Run_TCAV.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "kernelspec": {
      "display_name": "Python 3.7.7 64-bit ('tf1.15': conda)",
      "language": "python",
      "name": "python_defaultSpec_1594677056398"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.7.7-final"
    },
    "colab": {
      "name": "Run TCAV.ipynb",
      "provenance": []
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "v88fM4ciYYde"
      },
      "source": [
        "# Running TCAV"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "5-RlU8wNYYdj"
      },
      "source": [
        "This notebook walks you through an example application of the TCAV library step-by-step to understand which human interpretable concepts (e.g. stripes, dots, zigzags) are important to the image classifier GoogleNet's (a.k.a. Inception v1) prediction of Zebras."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "J1zIj8-aYYdj"
      },
      "source": [
        "## Install required packages\n",
        "\n",
        "To run through this notebook in the interim, you are encouraged to utilize a `virtualenv` or `conda` environment for installing and working with the required packages to avoid any dependency and compatability issues with different versions of packages."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "tags": [],
        "id": "sIHww5CuYYdk"
      },
      "source": [
        "%pip install tensorflow\n",
        "%pip install tcav"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "-NT91qkMYYdk"
      },
      "source": [
        "## Download example models and images\n",
        "\n",
        "Open a terminal and run the following commands:\n",
        "\n",
        "```\n",
        "cd tcav/tcav_examples/image_models/imagenet\n",
        "\n",
        "python download_and_make_datasets.py --source_dir=YOUR_PATH --number_of_images_per_folder=50 --number_of_random_folders=3\n",
        "```\n",
        "\n",
        "This script will download the following content into separate folders into a directory you specify with the `--source_dir=` argument:\n",
        "\n",
        "**Images**\n",
        "*  ImageNet images for the target Zebra class\n",
        "*  [Broden dataset](http://netdissect.csail.mit.edu/) images for three concepts (e.g. striped, dotted, zigzagged)\n",
        "*  Random ImageNet class images used by TCAV for hypothesis testing of important concepts\n",
        "\n",
        "**Models**\n",
        "*  [Inception 5h model](https://github.com/Hvass-Labs/TensorFlow-Tutorials/blob/master/inception5h.py)\n",
        "*  [Mobilenet V2 model](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "rlrLUu4zYYdl"
      },
      "source": [
        "## Import extensions and libraries"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "sS1ZjSZjYYdl"
      },
      "source": [
        "%load_ext autoreload\n",
        "%autoreload 2"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "U4yP9kDlYYdl"
      },
      "source": [
        "import tcav.activation_generator as act_gen\n",
        "import tcav.cav as cav\n",
        "import tcav.model  as model\n",
        "import tcav.tcav as tcav\n",
        "import tcav.utils as utils\n",
        "import tcav.utils_plot as utils_plot # utils_plot requires matplotlib\n",
        "import os \n",
        "import tensorflow as tf"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "yitpnXmEYYdm"
      },
      "source": [
        "## TCAV step-by-step\n",
        "\n",
        "You will walk through the following steps below:\n",
        "\n",
        "1. **Store example images in each folder** (you have this if you ran the above)\n",
        " * images for each concept\n",
        " * images for the class/labels of interest\n",
        " * random images that will be negative examples when learning CAVs (images that probably don't belong to any concepts)\n",
        "2. **Write a model wrapper** (below uses example from tcav/model.py)\n",
        " * an instance of  ModelWrapper abstract class (in model.py). This tells TCAV class (tcav.py) how to communicate with your model (e.g., getting internal tensors)\n",
        "3. **Retrieve model activations** (below uses example from tcav/activation_generator.py)\n",
        " * an instance of ActivationGeneratorInterface that tells TCAV class how to load example data and how to get activations from the model\n",
        "4. Run TCAV and visualize scores for important concepts."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "zjjLK3pBYYdm"
      },
      "source": [
        "## Step 1: Store concept and target class images to local folders\n",
        "\n",
        "... and tell TCAV where they are.\n",
        "\n",
        "**source_dir**: where images of concepts, target class and random images (negative samples when learning CAVs) live. Each should be a sub-folder within this directory.\n",
        "\n",
        "Note that random image directories can be in any name. In this example, we are using `random500_0`, `random500_1`,.. for an arbitrary reason. \n",
        "\n",
        "You need roughly 50-200 images per concept and target class (10-20 pictures also tend to work, but 200 is pretty safe).\n",
        "\n",
        "\n",
        "**cav_dir**: directory to store CAVs (`None` if you don't want to store)\n",
        "\n",
        "**target, concept**: names of the target class (that you want to investigate) and concepts (strings) - these are folder names in source_dir\n",
        "\n",
        "**bottlenecks**: list of bottleneck names (intermediate layers in your model) that you want to use for TCAV. These names are defined in the model wrapper below.\n",
        "\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "tags": [],
        "id": "EHabMnMNYYdm"
      },
      "source": [
        "print ('REMEMBER TO UPDATE YOUR_PATH (where images, models are)!')\n",
        "\n",
        "# This is the name of your model wrapper (InceptionV3 and GoogleNet are provided in model.py)\n",
        "model_to_run = 'GoogleNet'  \n",
        "user = 'beenkim'\n",
        "# the name of the parent directory that results are stored (only if you want to cache)\n",
        "project_name = 'tcav_class_test'\n",
