... .......... .......... .......... .......... 74% 6.98M 1s\n",
      " 12600K .......... .......... .......... .......... .......... 74% 6.84M 1s\n",
      " 12650K .......... .......... .......... .......... .......... 74% 6.63M 1s\n",
      " 12700K .......... .......... .......... .......... .......... 75% 6.98M 1s\n",
      " 12750K .......... .......... .......... .......... .......... 75% 5.17M 1s\n",
      " 12800K .......... .......... .......... .......... .......... 75% 6.58M 1s\n",
      " 12850K .......... .......... .......... .......... .......... 76% 7.33M 1s\n",
      " 12900K .......... .......... .......... .......... .......... 76% 6.03M 1s\n",
      " 12950K .......... .......... .......... .......... .......... 76% 7.39M 1s\n",
      " 13000K .......... .......... .......... .......... .......... 76% 6.84M 1s\n",
      " 13050K .......... .......... .......... .......... .......... 77% 6.81M 1s\n",
      " 13100K .......... .......... .......... .......... .......... 77% 8.00M 1s\n",
      " 13150K .......... .......... .......... .......... .......... 77% 5.07M 1s\n",
      " 13200K .......... .......... .......... .......... .......... 78% 5.99M 1s\n",
      " 13250K .......... .......... .......... .......... .......... 78% 6.86M 1s\n",
      " 13300K .......... .......... .......... .......... .......... 78% 6.76M 1s\n",
      " 13350K .......... .......... .......... .......... .......... 78% 7.03M 1s\n",
      " 13400K .......... .......... .......... .......... .......... 79% 6.61M 1s\n",
      " 13450K .......... .......... .......... .......... .......... 79% 7.41M 1s\n",
      " 13500K .......... .......... .......... .......... .......... 79% 6.54M 1s\n",
      " 13550K .......... .......... .......... .......... .......... 80% 4.85M 1s\n",
      " 13600K .......... .......... .......... .......... .......... 80% 6.72M 1s\n",
      " 13650K .......... .......... .......... .......... .......... 80% 7.45M 1s\n",
      " 13700K .......... .......... .......... .......... .......... 81% 6.79M 0s\n",
      " 13750K .......... .......... .......... .......... .......... 81% 6.72M 0s\n",
      " 13800K .......... .......... .......... .......... .......... 81% 6.99M 0s\n",
      " 13850K .......... .......... .......... .......... .......... 81% 6.97M 0s\n",
      " 13900K .......... .......... .......... .......... .......... 82% 6.44M 0s\n",
      " 13950K .......... .......... .......... .......... .......... 82% 5.10M 0s\n",
      " 14000K .......... .......... .......... .......... .......... 82% 6.88M 0s\n",
      " 14050K .......... .......... .......... .......... .......... 83% 7.09M 0s\n",
      " 14100K .......... .......... .......... .......... .......... 83% 6.64M 0s\n",
      " 14150K .......... .......... .......... .......... .......... 83% 7.28M 0s\n",
      " 14200K .......... .......... .......... .......... .......... 83% 6.38M 0s\n",
      " 14250K .......... .......... .......... .......... .......... 84% 7.02M 0s\n",
      " 14300K .......... .......... .......... .......... .......... 84% 6.37M 0s\n",
      " 14350K .......... .......... .......... .......... .......... 84% 5.36M 0s\n",
      " 14400K .......... .......... .......... .......... .......... 85% 6.84M 0s\n",
      " 14450K .......... .......... .......... .......... .......... 85% 6.61M 0s\n",
      " 14500K .......... .......... .......... .......... .......... 85% 6.80M 0s\n",
      " 14550K .......... .......... .......... .......... .......... 86% 6.95M 0s\n",
      " 14600K .......... .......... .......... .......... .......... 86% 7.72M 0s\n",
      " 14650K .......... .......... .......... .......... .......... 86% 6.24M 0s\n",
      " 14700K .......... .......... .......... .......... .......... 86% 6.65M 0s\n",
      " 14750K .......... .......... .......... .......... .......... 87% 5.08M 0s\n",
      " 14800K .......... .......... .......... .......... .......... 87% 6.94M 0s\n",
      " 14850K .......... .......... .......... .......... .......... 87% 6.97M 0s\n",
      " 14900K .......... .......... .......... .......... .......... 88% 6.70M 0s\n",
      " 14950K .......... .......... .......... .......... .......... 88% 6.64M 0s\n",
      " 15000K .......... .......... .......... .......... .......... 88% 7.15M 0s\n",
      " 15050K .......... .......... .......... .......... .......... 88% 6.69M 0s\n",
      " 15100K .......... .......... .......... .......... .......... 89% 6.72M 0s\n",
      " 15150K .......... .......... .......... .......... .......... 89% 5.11M 0s\n",
      " 15200K .......... .......... .......... .......... .......... 89% 6.85M 0s\n",
      " 15250K .......... .......... .......... .......... .......... 90% 6.82M 0s\n",
      " 15300K .......... .......... .......... .......... .......... 90% 7.09M 0s\n",
      " 15350K .......... .......... .......... .......... .......... 90% 6.62M 0s\n",
      " 15400K .......... .......... .......... .......... .......... 91% 6.79M 0s\n",
      " 15450K .......... .......... .......... .......... .......... 91% 7.19M 0s\n",
      " 15500K .......... .......... .......... .......... .......... 91% 6.46M 0s\n",
      " 15550K .......... .......... .......... .......... .......... 91% 5.67M 0s\n",
      " 15600K .......... .......... .......... .......... .......... 92% 6.08M 0s\n",
      " 15650K .......... .......... .......... .......... .......... 92% 6.79M 0s\n",
      " 15700K .......... .......... .......... .......... .......... 92% 6.95M 0s\n",
      " 15750K .......... .......... .......... .......... .......... 93% 6.54M 0s\n",
      " 15800K .......... .......... .......... .......... .......... 93% 7.94M 0s\n",
      " 15850K .......... .......... .......... .......... .......... 93% 6.12M 0s\n",
      " 15900K .......... .......... .......... .......... .......... 93% 6.92M 0s\n",
      " 15950K .......... .......... .......... .......... .......... 94% 5.21M 0s\n",
      " 16000K .......... .......... .......... .......... .......... 94% 6.80M 0s\n",
      " 16050K .......... .......... .......... .......... .......... 94% 6.49M 0s\n",
      " 16100K .......... .......... .......... .......... .......... 95% 6.96M 0s\n",
      " 16150K .......... .......... .......... .......... .......... 95% 6.53M 0s\n",
      " 16200K .......... .......... .......... .......... .......... 95% 6.87M 0s\n",
      " 16250K .......... .......... .......... .......... .......... 96% 7.59M 0s\n",
      " 16300K .......... .......... .......... .......... .......... 96% 6.29M 0s\n",
      " 16350K .......... .......... .......... .......... .......... 96% 5.26M 0s\n",
      " 16400K .......... .......... .......... .......... .......... 96% 7.62M 0s\n",
      " 16450K .......... .......... .......... .......... .......... 97% 5.84M 0s\n",
      " 16500K .......... .......... .......... .......... .......... 97% 7.04M 0s\n",
      " 16550K .......... .......... .......... .......... .......... 97% 7.79M 0s\n",
      " 16600K .......... .......... .......... .......... .......... 98% 6.33M 0s\n",
      " 16650K .......... .......... .......... .......... .......... 98% 6.46M 0s\n",
      " 16700K .......... .......... .......... .......... .......... 98% 7.94M 0s\n",
      " 16750K .......... .......... .......... .......... .......... 98% 4.81M 0s\n",
      " 16800K .......... .......... .......... .......... .......... 99% 6.69M 0s\n",
      " 16850K .......... .......... .......... .......... .......... 99% 5.90M 0s\n",
      " 16900K .......... .......... .......... .......... .......... 99% 7.84M 0s\n",
      " 16950K .......... ..........                                 100% 11.8M=2.6s\n",
      "\n",
      "2022-03-28 20:19:47 (6.37 MB/s) - `/tmp/m2nist.zip' saved [17378168/17378168]\n",
      "\n"
     ]
    }
   ],
   "source": [
    "# download zipped dataset\n",
    "!wget --no-check-certificate \\\n",
    "    https://storage.googleapis.com/laurencemoroney-blog.appspot.com/m2nist.zip \\\n",
    "    -O /tmp/m2nist.zip\n",
    "\n",
    "# find and extract to a local folder ('/tmp/training')\n",
    "local_zip = '/tmp/m2nist.zip'\n",
    "zip_ref = zipfile.ZipFile(local_zip, 'r')\n",
    "zip_ref.extractall('/tmp/training')\n",
    "zip_ref.close()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "xy17LYR7XJNa"
   },
   "source": [
    "## Load and Preprocess the Dataset"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "LXGMrWnkrvpK"
   },
   "source": [
    "This dataset can be easily preprocessed since it is available as **Numpy Array Files (.npy)**\n",
    "\n",
    "1. **combined.npy** has the image files containing the multiple MNIST digits. Each image is of size **64 x 84** (height x width, in pixels).\n",
    "\n",
    "2. **segmented.npy** has the corresponding segmentation masks. Each segmentation mask is also of size **64 x 84**.\n",
    "\n",
    "This dataset has **5000** samples and you can make appropriate training, validation, and test splits as required for the problem.\n",
    "\n",
    "With that, let's define a few utility functions for loading and preprocessing the dataset."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "id": "Jy_pw5I2-xLP"
   },
   "outputs": [],
   "source": [
    "BATCH_SIZE = 32\n",
    "\n",
    "def read_image_and_annotation(image, annotation):\n",
    "  '''\n",
    "  Casts the image and annotation to their expected data type and\n",
    "  normalizes the input image so that each pixel is in the range [-1, 1]\n",
    "\n",
    "  Args:\n",
    "    image (numpy array) -- input image\n",
    "    annotation (numpy array) -- ground truth label map\n",
    "\n",
    "  Returns:\n",
    "    preprocessed image-annotation pair\n",
    "  '''\n",
    "\n",
    "  image = tf.cast(image, dtype=tf.float32)\n",
    "  image = tf.reshape(image, (image.shape[0], image.shape[1], 1,))\n",
    "  annotation = tf.cast(annotation, dtype=tf.int32)\n",
    "  image = image / 127.5\n",
    "  image -= 1\n",
    "\n",
    "  return image, annotation\n",
    "\n",
    "\n",
    "def get_training_dataset(images, annos):\n",
    "  '''\n",
    "  Prepares shuffled batches of the training set.\n",
    "  \n",
    "  Args:\n",
    "    images (list of strings) -- paths to each image file in the train set\n",
    "    annos (list of strings) -- paths to each label map in the train set\n",
    "\n",
    "  Returns:\n",
    "    tf Dataset containing the preprocessed train set\n",
    "  '''\n",
    "  training_dataset = tf.data.Dataset.from_tensor_slices((images, annos))\n",
    "  training_dataset = training_dataset.map(read_image_and_annotation)\n",
    "\n",
    "  training_dataset = training_dataset.shuffle(512, reshuffle_each_iteration=True)\n",
    "  training_dataset = training_dataset.batch(BATCH_SIZE)\n",
    "  training_dataset = training_dataset.repeat()\n",
    "  training_dataset = training_dataset.prefetch(-1)\n",
    "\n",
    "  return training_dataset\n",
