3-ManFeatExt.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "7cd3a024-5f12-47ce-8511-c2bfae262628",
   "metadata": {},
   "source": [
    "# Tutorial 3: Manual Feature Extraction/Engineering"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ed665b16-22c8-4a4d-b4db-0837264fb837",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "58a99b7c-2540-4862-bcca-315eeb1c0803",
   "metadata": {},
   "source": [
    "### Introduction"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "21e5412c-eb3a-416a-81ff-ce5f77a0df88",
   "metadata": {},
   "source": [
    "In this notebook, we will extract/engineer features using traditional statistical methods, including:\n",
    "\n",
    "- discrete Laplacian\n",
    "- gaussian_gradient_magnitude\n",
    "- generic_gradient_magnitude"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8030153b-d68e-4f25-8774-4b9e4bb5bfef",
   "metadata": {},
   "source": [
    "First we import some libraries for data manipulation (pandas) mathematical operations (numpy), image processing (scipy) and plotting (pyplot) :"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "fe9d5526-086d-4d0e-b546-e81af265749a",
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt # for plotting data/graphs\n",
    "import numpy as np # For handling N-DIMENSIONAL ARRAYS\n",
    "import pandas as pd # for importing data into data frame format\n",
    "\n",
    "from scipy import ndimage, misc # ndimage: contains various functions for multidimensional image processing\n",
    "                               # misc: Various utilities that don’t have another home.\n",
    "import matplotlib.pyplot as plt # to draw images\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "20c7b065-2cd2-4d88-8329-a7b0e9591e1a",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7c7a52a2-6542-4535-9aad-0782ed0eb4a4",
   "metadata": {},
   "source": [
    "### Reading in data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "700ba1a6-c0cd-4bb9-ae83-74143f138397",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Import custom data processing function from galaxy_mnist.py (in this folder)\n",
    "from galaxy_mnist import GalaxyMNISTHighrez\n",
    "\n",
    "dataset_train = GalaxyMNISTHighrez(\n",
    "    root='data_import/data',\n",
    "    download=True,\n",
    "    train=True  # by default, or False for canonical test set\n",
    ")\n",
    "# for the testing data\n",
    "dataset_test = GalaxyMNISTHighrez(\n",
    "    root='data_import/data',\n",
    "    download=True,\n",
    "    train=False  # by default, or False for canonical test set\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c4639a4d-6649-48e5-8fec-47d8026828fa",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Define inputs and labels (targets) for training and testing\n",
    "images_train = dataset_train.data\n",
    "images_test = dataset_test.data\n",
    "labels_train = dataset_train.targets\n",
    "labels_test = dataset_test.targets\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a153a956-dcbf-4860-8a92-040e189d5b97",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "55f4d509-8977-4f8d-a158-8a0bfc75db0b",
   "metadata": {},
   "source": [
    "### Pre-processing "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f6decf49-9a17-43b2-b659-6a429440326b",
   "metadata": {},
   "outputs": [],
   "source": [
    "#  A predefined function to pre-process the data as we did in Tutorial 1\n",
    "# (source.pre.py is a file in this directory)\n",
    "from source.pre import pre_processing "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b2b9da92-41d1-41b7-a888-593598261aed",
   "metadata": {},
   "outputs": [],
   "source": [
    "# pre-processing(data, size) function takes two arguments\n",
    "# 1. data: the image data to be processed\n",
    "# 2. The width of resized images (images are square, so width=height).\n",
    "images_trainPre = pre_processing(images_train, 112)\n",
    "images_testPre = pre_processing(images_test, 112)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "887d6bae-5885-4ce3-83d7-74dbdaeb8b34",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Print number and size of pre-processed images\n",
    "images_trainPre.shape"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ba698cf5-b955-43a1-8dcd-e97f4088e2de",
   "metadata": {},
   "source": [
    "Displaying images after pre-processing"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8bb62536-7817-4f05-9110-263a0f57da10",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "rows = 1\n",
    "columns = 5\n",
    "for j in range(len(GalaxyMNISTHighrez.classes)):\n",
    "    fig = plt.figure(figsize=(8, 8))# Figure is 8 inches by 8 inches\n",
    "    for i in range (columns):    # Create images in each column\n",
    "        train_image = images_trainPre[(labels_train == j)][i]\n",