        "working_dir = \"/tmp/\" + user + '/' + project_name\n",
        "# where activations are stored (only if your act_gen_wrapper does so)\n",
        "activation_dir =  working_dir+ '/activations/'\n",
        "# where CAVs are stored. \n",
        "# You can say None if you don't wish to store any.\n",
        "cav_dir = working_dir + '/cavs/'\n",
        "# where the images live.\n",
        "\n",
        "# TODO: replace 'YOUR_PATH' with path to downloaded models and images. \n",
        "source_dir = ''\n",
        "bottlenecks = [ 'mixed4c']  # @param \n",
        "      \n",
        "utils.make_dir_if_not_exists(activation_dir)\n",
        "utils.make_dir_if_not_exists(working_dir)\n",
        "utils.make_dir_if_not_exists(cav_dir)\n",
        "\n",
        "# this is a regularizer penalty parameter for linear classifier to get CAVs. \n",
        "alphas = [0.1]   \n",
        "\n",
        "target = 'zebra'  \n",
        "concepts = [\"dotted\",\"striped\",\"zigzagged\"]  \n"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "u_DfQqsAYYdn"
      },
      "source": [
        "## Step 2: Write your model wrapper\n",
        "\n",
        "The next step is to tell TCAV how to communicate with your model. See `model.GoogleNetWrapper_public ` for details.\n",
        "\n",
        "You can define a subclass of ModelWrapper abstract class to do this. Let me walk you thru what each function does (tho they are pretty self-explanatory).  This wrapper includes a lot of the functions that you already have, for example, `get_prediction`.\n",
        "\n",
        "### 2.1: Tensors from the graph: bottleneck tensors and ends\n",
        "First, store your bottleneck tensors in `self.bottlenecks_tensors` as a dictionary. You only need bottlenecks that you are interested in running TCAV with. Similarly, fill in `self.ends` dictionary with `input`, `logit` and `prediction` tensors.\n",
        "\n",
        "### 2.2: Define loss\n",
        "Get your loss tensor, and assigned it to `self.loss`. This is what TCAV uses to take directional derivatives. \n",
        "\n",
        "While doing so, you would also want to set \n",
        "```python\n",
        "self.y_input \n",
        "```\n",
        "this simply is a tensorflow place holder for the target index in the logit layer (e.g., 0 index for a dog, 1 for a cat).\n",
        "For multi-class classification, typically something like this works:\n",
        "\n",
        "```python\n",
        "self.y_input = tf.placeholder(tf.int64, shape=[None])\n",
        "```\n",
        "\n",
        "For example, for a multiclass classifier, something like below would work. \n",
        "\n",
        "```python\n",
        "    # Construct gradient ops.\n",
        "    with g.as_default():\n",
        "      self.y_input = tf.placeholder(tf.int64, shape=[None])\n",
        "\n",
        "      self.pred = tf.expand_dims(self.ends['prediction'][0], 0)\n",
        "\n",
        "      self.loss = tf.reduce_mean(\n",
        "          tf.nn.softmax_cross_entropy_with_logits(\n",
        "              labels=tf.one_hot(self.y_input, len(self.labels)),\n",
        "              logits=self.pred))\n",
        "    self._make_gradient_tensors()\n",
        "```\n",
        "\n",
        "### 2.3: Call _make_gradient_tensors in __init__() of your wrapper\n",
        "```python\n",
        "_make_gradient_tensors()  \n",
        "```\n",
        "does what you expect - given the loss and bottleneck tensors defined above, it adds gradient tensors.\n",
        "\n",
        "### 2.4: Fill in labels, image shapes and a model name.\n",
        "Get the mapping from labels (strings) to indice in the logit layer (int) in a dictionary format.\n",
        "\n",
        "```python\n",
        "def id_to_label(self, idx)\n",
        "def label_to_id(self, label)\n",
        "```\n",
        "\n",
        "Set your input image shape at  `self.image_shape`\n",
        "\n",
        "\n",
        "Set your model name to `self.model_name`\n",
        "\n",
        "You are done with writing the model wrapper! See the two example model wrapers, InceptionV3 and Googlenet in `tcav/model.py`."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "tags": [],
        "id": "hH-YQiEIYYdn"
      },
      "source": [
        "# Create TensorFlow session.\n",
        "sess = utils.create_session()\n",
        "\n",
        "# GRAPH_PATH is where the trained model is stored.\n",
        "GRAPH_PATH = source_dir + \"/inception5h/tensorflow_inception_graph.pb\"\n",
        "# LABEL_PATH is where the labels are stored. Each line contains one class, and they are ordered with respect to their index in \n",
        "# the logit layer. (yes, id_to_label function in the model wrapper reads from this file.)\n",
        "# For example, imagenet_comp_graph_label_strings.txt looks like:\n",
        "# dummy                                                                                      \n",
        "# kit fox\n",
        "# English setter\n",
        "# Siberian husky ...\n",
        "\n",
        "LABEL_PATH = source_dir + \"/inception5h/imagenet_comp_graph_label_strings.txt\"\n",
        "\n",
        "mymodel = model.GoogleNetWrapper_public(sess,\n",
        "                                        GRAPH_PATH,\n",
        "                                        LABEL_PATH)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "yY5kXbVAYYdo"
      },
      "source": [
        "## Step 3: Implement a class that returns activations (maybe with caching!)\n",
        "\n",
        "Lastly, you will implement a class of the ActivationGenerationInterface which TCAV uses to load example data for a given concept or target, call into your model wrapper and return activations. I pulled out this logic outside of mymodel because this step often takes the longest. By making it modular, you can cache your activations and/or parallelize your computations, as I have done in `ActivationGeneratorBase.process_and_load_activations` in `activation_generator.py`.\n",
        "\n",
        "\n",