    "\n",
    "\n",
    "def get_validation_dataset(images, annos):\n",
    "  '''\n",
    "  Prepares batches of the validation set.\n",
    "  \n",
    "  Args:\n",
    "    images (list of strings) -- paths to each image file in the val set\n",
    "    annos (list of strings) -- paths to each label map in the val set\n",
    "\n",
    "  Returns:\n",
    "    tf Dataset containing the preprocessed validation set\n",
    "  '''\n",
    "  validation_dataset = tf.data.Dataset.from_tensor_slices((images, annos))\n",
    "  validation_dataset = validation_dataset.map(read_image_and_annotation)\n",
    "  validation_dataset = validation_dataset.batch(BATCH_SIZE)\n",
    "  validation_dataset = validation_dataset.repeat()\n",
    "\n",
    "  return validation_dataset\n",
    "\n",
    "\n",
    "def get_test_dataset(images, annos):\n",
    "  '''\n",
    "  Prepares batches of the test set.\n",
    "  \n",
    "  Args:\n",
    "    images (list of strings) -- paths to each image file in the test set\n",
    "    annos (list of strings) -- paths to each label map in the test set\n",
    "\n",
    "  Returns:\n",
    "    tf Dataset containing the preprocessed validation set\n",
    "  '''\n",
    "  test_dataset = tf.data.Dataset.from_tensor_slices((images, annos))\n",
    "  test_dataset = test_dataset.map(read_image_and_annotation)\n",
    "  test_dataset = test_dataset.batch(BATCH_SIZE, drop_remainder=True)\n",
    "\n",
    "  return test_dataset\n",
    "\n",
    "\n",
    "def load_images_and_segments():\n",
    "  '''\n",
    "  Loads the images and segments as numpy arrays from npy files \n",
    "  and makes splits for training, validation and test datasets.\n",
    "\n",
    "  Returns:\n",
    "    3 tuples containing the train, val, and test splits\n",
    "  '''\n",
    "\n",
    "  #Loads images and segmentation masks.\n",
    "  images = np.load('/tmp/training/combined.npy')\n",
    "  segments = np.load('/tmp/training/segmented.npy')\n",
    "\n",
    "  #Makes training, validation, test splits from loaded images and segmentation masks.\n",
    "  train_images, val_images, train_annos, val_annos = train_test_split(images, segments, test_size=0.2, shuffle=True)\n",
    "  val_images, test_images, val_annos, test_annos = train_test_split(val_images, val_annos, test_size=0.2, shuffle=True)\n",
    "\n",
    "  return (train_images, train_annos), (val_images, val_annos), (test_images, test_annos)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "iPHO1YbTACcu"
   },
   "source": [
    "You can now load the preprocessed dataset and define the training, validation, and test sets."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "id": "hIS70_um_Y7n",
    "lines_to_next_cell": 2
   },
   "outputs": [],
   "source": [
    "# Load Dataset\n",
    "train_slices, val_slices, test_slices = load_images_and_segments()\n",
    "\n",
    "# Create training, validation, test datasets.\n",
    "training_dataset = get_training_dataset(train_slices[0], train_slices[1])\n",
    "validation_dataset = get_validation_dataset(val_slices[0], val_slices[1])\n",
    "test_dataset = get_test_dataset(test_slices[0], test_slices[1])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "oKXJYZi7A0dF"
   },
   "source": [
    "## Let's Take a Look at the Dataset\n",
    "\n",
    "You may want to visually inspect the dataset before and after training. Like above, we've included utility functions to help show a few images as well as their annotations (i.e. labels)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "cellView": "both",
    "id": "d46YCbvPafbp"
   },
   "outputs": [],
   "source": [
    "# Visualization Utilities\n",
    "\n",
    "# there are 11 classes in the dataset: one class for each digit (0 to 9) plus the background class\n",
    "n_classes = 11\n",
    "\n",
    "# assign a random color for each class\n",
    "colors = [tuple(np.random.randint(256, size=3) / 255.0) for i in range(n_classes)]\n",
    "\n",
    "def fuse_with_pil(images):\n",
    "  '''\n",
    "  Creates a blank image and pastes input images\n",
    "\n",
    "  Args:\n",
    "    images (list of numpy arrays) - numpy array representations of the images to paste\n",
    "  \n",
    "  Returns:\n",
    "    PIL Image object containing the images\n",
    "  '''\n",
    "\n",
    "  widths = (image.shape[1] for image in images)\n",
    "  heights = (image.shape[0] for image in images)\n",
    "  total_width = sum(widths)\n",
    "  max_height = max(heights)\n",
    "\n",
    "  new_im = PIL.Image.new('RGB', (total_width, max_height))\n",
    "\n",
    "  x_offset = 0\n",
    "  for im in images:\n",
    "    pil_image = PIL.Image.fromarray(np.uint8(im))\n",
    "    new_im.paste(pil_image, (x_offset,0))\n",
    "    x_offset += im.shape[1]\n",
    "  \n",
    "  return new_im\n",
    "\n",
    "\n",
    "def give_color_to_annotation(annotation):\n",
    "  '''\n",
    "  Converts a 2-D annotation to a numpy array with shape (height, width, 3) where\n",
    "  the third axis represents the color channel. The label values are multiplied by\n",
    "  255 and placed in this axis to give color to the annotation\n",
    "\n",
    "  Args:\n",
    "    annotation (numpy array) - label map array\n",
    "  \n",
    "  Returns:\n",
    "    the annotation array with an additional color channel/axis\n",
    "  '''\n",
    "  seg_img = np.zeros( (annotation.shape[0],annotation.shape[1], 3) ).astype('float')\n",
    "  \n",
    "  for c in range(n_classes):\n",
    "    segc = (annotation == c)\n",
    "    seg_img[:,:,0] += segc*( colors[c][0] * 255.0)\n",
    "    seg_img[:,:,1] += segc*( colors[c][1] * 255.0)\n",
    "    seg_img[:,:,2] += segc*( colors[c][2] * 255.0)\n",
    "  \n",
    "  return seg_img\n",
    "\n",
    "\n",
    "def show_annotation_and_prediction(image, annotation, prediction, iou_list, dice_score_list):\n",
    "  '''\n",
    "  Displays the images with the ground truth and predicted label maps. Also overlays the metrics.\n",
    "\n",
    "  Args:\n",
    "    image (numpy array) -- the input image\n",
    "    annotation (numpy array) -- the ground truth label map\n",
    "    prediction (numpy array) -- the predicted label map\n",
    "    iou_list (list of floats) -- the IOU values for each class\n",
    "    dice_score_list (list of floats) -- the Dice Score for each class\n",
    "  '''\n",
    "\n",
    "  new_ann = np.argmax(annotation, axis=2)\n",
    "  true_img = give_color_to_annotation(new_ann)\n",
    "  pred_img = give_color_to_annotation(prediction)\n",
    "\n",
    "  image = image + 1\n",
    "  image = image * 127.5\n",
    "  image = np.reshape(image, (image.shape[0], image.shape[1],))\n",
    "  image = np.uint8(image)\n",
    "  images = [image, np.uint8(pred_img), np.uint8(true_img)]\n",
    "\n",
    "  metrics_by_id = [(idx, iou, dice_score) for idx, (iou, dice_score) in enumerate(zip(iou_list, dice_score_list)) if iou > 0.0 and idx < 10]\n",
    "  metrics_by_id.sort(key=lambda tup: tup[1], reverse=True)  # sorts in place\n",
    "\n",
    "  display_string_list = [\"{}: IOU: {} Dice Score: {}\".format(idx, iou, dice_score) for idx, iou, dice_score in metrics_by_id]\n",
    "  display_string = \"\\n\".join(display_string_list)\n",
    "\n",
    "  plt.figure(figsize=(15, 4))\n",
    "\n",
    "  for idx, im in enumerate(images):\n",
    "    plt.subplot(1, 3, idx+1)\n",
    "    if idx == 1:\n",
    "      plt.xlabel(display_string)\n",
    "    plt.xticks([])\n",
    "    plt.yticks([])\n",
    "    plt.imshow(im)\n",
    "\n",
    "\n",
    "def show_annotation_and_image(image, annotation):\n",
    "  '''\n",
    "  Displays the image and its annotation side by side\n",
    "\n",
    "  Args:\n",
    "    image (numpy array) -- the input image\n",
    "    annotation (numpy array) -- the label map\n",
    "  '''\n",
    "  new_ann = np.argmax(annotation, axis=2)\n",
    "  seg_img = give_color_to_annotation(new_ann)\n",
    "  \n",
    "  image = image + 1\n",
    "  image = image * 127.5\n",
    "  image = np.reshape(image, (image.shape[0], image.shape[1],))\n",
    "\n",
    "  image = np.uint8(image)\n",
    "  images = [image, seg_img]\n",
    "  \n",
    "  images = [image, seg_img]\n",
    "  fused_img = fuse_with_pil(images)\n",
    "  plt.imshow(fused_img)\n",
    "\n",
    "\n",
    "def list_show_annotation(dataset, num_images):\n",
    "  '''\n",
    "  Displays images and its annotations side by side\n",
    "\n",
    "  Args:\n",
    "    dataset (tf Dataset) -- batch of images and annotations\n",
    "    num_images (int) -- number of images to display\n",
    "  '''\n",
    "  ds = dataset.unbatch()\n",
    "\n",
    "  plt.figure(figsize=(20, 15))\n",
    "  plt.title(\"Images And Annotations\")\n",
    "  plt.subplots_adjust(bottom=0.1, top=0.9, hspace=0.05)\n",
    "\n",
    "  for idx, (image, annotation) in enumerate(ds.take(num_images)):\n",
    "    plt.subplot(5, 5, idx + 1)\n",
    "    plt.yticks([])\n",
    "    plt.xticks([])\n",
    "    show_annotation_and_image(image.numpy(), annotation.numpy())\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "hEfyChmKEFKe"
   },
   "source": [
    "You can view a subset of the images from the dataset with the `list_show_annotation()` function defined above. Run the cells below to see the image on the left and its pixel-wise ground truth label map on the right."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 282
    },
    "id": "AFO_hIhLWYT4",
    "outputId": "c32e963d-3c90-4bf7-9e58-bf771433abc5"
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1440x1080 with 10 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# get 10 images from the training set\n",
    "list_show_annotation(training_dataset, 10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 282
    },
    "id": "bdgVkp8wZua0",
    "outputId": "de0a7766-37f0-4706-e940-c36f329c2242"
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1440x1080 with 10 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# get 10 images from the validation set\n",
    "list_show_annotation(validation_dataset, 10)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "KkpgIcE2FeKI"
   },
   "source": [
    "You see from the images above the colors assigned to each class (i.e 0 to 9 plus the background). If you don't like these colors, feel free to rerun the cell where `colors` is defined to get another set of random colors. Alternatively, you can assign the RGB values for each class instead of relying on random values."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "QFv2k8xabRb8"
   },
   "source": [
    "## Define the Model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "Q_ylpyqJDQiF"
   },
   "source": [
    "As discussed in the lectures, the image segmentation model will have two paths:\n",
    "\n",
