SVM-RBF sensitivity to translation.ipynb

svm-rbf-sensitivity-to-translation.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "SVM-RBF sensitivity to translation.ipynb",
      "provenance": [],
      "collapsed_sections": [],
      "toc_visible": true,
      "authorship_tag": "ABX9TyOF2Z+MPhxOHaV1DPw1vbgU",
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/gist/shadiakiki1986/689980135fe9dde1d892127bde40a5a1/svm-rbf-sensitivity-to-translation.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "8GaAZFCMWaw4"
      },
      "source": [
        "Testing if RBF is invariant to translation:\n",
        "\n",
        "- sklearn issue: https://github.com/scikit-learn/scikit-learn/issues/18432\n",
        "- gist: https://gist.github.com/xtomasch/84d1d8574ef51eb8d42e77560d647e06\n",
        "\n",
        "\n",
        "Uses my digits dataset with jitter: https://github.com/shadiakiki1986/mnist-digits-jitter\n",
        "\n",
        "Published as gist at https://gist.github.com/shadiakiki1986/689980135fe9dde1d892127bde40a5a1"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "HGkCzTmwXS4o"
      },
      "source": [
        "# get data"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "wxWb-jzbXTdv"
      },
      "source": [
        "# first the original data\n",
        "from sklearn.datasets import load_digits\n",
        "digits = load_digits()"
      ],
      "execution_count": 1,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "RrAkpWbdfPMu",
        "outputId": "0ac4e1d5-748f-407d-95ce-622f09c8e8db"
      },
      "source": [
        "# then the padded and jittered data\n",
        "!git clone https://github.com/shadiakiki1986/mnist-digits-jitter\n",
        "\n",
        "# Update: no need to gunzip since np.loadtxt can automatically do it\n",
        "#!gunzip mnist-digits-jitter/digits_padded.csv.gz\n",
        "#!gunzip mnist-digits-jitter/digits_jitter.csv.gz"
      ],
      "execution_count": 2,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Cloning into 'mnist-digits-jitter'...\n",
            "remote: Enumerating objects: 47, done.\u001b[K\n",
            "remote: Counting objects: 100% (47/47), done.\u001b[K\n",
            "remote: Compressing objects: 100% (44/44), done.\u001b[K\n",
            "remote: Total 47 (delta 20), reused 8 (delta 1), pack-reused 0\u001b[K\n",
            "Unpacking objects: 100% (47/47), done.\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "66oFvhVEfZMk",
        "outputId": "db7a154b-a160-47c1-ee81-502f210d5e0b"
      },
      "source": [
        "import numpy as np\n",
        "# np.loadtxt can decompress the files on read\n",
        "digpad = {\"data\": np.loadtxt(\"mnist-digits-jitter/digits_padded.csv.gz\", delimiter=\",\", dtype=int)}\n",
        "digjit = {\"data\": np.loadtxt(\"mnist-digits-jitter/digits_jitter.csv.gz\", delimiter=\",\", dtype=int)}\n",
        "digpad[\"data\"].shape, digjit[\"data\"].shape"
      ],
      "execution_count": 3,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "((1797, 225), (1797, 225))"
            ]
          },
          "metadata": {},
          "execution_count": 3
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "nBjvLW_9f3YM"
      },
      "source": [
        "# convert data to image (not flat)\n",
        "#def im2data(digxxx_img):\n",
        "#  return np.vstack([img.reshape((-1,1)).squeeze() for img in digxxx_img])\n",
        "\n",
        "def data2im(digxxx_data):\n",
        "  s = int(digxxx_data.shape[1]**.5)\n",
        "  l = digxxx_data.reshape((-1,s,s))\n",
        "  return l\n",
        "\n",
        "digjit[\"images\"] = data2im(digjit[\"data\"])\n",
        "digpad[\"images\"] = data2im(digpad[\"data\"])\n",
        "\n",
        "assert digjit[\"data\"].shape == (1797, 225)\n",
        "assert digpad[\"data\"].shape == (1797, 225)\n",
        "assert digjit[\"images\"].shape == (1797, 15, 15)\n",
        "assert digpad[\"images\"].shape == (1797, 15, 15)"
      ],
      "execution_count": 4,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "N8Ukyc9Twqve"
      },
      "source": [
        "# run svm and knn"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "VIT0GtY0ww5a",
        "outputId": "2630bfe0-c5d1-4bfb-f92d-4d5094d7dfbb"
      },
      "source": [
        "from sklearn.neighbors import KNeighborsClassifier\n",
        "from sklearn import metrics, svm, model_selection\n",
        "import statistics\n",
        "\n",
        "clf_l = [\n",
        "  # n neighbors = 3 is better than any of 1,2,4,5\n",
        "  (\"KNN\", KNeighborsClassifier(n_neighbors=3)),\n",
        "  # default is kernel=rbf\n",
        "  # Gamme 1e-3 is better than any of (1e-5,1e-4,1e-2,1e-1)\n",
        "  # Try linear, poly, RBF as xtomasch gist\n",
        "  # https://gist.github.com/xtomasch/84d1d8574ef51eb8d42e77560d647e06\n",
        "  (\"SVM linear\", svm.SVC(kernel=\"linear\")),\n",
        "  (\"SVM poly\", svm.SVC(kernel=\"poly\")),\n",
        "  (\"SVM RBF\", svm.SVC(kernel=\"rbf\", gamma=0.001)),\n",
        "]\n",
        "\n",
        "X_l = [\n",
        "       (\"no jitter\",    digpad[\"data\"]),\n",
        "       (\"with jitter\", digjit[\"data\"]),\n",
        "      ]\n",
        "\n",
        "\n",
        "for X_name, X_i in X_l:\n",
        "  for clf_name, clf_i in clf_l:\n",
        "    results = model_selection.cross_val_score(clf_i, X_i, digits.target)\n",
        "    print(f\"{clf_name}, {X_name}: {statistics.mean(results).round(2)}\")"
      ],
      "execution_count": 5,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "KNN, no jitter: 0.97\n",
            "SVM linear, no jitter: 0.95\n",
            "SVM poly, no jitter: 0.96\n",
            "SVM RBF, no jitter: 0.97\n",
            "KNN, with jitter: 0.35\n",
            "SVM linear, with jitter: 0.1\n",
            "SVM poly, with jitter: 0.27\n",
            "SVM RBF, with jitter: 0.32\n"
          ],
          "name": "stdout"
        }
      ]
    }
  ]
}