        "The `process_and_load_activations` method of the activation generator must return a dictionary of activations that has concept or target name as  a first key, and the bottleneck name as a second key. So something like:\n",
        "\n",
        "```python\n",
        "{concept1: {bottleneck1: [[0.2, 0.1, ....]]},\n",
        "concept2: {bottleneck1: [[0.1, 0.02, ....]]},\n",
        "target1: {bottleneck1: [[0.02, 0.99, ....]]}\n",
        "```\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ZmSyFxQbYYdo"
      },
      "source": [
        "act_generator = act_gen.ImageActivationGenerator(mymodel, source_dir, activation_dir, max_examples=100)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "uABCWhp8YYdo"
      },
      "source": [
        "## Step 4: Run TCAV and visualize concept importance\n",
        "\n",
        "You are now ready to run TCAV! Let's do it.\n",
        "\n",
        "**num_random_exp**: number of experiments to confirm meaningful concept direction. TCAV will search for this many folders named `random500_0`, `random500_1`, etc. You can alternatively set the `random_concepts` keyword to be a list of folders of random concepts. Run at least 10-20 for meaningful tests. \n",
        "\n",
        "**random_counterpart**: as well as the above, you can optionally supply a single folder with random images as the \"positive set\" for statistical testing. Reduces computation time at the cost of less reliable random TCAV scores.\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": false,
        "tags": [],
        "id": "F2FVOGSvYYdp"
      },
      "source": [
        "import absl\n",
        "absl.logging.set_verbosity(0)\n",
        "num_random_exp=3\n",
        "## only running num_random_exp = 10 to save some time. The paper number are reported for 500 random runs. \n",
        "mytcav = tcav.TCAV(sess,\n",
        "                   target,\n",
        "                   concepts,\n",
        "                   bottlenecks,\n",
        "                   act_generator,\n",
        "                   alphas,\n",
        "                   cav_dir=cav_dir,\n",
        "                   num_random_exp=num_random_exp)#10)\n",
        "print ('This may take a while... Go get coffee!')\n",
        "results = mytcav.run(run_parallel=False)\n",
        "print ('done!')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": false,
        "tags": [],
        "id": "hjKVKa80YYdp",
        "outputId": "ff3cbebe-4edd-4342-cede-24dc10ce9d99"
      },
      "source": [
        "utils_plot.plot_results(results, num_random_exp=num_random_exp)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Class = zebra\n",
            "  Concept = dotted\n",
            "    Bottleneck = mixed4c. TCAV Score = 0.56 (+- 0.12), random was 0.54 (+- 0.16). p-val = 0.917 (not significant)\n",
            "  Concept = striped\n",
            "    Bottleneck = mixed4c. TCAV Score = 0.89 (+- 0.09), random was 0.54 (+- 0.16). p-val = 0.018 (significant)\n",
            "  Concept = zigzagged\n",
            "    Bottleneck = mixed4c. TCAV Score = 0.84 (+- 0.10), random was 0.54 (+- 0.16). p-val = 0.036 (significant)\n",
            "{'mixed4c': {'bn_vals': [0.01, 0.888888888888889, 0.8376068376068376], 'bn_stds': [0, 0.08716272672808177, 0.09669836323918596], 'significant': [False, True, True]}}\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ],
            "image/svg+xml": "<?xml version=\"1.0\" encoding=\"utf-8\" standalone=\"no\"?>\n<!DOCTYPE svg PUBLIC \"-//W3C//DTD SVG 1.1//EN\"\n  \"http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd\">\n<!-- Created with matplotlib (https://matplotlib.org/) -->\n<svg height=\"279.59625pt\" version=\"1.1\" viewBox=\"0 0 423.248037 279.59625\" width=\"423.248037pt\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\">\n <defs>\n  <style type=\"text/css\">\n*{stroke-linecap:butt;stroke-linejoin:round;}\n  </style>\n </defs>\n <g id=\"figure_1\">\n  <g id=\"patch_1\">\n   <path d=\"M 0 279.59625 \nL 423.248037 279.59625 \nL 423.248037 0 \nL 0 0 \nz\n\" style=\"fill:none;\"/>\n  </g>\n  <g id=\"axes_1\">\n   <g id=\"patch_2\">\n    <path d=\"M 43.78125 255.718125 \nL 407.086932 255.718125 \nL 407.086932 22.318125 \nL 43.78125 22.318125 \nz\n\" style=\"fill:#ffffff;\"/>\n   </g>\n   <g id=\"patch_3\">\n    <path clip-path=\"url(#p4693219472)\" d=\"M 60.295145 255.718125 \nL 126.350723 255.718125 \nL 126.350723 253.440728 \nL 60.295145 253.440728 \nz\n\" style=\"fill:#1f77b4;\"/>\n   </g>\n   <g id=\"patch_4\">\n    <path clip-path=\"url(#p4693219472)\" d=\"M 192.406302 255.718125 \nL 258.46188 255.718125 \nL 258.46188 53.282825 \nL 192.406302 53.282825 \nz\n\" style=\"fill:#1f77b4;\"/>\n   </g>\n   <g id=\"patch_5\">\n    <path clip-path=\"url(#p4693219472)\" d=\"M 324.517459 255.718125 \nL 390.573037 255.718125 \nL 390.573037 64.961785 \nL 324.517459 64.961785 \nz\n\" style=\"fill:#1f77b4;\"/>\n   </g>\n   <g id=\"matplotlib.axis_1\">\n    <g id=\"xtick_1\">\n     <g id=\"line2d_1\">\n      <defs>\n       <path d=\"M 0 0 \nL 0 3.5 \n\" id=\"m4d347a0db9\" style=\"stroke:#000000;stroke-width:0.8;\"/>\n      </defs>\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"126.350723\" xlink:href=\"#m4d347a0db9\" y=\"255.718125\"/>\n      </g>\n     </g>\n     <g id=\"text_1\">\n      <!-- dotted -->\n      <defs>\n       <path d=\"M 45.40625 46.390625 \nL 45.40625 75.984375 \nL 54.390625 75.984375 \nL 54.390625 0 \nL 45.40625 0 \nL 45.40625 8.203125 \nQ 42.578125 3.328125 38.25 0.953125 \nQ 33.9375 -1.421875 27.875 -1.421875 \nQ 17.96875 -1.421875 11.734375 6.484375 \nQ 5.515625 14.40625 5.515625 27.296875 \nQ 5.515625 40.1875 11.734375 48.09375 \nQ 17.96875 56 27.875 56 \nQ 33.9375 56 38.25 53.625 \nQ 42.578125 51.265625 45.40625 46.390625 \nz\nM 14.796875 27.296875 \nQ 14.796875 17.390625 18.875 11.75 \nQ 22.953125 6.109375 30.078125 6.109375 \nQ 37.203125 6.109375 41.296875 11.75 \nQ 45.40625 17.390625 45.40625 27.296875 \nQ 45.40625 37.203125 41.296875 42.84375 \nQ 37.203125 48.484375 30.078125 48.484375 \nQ 22.953125 48.484375 18.875 42.84375 \nQ 14.796875 37.203125 14.796875 27.296875 \nz\n\" id=\"DejaVuSans-100\"/>\n       <path d=\"M 30.609375 48.390625 \nQ 23.390625 48.390625 19.1875 42.75 \nQ 14.984375 37.109375 14.984375 27.296875 \nQ 