    "1. **Downsampling Path** - This part of the network extracts the features in the image. This is done through a series of convolution and pooling layers. The final output is a reduced image (because of the pooling layers) with the extracted features. You will build a custom CNN from scratch for this path.\n",
    "\n",
    "2. **Upsampling Path** - This takes the output of the downsampling path and generates the predictions while also converting the image back to its original size. You will use an FCN-8 decoder for this path."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "BHlBUZvsDybt"
   },
   "source": [
    "### Define the Basic Convolution Block"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "T-XO9xkN1OR3"
   },
   "source": [
    "<a name='exercise-1'></a>\n",
    "\n",
    "#### **Exercise 1**\n",
    "\n",
    "Please complete the function below to build the basic convolution block for our CNN. This will have two [Conv2D](https://keras.io/api/layers/convolution_layers/convolution2d/) layers each followed by a [LeakyReLU](https://keras.io/api/layers/activation_layers/leaky_relu/), then [max pooled](https://keras.io/api/layers/pooling_layers/max_pooling2d/) and [batch-normalized](https://keras.io/api/layers/normalization_layers/batch_normalization/). Use the functional syntax to stack these layers.\n",
    "\n",
    "$$Input -> Conv2D -> LeakyReLU -> Conv2D -> LeakyReLU -> MaxPooling2D -> BatchNormalization$$\n",
    "\n",
    "When defining the Conv2D layers, note that our data inputs will have the 'channels' dimension last. You may want to check the `data_format` argument in the [docs](https://keras.io/api/layers/convolution_layers/convolution2d/) regarding this. Take note of the `padding` argument too like you did in the ungraded labs.\n",
    "\n",
    "Lastly, to use the `LeakyReLU` activation, you **do not** need to nest it inside an `Activation` layer (e.g. `x = tf.keras.layers.Activation(tf.keras.layers.LeakyReLU()(x)`). You can simply stack the layer directly instead (e.g. `x = tf.keras.layers.LeakyReLU()(x)`)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "id": "azEEVytHR0Kn"
   },
   "outputs": [],
   "source": [
    "# parameter describing where the channel dimension is found in our dataset\n",
    "IMAGE_ORDERING = 'channels_last'\n",
    "\n",
    "def conv_block(input, filters, kernel_size, pooling_size, pool_strides):\n",
    "  '''\n",
    "  Args:\n",
    "    input (tensor) -- batch of images or features\n",
    "    filters (int) -- number of filters of the Conv2D layers\n",
    "    kernel_size (int) -- kernel_size setting of the Conv2D layers\n",
    "    pooling_size (int) -- pooling size of the MaxPooling2D layers\n",
    "    pool_strides (int) -- strides setting of the MaxPooling2D layers\n",
    "  \n",
    "  Returns:\n",
    "    (tensor) max pooled and batch-normalized features of the input \n",
    "  '''\n",
    "  ### START CODE HERE ###\n",
    "  # use the functional syntax to stack the layers as shown in the diagram above\n",
    "  x = tf.keras.layers.Conv2D(filters, kernel_size, padding='same', data_format=IMAGE_ORDERING)(input)\n",
    "  x = tf.keras.layers.LeakyReLU()(x)\n",
    "  x = tf.keras.layers.Conv2D(filters, kernel_size, padding='same', data_format=IMAGE_ORDERING)(x)\n",
    "  x = tf.keras.layers.LeakyReLU()(x)\n",
    "  x = tf.keras.layers.MaxPooling2D()(x)\n",
    "  x = tf.keras.layers.BatchNormalization()(x)\n",
    "  ### END CODE HERE ###\n",
    "\n",
    "  return x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "uGtlHuM6KCRg",
    "outputId": "c029f872-b6b2-4d42-d69a-ca9e7b81bcf4"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"model\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " input_1 (InputLayer)        [(None, 64, 84, 1)]       0         \n",
      "                                                                 \n",
      " conv2d (Conv2D)             (None, 64, 84, 32)        320       \n",
      "                                                                 \n",
      " leaky_re_lu (LeakyReLU)     (None, 64, 84, 32)        0         \n",
      "                                                                 \n",
      " conv2d_1 (Conv2D)           (None, 64, 84, 32)        9248      \n",
      "                                                                 \n",
      " leaky_re_lu_1 (LeakyReLU)   (None, 64, 84, 32)        0         \n",
      "                                                                 \n",
      " max_pooling2d (MaxPooling2D  (None, 32, 42, 32)       0         \n",
      " )                                                               \n",
      "                                                                 \n",
      " batch_normalization (BatchN  (None, 32, 42, 32)       128       \n",
      " ormalization)                                                   \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 9,696\n",
      "Trainable params: 9,632\n",
      "Non-trainable params: 64\n",
      "_________________________________________________________________\n",
      "None\n"
     ]
    }
   ],
   "source": [
    "# TEST CODE:\n",
    "\n",
    "test_input = tf.keras.layers.Input(shape=(64,84, 1))\n",
    "test_output = conv_block(test_input, 32, 3, 2, 2)\n",
    "test_model = tf.keras.Model(inputs=test_input, outputs=test_output)\n",
    "\n",
    "print(test_model.summary())\n",
    "\n",
    "# free up test resources\n",
    "del test_input, test_output, test_model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "Mmf_ZaMP2rV3"
   },
   "source": [
    "**Expected Output**:\n",
    "\n",
    "Please pay attention to the *(type)* and *Output Shape* columns. The *Layer* name beside the type may be different depending on how many times you ran the cell (e.g. `input_7` can be `input_1`)\n",
    "\n",
    "```txt\n",
    "Model: \"functional_1\"\n",
    "_________________________________________________________________\n",
    "Layer (type)                 Output Shape              Param #   \n",
    "=================================================================\n",
    "input_1 (InputLayer)         [(None, 64, 84, 1)]       0         \n",
    "_________________________________________________________________\n",
    "conv2d (Conv2D)              (None, 64, 84, 32)        320       \n",
    "_________________________________________________________________\n",
    "leaky_re_lu (LeakyReLU)      (None, 64, 84, 32)        0         \n",
    "_________________________________________________________________\n",
    "conv2d_1 (Conv2D)            (None, 64, 84, 32)        9248      \n",
    "_________________________________________________________________\n",
    "leaky_re_lu_1 (LeakyReLU)    (None, 64, 84, 32)        0         \n",
    "_________________________________________________________________\n",
    "max_pooling2d (MaxPooling2D) (None, 32, 42, 32)        0         \n",
    "_________________________________________________________________\n",
    "batch_normalization (BatchNo (None, 32, 42, 32)        128       \n",
    "=================================================================\n",
    "Total params: 9,696\n",
    "Trainable params: 9,632\n",
    "Non-trainable params: 64\n",
    "_________________________________________________________________\n",
    "None\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "4-jJbC91EXTV"
   },
   "source": [
    "### Define the Downsampling Path"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "e2ueOTmc3HxJ"
   },
   "source": [
    "<a name='exercise-2'></a>\n",
    "\n",
    "#### **Exercise 2**\n",
    "\n",
    "Now that we've defined the building block of our encoder, you can now build the downsampling path. Please complete the function below to create the encoder. This should chain together five convolution building blocks to create a feature extraction CNN minus the fully connected layers.\n",
    "\n",
    "*Notes*: \n",
    "1. To optimize processing or to make the output dimensions of each layer easier to work with, it is sometimes advisable to apply some zero-padding to the input image. With the boilerplate code we have provided below, we have padded the input width to 96 pixels using the [ZeroPadding2D layer](https://keras.io/api/layers/reshaping_layers/zero_padding2d/). This works well if you're going to use the first ungraded lab of this week as reference. This is not required however. You can remove it later and see how it will affect your parameters. For instance, you might need to pass in a non-square kernel size to the decoder in Exercise 3 (e.g. `(4,5)`) to match the output dimensions of Exercise 2. \n",
    "\n",
    "2. We recommend keeping the pool size and stride parameters constant at 2."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "id": "F2VNB99LRwQr"
   },
   "outputs": [],
   "source": [
    "def FCN8(input_height=64, input_width=84):\n",
    "    '''\n",
    "    Defines the downsampling path of the image segmentation model.\n",
    "\n",
    "    Args:\n",
    "      input_height (int) -- height of the images\n",
    "      width (int) -- width of the images\n",
    "\n",
    "    Returns:\n",
    "    (tuple of tensors, tensor)\n",
    "      tuple of tensors -- features extracted at blocks 3 to 5\n",
    "      tensor -- copy of the input\n",
    "    '''\n",
    "   \n",
    "    img_input = tf.keras.layers.Input(shape=(input_height,input_width, 1))\n",
    "\n",
    "    ### START CODE HERE ###\n",
    "    \n",
    "    # pad the input image width to 96 pixels\n",
    "    x = tf.keras.layers.ZeroPadding2D(((0, 0), (0, 96-input_width)))(img_input)\n",
    "    \n",
    "    # Block 1\n",
    "    x = conv_block(x, filters = 32, kernel_size =(3, 3), pooling_size = (2, 2), pool_strides = (2, 2))\n",
    "\n",
    "    # Block 2\n",
    "    x = conv_block(x, filters = 64, kernel_size =(3, 3), pooling_size = (2, 2), pool_strides = (2, 2))\n",
    "\n",
    "    # Block 3\n",
    "    x = conv_block(x, filters = 128, kernel_size =(3, 3), pooling_size = (2, 2), pool_strides = (2, 2))\n",
    "    # save the feature map at this stage\n",
    "    f3 = x\n",
    "\n",
    "    # Block 4\n",
    "    x = conv_block(x, filters = 256, kernel_size =(3, 3), pooling_size = (2, 2), pool_strides = (2, 2))\n",
    "    # save the feature map at this stage\n",
    "    f4 = x\n",
    "\n",
    "    # Block 5\n",
    "    x = conv_block(x, filters = 256, kernel_size =(3, 3), pooling_size = (2, 2), pool_strides = (2, 2))\n",
    "    # save the feature map at this stage\n",
    "    f5 = x\n",
    "\n",
    "    ### END CODE HERE ###\n",
    "  \n",
    "    return (f3, f4, f5), img_input"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "HVQm1W0CNICS",
    "outputId": "8487e2dd-fa4f-402d-f2a8-b650ac069b7e"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"model_1\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " input_2 (InputLayer)        [(None, 64, 84, 1)]       0         \n",
      "                                                                 \n",
