Evaluation_OOD.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "accelerator": "GPU",
    "colab": {
      "name": "Evaluation_OOD.ipynb",
      "provenance": [],
      "machine_shape": "hm",
      "include_colab_link": true
    },
    "environment": {
      "name": "tf2-gpu.2-4.mnightly-2021-01-20-debian-10-test",
      "type": "gcloud",
      "uri": "gcr.io/deeplearning-platform-release/tf2-gpu.2-4:mnightly-2021-01-20-debian-10-test"
    },
    "kernelspec": {
      "display_name": "Python 3",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.8.6"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/github/sayakpaul/Generalized-ODIN-TF/blob/main/Evaluation_OOD.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "heG-9quwoktf"
      },
      "source": [
        "## Setup"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "rhDWmAR9OERr"
      },
      "source": [
        "# Grab the initial model weights\n",
        "!wget -q https://github.com/sayakpaul/Generalized-ODIN-TF/releases/download/v1.0.0/models.tar.gz\n",
        "!tar xf models.tar.gz"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "QDpKFhBfSrR8"
      },
      "source": [
        "!git clone https://github.com/sayakpaul/Generalized-ODIN-TF"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "wJ4sAReHPETA"
      },
      "source": [
        "import sys\n",
        "sys.path.append(\"Generalized-ODIN-TF\")\n",
        "from scripts import metrics\n",
        "\n",
        "import tensorflow as tf\n",
        "import tensorflow_datasets as tfds\n",
        "\n",
        "from tqdm.notebook import tqdm\n",
        "import matplotlib.pyplot as plt\n",
        "import numpy as np\n",
        "\n",
        "tf.random.set_seed(42)\n",
        "np.random.seed(42)\n",
        "tfds.disable_progress_bar()"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "0UWypWRHopbU"
      },
      "source": [
        "## Load the pre-trained model"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "Y1laPh_dPKHs",
        "outputId": "9108dfe2-292a-4ed3-837b-223bd8ca2ad4"
      },
      "source": [
        "rn_odin_model = tf.keras.models.load_model(\"odin_rn_model\")\n",
        "rn_baseline_model = tf.keras.models.load_model(\"rn_baseline_model\")\n",
        "print(f\"Pre-trained models loaded.\")"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Pre-trained models loaded.\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "rVV1WGLcorWy"
      },
      "source": [
        "## Load OOD dataset -- SVHN dataset\n",
        "\n",
        "The purpose of this work is to test a model's ability to detect out-of-distribution (OOD) samples. For that purpose, we are using the [SVHN dataset](https://http://ufldl.stanford.edu/housenumbers/) which is an OOD dataset in this case. "
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 551
        },
        "id": "iI_DR5sPPnSN",
        "outputId": "dd3c83d5-84f2-4a67-c716-4467efc5596e"
      },
      "source": [
        "svhn_test, info = tfds.load(\"svhn_cropped\", split=\"test\", \n",
        "                      as_supervised=True, with_info=True)\n",
        "tfds.show_examples(svhn_test, info)\n",
        "print(f\"Total examples in the test set of SVHN: {tf.data.experimental.cardinality(svhn_test).numpy()}\")"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 648x648 with 9 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        },
        {
          "output_type": "stream",
          "text": [
            "Total examples in the test set of SVHN: 26032\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "RiEoAkopowcY"
      },
      "source": [
        "## Prepare the data loaders"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "6tzvQDiJSf8Z"
      },
      "source": [
        "BATCH_SIZE = 128\n",
        "AUTO = tf.data.AUTOTUNE"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "njjGGYBwQE8d"
      },
      "source": [
        "# SVHN test set\n",
        "svhn_test = (svhn_test\n",
        "             .map(lambda x, y: (tf.cast(x, tf.float32), y), \n",
        "                  num_parallel_calls=AUTO))\n",
        "\n",
        "# CIFAR10 test set\n",
        "(_, _), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()\n",
        "x_test = x_test.astype(\"float32\")"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "7R7TjV1XSmJ9"
      },
      "source": [
        "ds = tf.data.Dataset.from_tensor_slices((x_test, y_test))\n",
        "cifar10_test = ds.skip(1000).take(len(x_test) - 1000)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "hfEaWU0xSmJ-",
        "outputId": "51640be6-f9d0-445b-81e2-bb87bce3d757"
      },
      "source": [
        "NUM_SAMPLE_MAPPING = {\n",
        "    \"svhn\": tf.data.experimental.cardinality(svhn_test).numpy(),\n",
        "    \"cifar10_test\": tf.data.experimental.cardinality(cifar10_test).numpy()\n",
        "}\n",
        "NUM_SAMPLE_MAPPING"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "{'svhn': 26032, 'cifar10_test': 9000}"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 8
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Xv-nA5hkSmJ-"
      },
      "source": [
        "We are skipping 1000 samples because we used them to calculate the perturbation magnitude (refer to [this notebook](https://git.io/JOTgx)). "
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "EZttgyK6o1U5"
      },
      "source": [
        "## Define our `scorer`"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "WzUie-5-8jAZ"
      },
      "source": [
        "# Let's define our model to obtain scores.\n",
        "scorer = tf.keras.Model(rn_odin_model.input, rn_odin_model.layers[-3].output)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "XtilNHIPo6WX"
      },
      "source": [
        "## Perturbing utility\n",
        "\n",