14.984375 17.484375 19.15625 11.84375 \nQ 23.34375 6.203125 30.609375 6.203125 \nQ 37.796875 6.203125 41.984375 11.859375 \nQ 46.1875 17.53125 46.1875 27.296875 \nQ 46.1875 37.015625 41.984375 42.703125 \nQ 37.796875 48.390625 30.609375 48.390625 \nz\nM 30.609375 56 \nQ 42.328125 56 49.015625 48.375 \nQ 55.71875 40.765625 55.71875 27.296875 \nQ 55.71875 13.875 49.015625 6.21875 \nQ 42.328125 -1.421875 30.609375 -1.421875 \nQ 18.84375 -1.421875 12.171875 6.21875 \nQ 5.515625 13.875 5.515625 27.296875 \nQ 5.515625 40.765625 12.171875 48.375 \nQ 18.84375 56 30.609375 56 \nz\n\" id=\"DejaVuSans-111\"/>\n       <path d=\"M 18.3125 70.21875 \nL 18.3125 54.6875 \nL 36.8125 54.6875 \nL 36.8125 47.703125 \nL 18.3125 47.703125 \nL 18.3125 18.015625 \nQ 18.3125 11.328125 20.140625 9.421875 \nQ 21.96875 7.515625 27.59375 7.515625 \nL 36.8125 7.515625 \nL 36.8125 0 \nL 27.59375 0 \nQ 17.1875 0 13.234375 3.875 \nQ 9.28125 7.765625 9.28125 18.015625 \nL 9.28125 47.703125 \nL 2.6875 47.703125 \nL 2.6875 54.6875 \nL 9.28125 54.6875 \nL 9.28125 70.21875 \nz\n\" id=\"DejaVuSans-116\"/>\n       <path d=\"M 56.203125 29.59375 \nL 56.203125 25.203125 \nL 14.890625 25.203125 \nQ 15.484375 15.921875 20.484375 11.0625 \nQ 25.484375 6.203125 34.421875 6.203125 \nQ 39.59375 6.203125 44.453125 7.46875 \nQ 49.3125 8.734375 54.109375 11.28125 \nL 54.109375 2.78125 \nQ 49.265625 0.734375 44.1875 -0.34375 \nQ 39.109375 -1.421875 33.890625 -1.421875 \nQ 20.796875 -1.421875 13.15625 6.1875 \nQ 5.515625 13.8125 5.515625 26.8125 \nQ 5.515625 40.234375 12.765625 48.109375 \nQ 20.015625 56 32.328125 56 \nQ 43.359375 56 49.78125 48.890625 \nQ 56.203125 41.796875 56.203125 29.59375 \nz\nM 47.21875 32.234375 \nQ 47.125 39.59375 43.09375 43.984375 \nQ 39.0625 48.390625 32.421875 48.390625 \nQ 24.90625 48.390625 20.390625 44.140625 \nQ 15.875 39.890625 15.1875 32.171875 \nz\n\" id=\"DejaVuSans-101\"/>\n      </defs>\n      <g transform=\"translate(109.946036 270.316563)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-100\"/>\n       <use x=\"63.476562\" xlink:href=\"#DejaVuSans-111\"/>\n       <use x=\"124.658203\" xlink:href=\"#DejaVuSans-116\"/>\n       <use x=\"163.867188\" xlink:href=\"#DejaVuSans-116\"/>\n       <use x=\"203.076172\" xlink:href=\"#DejaVuSans-101\"/>\n       <use x=\"264.599609\" xlink:href=\"#DejaVuSans-100\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"xtick_2\">\n     <g id=\"line2d_2\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"258.46188\" xlink:href=\"#m4d347a0db9\" y=\"255.718125\"/>\n      </g>\n     </g>\n     <g id=\"text_2\">\n      <!-- striped -->\n      <defs>\n       <path d=\"M 44.28125 53.078125 \nL 44.28125 44.578125 \nQ 40.484375 46.53125 36.375 47.5 \nQ 32.28125 48.484375 27.875 48.484375 \nQ 21.1875 48.484375 17.84375 46.4375 \nQ 14.5 44.390625 14.5 40.28125 \nQ 14.5 37.15625 16.890625 35.375 \nQ 19.28125 33.59375 26.515625 31.984375 \nL 29.59375 31.296875 \nQ 39.15625 29.25 43.1875 25.515625 \nQ 47.21875 21.78125 47.21875 15.09375 \nQ 47.21875 7.46875 41.1875 3.015625 \nQ 35.15625 -1.421875 24.609375 -1.421875 \nQ 20.21875 -1.421875 15.453125 -0.5625 \nQ 10.6875 0.296875 5.421875 2 \nL 5.421875 11.28125 \nQ 10.40625 8.6875 15.234375 7.390625 \nQ 20.0625 6.109375 24.8125 6.109375 \nQ 31.15625 6.109375 34.5625 8.28125 \nQ 37.984375 10.453125 37.984375 14.40625 \nQ 37.984375 18.0625 35.515625 20.015625 \nQ 33.0625 21.96875 24.703125 23.78125 \nL 21.578125 24.515625 \nQ 13.234375 26.265625 9.515625 29.90625 \nQ 5.8125 33.546875 5.8125 39.890625 \nQ 5.8125 47.609375 11.28125 51.796875 \nQ 16.75 56 26.8125 56 \nQ 31.78125 56 36.171875 55.265625 \nQ 40.578125 54.546875 44.28125 53.078125 \nz\n\" id=\"DejaVuSans-115\"/>\n       <path d=\"M 41.109375 46.296875 \nQ 39.59375 47.171875 37.8125 47.578125 \nQ 36.03125 48 33.890625 48 \nQ 26.265625 48 22.1875 43.046875 \nQ 18.109375 38.09375 18.109375 28.8125 \nL 18.109375 0 \nL 9.078125 0 \nL 9.078125 54.6875 \nL 18.109375 54.6875 \nL 18.109375 46.1875 \nQ 20.953125 51.171875 25.484375 53.578125 \nQ 30.03125 56 36.53125 56 \nQ 37.453125 56 38.578125 55.875 \nQ 39.703125 55.765625 41.0625 55.515625 \nz\n\" id=\"DejaVuSans-114\"/>\n       <path d=\"M 9.421875 54.6875 \nL 18.40625 54.6875 \nL 18.40625 0 \nL 9.421875 0 \nz\nM 9.421875 75.984375 \nL 18.40625 75.984375 \nL 18.40625 64.59375 \nL 9.421875 64.59375 \nz\n\" id=\"DejaVuSans-105\"/>\n       <path d=\"M 18.109375 8.203125 \nL 18.109375 -20.796875 \nL 9.078125 -20.796875 \nL 9.078125 54.6875 \nL 18.109375 54.6875 \nL 18.109375 46.390625 \nQ 20.953125 51.265625 25.265625 53.625 \nQ 29.59375 56 35.59375 56 \nQ 45.5625 56 51.78125 48.09375 \nQ 58.015625 40.1875 58.015625 27.296875 \nQ 58.015625 14.40625 51.78125 6.484375 \nQ 45.5625 -1.421875 35.59375 -1.421875 \nQ 29.59375 -1.421875 25.265625 0.953125 \nQ 20.953125 3.328125 18.109375 8.203125 \nz\nM 48.6875 27.296875 \nQ 48.6875 37.203125 44.609375 42.84375 \nQ 40.53125 48.484375 33.40625 48.484375 \nQ 26.265625 48.484375 22.1875 42.84375 \nQ 18.109375 37.203125 18.109375 27.296875 \nQ 18.109375 17.390625 22.1875 11.75 \nQ 26.265625 6.109375 33.40625 6.109375 \nQ 40.53125 6.109375 44.609375 11.75 \nQ 48.6875 17.390625 48.6875 27.296875 \nz\n\" id=\"DejaVuSans-112\"/>\n      </defs>\n      <g transform=\"translate(241.027505 270.316563)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-115\"/>\n       <use x=\"52.099609\" xlink:href=\"#DejaVuSans-116\"/>\n       <use x=\"91.308594\" xlink:href=\"#DejaVuSans-114\"/>\n       <use x=\"132.421875\" xlink:href=\"#DejaVuSans-105\"/>\n       <use x=\"160.205078\" xlink:href=\"#DejaVuSans-112\"/>\n       <use x=\"223.681641\" xlink:href=\"#DejaVuSans-101\"/>\n       <use x=\"285.205078\" xlink:href=\"#DejaVuSans-100\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"xtick_3\">\n     <g id=\"line2d_3\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"390.573037\" xlink:href=\"#m4d347a0db9\" y=\"255.718125\"/>\n      </g>\n     </g>\n     <g id=\"text_3\">\n      <!