    "        fig.add_subplot(rows, columns, i+1)\n",
    "        plt.imshow(train_image*255,cmap='gray', vmin=0, vmax=255) \n",
    "                            # we have to multiply the image by 255 to restore the original values\n",
    "    print(\"label: \"+str(GalaxyMNISTHighrez.classes[j]))\n",
    "    plt.tight_layout()\n",
    "    plt.show() "
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7444a5c0-dd53-49df-bcd0-fa56e2f874d3",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "12cd1b55-87ec-4794-925a-c9bcf919c35f",
   "metadata": {},
   "source": [
    "### Feature Engineering\n",
    "\n",
    "In this section, we will try to use manual tools to engineer the input images. The aim is to highlight features in the images that can be detected easily by classical machine learning tools. Notice that this step is being done automatically in deep learning tools.\n",
    "More information on these tools can be found [here](https://docs.scipy.org/doc/scipy/reference/ndimage.html)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ccc59a2f-14d3-47db-90d0-65454e41db44",
   "metadata": {},
   "source": [
    "#### Discrete Laplacian [(documentation)](https://docs.scipy.org/doc/scipy/reference/generated/scipy.ndimage.laplace.html)\n",
    "The discrete Laplacian transforms the image to highlight the image regions which are darkest and brightest.\n",
    "\n",
    "In the following code we display images of galaxies of different types before and after application of the discrete Laplacian."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "91913044-6f3c-475f-8fc8-c6ce791b5488",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "rows = 1\n",
    "columns = 5\n",
    "for j in range(len(GalaxyMNISTHighrez.classes)):\n",
    "    fig = plt.figure(figsize=(8, 8))# Figure is 8 inches by 8 inches\n",
    "    for i in range (columns):    # Create images in each column\n",
    "        train_image = images_trainPre[(labels_train == j)][i]\n",
    "        fig.add_subplot(rows, columns, i+1)\n",
    "        plt.imshow(train_image*255,cmap='gray', vmin=0, vmax=255) \n",
    "                            # we have to multiply the image by 255 to restore the original values\n",
    "    print(\"Original: \"+str(GalaxyMNISTHighrez.classes[j]))\n",
    "    plt.tight_layout()\n",
    "    plt.show() \n",
    "    \n",
    "    fig = plt.figure(figsize=(8, 8))# Figure is 8 inches by 8 inches\n",
    "    for i in range (columns):    # Create images in each column\n",
    "        pre_image = ndimage.laplace(images_trainPre[(labels_train == j)][i] ,mode = 'nearest'  )*255*200 # can we highlight the difference better than this\n",
    "        fig.add_subplot(rows, columns, i+1)\n",
    "        plt.imshow(pre_image,cmap='gray', vmin=0, vmax=255) \n",
    "                            # we have to multiply the image by 255 to restore the original values\n",
    "    print(\"pre_image: \"+str(GalaxyMNISTHighrez.classes[j]))\n",
    "    plt.tight_layout()\n",
    "    plt.show() \n",
    "    print(\"        ______________________________________________________________________________     \")\n",
    "    "
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1f786da5",
   "metadata": {},
   "source": [
    "The above images show that the discrete Laplacian filtering does not do a good job of distinguishing between smooth_cigar and edge_on_disk.  However, it does a great job of bringing out the spiral arms in unbarred_spiral."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "428e4168-cabf-4b4f-997a-3096c06d3e1f",
   "metadata": {},
   "source": [
    "### Other filtering options \n",
    "Two other filters that can be used are:\n",
    "* gaussian_gradient_magnitude[(documentation](https://docs.scipy.org/doc/scipy/reference/generated/scipy.ndimage.gaussian_gradient_magnitude.html#scipy.ndimage.gaussian_gradient_magnitude): This filter brings out edges (i.e. rapid transitions from light to dark)\n",
    "* generic_gradient_magnitude[(documentation)](https://docs.scipy.org/doc/scipy/reference/generated/scipy.ndimage.generic_gradient_magnitude.html): Similar to Gaussian gradient magnitude\n",
    "\n",
    "### Exercise:  \n",
    "Implement these two filters and display images (you may reuse and modify the code that was used for discrete Laplacian)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "db0fad9d-9da3-454b-9187-201f8c1b55f9",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ca031aad-2a6d-42a2-bc71-ce97eabd391c",
   "metadata": {},
   "source": [
    "#### **_Save the data that you think will give the best performance_**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c80fb440-3e58-4562-adc3-41e001c18acc",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "gal_ker",
   "language": "python",
   "name": "gal_ker"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}