colab-mnist-tutorial.ipynb

colab-mnist-tutorial.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "colab-mnist-tutorial.ipynb",
      "version": "0.3.2",
      "provenance": [],
      "collapsed_sections": [],
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/gist/Gijs-Koot/686587b4bb95013d381b7f4d552b54b5/colab-mnist-tutorial.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "metadata": {
        "id": "yGtbkw5cIHqM",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## this imports some software that we need for the tutorial\n",
        "\n",
        "from PIL import Image, ImageChops, ImageMath\n",
        "import matplotlib.pyplot as plt\n",
        "import numpy as np"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "GLL2iSrMIYnT",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## load the dataset, download from google if needed\n",
        "\n",
        "from tensorflow.keras.datasets import mnist\n",
        "\n",
        "(train_images, train_labels), (test_images, test_labels) = mnist.load_data()\n",
        "\n",
        "print(f\"Dimensions of train data: {train_images.shape}\")\n",
        "print(f\"Dimensions of test data: {test_images.shape}\")"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "SYQxjy65L38R",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## this code shows some of the images in our trainset\n",
        "\n",
        "fig, axes = plt.subplots(10, 10, figsize=(10, 10), sharex=True, sharey=True)\n",
        "\n",
        "for i in range(10):\n",
        "  for j in range(10):\n",
        "    axes[i,j].imshow(train_images[i * 10 + j])"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "78hXplpuVWPg",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## these are the labels\n",
        "\n",
        "train_labels[:100]"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "E-SRbU_DUJyk",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## to get a single training (or test) instance, you can index with a number\n",
        "\n",
        "train_images[1]"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "KpvBXHS8Tn0x",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## this is an array with pixel values, with the code below you can see the image\n",
        "plt.imshow(train_images[230])"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "rSC-UA-OUp-1",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## for training we want the pixel values to be between 0 and 1, and instead of a matrix, we want a single vector of length 784\n",
        "\n",
        "X_train = train_images.reshape(60000, -1) / 255\n",
        "X_test = test_images.reshape(10000, -1) / 255\n",
        "\n",
        "X_train[230]"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "TIfbHm6sVp-e",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## another way to look at the training data - no problem for a computer but no human would be able to recognize a seven in this image\n",
        "\n",
        "plt.plot(np.arange(784), X_train[230])"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "MBWgRr-BLwPZ",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## this is the point of the tutorial\n",
        "## creating a neural network is easy and very concrete\n",
        "\n",
        "from tensorflow.keras import models, layers\n",
        "\n",
        "model = models.Sequential()\n",
        "\n",
        "## we have one hidden layer with 512 \"neurons\"\n",
        "model.add(layers.Dense(512, activation='relu', input_shape=(784,)))\n",
        "\n",
        "## output has 10 neurons, corresponding to 10 digits\n",
        "model.add(layers.Dense(10, activation='softmax'))"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "U6QKajKBJWD0",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## with this line, we tell how to optimize and which metrics to track\n",
        "## the model is actually compiled to run on the GPU\n",
        "\n",
        "model.compile(optimizer='rmsprop', loss='sparse_categorical_crossentropy', metrics=['accuracy'])"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "R5yIvmCfhFea",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "model.summary()"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "Vpwn_y1OPGsA",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## this line starts the training\n",
        "## you can see the error on the trainset and the error on the testset\n",
        "## after how many epochs does the model start overfitting?\n",
        "\n",
        "hist = model.fit(X_train, train_labels, epochs=10, batch_size=32, validation_data=(X_test, test_labels))"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "dUiuo9ngYjVu",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## reported accuracy is 98%, let's see\n",
        "\n",
        "test_sample = X_test[100]\n",
        "\n",
        "plt.imshow(test_sample.reshape(28, 28))"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "7VIBiWaiY2ND",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "test_labels[100]"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "XH1K22CtaCuf",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## a technical challenge, the model outputs probabilities, not the actual label\n",
        "\n",
        "model.predict(X_test[100].reshape(1,-1))"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "WWdnr1umZCux",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## here is a helper function to convert dimensions and translate the probabilities\n",
        "\n",
        "def get_model_prediction(model, sample):\n",
        "  sample = np.asarray(sample)\n",
        "  if sample.shape in [(28, 28), (784,)]:\n",
        "    p = model.predict(sample.reshape(1, -1))\n",
        "  else:\n",
        "    raise Exception(f\"Wrong input dimension: {sample.shape}\")\n",
        "  return np.argmax(p)"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "R5U-8nyRZrj6",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "get_model_prediction(model, X_test[100])"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "QUDwb45hcooh",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "## here is a function that allows you to upload files\n",
        "\n",
        "from google.colab import files\n",
        "\n",
        "def upload_image():\n",
        "\n",
        "  uploaded = files.upload()\n",
        "  arr = np.asarray(Image.open(list(uploaded.keys())[0]))\n",
        "  arr = arr.max(axis = 2)\n",
        "  return 1 - arr / arr.max()"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "j3gLXn9Kds3I",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "img = upload_image()"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "dysZbxEVfWc4",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "plt.imshow(img)"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "od3hI-85eXnI",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "get_model_prediction(model, img)"
      ],
      "execution_count": 0,
      "outputs": []
    }
  ]
}