-- zigzagged -->\n      <defs>\n       <path d=\"M 5.515625 54.6875 \nL 48.1875 54.6875 \nL 48.1875 46.484375 \nL 14.40625 7.171875 \nL 48.1875 7.171875 \nL 48.1875 0 \nL 4.296875 0 \nL 4.296875 8.203125 \nL 38.09375 47.515625 \nL 5.515625 47.515625 \nz\n\" id=\"DejaVuSans-122\"/>\n       <path d=\"M 45.40625 27.984375 \nQ 45.40625 37.75 41.375 43.109375 \nQ 37.359375 48.484375 30.078125 48.484375 \nQ 22.859375 48.484375 18.828125 43.109375 \nQ 14.796875 37.75 14.796875 27.984375 \nQ 14.796875 18.265625 18.828125 12.890625 \nQ 22.859375 7.515625 30.078125 7.515625 \nQ 37.359375 7.515625 41.375 12.890625 \nQ 45.40625 18.265625 45.40625 27.984375 \nz\nM 54.390625 6.78125 \nQ 54.390625 -7.171875 48.1875 -13.984375 \nQ 42 -20.796875 29.203125 -20.796875 \nQ 24.46875 -20.796875 20.265625 -20.09375 \nQ 16.0625 -19.390625 12.109375 -17.921875 \nL 12.109375 -9.1875 \nQ 16.0625 -11.328125 19.921875 -12.34375 \nQ 23.78125 -13.375 27.78125 -13.375 \nQ 36.625 -13.375 41.015625 -8.765625 \nQ 45.40625 -4.15625 45.40625 5.171875 \nL 45.40625 9.625 \nQ 42.625 4.78125 38.28125 2.390625 \nQ 33.9375 0 27.875 0 \nQ 17.828125 0 11.671875 7.65625 \nQ 5.515625 15.328125 5.515625 27.984375 \nQ 5.515625 40.671875 11.671875 48.328125 \nQ 17.828125 56 27.875 56 \nQ 33.9375 56 38.28125 53.609375 \nQ 42.625 51.21875 45.40625 46.390625 \nL 45.40625 54.6875 \nL 54.390625 54.6875 \nz\n\" id=\"DejaVuSans-103\"/>\n       <path d=\"M 34.28125 27.484375 \nQ 23.390625 27.484375 19.1875 25 \nQ 14.984375 22.515625 14.984375 16.5 \nQ 14.984375 11.71875 18.140625 8.90625 \nQ 21.296875 6.109375 26.703125 6.109375 \nQ 34.1875 6.109375 38.703125 11.40625 \nQ 43.21875 16.703125 43.21875 25.484375 \nL 43.21875 27.484375 \nz\nM 52.203125 31.203125 \nL 52.203125 0 \nL 43.21875 0 \nL 43.21875 8.296875 \nQ 40.140625 3.328125 35.546875 0.953125 \nQ 30.953125 -1.421875 24.3125 -1.421875 \nQ 15.921875 -1.421875 10.953125 3.296875 \nQ 6 8.015625 6 15.921875 \nQ 6 25.140625 12.171875 29.828125 \nQ 18.359375 34.515625 30.609375 34.515625 \nL 43.21875 34.515625 \nL 43.21875 35.40625 \nQ 43.21875 41.609375 39.140625 45 \nQ 35.0625 48.390625 27.6875 48.390625 \nQ 23 48.390625 18.546875 47.265625 \nQ 14.109375 46.140625 10.015625 43.890625 \nL 10.015625 52.203125 \nQ 14.9375 54.109375 19.578125 55.046875 \nQ 24.21875 56 28.609375 56 \nQ 40.484375 56 46.34375 49.84375 \nQ 52.203125 43.703125 52.203125 31.203125 \nz\n\" id=\"DejaVuSans-97\"/>\n      </defs>\n      <g transform=\"translate(365.098037 270.316563)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-122\"/>\n       <use x=\"52.490234\" xlink:href=\"#DejaVuSans-105\"/>\n       <use x=\"80.273438\" xlink:href=\"#DejaVuSans-103\"/>\n       <use x=\"143.75\" xlink:href=\"#DejaVuSans-122\"/>\n       <use x=\"196.240234\" xlink:href=\"#DejaVuSans-97\"/>\n       <use x=\"257.519531\" xlink:href=\"#DejaVuSans-103\"/>\n       <use x=\"320.996094\" xlink:href=\"#DejaVuSans-103\"/>\n       <use x=\"384.472656\" xlink:href=\"#DejaVuSans-101\"/>\n       <use x=\"445.996094\" xlink:href=\"#DejaVuSans-100\"/>\n      </g>\n     </g>\n    </g>\n   </g>\n   <g id=\"matplotlib.axis_2\">\n    <g id=\"ytick_1\">\n     <g id=\"line2d_4\">\n      <defs>\n       <path d=\"M 0 0 \nL -3.5 0 \n\" id=\"m1eb8253b76\" style=\"stroke:#000000;stroke-width:0.8;\"/>\n      </defs>\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"43.78125\" xlink:href=\"#m1eb8253b76\" y=\"255.718125\"/>\n      </g>\n     </g>\n     <g id=\"text_4\">\n      <!-- 0.0 -->\n      <defs>\n       <path d=\"M 31.78125 66.40625 \nQ 24.171875 66.40625 20.328125 58.90625 \nQ 16.5 51.421875 16.5 36.375 \nQ 16.5 21.390625 20.328125 13.890625 \nQ 24.171875 6.390625 31.78125 6.390625 \nQ 39.453125 6.390625 43.28125 13.890625 \nQ 47.125 21.390625 47.125 36.375 \nQ 47.125 51.421875 43.28125 58.90625 \nQ 39.453125 66.40625 31.78125 66.40625 \nz\nM 31.78125 74.21875 \nQ 44.046875 74.21875 50.515625 64.515625 \nQ 56.984375 54.828125 56.984375 36.375 \nQ 56.984375 17.96875 50.515625 8.265625 \nQ 44.046875 -1.421875 31.78125 -1.421875 \nQ 19.53125 -1.421875 13.0625 8.265625 \nQ 6.59375 17.96875 6.59375 36.375 \nQ 6.59375 54.828125 13.0625 64.515625 \nQ 19.53125 74.21875 31.78125 74.21875 \nz\n\" id=\"DejaVuSans-48\"/>\n       <path d=\"M 10.6875 12.40625 \nL 21 12.40625 \nL 21 0 \nL 10.6875 0 \nz\n\" id=\"DejaVuSans-46\"/>\n      </defs>\n      <g transform=\"translate(20.878125 259.517344)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-46\"/>\n       <use x=\"95.410156\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"ytick_2\">\n     <g id=\"line2d_5\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"43.78125\" xlink:href=\"#m1eb8253b76\" y=\"210.170183\"/>\n      </g>\n     </g>\n     <g id=\"text_5\">\n      <!-- 0.2 -->\n      <defs>\n       <path d=\"M 19.1875 8.296875 \nL 53.609375 8.296875 \nL 53.609375 0 \nL 7.328125 0 \nL 7.328125 8.296875 \nQ 12.9375 14.109375 22.625 23.890625 \nQ 32.328125 33.6875 34.8125 36.53125 \nQ 39.546875 41.84375 41.421875 45.53125 \nQ 43.3125 49.21875 43.3125 52.78125 \nQ 43.3125 58.59375 39.234375 62.25 \nQ 35.15625 65.921875 28.609375 65.921875 \nQ 23.96875 65.921875 18.8125 64.3125 \nQ 13.671875 62.703125 7.8125 59.421875 \nL 7.8125 69.390625 \nQ 13.765625 71.78125 18.9375 73 \nQ 24.125 74.21875 28.421875 74.21875 \nQ 39.75 74.21875 46.484375 68.546875 \nQ 53.21875 62.890625 53.21875 53.421875 \nQ 53.21875 48.921875 51.53125 44.890625 \nQ 49.859375 40.875 45.40625 35.40625 \nQ 44.1875 33.984375 37.640625 27.21875 \nQ 31.109375 20.453125 19.1875 8.296875 \nz\n\" id=\"DejaVuSans-50\"/>\n      </defs>\n      <g transform=\"translate(20.878125 213.969401)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-46\"/>\n       <use x=\"95.410156\" xlink:href=\"#DejaVuSans-50\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"ytick_3\">\n     <g id=\"line2d_6\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"43.78125\" xlink:href=\"#m1eb8253b76\" y=\"164.62224\"/>\n      </g>\n     </g>\n     <g id=\"text_6\">\n      <!-- 0.4 -->\n      <defs>\n       <path d=\"M 37.796875 64.3125 \nL 12.890625 25.390625 \nL 37.796875 25.390625 \nz\nM 35.203125 72.90625 \nL 47.609375 72.90625 \nL 47.609375 25.390625 \nL 58.015625 25.390625 \nL 58.015625 17.1875 \nL 47.609375 17.1875 \nL 47.609375 0 \nL 37.796875 0 \nL 37.796875 17.1875 \nL 4.890625 17.1875 \nL 4.890625 26.703125 \nz\n\" id=\"DejaVuSans-52\"/>\n      </defs>\n      <g transform=\"translate(20.878125 168.421459)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-46\"/>\n       <use x=\"95.410156\" xlink:href=\"#DejaVuSans-52\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"ytick_4\">\n     <g id=\"line2d_7\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"43.78125\" xlink:href=\"#m1eb8253b76\" y=\"119.074298\"/>\n      </g>\n     </g>\n     <g id=\"text_7\">\n      <!