      " zero_padding2d (ZeroPadding  (None, 64, 96, 1)        0         \n",
      " 2D)                                                             \n",
      "                                                                 \n",
      " conv2d_2 (Conv2D)           (None, 64, 96, 32)        320       \n",
      "                                                                 \n",
      " leaky_re_lu_2 (LeakyReLU)   (None, 64, 96, 32)        0         \n",
      "                                                                 \n",
      " conv2d_3 (Conv2D)           (None, 64, 96, 32)        9248      \n",
      "                                                                 \n",
      " leaky_re_lu_3 (LeakyReLU)   (None, 64, 96, 32)        0         \n",
      "                                                                 \n",
      " max_pooling2d_1 (MaxPooling  (None, 32, 48, 32)       0         \n",
      " 2D)                                                             \n",
      "                                                                 \n",
      " batch_normalization_1 (Batc  (None, 32, 48, 32)       128       \n",
      " hNormalization)                                                 \n",
      "                                                                 \n",
      " conv2d_4 (Conv2D)           (None, 32, 48, 64)        18496     \n",
      "                                                                 \n",
      " leaky_re_lu_4 (LeakyReLU)   (None, 32, 48, 64)        0         \n",
      "                                                                 \n",
      " conv2d_5 (Conv2D)           (None, 32, 48, 64)        36928     \n",
      "                                                                 \n",
      " leaky_re_lu_5 (LeakyReLU)   (None, 32, 48, 64)        0         \n",
      "                                                                 \n",
      " max_pooling2d_2 (MaxPooling  (None, 16, 24, 64)       0         \n",
      " 2D)                                                             \n",
      "                                                                 \n",
      " batch_normalization_2 (Batc  (None, 16, 24, 64)       256       \n",
      " hNormalization)                                                 \n",
      "                                                                 \n",
      " conv2d_6 (Conv2D)           (None, 16, 24, 128)       73856     \n",
      "                                                                 \n",
      " leaky_re_lu_6 (LeakyReLU)   (None, 16, 24, 128)       0         \n",
      "                                                                 \n",
      " conv2d_7 (Conv2D)           (None, 16, 24, 128)       147584    \n",
      "                                                                 \n",
      " leaky_re_lu_7 (LeakyReLU)   (None, 16, 24, 128)       0         \n",
      "                                                                 \n",
      " max_pooling2d_3 (MaxPooling  (None, 8, 12, 128)       0         \n",
      " 2D)                                                             \n",
      "                                                                 \n",
      " batch_normalization_3 (Batc  (None, 8, 12, 128)       512       \n",
      " hNormalization)                                                 \n",
      "                                                                 \n",
      " conv2d_8 (Conv2D)           (None, 8, 12, 256)        295168    \n",
      "                                                                 \n",
      " leaky_re_lu_8 (LeakyReLU)   (None, 8, 12, 256)        0         \n",
      "                                                                 \n",
      " conv2d_9 (Conv2D)           (None, 8, 12, 256)        590080    \n",
      "                                                                 \n",
      " leaky_re_lu_9 (LeakyReLU)   (None, 8, 12, 256)        0         \n",
      "                                                                 \n",
      " max_pooling2d_4 (MaxPooling  (None, 4, 6, 256)        0         \n",
      " 2D)                                                             \n",
      "                                                                 \n",
      " batch_normalization_4 (Batc  (None, 4, 6, 256)        1024      \n",
      " hNormalization)                                                 \n",
      "                                                                 \n",
      " conv2d_10 (Conv2D)          (None, 4, 6, 256)         590080    \n",
      "                                                                 \n",
      " leaky_re_lu_10 (LeakyReLU)  (None, 4, 6, 256)         0         \n",
      "                                                                 \n",
      " conv2d_11 (Conv2D)          (None, 4, 6, 256)         590080    \n",
      "                                                                 \n",
      " leaky_re_lu_11 (LeakyReLU)  (None, 4, 6, 256)         0         \n",
      "                                                                 \n",
      " max_pooling2d_5 (MaxPooling  (None, 2, 3, 256)        0         \n",
      " 2D)                                                             \n",
      "                                                                 \n",
      " batch_normalization_5 (Batc  (None, 2, 3, 256)        1024      \n",
      " hNormalization)                                                 \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 2,354,784\n",
      "Trainable params: 2,353,312\n",
      "Non-trainable params: 1,472\n",
      "_________________________________________________________________\n",
      "None\n"
     ]
    }
   ],
   "source": [
    "# TEST CODE:\n",
    "\n",
    "test_convs, test_img_input = FCN8()\n",
    "test_model = tf.keras.Model(inputs=test_img_input, outputs=[test_convs, test_img_input])\n",
    "\n",
    "print(test_model.summary())\n",
    "\n",
    "del test_convs, test_img_input, test_model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "sxklZe_D3jOI"
   },
   "source": [
    "**Expected Output**:\n",
    "\n",
    "You should see the layers of your `conv_block()` being repeated 5 times like the output below.\n",
    "\n",
    "```txt\n",
    "Model: \"functional_3\"\n",
    "_________________________________________________________________\n",
    "Layer (type)                 Output Shape              Param #   \n",
    "=================================================================\n",
    "input_3 (InputLayer)         [(None, 64, 84, 1)]       0         \n",
    "_________________________________________________________________\n",
    "zero_padding2d (ZeroPadding2 (None, 64, 96, 1)         0         \n",
    "_________________________________________________________________\n",
    "conv2d_2 (Conv2D)            (None, 64, 96, 32)        320       \n",
    "_________________________________________________________________\n",
    "leaky_re_lu_2 (LeakyReLU)    (None, 64, 96, 32)        0         \n",
    "_________________________________________________________________\n",
    "conv2d_3 (Conv2D)            (None, 64, 96, 32)        9248      \n",
    "_________________________________________________________________\n",
    "leaky_re_lu_3 (LeakyReLU)    (None, 64, 96, 32)        0         \n",
    "_________________________________________________________________\n",
    "max_pooling2d_1 (MaxPooling2 (None, 32, 48, 32)        0         \n",
    "_________________________________________________________________\n",
    "batch_normalization_1 (Batch (None, 32, 48, 32)        128       \n",
    "_________________________________________________________________\n",
    "conv2d_4 (Conv2D)            (None, 32, 48, 64)        18496     \n",
    "_________________________________________________________________\n",
    "leaky_re_lu_4 (LeakyReLU)    (None, 32, 48, 64)        0         \n",
    "_________________________________________________________________\n",
    "conv2d_5 (Conv2D)            (None, 32, 48, 64)        36928     \n",
    "_________________________________________________________________\n",
    "leaky_re_lu_5 (LeakyReLU)    (None, 32, 48, 64)        0         \n",
    "_________________________________________________________________\n",
    "max_pooling2d_2 (MaxPooling2 (None, 16, 24, 64)        0         \n",
    "_________________________________________________________________\n",
    "batch_normalization_2 (Batch (None, 16, 24, 64)        256       \n",
    "_________________________________________________________________\n",
    "conv2d_6 (Conv2D)            (None, 16, 24, 128)       73856     \n",
    "_________________________________________________________________\n",
    "leaky_re_lu_6 (LeakyReLU)    (None, 16, 24, 128)       0         \n",
    "_________________________________________________________________\n",
    "conv2d_7 (Conv2D)            (None, 16, 24, 128)       147584    \n",
    "_________________________________________________________________\n",
    "leaky_re_lu_7 (LeakyReLU)    (None, 16, 24, 128)       0         \n",
    "_________________________________________________________________\n",
    "max_pooling2d_3 (MaxPooling2 (None, 8, 12, 128)        0         \n",
    "_________________________________________________________________\n",
    "batch_normalization_3 (Batch (None, 8, 12, 128)        512       \n",
    "_________________________________________________________________\n",
    "conv2d_8 (Conv2D)            (None, 8, 12, 256)        295168    \n",
    "_________________________________________________________________\n",
    "leaky_re_lu_8 (LeakyReLU)    (None, 8, 12, 256)        0         \n",
    "_________________________________________________________________\n",
    "conv2d_9 (Conv2D)            (None, 8, 12, 256)        590080    \n",
    "_________________________________________________________________\n",
    "leaky_re_lu_9 (LeakyReLU)    (None, 8, 12, 256)        0         \n",
    "_________________________________________________________________\n",
    "max_pooling2d_4 (MaxPooling2 (None, 4, 6, 256)         0         \n",
    "_________________________________________________________________\n",
    "batch_normalization_4 (Batch (None, 4, 6, 256)         1024      \n",
    "_________________________________________________________________\n",
    "conv2d_10 (Conv2D)           (None, 4, 6, 256)         590080    \n",
    "_________________________________________________________________\n",
    "leaky_re_lu_10 (LeakyReLU)   (None, 4, 6, 256)         0         \n",
    "_________________________________________________________________\n",
    "conv2d_11 (Conv2D)           (None, 4, 6, 256)         590080    \n",
    "_________________________________________________________________\n",