        "As per equation 9 of the paper, we will be perturbing the input images with $\\epsilon$ which was calculated in [this notebook](https://git.io/JOTgx). The value we derived is **0.04**. "
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "R1u32-NbTHx1"
      },
      "source": [
        "def perturb_images(images, model, epsilon=0.04):\n",
        "\n",
        "    test_ds_var = tf.Variable(images, trainable=True)\n",
        "\n",
        "    with tf.GradientTape() as tape:\n",
        "        # Calculate the scores.\n",
        "        tape.watch(test_ds_var)\n",
        "        logits = model(test_ds_var, training=False)\n",
        "        loss = tf.reduce_max(logits, axis=1)\n",
        "        loss = -tf.reduce_mean(loss)\n",
        "\n",
        "    # Calculate the gradients of the scores with respect to the inputs.\n",
        "    gradients = tape.gradient(loss, test_ds_var)\n",
        "    gradients = tf.math.greater_equal(gradients, 0)\n",
        "    gradients = tf.cast(gradients, tf.float32)\n",
        "    gradients = (gradients - 0.5) * 2\n",
        "\n",
        "    # Perturb the inputs and derive new mean score.\n",
        "    # test_ds_var.assign_add(epsilon * gradients)\n",
        "    static_tensor = tf.convert_to_tensor(test_ds_var)\n",
        "    static_tensor = static_tensor - epsilon * gradients\n",
        "    static_tensor = tf.clip_by_value(static_tensor, 0., 255.)\n",
        "    \n",
        "    return static_tensor"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Ua1UDZojp8PU"
      },
      "source": [
        "## Model evalution with AUROC & TNR@TPR95\n",
        "\n",
        "The authors use two metrics for evaluation -- **AUROC** & **TNR@TPR95**. \n",
        "\n",
        "Here are the steps we will roughly be following:\n",
        "\n",
        "1. Perturb the images using `perturb_images()`. For this, we will be using `scorer`. \n",
        "2. Pass those images to `model` to obtain predictions. From the paper - \"At testing time, the class prediction can be made by either calculating $\\arg \\max _{i} f_{i}(\\boldsymbol{x})$ or\n",
        "$\\arg \\max _{i} h_{i}(\\boldsymbol{x})$ (both will give the same predictions).\"\n",
        "3. Use the original labels of the dataset under evaluation and the prediction scores from (2) and calculate the metrics. \n",
        "\n",
        "For evaluating the baseline model, we won't be adding perturbations. "
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "pi3fwtsJSmJ_"
      },
      "source": [
        "# We first perturb the images from both of the datasets.\n",
        "\n",
        "def create_pertubed_ds(ds, model, ds_name):\n",
        "    num_samples = NUM_SAMPLE_MAPPING[ds_name]\n",
        "    perturbed_images = np.zeros((num_samples, 32, 32, 3))\n",
        "    unchanged_labels = np.zeros((num_samples, ))\n",
        "    \n",
        "    for i, (image, label) in enumerate(tqdm(ds)):\n",
        "        image = tf.expand_dims(image, 0)\n",
        "        perturbed_images[i] = perturb_images(image, model)\n",
        "        unchanged_labels[i] = label\n",
        "    \n",
        "    perturbed_ds = tf.data.Dataset.from_tensor_slices((perturbed_images, unchanged_labels))\n",
        "    return perturbed_ds"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "referenced_widgets": [
            "27ebbdc62de84d708fa25185b19b479b",
            "d0578b19efd143959e95d922ab194d08"
          ]
        },
        "id": "BK03CmlHSmJ_",
        "outputId": "9478ef8a-1cf0-4b1c-a96e-70547d9bf5ca"
      },
      "source": [
        "perturbed_svhn = create_pertubed_ds(svhn_test, scorer, \"svhn\")\n",
        "pertubed_cifar10 = create_pertubed_ds(cifar10_test, scorer, \"cifar10_test\")"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "application/vnd.jupyter.widget-view+json": {
              "model_id": "27ebbdc62de84d708fa25185b19b479b",
              "version_major": 2,
              "version_minor": 0
            },
            "text/plain": [
              "  0%|          | 0/26032 [00:00<?, ?it/s]"
            ]
          },
          "metadata": {
            "tags": []
          }
        },
        {
          "output_type": "display_data",
          "data": {
            "application/vnd.jupyter.widget-view+json": {
              "model_id": "d0578b19efd143959e95d922ab194d08",
              "version_major": 2,
              "version_minor": 0
            },
            "text/plain": [
              "  0%|          | 0/9000 [00:00<?, ?it/s]"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "4mVBhOgBSmKA"
      },
      "source": [
        "# We create new data loaders.\n",
        "perturbed_svhn = perturbed_svhn.batch(BATCH_SIZE)\n",
        "pertubed_cifar10 = pertubed_cifar10.batch(BATCH_SIZE)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "VLzU6GMqSmKA",
        "outputId": "4fff73ec-dc09-4ae6-f274-b178726c2d1d"
      },
      "source": [
        "# Calculate AUROC.\n",
        "print(metrics.metric(pertubed_cifar10, perturbed_svhn, rn_odin_model, stype=\"generalized_odin\"))"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "{'generalized_odin': {'TNR': 0.5418715427166565, 'AUROC': 0.9215311326230964}}\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "wvxQGujtSmKA",
        "outputId": "91632ebe-02ce-4aa4-82be-7552ef699202"
      },
      "source": [
        "# Baseline AUROC (with perturbation).\n",
        "cifar10_test = pertubed_cifar10.batch(BATCH_SIZE)\n",
        "print(metrics.metric(pertubed_cifar10, perturbed_svhn, rn_baseline_model, stype=\"baseline\"))"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "{'baseline': {'TNR': 0.4053088506453596, 'AUROC': 0.9114120185412826}}\n"
          ],
          "name": "stdout"
        }
      ]
    }
  ]
}