-- 0.6 -->\n      <defs>\n       <path d=\"M 33.015625 40.375 \nQ 26.375 40.375 22.484375 35.828125 \nQ 18.609375 31.296875 18.609375 23.390625 \nQ 18.609375 15.53125 22.484375 10.953125 \nQ 26.375 6.390625 33.015625 6.390625 \nQ 39.65625 6.390625 43.53125 10.953125 \nQ 47.40625 15.53125 47.40625 23.390625 \nQ 47.40625 31.296875 43.53125 35.828125 \nQ 39.65625 40.375 33.015625 40.375 \nz\nM 52.59375 71.296875 \nL 52.59375 62.3125 \nQ 48.875 64.0625 45.09375 64.984375 \nQ 41.3125 65.921875 37.59375 65.921875 \nQ 27.828125 65.921875 22.671875 59.328125 \nQ 17.53125 52.734375 16.796875 39.40625 \nQ 19.671875 43.65625 24.015625 45.921875 \nQ 28.375 48.1875 33.59375 48.1875 \nQ 44.578125 48.1875 50.953125 41.515625 \nQ 57.328125 34.859375 57.328125 23.390625 \nQ 57.328125 12.15625 50.6875 5.359375 \nQ 44.046875 -1.421875 33.015625 -1.421875 \nQ 20.359375 -1.421875 13.671875 8.265625 \nQ 6.984375 17.96875 6.984375 36.375 \nQ 6.984375 53.65625 15.1875 63.9375 \nQ 23.390625 74.21875 37.203125 74.21875 \nQ 40.921875 74.21875 44.703125 73.484375 \nQ 48.484375 72.75 52.59375 71.296875 \nz\n\" id=\"DejaVuSans-54\"/>\n      </defs>\n      <g transform=\"translate(20.878125 122.873516)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-46\"/>\n       <use x=\"95.410156\" xlink:href=\"#DejaVuSans-54\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"ytick_5\">\n     <g id=\"line2d_8\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"43.78125\" xlink:href=\"#m1eb8253b76\" y=\"73.526355\"/>\n      </g>\n     </g>\n     <g id=\"text_8\">\n      <!-- 0.8 -->\n      <defs>\n       <path d=\"M 31.78125 34.625 \nQ 24.75 34.625 20.71875 30.859375 \nQ 16.703125 27.09375 16.703125 20.515625 \nQ 16.703125 13.921875 20.71875 10.15625 \nQ 24.75 6.390625 31.78125 6.390625 \nQ 38.8125 6.390625 42.859375 10.171875 \nQ 46.921875 13.96875 46.921875 20.515625 \nQ 46.921875 27.09375 42.890625 30.859375 \nQ 38.875 34.625 31.78125 34.625 \nz\nM 21.921875 38.8125 \nQ 15.578125 40.375 12.03125 44.71875 \nQ 8.5 49.078125 8.5 55.328125 \nQ 8.5 64.0625 14.71875 69.140625 \nQ 20.953125 74.21875 31.78125 74.21875 \nQ 42.671875 74.21875 48.875 69.140625 \nQ 55.078125 64.0625 55.078125 55.328125 \nQ 55.078125 49.078125 51.53125 44.71875 \nQ 48 40.375 41.703125 38.8125 \nQ 48.828125 37.15625 52.796875 32.3125 \nQ 56.78125 27.484375 56.78125 20.515625 \nQ 56.78125 9.90625 50.3125 4.234375 \nQ 43.84375 -1.421875 31.78125 -1.421875 \nQ 19.734375 -1.421875 13.25 4.234375 \nQ 6.78125 9.90625 6.78125 20.515625 \nQ 6.78125 27.484375 10.78125 32.3125 \nQ 14.796875 37.15625 21.921875 38.8125 \nz\nM 18.3125 54.390625 \nQ 18.3125 48.734375 21.84375 45.5625 \nQ 25.390625 42.390625 31.78125 42.390625 \nQ 38.140625 42.390625 41.71875 45.5625 \nQ 45.3125 48.734375 45.3125 54.390625 \nQ 45.3125 60.0625 41.71875 63.234375 \nQ 38.140625 66.40625 31.78125 66.40625 \nQ 25.390625 66.40625 21.84375 63.234375 \nQ 18.3125 60.0625 18.3125 54.390625 \nz\n\" id=\"DejaVuSans-56\"/>\n      </defs>\n      <g transform=\"translate(20.878125 77.325574)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-46\"/>\n       <use x=\"95.410156\" xlink:href=\"#DejaVuSans-56\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"ytick_6\">\n     <g id=\"line2d_9\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"43.78125\" xlink:href=\"#m1eb8253b76\" y=\"27.978413\"/>\n      </g>\n     </g>\n     <g id=\"text_9\">\n      <!-- 1.0 -->\n      <defs>\n       <path d=\"M 12.40625 8.296875 \nL 28.515625 8.296875 \nL 28.515625 63.921875 \nL 10.984375 60.40625 \nL 10.984375 69.390625 \nL 28.421875 72.90625 \nL 38.28125 72.90625 \nL 38.28125 8.296875 \nL 54.390625 8.296875 \nL 54.390625 0 \nL 12.40625 0 \nz\n\" id=\"DejaVuSans-49\"/>\n      </defs>\n      <g transform=\"translate(20.878125 31.777631)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-49\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-46\"/>\n       <use x=\"95.410156\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"text_10\">\n     <!-- TCAV Score -->\n     <defs>\n      <path d=\"M -0.296875 72.90625 \nL 61.375 72.90625 \nL 61.375 64.59375 \nL 35.5 64.59375 \nL 35.5 0 \nL 25.59375 0 \nL 25.59375 64.59375 \nL -0.296875 64.59375 \nz\n\" id=\"DejaVuSans-84\"/>\n      <path d=\"M 64.40625 67.28125 \nL 64.40625 56.890625 \nQ 59.421875 61.53125 53.78125 63.8125 \nQ 48.140625 66.109375 41.796875 66.109375 \nQ 29.296875 66.109375 22.65625 58.46875 \nQ 16.015625 50.828125 16.015625 36.375 \nQ 16.015625 21.96875 22.65625 14.328125 \nQ 29.296875 6.6875 41.796875 6.6875 \nQ 48.140625 6.6875 53.78125 8.984375 \nQ 59.421875 11.28125 64.40625 15.921875 \nL 64.40625 5.609375 \nQ 59.234375 2.09375 53.4375 0.328125 \nQ 47.65625 -1.421875 41.21875 -1.421875 \nQ 24.65625 -1.421875 15.125 8.703125 \nQ 5.609375 18.84375 5.609375 36.375 \nQ 5.609375 53.953125 15.125 64.078125 \nQ 24.65625 74.21875 41.21875 74.21875 \nQ 47.75 74.21875 53.53125 72.484375 \nQ 59.328125 70.75 64.40625 67.28125 \nz\n\" id=\"DejaVuSans-67\"/>\n      <path d=\"M 34.1875 63.1875 \nL 20.796875 26.90625 \nL 47.609375 26.90625 \nz\nM 28.609375 72.90625 \nL 39.796875 72.90625 \nL 67.578125 0 \nL 57.328125 0 \nL 50.6875 18.703125 \nL 17.828125 18.703125 \nL 11.1875 0 \nL 0.78125 0 \nz\n\" id=\"DejaVuSans-65\"/>\n      <path d=\"M 28.609375 0 \nL 0.78125 72.90625 \nL 11.078125 72.90625 \nL 34.1875 11.53125 \nL 57.328125 72.90625 \nL 67.578125 72.90625 \nL 39.796875 0 \nz\n\" id=\"DejaVuSans-86\"/>\n      <path id=\"DejaVuSans-32\"/>\n      <path d=\"M 53.515625 70.515625 \nL 53.515625 60.890625 \nQ 47.90625 63.578125 42.921875 64.890625 \nQ 37.9375 66.21875 33.296875 66.21875 \nQ 25.25 66.21875 20.875 63.09375 \nQ 16.5 59.96875 16.5 54.203125 \nQ 16.5 49.359375 19.40625 46.890625 \nQ 22.3125 44.4375 30.421875 42.921875 \nL 36.375 41.703125 \nQ 47.40625 39.59375 52.65625 34.296875 \nQ 57.90625 29 57.90625 20.125 \nQ 57.90625 9.515625 50.796875 4.046875 \nQ 43.703125 -1.421875 29.984375 -1.421875 \nQ 24.8125 -1.421875 18.96875 -0.25 \nQ 13.140625 0.921875 6.890625 3.21875 \nL 6.890625 13.375 \nQ 12.890625 10.015625 18.65625 8.296875 \nQ 24.421875 6.59375 29.984375 6.59375 \nQ 38.421875 6.59375 43.015625 9.90625 \nQ 47.609375 13.234375 47.609375 19.390625 \nQ 47.609375 24.75 44.3125 27.78125 \nQ 41.015625 30.8125 33.5 32.328125 \nL 27.484375 33.5 \nQ 16.453125 35.6875 11.515625 40.375 \nQ 6.59375 45.0625 6.59375 53.421875 \nQ 6.59375 63.09375 13.40625 68.65625 \nQ 20.21875 74.21875 32.171875 74.21875 \nQ 37.3125 74.21875 42.625 73.28125 \nQ 47.953125 72.359375 53.515625 70.515625 \nz\n\" id=\"DejaVuSans-83\"/>\n      <path d=\"M 48.78125 52.59375 \nL 48.78125 44.1875 \nQ 44.96875 46.296875 41.140625 47.34375 \nQ 37.3125 48.390625 33.40625 48.390625 \nQ 24.65625 48.390625 19.8125 42.84375 \nQ 14.984375 37.3125 14.984375 27.296875 \nQ 14.984375 17.28125 19.8125 11.734375 \nQ 24.65625 6.203125 33.40625 6.203125 \nQ 37.3125 6.203125 41.140625 7.25 \nQ 44.96875 8.296875 48.78125 10.40625 \nL 48.78125 2.09375 \nQ 45.015625 0.34375 40.984375 -0.53125 \nQ 36.96875 -1.421875 32.421875 -1.421875 \nQ 20.0625 -1.421875 12.78125 6.34375 \nQ 5.515625 14.109375 5.515625 27.296875 \nQ 5.515625 40.671875 12.859375 48.328125 \nQ 20.21875 56 33.015625 56 \nQ 37.15625 56 41.109375 55.140625 \nQ 45.0625 54.296875 48.78125 52.59375 \nz\n\" id=\"DejaVuSans-99\"/>\n     </defs>\n     <g transform=\"translate(14.798438 167.382969)rotate(-90)scale(0.1 -0.1)\">\n      <use xlink:href=\"#DejaVuSans-84\"/>\n      <use x=\"55.208984\" xlink:href=\"#DejaVuSans-67\"/>\n      <use x=\"125.033203\" xlink:href=\"#DejaVuSans-65\"/>\n      <use x=\"187.066406\" xlink:href=\"#DejaVuSans-86\"/>\n      <use x=\"255.474609\" xlink:href=\"#DejaVuSans-32\"/>\n      <use x=\"287.261719\" xlink:href=\"#DejaVuSans-83\"/>\n      <use x=\"350.738281\" xlink:href=\"#DejaVuSans-99\"/>\n      <use x=\"405.71875\" xlink:href=\"#DejaVuSans-111\"/>\n      <use x=\"466.900391\" xlink:href=\"#DejaVuSans-114\"/>\n      <use x=\"505.763672\" xlink:href=\"#DejaVuSans-101\"/>\n     </g>\n    </g>\n   </g>\n   <g id=\"LineCollection_1\">\n    <path clip-path=\"url(#p4693219472)\" d=\"M 93.322934 253.440728 \nL 93.322934 253.440728 \n\" style=\"fill:none;stroke:#000000;stroke-width:1.5;\"/>\n    <path clip-path=\"url(#p4693219472)\" d=\"M 225.434091 73.133239 \nL 225.434091 33.432411 \n\" style=\"fill:none;stroke:#000000;stroke-width:1.5;\"/>\n    <path clip-path=\"url(#p4693219472)\" d=\"M 357.545248 86.983842 \nL 357.545248 42.939727 \n\" style=\"fill:none;stroke:#000000;stroke-width:1.5;\"/>\n   </g>\n   <g id=\"patch_6\">\n    <path d=\"M 43.78125 255.718125 \nL 43.78125 22.318125 \n\" style=\"fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;\"/>\n   </g>\n   <g id=\"patch_7\">\n    <path d=\"M 407.086932 255.718125 \nL 407.086932 22.318125 \n\" style=\"fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;\"/>\n   </g>\n   <g id=\"patch_8\">\n    <path d=\"M 43.78125 255.718125 \nL 407.086932 255.718125 \n\" style=\"fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;\"/>\n   </g>\n   <g id=\"patch_9\">\n    <path d=\"M 43.78125 22.318125 \nL 407.086932 22.318125 \n\" style=\"fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;\"/>\n   </g>\n   <g id=\"text_11\">\n    <!-- * -->\n    <defs>\n     <path d=\"M 50.296875 59.421875 \nL 34.1875 50.984375 \nL 50.296875 42.484375 \nL 46.578125 35.59375 \nL 30.328125 44.578125 \nL 30.328125 27.78125 \nL 22.015625 27.78125 \nL 22.015625 44.578125 \nL 5.71875 35.59375 \nL 2 42.484375 \nL 18.3125 50.984375 \nL 2 59.421875 \nL 5.71875 66.3125 \nL 22.015625 57.421875 \nL 22.015625 74.21875 \nL 30.328125 74.21875 \nL 30.328125 57.421875 \nL 46.578125 66.3125 \nz\n\" id=\"DejaVuSans-Bold-42\"/>\n    </defs>\n    <g style=\"fill:#1f77b4;\" transform=\"translate(74.449911 253.440728)scale(0.16 -0.16)\">\n     <use xlink:href=\"#DejaVuSans-Bold-42\"/>\n    </g>\n   </g>\n   <g id=\"text_12\">\n    <!-- TCAV Scores for each concept and bottleneck -->\n    <defs>\n     <path d=\"M 37.109375 75.984375 \nL 37.109375 68.5 \nL 28.515625 68.5 \nQ 23.6875 68.5 21.796875 66.546875 \nQ 19.921875 64.59375 19.921875 59.515625 \nL 19.921875 54.6875 \nL 34.71875 54.6875 \nL 34.71875 47.703125 \nL 19.921875 47.703125 \nL 19.921875 0 \nL 10.890625 0 \nL 10.890625 47.703125 \nL 2.296875 47.703125 \nL 2.296875 54.6875 \nL 10.890625 54.6875 \nL 10.890625 58.5 \nQ 10.890625 67.625 15.140625 71.796875 \nQ 19.390625 75.984375 28.609375 75.984375 \nz\n\" id=\"DejaVuSans-102\"/>\n     <path d=\"M 54.890625 33.015625 \nL 54.890625 0 \nL 45.90625 0 \nL 45.90625 32.71875 \nQ 45.90625 40.484375 42.875 44.328125 \nQ 39.84375 48.1875 33.796875 48.1875 \nQ 26.515625 48.1875 22.3125 43.546875 \nQ 18.109375 38.921875 18.109375 30.90625 \nL 18.109375 0 \nL 9.078125 0 \nL 9.078125 75.984375 \nL 18.109375 75.984375 \nL 18.109375 46.1875 \nQ 21.34375 51.125 25.703125 53.5625 \nQ 30.078125 56 35.796875 56 \nQ 45.21875 56 50.046875 50.171875 \nQ 54.890625 44.34375 54.890625 33.015625 \nz\n\" id=\"DejaVuSans-104\"/>\n     <path d=\"M 54.890625 33.015625 \nL 54.890625 0 \nL 45.90625 0 \nL 45.90625 32.71875 \nQ 45.90625 40.484375 42.875 44.328125 \nQ 39.84375 48.1875 33.796875 48.1875 \nQ 26.515625 48.1875 22.3125 43.546875 \nQ 18.109375 38.921875 18.109375 30.90625 \nL 18.109375 0 \nL 9.078125 0 \nL 9.078125 54.6875 \nL 18.109375 54.6875 \nL 18.109375 46.1875 \nQ 21.34375 51.125 25.703125 53.5625 \nQ 30.078125 56 35.796875 56 \nQ 45.21875 56 50.046875 50.171875 \nQ 54.890625 44.34375 54.890625 33.015625 \nz\n\" id=\"DejaVuSans-110\"/>\n     <path d=\"M 48.6875 27.296875 \nQ 48.6875 37.203125 44.609375 42.84375 \nQ 40.53125 48.484375 33.40625 48.484375 \nQ 26.265625 48.484375 22.1875 42.84375 \nQ 18.109375 37.203125 18.109375 27.296875 \nQ 18.109375 17.390625 22.1875 11.75 \nQ 26.265625 6.109375 33.40625 6.109375 \nQ 40.53125 6.109375 44.609375 11.75 \nQ 48.6875 17.390625 48.6875 27.296875 \nz\nM 18.109375 46.390625 \nQ 20.953125 51.265625 25.265625 53.625 \nQ 29.59375 56 35.59375 