05_1_GNN_XAI(GNNExplainer).ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "05.1: GNN_XAI(GNNExplainer).ipynb",
      "provenance": [],
      "collapsed_sections": [
        "2sKRIEElgvHr",
        "U_9vH8Kcd_Ls",
        "Q7PRhFgs8guL",
        "n3pq7VnVf-Ln",
        "ggcoaQiDhpdU",
        "vyYKjY2JiQrX"
      ],
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/github/gulabpatel/Graph_Neural_Network/blob/main/05.1%3A%20GNN_XAI(GNNExplainer).ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "8y1rO9ggdybc"
      },
      "source": [
        "# How to explain Graph Neural Networks using GNNExplainer"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "408-5a1_gWqN"
      },
      "source": [
        "Video walkthrough the code : https://www.youtube.com/watch?v=aFnlmzFh4iQ"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "2sKRIEElgvHr"
      },
      "source": [
        "## Install Pytorch Geometric"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "KbFF_8P4gg2W",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "b518858b-400b-4662-a852-a1a8002baa19"
      },
      "source": [
        "import torch\n",
        "torch.manual_seed(42)\n",
        "from IPython.display import clear_output \n",
        "torch_version = torch.__version__\n",
        "print(\"Torch version: \", torch_version)\n",
        "pytorch_version = f\"torch-{torch.__version__}.html\"\n",
        "!pip install --no-index torch-scatter -f https://pytorch-geometric.com/whl/$pytorch_version\n",
        "!pip install --no-index torch-sparse -f https://pytorch-geometric.com/whl/$pytorch_version\n",
        "!pip install --no-index torch-cluster -f https://pytorch-geometric.com/whl/$pytorch_version\n",
        "!pip install --no-index torch-spline-conv -f https://pytorch-geometric.com/whl/$pytorch_version\n",
        "!pip install torch-geometric\n",
        "clear_output()\n",
        "print(\"Done.\")"
      ],
      "execution_count": 1,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Done.\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "U_9vH8Kcd_Ls"
      },
      "source": [
        "## Twitch Streamer Dataset\n",
        "\n",
        "- Twitch user-user networks of gamers who stream in a certain language\n",
        "- Nodes are the users themselves and the links are mutual friendships between them\n",
        "- These social networks were collected in May 2018"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "5s-QUsvneD7F",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "263c138b-15c4-4e10-c06f-2196cf75267e"
      },
      "source": [
        "from torch_geometric.datasets import Twitch\n",
        "# Dataset source: https://github.com/benedekrozemberczki/datasets#twitch-social-networks\n",
        "graph = Twitch(root=\".\", name=\"EN\")[0]\n",
        "graph"
      ],
      "execution_count": 2,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stderr",
          "text": [
            "Downloading https://graphmining.ai/datasets/ptg/twitch/EN.npz\n",
            "Processing...\n",
            "Done!\n"
          ]
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "Data(x=[7126, 128], edge_index=[2, 77774], y=[7126])"
            ]
          },
          "metadata": {},
          "execution_count": 2
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "UJFxP3DWQKMs"
      },
      "source": [
        "### The node features \n",
        "- Extracted based on the games played and liked, location and streaming habits\n",
        "- These are embeddings, which cannot be interpreted directly"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "ODF6RUtyoj_J",
        "outputId": "da6d6aef-4e02-468c-f0f7-04d8d2221d4f"
      },
      "source": [
        "graph.x"
      ],
      "execution_count": 3,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "tensor([[-0.2367, -0.2307, -0.1605,  ...,  1.3203, -0.2558, -0.1839],\n",
              "        [-0.2367, -0.2150, -0.1605,  ...,  1.3572, -0.2538, -0.1685],\n",
              "        [-0.2367, -0.2043, -0.0994,  ..., -0.6348, -0.2558, -0.1839],\n",
              "        ...,\n",
              "        [-0.2193, -0.2307, -0.1605,  ..., -0.6348, -0.2558, -0.1839],\n",
              "        [-0.2367, -0.2307, -0.1605,  ..., -0.6348, -0.2558, -0.1733],\n",
              "        [ 4.6088, -0.2307, -0.1605,  ...,  1.2370, -0.2558, -0.1839]])"
            ]
          },
          "metadata": {},
          "execution_count": 3
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "G9qdMzWB7gny"
      },
      "source": [
        "### A binary node classification task\n",