    "leaky_re_lu_11 (LeakyReLU)   (None, 4, 6, 256)         0         \n",
    "_________________________________________________________________\n",
    "max_pooling2d_5 (MaxPooling2 (None, 2, 3, 256)         0         \n",
    "_________________________________________________________________\n",
    "batch_normalization_5 (Batch (None, 2, 3, 256)         1024      \n",
    "=================================================================\n",
    "Total params: 2,354,784\n",
    "Trainable params: 2,353,312\n",
    "Non-trainable params: 1,472\n",
    "_________________________________________________________________\n",
    "None\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "LbjYEQU8Eq-T"
   },
   "source": [
    "### Define the FCN-8 decoder"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "Ux5nAU155E5O"
   },
   "source": [
    "<a name='exercise-3'></a>\n",
    "\n",
    "#### **Exercise 3**\n",
    "\n",
    "Now you can define the upsampling path taking the outputs of convolutions at each stage as arguments. This will be very similar to what you did in the ungraded lab (VGG16-FCN8-CamVid) so you can refer to it if you need a refresher. \n",
    "* Note: remember to set the `data_format` parameter for the Conv2D layers. \n",
    "\n",
    "Here is also the diagram you saw in class on how it should work:\n",
    "\n",
    "<img src='https://drive.google.com/uc?export=view&id=1lrqB4YegV8jXWNfyYAaeuFlwXIc54aRP' alt='fcn-8'>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "id": "giYEct_Se5Xj"
   },
   "outputs": [],
   "source": [
    "def fcn8_decoder(convs, n_classes):\n",
    "  # features from the encoder stage\n",
    "  f3, f4, f5 = convs\n",
    "\n",
    "  # number of filters\n",
    "  n = 512\n",
    "\n",
    "  # add convolutional layers on top of the CNN extractor.\n",
    "  o = tf.keras.layers.Conv2D(n , (7 , 7) , activation='relu' , padding='same', name=\"conv6\", data_format=IMAGE_ORDERING)(f5)\n",
    "  o = tf.keras.layers.Dropout(0.5)(o)\n",
    "\n",
    "  o = tf.keras.layers.Conv2D(n , (1 , 1) , activation='relu' , padding='same', name=\"conv7\", data_format=IMAGE_ORDERING)(o)\n",
    "  o = tf.keras.layers.Dropout(0.5)(o)\n",
    "\n",
    "  o = tf.keras.layers.Conv2D(n_classes,  (1, 1), activation='relu' , padding='same', data_format=IMAGE_ORDERING)(o)\n",
    "\n",
    "    \n",
    "  ### START CODE HERE ###\n",
    "\n",
    "  # Upsample `o` above and crop any extra pixels introduced\n",
    "  o = tf.keras.layers.Conv2DTranspose(n_classes, kernel_size = (4, 4), strides = (2, 2), use_bias = False)(o)\n",
    "  o = tf.keras.layers.Cropping2D(cropping = (1,1))(o)\n",
    "\n",
    "  # load the pool 4 prediction and do a 1x1 convolution to reshape it to the same shape of `o` above\n",
    "  o2 = f4\n",
    "  o2 = tf.keras.layers.Conv2D(n_classes , ( 1 , 1 ) , activation='relu' , padding='same', data_format=IMAGE_ORDERING)(o2)\n",
    "\n",
    "  # add the results of the upsampling and pool 4 prediction\n",
    "  o = tf.keras.layers.Add()([o, o2])\n",
    "\n",
    "  # upsample the resulting tensor of the operation you just did\n",
    "  o = tf.keras.layers.Conv2DTranspose(n_classes, kernel_size = (4, 4), strides = (2, 2), use_bias = False)(o)\n",
    "  o = tf.keras.layers.Cropping2D(cropping = (1,1))(o)\n",
    "\n",
    "  # load the pool 3 prediction and do a 1x1 convolution to reshape it to the same shape of `o` above\n",
    "  o2 = f3\n",
    "  o2 = tf.keras.layers.Conv2D(n_classes , ( 1 , 1 ) , activation='relu' , padding='same', data_format=IMAGE_ORDERING)(o2)\n",
    "\n",
    "  # add the results of the upsampling and pool 3 prediction\n",
    "  o = tf.keras.layers.Add()([o, o2])\n",
    "\n",
    "  # upsample up to the size of the original image\n",
    "  o = tf.keras.layers.Conv2DTranspose(n_classes, kernel_size = (8, 8), strides = (8, 8), use_bias = False)(o)\n",
    "  o = tf.keras.layers.Cropping2D(((0, 0), (0, 96-84)))(o)\n",
    "\n",
    "  # append a sigmoid activation\n",
    "  o = (tf.keras.layers.Activation('sigmoid'))(o)\n",
    "  ### END CODE HERE ###\n",
    "\n",
    "  return o"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "lQHioDlR5K1_",
    "outputId": "b0dd13d1-9af0-4999-ac39-202cd24520e9"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(None, 64, 84, 11)\n"
     ]
    }
   ],
   "source": [
    "# TEST CODE\n",
    "\n",
    "test_convs, test_img_input = FCN8()\n",
    "test_fcn8_decoder = fcn8_decoder(test_convs, 11)\n",
    "\n",
    "print(test_fcn8_decoder.shape)\n",
    "\n",
    "del test_convs, test_img_input, test_fcn8_decoder"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "C1qTwTX-5fwH"
   },
   "source": [
    "**Expected Output:**\n",
    "\n",
    "```txt\n",
    "(None, 64, 84, 11)\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "cJOhQz86Qk6n"
   },
   "source": [
    "### Define the Complete Model\n",
    "\n",
    "The downsampling and upsampling paths can now be combined as shown below."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "id": "9EJEf484312h",
    "lines_to_next_cell": 2
   },
   "outputs": [],
   "source": [
    "# start the encoder using the default input size 64 x 84\n",
    "convs, img_input = FCN8()\n",
    "\n",
    "# pass the convolutions obtained in the encoder to the decoder\n",
    "dec_op = fcn8_decoder(convs, n_classes)\n",
    "\n",
    "# define the model specifying the input (batch of images) and output (decoder output)\n",
    "model = tf.keras.Model(inputs = img_input, outputs = dec_op)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "2GAenp1M4gXx",
    "outputId": "ba457ed0-866b-4774-c9e4-68a39984e9c5"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"model_2\"\n",
      "__________________________________________________________________________________________________\n",
      " Layer (type)                   Output Shape         Param #     Connected to                     \n",
      "==================================================================================================\n",
      " input_4 (InputLayer)           [(None, 64, 84, 1)]  0           []                               \n",
      "                                                                                                  \n",
      " zero_padding2d_2 (ZeroPadding2  (None, 64, 96, 1)   0           ['input_4[0][0]']                \n",
      " D)                                                                                               \n",
      "                                                                                                  \n",
      " conv2d_25 (Conv2D)             (None, 64, 96, 32)   320         ['zero_padding2d_2[0][0]']       \n",
      "                                                                                                  \n",
      " leaky_re_lu_22 (LeakyReLU)     (None, 64, 96, 32)   0           ['conv2d_25[0][0]']              \n",
      "                                                                                                  \n",
      " conv2d_26 (Conv2D)             (None, 64, 96, 32)   9248        ['leaky_re_lu_22[0][0]']         \n",
      "                                                                                                  \n",
      " leaky_re_lu_23 (LeakyReLU)     (None, 64, 96, 32)   0           ['conv2d_26[0][0]']              \n",
      "                                                                                                  \n",
      " max_pooling2d_11 (MaxPooling2D  (None, 32, 48, 32)  0           ['leaky_re_lu_23[0][0]']         \n",
      " )                                                                                                \n",
      "                                                                                                  \n",
      " batch_normalization_11 (BatchN  (None, 32, 48, 32)  128         ['max_pooling2d_11[0][0]']       \n",
      " ormalization)                                                                                    \n",
      "                                                                                                  \n",
      " conv2d_27 (Conv2D)             (None, 32, 48, 64)   18496       ['batch_normalization_11[0][0]'] \n",
      "                                                                                                  \n",
      " leaky_re_lu_24 (LeakyReLU)     (None, 32, 48, 64)   0           ['conv2d_27[0][0]']              \n",
      "                                                                                                  \n",
      " conv2d_28 (Conv2D)             (None, 32, 48, 64)   36928       ['leaky_re_lu_24[0][0]']         \n",
      "                                                                                                  \n",
      " leaky_re_lu_25 (LeakyReLU)     (None, 32, 48, 64)   0           ['conv2d_28[0][0]']              \n",
      "                                                                                                  \n",
      " max_pooling2d_12 (MaxPooling2D  (None, 16, 24, 64)  0           ['leaky_re_lu_25[0][0]']         \n",
      " )                                                                                                \n",
      "                                                                                                  \n",
      " batch_normalization_12 (BatchN  (None, 16, 24, 64)  256         ['max_pooling2d_12[0][0]']       \n",
      " ormalization)                                                                                    \n",
      "                                                                                                  \n",
      " conv2d_29 (Conv2D)             (None, 16, 24, 128)  73856       ['batch_normalization_12[0][0]'] \n",
      "                                                                                                  \n",
      " leaky_re_lu_26 (LeakyReLU)     (None, 16, 24, 128)  0           ['conv2d_29[0][0]']              \n",
      "                                                                                                  \n",
      " conv2d_30 (Conv2D)             (None, 16, 24, 128)  147584      ['leaky_re_lu_26[0][0]']         \n",
      "                                                                                                  \n",
      " leaky_re_lu_27 (LeakyReLU)     (None, 16, 24, 128)  0           ['conv2d_30[0][0]']              \n",
      "                                                                                                  \n",
      " max_pooling2d_13 (MaxPooling2D  (None, 8, 12, 128)  0           ['leaky_re_lu_27[0][0]']         \n",
      " )                                                                                                \n",
      "                                                                                                  \n",
      " batch_normalization_13 (BatchN  (None, 8, 12, 128)  512         ['max_pooling2d_13[0][0]']       \n",