56 \nQ 45.5625 56 51.78125 48.09375 \nQ 58.015625 40.1875 58.015625 27.296875 \nQ 58.015625 14.40625 51.78125 6.484375 \nQ 45.5625 -1.421875 35.59375 -1.421875 \nQ 29.59375 -1.421875 25.265625 0.953125 \nQ 20.953125 3.328125 18.109375 8.203125 \nL 18.109375 0 \nL 9.078125 0 \nL 9.078125 75.984375 \nL 18.109375 75.984375 \nz\n\" id=\"DejaVuSans-98\"/>\n     <path d=\"M 9.421875 75.984375 \nL 18.40625 75.984375 \nL 18.40625 0 \nL 9.421875 0 \nz\n\" id=\"DejaVuSans-108\"/>\n     <path d=\"M 9.078125 75.984375 \nL 18.109375 75.984375 \nL 18.109375 31.109375 \nL 44.921875 54.6875 \nL 56.390625 54.6875 \nL 27.390625 29.109375 \nL 57.625 0 \nL 45.90625 0 \nL 18.109375 26.703125 \nL 18.109375 0 \nL 9.078125 0 \nz\n\" id=\"DejaVuSans-107\"/>\n    </defs>\n    <g transform=\"translate(88.991278 16.318125)scale(0.12 -0.12)\">\n     <use xlink:href=\"#DejaVuSans-84\"/>\n     <use x=\"55.208984\" xlink:href=\"#DejaVuSans-67\"/>\n     <use x=\"125.033203\" xlink:href=\"#DejaVuSans-65\"/>\n     <use x=\"187.066406\" xlink:href=\"#DejaVuSans-86\"/>\n     <use x=\"255.474609\" xlink:href=\"#DejaVuSans-32\"/>\n     <use x=\"287.261719\" xlink:href=\"#DejaVuSans-83\"/>\n     <use x=\"350.738281\" xlink:href=\"#DejaVuSans-99\"/>\n     <use x=\"405.71875\" xlink:href=\"#DejaVuSans-111\"/>\n     <use x=\"466.900391\" xlink:href=\"#DejaVuSans-114\"/>\n     <use x=\"505.763672\" xlink:href=\"#DejaVuSans-101\"/>\n     <use x=\"567.287109\" xlink:href=\"#DejaVuSans-115\"/>\n     <use x=\"619.386719\" xlink:href=\"#DejaVuSans-32\"/>\n     <use x=\"651.173828\" xlink:href=\"#DejaVuSans-102\"/>\n     <use x=\"686.378906\" xlink:href=\"#DejaVuSans-111\"/>\n     <use x=\"747.560547\" xlink:href=\"#DejaVuSans-114\"/>\n     <use x=\"788.673828\" xlink:href=\"#DejaVuSans-32\"/>\n     <use x=\"820.460938\" xlink:href=\"#DejaVuSans-101\"/>\n     <use x=\"881.984375\" xlink:href=\"#DejaVuSans-97\"/>\n     <use x=\"943.263672\" xlink:href=\"#DejaVuSans-99\"/>\n     <use x=\"998.244141\" xlink:href=\"#DejaVuSans-104\"/>\n     <use x=\"1061.623047\" xlink:href=\"#DejaVuSans-32\"/>\n     <use x=\"1093.410156\" xlink:href=\"#DejaVuSans-99\"/>\n     <use x=\"1148.390625\" xlink:href=\"#DejaVuSans-111\"/>\n     <use x=\"1209.572266\" xlink:href=\"#DejaVuSans-110\"/>\n     <use x=\"1272.951172\" xlink:href=\"#DejaVuSans-99\"/>\n     <use x=\"1327.931641\" xlink:href=\"#DejaVuSans-101\"/>\n     <use x=\"1389.455078\" xlink:href=\"#DejaVuSans-112\"/>\n     <use x=\"1452.931641\" xlink:href=\"#DejaVuSans-116\"/>\n     <use x=\"1492.140625\" xlink:href=\"#DejaVuSans-32\"/>\n     <use x=\"1523.927734\" xlink:href=\"#DejaVuSans-97\"/>\n     <use x=\"1585.207031\" xlink:href=\"#DejaVuSans-110\"/>\n     <use x=\"1648.585938\" xlink:href=\"#DejaVuSans-100\"/>\n     <use x=\"1712.0625\" xlink:href=\"#DejaVuSans-32\"/>\n     <use x=\"1743.849609\" xlink:href=\"#DejaVuSans-98\"/>\n     <use x=\"1807.326172\" xlink:href=\"#DejaVuSans-111\"/>\n     <use x=\"1868.507812\" xlink:href=\"#DejaVuSans-116\"/>\n     <use x=\"1907.716797\" xlink:href=\"#DejaVuSans-116\"/>\n     <use x=\"1946.925781\" xlink:href=\"#DejaVuSans-108\"/>\n     <use x=\"1974.708984\" xlink:href=\"#DejaVuSans-101\"/>\n     <use x=\"2036.232422\" xlink:href=\"#DejaVuSans-110\"/>\n     <use x=\"2099.611328\" xlink:href=\"#DejaVuSans-101\"/>\n     <use x=\"2161.134766\" xlink:href=\"#DejaVuSans-99\"/>\n     <use x=\"2216.115234\" xlink:href=\"#DejaVuSans-107\"/>\n    </g>\n   </g>\n   <g id=\"legend_1\">\n    <g id=\"patch_10\">\n     <path d=\"M 325.599432 44.99625 \nL 400.086932 44.99625 \nQ 402.086932 44.99625 402.086932 42.99625 \nL 402.086932 29.318125 \nQ 402.086932 27.318125 400.086932 27.318125 \nL 325.599432 27.318125 \nQ 323.599432 27.318125 323.599432 29.318125 \nL 323.599432 42.99625 \nQ 323.599432 44.99625 325.599432 44.99625 \nz\n\" style=\"fill:#ffffff;opacity:0.8;stroke:#cccccc;stroke-linejoin:miter;\"/>\n    </g>\n    <g id=\"patch_11\">\n     <path d=\"M 327.599432 38.916562 \nL 347.599432 38.916562 \nL 347.599432 31.916562 \nL 327.599432 31.916562 \nz\n\" style=\"fill:#1f77b4;\"/>\n    </g>\n    <g id=\"text_13\">\n     <!-- mixed4c -->\n     <defs>\n      <path d=\"M 52 44.1875 \nQ 55.375 50.25 60.0625 53.125 \nQ 64.75 56 71.09375 56 \nQ 79.640625 56 84.28125 50.015625 \nQ 88.921875 44.046875 88.921875 33.015625 \nL 88.921875 0 \nL 79.890625 0 \nL 79.890625 32.71875 \nQ 79.890625 40.578125 77.09375 44.375 \nQ 74.3125 48.1875 68.609375 48.1875 \nQ 61.625 48.1875 57.5625 43.546875 \nQ 53.515625 38.921875 53.515625 30.90625 \nL 53.515625 0 \nL 44.484375 0 \nL 44.484375 32.71875 \nQ 44.484375 40.625 41.703125 44.40625 \nQ 38.921875 48.1875 33.109375 48.1875 \nQ 26.21875 48.1875 22.15625 43.53125 \nQ 18.109375 38.875 18.109375 30.90625 \nL 18.109375 0 \nL 9.078125 0 \nL 9.078125 54.6875 \nL 18.109375 54.6875 \nL 18.109375 46.1875 \nQ 21.1875 51.21875 25.484375 53.609375 \nQ 29.78125 56 35.6875 56 \nQ 41.65625 56 45.828125 52.96875 \nQ 50 49.953125 52 44.1875 \nz\n\" id=\"DejaVuSans-109\"/>\n      <path d=\"M 54.890625 54.6875 \nL 35.109375 28.078125 \nL 55.90625 0 \nL 45.3125 0 \nL 29.390625 21.484375 \nL 13.484375 0 \nL 2.875 0 \nL 24.125 28.609375 \nL 4.6875 54.6875 \nL 15.28125 54.6875 \nL 29.78125 35.203125 \nL 44.28125 54.6875 \nz\n\" id=\"DejaVuSans-120\"/>\n     </defs>\n     <g transform=\"translate(355.599432 38.916562)scale(0.1 -0.1)\">\n      <use xlink:href=\"#DejaVuSans-109\"/>\n      <use x=\"97.412109\" xlink:href=\"#DejaVuSans-105\"/>\n      <use x=\"125.195312\" xlink:href=\"#DejaVuSans-120\"/>\n      <use x=\"181.25\" xlink:href=\"#DejaVuSans-101\"/>\n      <use x=\"242.773438\" xlink:href=\"#DejaVuSans-100\"/>\n      <use x=\"306.25\" xlink:href=\"#DejaVuSans-52\"/>\n      <use x=\"369.873047\" xlink:href=\"#DejaVuSans-99\"/>\n     </g>\n    </g>\n   </g>\n  </g>\n </g>\n <defs>\n  <clipPath id=\"p4693219472\">\n   <rect height=\"233.4\" width=\"363.305682\" x=\"43.78125\" y=\"22.318125\"/>\n  </clipPath>\n </defs>\n</svg>\n",
            "image/png": "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\n"
          },
          "metadata": {
            "tags": [],
            "needs_background": "light"
          }
        }
      ]
    }
  ]
}