      " ormalization)                                                                                    \n",
      "                                                                                                  \n",
      " conv2d_31 (Conv2D)             (None, 8, 12, 256)   295168      ['batch_normalization_13[0][0]'] \n",
      "                                                                                                  \n",
      " leaky_re_lu_28 (LeakyReLU)     (None, 8, 12, 256)   0           ['conv2d_31[0][0]']              \n",
      "                                                                                                  \n",
      " conv2d_32 (Conv2D)             (None, 8, 12, 256)   590080      ['leaky_re_lu_28[0][0]']         \n",
      "                                                                                                  \n",
      " leaky_re_lu_29 (LeakyReLU)     (None, 8, 12, 256)   0           ['conv2d_32[0][0]']              \n",
      "                                                                                                  \n",
      " max_pooling2d_14 (MaxPooling2D  (None, 4, 6, 256)   0           ['leaky_re_lu_29[0][0]']         \n",
      " )                                                                                                \n",
      "                                                                                                  \n",
      " batch_normalization_14 (BatchN  (None, 4, 6, 256)   1024        ['max_pooling2d_14[0][0]']       \n",
      " ormalization)                                                                                    \n",
      "                                                                                                  \n",
      " conv2d_33 (Conv2D)             (None, 4, 6, 256)    590080      ['batch_normalization_14[0][0]'] \n",
      "                                                                                                  \n",
      " leaky_re_lu_30 (LeakyReLU)     (None, 4, 6, 256)    0           ['conv2d_33[0][0]']              \n",
      "                                                                                                  \n",
      " conv2d_34 (Conv2D)             (None, 4, 6, 256)    590080      ['leaky_re_lu_30[0][0]']         \n",
      "                                                                                                  \n",
      " leaky_re_lu_31 (LeakyReLU)     (None, 4, 6, 256)    0           ['conv2d_34[0][0]']              \n",
      "                                                                                                  \n",
      " max_pooling2d_15 (MaxPooling2D  (None, 2, 3, 256)   0           ['leaky_re_lu_31[0][0]']         \n",
      " )                                                                                                \n",
      "                                                                                                  \n",
      " batch_normalization_15 (BatchN  (None, 2, 3, 256)   1024        ['max_pooling2d_15[0][0]']       \n",
      " ormalization)                                                                                    \n",
      "                                                                                                  \n",
      " conv6 (Conv2D)                 (None, 2, 3, 512)    6423040     ['batch_normalization_15[0][0]'] \n",
      "                                                                                                  \n",
      " dropout_2 (Dropout)            (None, 2, 3, 512)    0           ['conv6[0][0]']                  \n",
      "                                                                                                  \n",
      " conv7 (Conv2D)                 (None, 2, 3, 512)    262656      ['dropout_2[0][0]']              \n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "                                                                                                  \n",
      " dropout_3 (Dropout)            (None, 2, 3, 512)    0           ['conv7[0][0]']                  \n",
      "                                                                                                  \n",
      " conv2d_35 (Conv2D)             (None, 2, 3, 11)     5643        ['dropout_3[0][0]']              \n",
      "                                                                                                  \n",
      " conv2d_transpose_3 (Conv2DTran  (None, 6, 8, 11)    1936        ['conv2d_35[0][0]']              \n",
      " spose)                                                                                           \n",
      "                                                                                                  \n",
      " cropping2d_3 (Cropping2D)      (None, 4, 6, 11)     0           ['conv2d_transpose_3[0][0]']     \n",
      "                                                                                                  \n",
      " conv2d_36 (Conv2D)             (None, 4, 6, 11)     2827        ['batch_normalization_14[0][0]'] \n",
      "                                                                                                  \n",
      " add_2 (Add)                    (None, 4, 6, 11)     0           ['cropping2d_3[0][0]',           \n",
      "                                                                  'conv2d_36[0][0]']              \n",
      "                                                                                                  \n",
      " conv2d_transpose_4 (Conv2DTran  (None, 10, 14, 11)  1936        ['add_2[0][0]']                  \n",
      " spose)                                                                                           \n",
      "                                                                                                  \n",
      " cropping2d_4 (Cropping2D)      (None, 8, 12, 11)    0           ['conv2d_transpose_4[0][0]']     \n",
      "                                                                                                  \n",
      " conv2d_37 (Conv2D)             (None, 8, 12, 11)    1419        ['batch_normalization_13[0][0]'] \n",
      "                                                                                                  \n",
      " add_3 (Add)                    (None, 8, 12, 11)    0           ['cropping2d_4[0][0]',           \n",
      "                                                                  'conv2d_37[0][0]']              \n",
      "                                                                                                  \n",
      " conv2d_transpose_5 (Conv2DTran  (None, 64, 96, 11)  7744        ['add_3[0][0]']                  \n",
      " spose)                                                                                           \n",
      "                                                                                                  \n",
      " cropping2d_5 (Cropping2D)      (None, 64, 84, 11)   0           ['conv2d_transpose_5[0][0]']     \n",
      "                                                                                                  \n",
      " activation_1 (Activation)      (None, 64, 84, 11)   0           ['cropping2d_5[0][0]']           \n",
      "                                                                                                  \n",
      "==================================================================================================\n",
      "Total params: 9,061,985\n",
      "Trainable params: 9,060,513\n",
      "Non-trainable params: 1,472\n",
      "__________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "wAAXygZtbZmu"
   },
   "source": [
    "## Compile the Model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "sC-9m1R_5tjV"
   },
   "source": [
    "<a name='exercise-4'></a>\n",
    "\n",
    "### **Exercise 4**\n",
    "\n",
    "Compile the model using an appropriate loss, optimizer, and metric.\n",
    "\n",
    "_**Note:** There is a current issue with the grader accepting certain loss functions. We will be upgrading it but while in progress, please use this syntax:_\n",
    "\n",
    "```\n",
    "loss='<loss string name>'\n",
    "```\n",
    "\n",
    "*instead of:*\n",
    "\n",
    "```\n",
    "loss=tf.keras.losses.<StringCassName>\n",
    "```\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {
    "id": "ZpWpp8h4g_rE"
   },
   "outputs": [],
   "source": [
    "### START CODE HERE ###\n",
    "model.compile(loss='categorical_crossentropy', optimizer='Adam', metrics=['accuracy'])\n",
    "### END CODE HERE ###"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "510v0aVDXv1f"
   },
   "source": [
    "## Model Training"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "1udK8qps6PcG"
   },
   "source": [
    "<a name='exercise-5'></a>\n",
    "\n",
    "### **Exercise 5**\n",
    "\n",
    "You can now train the model. Set the number of epochs and observe the metrics returned at each iteration. You can also terminate the cell execution if you think your model is performing well already."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "8HoZwpGWhMB-",
    "outputId": "6c396492-ba38-40c6-d07b-f531b1f3232c"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/70\n",
      "125/125 [==============================] - 83s 655ms/step - loss: 0.2072 - accuracy: 0.9428 - val_loss: 0.1810 - val_accuracy: 0.9444\n",
      "Epoch 2/70\n",
      "125/125 [==============================] - 82s 655ms/step - loss: 0.1343 - accuracy: 0.9541 - val_loss: 0.1882 - val_accuracy: 0.9448\n",
      "Epoch 3/70\n",
      "125/125 [==============================] - 82s 657ms/step - loss: 0.0803 - accuracy: 0.9713 - val_loss: 0.1165 - val_accuracy: 0.9609\n",
      "Epoch 4/70\n",
      "125/125 [==============================] - 79s 634ms/step - loss: 0.0596 - accuracy: 0.9777 - val_loss: 0.0590 - val_accuracy: 0.9781\n",
      "Epoch 5/70\n",
      "125/125 [==============================] - 78s 628ms/step - loss: 0.0494 - accuracy: 0.9809 - val_loss: 0.0630 - val_accuracy: 0.9773\n",
      "Epoch 6/70\n",
      "125/125 [==============================] - 79s 635ms/step - loss: 0.0428 - accuracy: 0.9832 - val_loss: 0.0441 - val_accuracy: 0.9828\n",
      "Epoch 7/70\n",
      "125/125 [==============================] - 79s 632ms/step - loss: 0.0380 - accuracy: 0.9849 - val_loss: 0.0395 - val_accuracy: 0.9848\n",
      "Epoch 8/70\n",
      "125/125 [==============================] - 81s 649ms/step - loss: 0.0350 - accuracy: 0.9861 - val_loss: 0.0383 - val_accuracy: 0.9851\n",
      "Epoch 9/70\n",
      "125/125 [==============================] - 86s 687ms/step - loss: 0.0325 - accuracy: 0.9870 - val_loss: 0.0366 - val_accuracy: 0.9857\n",
      "Epoch 10/70\n",
      "125/125 [==============================] - 86s 688ms/step - loss: 0.0309 - accuracy: 0.9876 - val_loss: 0.0385 - val_accuracy: 0.9851\n",
      "Epoch 11/70\n",
      "125/125 [==============================] - 86s 689ms/step - loss: 0.0292 - accuracy: 0.9883 - val_loss: 0.0381 - val_accuracy: 0.9861\n",
      "Epoch 12/70\n",
      "125/125 [==============================] - 86s 691ms/step - loss: 0.0279 - accuracy: 0.9889 - val_loss: 0.0328 - val_accuracy: 0.9875\n",
      "Epoch 13/70\n",
      "125/125 [==============================] - 87s 696ms/step - loss: 0.0262 - accuracy: 0.9895 - val_loss: 0.0354 - val_accuracy: 0.9865\n",
      "Epoch 14/70\n",
      "125/125 [==============================] - 86s 691ms/step - loss: 0.0265 - accuracy: 0.9894 - val_loss: 0.0324 - val_accuracy: 0.9875\n",
      "Epoch 15/70\n",
      "125/125 [==============================] - 88s 701ms/step - loss: 0.0248 - accuracy: 0.9900 - val_loss: 0.0359 - val_accuracy: 0.9871\n",
      "Epoch 16/70\n",
      "125/125 [==============================] - 90s 717ms/step - loss: 0.0249 - accuracy: 0.9901 - val_loss: 0.0434 - val_accuracy: 0.9854\n",
      "Epoch 17/70\n",
      "125/125 [==============================] - 91s 729ms/step - loss: 0.0243 - accuracy: 0.9903 - val_loss: 0.0287 - val_accuracy: 0.9892\n",
      "Epoch 18/70\n",
      "125/125 [==============================] - 87s 697ms/step - loss: 0.0223 - accuracy: 0.9910 - val_loss: 0.0295 - val_accuracy: 0.9889\n",
      "Epoch 19/70\n",
      "125/125 [==============================] - 85s 684ms/step - loss: 0.0215 - accuracy: 0.9913 - val_loss: 0.0266 - val_accuracy: 0.9898\n",
      "Epoch 20/70\n",
      "125/125 [==============================] - 86s 689ms/step - loss: 0.0212 - accuracy: 0.9915 - val_loss: 0.0274 - val_accuracy: 0.9897\n",
      "Epoch 21/70\n",
      "125/125 [==============================] - 86s 689ms/step - loss: 0.0209 - accuracy: 0.9916 - val_loss: 0.0264 - val_accuracy: 0.9898\n",
      "Epoch 22/70\n",
      "125/125 [==============================] - 86s 686ms/step - loss: 0.0222 - accuracy: 0.9912 - val_loss: 0.0406 - val_accuracy: 0.9865\n",
      "Epoch 23/70\n",
      "125/125 [==============================] - 86s 687ms/step - loss: 0.0265 - accuracy: 0.9896 - val_loss: 0.0365 - val_accuracy: 0.9879\n",
      "Epoch 24/70\n",
      "125/125 [==============================] - 86s 686ms/step - loss: 0.0214 - accuracy: 0.9915 - val_loss: 0.0266 - val_accuracy: 0.9900\n",
      "Epoch 25/70\n",
      "125/125 [==============================] - 86s 687ms/step - loss: 0.0197 - accuracy: 0.9921 - val_loss: 0.0256 - val_accuracy: 0.9902\n",
      "Epoch 26/70\n",
      "125/125 [==============================] - 86s 686ms/step - loss: 0.0191 - accuracy: 0.9924 - val_loss: 0.0242 - val_accuracy: 0.9907\n",
      "Epoch 27/70\n",
      "125/125 [==============================] - 85s 681ms/step - loss: 0.0188 - accuracy: 0.9925 - val_loss: 0.0303 - val_accuracy: 0.9883\n",
      "Epoch 28/70\n",
      "125/125 [==============================] - 86s 688ms/step - loss: 0.0186 - accuracy: 0.9926 - val_loss: 0.0238 - val_accuracy: 0.9909\n",
      "Epoch 29/70\n",
      "125/125 [==============================] - 86s 684ms/step - loss: 0.0179 - accuracy: 0.9929 - val_loss: 0.0239 - val_accuracy: 0.9910\n",
      "Epoch 30/70\n",
      "125/125 [==============================] - 86s 687ms/step - loss: 0.0179 - accuracy: 0.9929 - val_loss: 0.0254 - val_accuracy: 0.9903\n",
      "Epoch 31/70\n",
      "125/125 [==============================] - 86s 687ms/step - loss: 0.0234 - accuracy: 0.9909 - val_loss: 0.0515 - val_accuracy: 0.9839\n",
      "Epoch 32/70\n",
      "125/125 [==============================] - 85s 682ms/step - loss: 0.0219 - accuracy: 0.9914 - val_loss: 0.0280 - val_accuracy: 0.9900\n",
      "Epoch 33/70\n",
      "125/125 [==============================] - 86s 686ms/step - loss: 0.0182 - accuracy: 0.9927 - val_loss: 0.0267 - val_accuracy: 0.9901\n",
      "Epoch 34/70\n",
      "125/125 [==============================] - 85s 681ms/step - loss: 0.0174 - accuracy: 0.9931 - val_loss: 0.0227 - val_accuracy: 0.9913\n",
      "Epoch 35/70\n",
      "125/125 [==============================] - 85s 684ms/step - loss: 0.0169 - accuracy: 0.9933 - val_loss: 0.0258 - val_accuracy: 0.9903\n",
      "Epoch 36/70\n",
      "125/125 [==============================] - 85s 678ms/step - loss: 0.0167 - accuracy: 0.9933 - val_loss: 0.0240 - val_accuracy: 0.9911\n",
      "Epoch 37/70\n",
      "125/125 [==============================] - 85s 683ms/step - loss: 0.0163 - accuracy: 0.9935 - val_loss: 0.0229 - val_accuracy: 0.9914\n",
      "Epoch 38/70\n",
      "125/125 [==============================] - 86s 687ms/step - loss: 0.0163 - accuracy: 0.9935 - val_loss: 0.0233 - val_accuracy: 0.9913\n",
      "Epoch 39/70\n",
      "125/125 [==============================] - 85s 684ms/step - loss: 0.0166 - accuracy: 0.9934 - val_loss: 0.0225 - val_accuracy: 0.9915\n",
      "Epoch 40/70\n",
      "125/125 [==============================] - 85s 683ms/step - loss: 0.0158 - accuracy: 0.9937 - val_loss: 0.0221 - val_accuracy: 0.9916\n",
      "Epoch 41/70\n",
      "125/125 [==============================] - 85s 679ms/step - loss: 0.0156 - accuracy: 0.9938 - val_loss: 0.0247 - val_accuracy: 0.9911\n",
      "Epoch 42/70\n",
      "125/125 [==============================] - 85s 681ms/step - loss: 0.0155 - accuracy: 0.9938 - val_loss: 0.0225 - val_accuracy: 0.9916\n",
      "Epoch 43/70\n",
      "125/125 [==============================] - 85s 681ms/step - loss: 0.0153 - accuracy: 0.9939 - val_loss: 0.0229 - val_accuracy: 0.9916\n",
      "Epoch 44/70\n",
      "125/125 [==============================] - 86s 689ms/step - loss: 0.0261 - accuracy: 0.9901 - val_loss: 0.0367 - val_accuracy: 0.9887\n",
      "Epoch 45/70\n",
      "125/125 [==============================] - 86s 691ms/step - loss: 0.0183 - accuracy: 0.9927 - val_loss: 0.0231 - val_accuracy: 0.9913\n",
      "Epoch 46/70\n",
      "125/125 [==============================] - 86s 686ms/step - loss: 0.0161 - accuracy: 0.9936 - val_loss: 0.0231 - val_accuracy: 0.9913\n",
      "Epoch 47/70\n",
      "125/125 [==============================] - 85s 684ms/step - loss: 0.0153 - accuracy: 0.9939 - val_loss: 0.0233 - val_accuracy: 0.9912\n",
      "Epoch 48/70\n",
      "125/125 [==============================] - 86s 684ms/step - loss: 0.0151 - accuracy: 0.9940 - val_loss: 0.0213 - val_accuracy: 0.9919\n",
      "Epoch 49/70\n",
      "125/125 [==============================] - 86s 685ms/step - loss: 0.0146 - accuracy: 0.9942 - val_loss: 0.0236 - val_accuracy: 0.9913\n",
      "Epoch 50/70\n",
      "125/125 [==============================] - 85s 682ms/step - loss: 0.0145 - accuracy: 0.9943 - val_loss: 0.0213 - val_accuracy: 0.9918\n",
      "Epoch 51/70\n",
      "125/125 [==============================] - 85s 683ms/step - loss: 0.0145 - accuracy: 0.9943 - val_loss: 0.0223 - val_accuracy: 0.9917\n",
      "Epoch 52/70\n",
      "125/125 [==============================] - 86s 687ms/step - loss: 0.0142 - accuracy: 0.9943 - val_loss: 0.0214 - val_accuracy: 0.9921\n",
      "Epoch 53/70\n",
      "125/125 [==============================] - 86s 688ms/step - loss: 0.0141 - accuracy: 0.9944 - val_loss: 0.0234 - val_accuracy: 0.9916\n",
      "Epoch 54/70\n",
      "125/125 [==============================] - 89s 716ms/step - loss: 0.0143 - accuracy: 0.9943 - val_loss: 0.0225 - val_accuracy: 0.9919\n",
      "Epoch 55/70\n",
      "125/125 [==============================] - 87s 699ms/step - loss: 0.0139 - accuracy: 0.9945 - val_loss: 0.0217 - val_accuracy: 0.9919\n",
      "Epoch 56/70\n",
      "125/125 [==============================] - 89s 714ms/step - loss: 0.0137 - accuracy: 0.9945 - val_loss: 0.0222 - val_accuracy: 0.9918\n",
      "Epoch 57/70\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "125/125 [==============================] - 86s 692ms/step - loss: 0.0137 - accuracy: 0.9946 - val_loss: 0.0217 - val_accuracy: 0.9920\n",
      "Epoch 58/70\n",
      "125/125 [==============================] - 89s 715ms/step - loss: 0.0135 - accuracy: 0.9946 - val_loss: 0.0216 - val_accuracy: 0.9920\n",
      "Epoch 59/70\n",
      "125/125 [==============================] - 88s 703ms/step - loss: 0.0135 - accuracy: 0.9946 - val_loss: 0.0218 - val_accuracy: 0.9920\n",
      "Epoch 60/70\n",
      "125/125 [==============================] - 90s 717ms/step - loss: 0.0138 - accuracy: 0.9945 - val_loss: 0.0217 - val_accuracy: 0.9922\n",
      "Epoch 61/70\n",
      "125/125 [==============================] - 90s 717ms/step - loss: 0.0133 - accuracy: 0.9947 - val_loss: 0.0254 - val_accuracy: 0.9912\n",
      "Epoch 62/70\n",
      "125/125 [==============================] - 89s 709ms/step - loss: 0.0131 - accuracy: 0.9948 - val_loss: 0.0228 - val_accuracy: 0.9920\n",
      "Epoch 63/70\n",
      "125/125 [==============================] - 89s 715ms/step - loss: 0.0131 - accuracy: 0.9948 - val_loss: 0.0247 - val_accuracy: 0.9910\n",
      "Epoch 64/70\n",
      "125/125 [==============================] - 87s 698ms/step - loss: 0.0130 - accuracy: 0.9948 - val_loss: 0.0219 - val_accuracy: 0.9921\n",
      "Epoch 65/70\n",
      "125/125 [==============================] - 88s 701ms/step - loss: 0.0129 - accuracy: 0.9949 - val_loss: 0.0217 - val_accuracy: 0.9921\n",
      "Epoch 66/70\n",
      "125/125 [==============================] - 90s 719ms/step - loss: 0.0138 - accuracy: 0.9945 - val_loss: 0.0236 - val_accuracy: 0.9920\n",
      "Epoch 67/70\n",
      "125/125 [==============================] - 89s 712ms/step - loss: 0.0129 - accuracy: 0.9949 - val_loss: 0.0214 - val_accuracy: 0.9923\n",
      "Epoch 68/70\n",
      "125/125 [==============================] - 87s 697ms/step - loss: 0.0127 - accuracy: 0.9949 - val_loss: 0.0221 - val_accuracy: 0.9923\n",
      "Epoch 69/70\n",
      "125/125 [==============================] - 86s 688ms/step - loss: 0.0128 - accuracy: 0.9949 - val_loss: 0.0219 - val_accuracy: 0.9922\n",
      "Epoch 70/70\n",
      "125/125 [==============================] - 86s 692ms/step - loss: 0.0126 - accuracy: 0.9950 - val_loss: 0.0214 - val_accuracy: 0.9924\n"
     ]
    }
   ],
   "source": [
    "# OTHER THAN SETTING THE EPOCHS NUMBER, DO NOT CHANGE ANY OTHER CODE\n",
    "\n",
    "### START CODE HERE ###\n",
    "EPOCHS = 70\n",
    "### END CODE HERE ###\n",
    "\n",
    "steps_per_epoch = 4000//BATCH_SIZE\n",
    "validation_steps = 800//BATCH_SIZE\n",
    "test_steps = 200//BATCH_SIZE\n",
    "\n",
    "\n",
    "history = model.fit(training_dataset,\n",
    "                    steps_per_epoch=steps_per_epoch, validation_data=validation_dataset, validation_steps=validation_steps, epochs=EPOCHS)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "BLNzLB3peeNG"
   },
   "source": [
    "**Expected Output:**\n",
    "\n",
    "The losses should generally be decreasing and the accuracies should generally be increasing. For example, observing the first 4 epochs should output something similar:\n",
    "\n",
    "```txt\n",
    "Epoch 1/70\n",
    "125/125 [==============================] - 6s 50ms/step - loss: 0.5542 - accuracy: 0.8635 - val_loss: 0.5335 - val_accuracy: 0.9427\n",
    "Epoch 2/70\n",
    "125/125 [==============================] - 6s 47ms/step - loss: 0.2315 - accuracy: 0.9425 - val_loss: 0.3362 - val_accuracy: 0.9427\n",
    "Epoch 3/70\n",
    "125/125 [==============================] - 6s 47ms/step - loss: 0.2118 - accuracy: 0.9426 - val_loss: 0.2592 - val_accuracy: 0.9427\n",
    "Epoch 4/70\n",
    "125/125 [==============================] - 6s 47ms/step - loss: 0.1782 - accuracy: 0.9431 - val_loss: 0.1770 - val_accuracy: 0.9432\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "eih-Q7GoXzJe"
   },
   "source": [
    "## Model Evaluation"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "2bTkaFM2X1gr"
   },
   "source": [
    "### Make Predictions\n",
    "\n",
    "Let's get the predictions using our test dataset as input and print the shape."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {
    "id": "zENjQuK0luH5"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(192, 64, 84, 11)\n"
     ]
    }
   ],
   "source": [
    "results = model.predict(test_dataset, steps=test_steps)\n",
    "\n",
    "print(results.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "IReKPT_DZHjA"
   },
   "source": [
    "As you can see, the resulting shape is `(192, 64, 84, 11)`. This means that for each of the 192 images that we have in our test set, there are 11 predictions generated (i.e. one for each class: 0 to 1 plus background)."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "OBPBqnHyaSaG"
   },
   "source": [
    "Thus, if you want to see the *probability* of the upper leftmost pixel of the 1st image belonging to class 0, then you can print something like `results[0,0,0,0]`. If you want the probability of the same pixel at class 10, then do `results[0,0,0,10]`.\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {
    "id": "jwFiR9WAf0Av"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.04677275\n",
      "0.9979538\n"
     ]
    }
   ],
   "source": [
    "print(results[0,0,0,0])\n",
    "print(results[0,0,0,10])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "wKDsqzKEf28V"
   },
   "source": [
    "What we're interested in is to get the *index* of the highest probability of each of these 11 slices and combine them in a single image. We can do that by getting the [argmax](https://numpy.org/doc/stable/reference/generated/numpy.argmax.html) at this axis."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {
    "id": "0_Uj_uuV9TQt"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(192, 64, 84)\n"
     ]
    }
   ],
   "source": [
    "results = np.argmax(results, axis=3)\n",
    "\n",
    "print(results.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "ClaGbOLhdKD2"
   },
   "source": [
    "The new array generated per image now only specifies the indices of the class with the highest probability. Let's see the output class of the upper most left pixel. As you might have observed earlier when you inspected the dataset, the upper left corner is usually just part of the background (class 10). The actual digits are written somewhere in the middle parts of the image."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {
    "id": "aBeBwvHQd2pZ"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "10\n",
      "[[10 10 10 ... 10 10 10]\n",
      " [10 10 10 ... 10 10 10]\n",
      " [10 10 10 ... 10 10 10]\n",
      " ...\n",
      " [10 10 10 ... 10 10 10]\n",
      " [10 10 10 ... 10 10 10]\n",
      " [10 10 10 ... 10 10 10]]\n"
     ]
    }
   ],
   "source": [
    "print(results[0,0,0])\n",
    "\n",
    "# prediction map for image 0\n",
    "print(results[0,:,:])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "V3FxyNgrg7IE"
   },
   "source": [
    "We will use this `results` array when we evaluate our predictions."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "hpKDUuAWX5Pj"
   },
   "source": [
    "### Metrics\n",
    "\n",
    "We showed in the lectures two ways to evaluate your predictions. The *intersection over union (IOU)* and the *dice score*. Recall that:\n",
    "\n",
    "$$IOU = \\frac{area\\_of\\_overlap}{area\\_of\\_union}$$\n",
    "<br>\n",
    "$$Dice Score = 2 * \\frac{area\\_of\\_overlap}{combined\\_area}$$\n",
    "\n",
    "The code below does that for you as you've also seen in the ungraded lab. A small smoothing factor is introduced in the denominators to prevent possible division by zero."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {
    "id": "gKTpLmly_RXb"
   },
   "outputs": [],
   "source": [
    "def class_wise_metrics(y_true, y_pred):\n",
    "  '''\n",
    "  Computes the class-wise IOU and Dice Score.\n",
    "\n",
    "  Args:\n",
    "    y_true (tensor) - ground truth label maps\n",
    "    y_pred (tensor) - predicted label maps\n",
    "  '''\n",
    "  class_wise_iou = []\n",
    "  class_wise_dice_score = []\n",
    "\n",
    "  smoothing_factor = 0.00001\n",
    "\n",
    "  for i in range(n_classes):\n",
    "    intersection = np.sum((y_pred == i) * (y_true == i))\n",
    "    y_true_area = np.sum((y_true == i))\n",
    "    y_pred_area = np.sum((y_pred == i))\n",
    "    combined_area = y_true_area + y_pred_area\n",
    "    \n",
    "    iou = (intersection) / (combined_area - intersection + smoothing_factor)\n",
    "    class_wise_iou.append(iou)\n",
    "    \n",
    "    dice_score =  2 * ((intersection) / (combined_area + smoothing_factor))\n",
    "    class_wise_dice_score.append(dice_score)\n",
    "\n",
    "  return class_wise_iou, class_wise_dice_score\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "FfWPwM4ZhHjE"
   },
   "source": [
    "### Visualize Predictions"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {
    "id": "Hkbsk_P1fpRM",
    "lines_to_next_cell": 2
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1080x288 with 3 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# place a number here between 0 to 191 to pick an image from the test set\n",
    "integer_slider = 105\n",
    "\n",
    "ds = test_dataset.unbatch()\n",
    "ds = ds.batch(200)\n",
    "images = []\n",
    "\n",
    "y_true_segments = []\n",
    "for image, annotation in ds.take(2):\n",
    "  y_true_segments = annotation\n",
    "  images = image\n",
    "  \n",
    "  \n",
    "iou, dice_score = class_wise_metrics(np.argmax(y_true_segments[integer_slider], axis=2), results[integer_slider])  \n",
    "show_annotation_and_prediction(image[integer_slider], annotation[integer_slider], results[integer_slider], iou, dice_score)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "EiG9K4t6X9iZ"
   },
   "source": [
    "### Compute IOU Score and Dice Score of your model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {
    "id": "2706boF0CNNS",
    "lines_to_next_cell": 2
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Digit 0: IOU: 0.9024848879751715 Dice Score: 0.9487432921852984\n",
      "Digit 1: IOU: 0.9205726138984701 Dice Score: 0.9586439036323108\n",
      "Digit 2: IOU: 0.871414680516025 Dice Score: 0.9312897772884208\n",
      "Digit 3: IOU: 0.8669015981744354 Dice Score: 0.9287062574933163\n",
      "Digit 4: IOU: 0.8256835114735982 Dice Score: 0.9045198757446725\n",
      "Digit 5: IOU: 0.8600545935212838 Dice Score: 0.9247627424667227\n",
      "Digit 6: IOU: 0.86232495214725 Dice Score: 0.9260735632124719\n",
      "Digit 7: IOU: 0.8692963734130141 Dice Score: 0.9300787031708924\n",
      "Digit 8: IOU: 0.8549293361911641 Dice Score: 0.9217918111604628\n",
      "Digit 9: IOU: 0.8464532859768552 Dice Score: 0.9168423511229465\n",
      "\n",
      "Grade is 86.80115833287267\n",
      "You passed!\n"
     ]
    }
   ],
   "source": [
    "cls_wise_iou, cls_wise_dice_score = class_wise_metrics(np.argmax(y_true_segments, axis=3), results)\n",
    "\n",
    "average_iou = 0.0\n",
    "for idx, (iou, dice_score) in enumerate(zip(cls_wise_iou[:-1], cls_wise_dice_score[:-1])):\n",
    "  print(\"Digit {}: IOU: {} Dice Score: {}\".format(idx, iou, dice_score)) \n",
    "  average_iou += iou\n",
    "\n",
    "grade = average_iou * 10\n",
    "\n",
    "print(\"\\nGrade is \" + str(grade))\n",
    "\n",
    "PASSING_GRADE = 60\n",
    "if (grade>PASSING_GRADE):\n",
    "  print(\"You passed!\")\n",
    "else:\n",
    "  print(\"You failed. Please check your model and re-train\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "vvw0HLY2kV3w"
   },
   "source": [
    "## Save the Model\n",
    "\n",
    "Once you're satisfied with the results, you will need to save your model so you can upload it to the grader in the Coursera classroom. After running the cell below, please look for `student_model.h5` in the File Explorer on the left and download it. Then go back to the Coursera classroom and upload it to the Lab item that points to the autograder of Week 3."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {
    "id": "ULCfGHEKkaO0"
   },
   "outputs": [],
   "source": [
    "model.save(\"model.h5\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "id": "QbVfJ2aLONqw"
   },
   "outputs": [
    {
     "ename": "ModuleNotFoundError",
     "evalue": "No module named 'google.colab'",
     "output_type": "error",
     "traceback": [
      "\u001b[1;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[1;31mModuleNotFoundError\u001b[0m                       Traceback (most recent call last)",
      "\u001b[1;32m~\\AppData\\Local\\Temp/ipykernel_30268/469455449.py\u001b[0m in \u001b[0;36m<module>\u001b[1;34m\u001b[0m\n\u001b[0;32m      1\u001b[0m \u001b[1;31m# You can also use this cell as a shortcut for downloading your model\u001b[0m\u001b[1;33m\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[1;32m----> 2\u001b[1;33m \u001b[1;32mfrom\u001b[0m \u001b[0mgoogle\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mcolab\u001b[0m \u001b[1;32mimport\u001b[0m \u001b[0mfiles\u001b[0m\u001b[1;33m\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0m\u001b[0;32m      3\u001b[0m \u001b[0mfiles\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mdownload\u001b[0m\u001b[1;33m(\u001b[0m\u001b[1;34m\"model.h5\"\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n",
      "\u001b[1;31mModuleNotFoundError\u001b[0m: No module named 'google.colab'"
     ]
    }
   ],
   "source": [
    "# # You can also use this cell as a shortcut for downloading your model\n",
    "# from google.colab import files\n",
    "# files.download(\"model.h5\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "dmyvoAynkgVw"
   },
   "source": [
    "**Congratulations on completing this assignment on image segmentation!**"
   ]
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "collapsed_sections": [],
   "include_colab_link": true,
   "name": "Copy of C3W3_Assignment.ipynb",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.9.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}