        "- Predict if a streamer uses explicit language"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "4InDfqFDol8V",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "714ba3eb-2529-4d54-a8a3-9e1ce54fa327"
      },
      "source": [
        "graph.y"
      ],
      "execution_count": 4,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "tensor([1, 0, 1,  ..., 0, 0, 1])"
            ]
          },
          "metadata": {},
          "execution_count": 4
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 333
        },
        "id": "w_Ra14AgKvO2",
        "outputId": "76171f31-1417-4a3e-cb03-b1f02b1092f3"
      },
      "source": [
        "import pandas as pd\n",
        "from pylab import rcParams\n",
        "rcParams['figure.figsize'] = 5, 5\n",
        "df = pd.DataFrame(graph.y.numpy(), columns=[\"explicit_language\"])\n",
        "df['explicit_language'].value_counts().plot(kind='bar')"
      ],
      "execution_count": 5,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<matplotlib.axes._subplots.AxesSubplot at 0x7fd602238610>"
            ]
          },
          "metadata": {},
          "execution_count": 5
        },
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 360x360 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Q7PRhFgs8guL"
      },
      "source": [
        "## A simple model"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "P9LXRORZ7xUy",
        "outputId": "248709df-31eb-4458-dccc-3ef98e50bd3e"
      },
      "source": [
        "from torch.nn import Linear\n",
        "import torch.nn.functional as F \n",
        "from torch_geometric.nn import GATConv\n",
        "embedding_size = 128\n",
        "\n",
        "class GNN(torch.nn.Module):\n",
        "    def __init__(self):\n",
        "        # Init parent\n",
        "        super(GNN, self).__init__()\n",
        "\n",
        "        # GCN layers\n",
        "        self.initial_conv = GATConv(graph.num_features, embedding_size)\n",
        "        self.conv1 = GATConv(embedding_size, embedding_size)\n",
        "\n",
        "        # Output layer\n",
        "        self.out = Linear(embedding_size, 1)\n",
        "\n",
        "    def forward(self, x, edge_index):\n",
        "        emb = F.relu(self.initial_conv(x, edge_index))\n",
        "        emb = F.relu(self.conv1(emb, edge_index))\n",
        "        return self.out(emb)\n",
        "\n",
        "model = GNN()\n",
        "print(model)\n",
        "print(\"Number of parameters: \", sum(p.numel() for p in model.parameters()))"
      ],
      "execution_count": 6,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "GNN(\n",
            "  (initial_conv): GATConv(128, 128, heads=1)\n",
            "  (conv1): GATConv(128, 128, heads=1)\n",
            "  (out): Linear(in_features=128, out_features=1, bias=True)\n",
            ")\n",
            "Number of parameters:  33665\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "wGHWi2qn_j7V"
      },
      "source": [
        "### Add train and test masks\n",
        "- Alternative: Transform + RandomNodeSplit"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "shTDGKlt_hsE",
        "outputId": "237563be-7cfd-4b5e-e525-69682b45f2e9"
      },
      "source": [
        "# Add train and test masks\n",
        "num_nodes = graph.x.shape[0]\n",
        "ones = torch.ones(num_nodes)\n",
        "ones[4000:] = 0\n",
        "graph.train_mask = ones.bool()\n",
        "graph.test_mask = ~graph.train_mask.bool()\n",
        "\n",
        "print(\"Train nodes: \", sum(graph.train_mask))\n",
        "print(\"Test nodes: \", sum(graph.test_mask))"
      ],
      "execution_count": 7,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Train nodes:  tensor(4000)\n",
            "Test nodes:  tensor(3126)\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "oqChlv_Pui86",
        "outputId": "5b9da535-fdc0-4475-f062-f88212fd7676"
      },
      "source": [
        "graph.train_mask"
      ],
      "execution_count": 8,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "tensor([ True,  True,  True,  ..., False, False, False])"
            ]
          },
          "metadata": {},
          "execution_count": 8
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "vF4d8N76uld-",
        "outputId": "f8b4558f-7dab-4286-8979-08a7cb0dee77"
      },
      "source": [
        "graph.test_mask"
      ],
      "execution_count": 9,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "tensor([False, False, False,  ...,  True,  True,  True])"
            ]
          },
          "metadata": {},
          "execution_count": 9
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "n3pq7VnVf-Ln"
      },
      "source": [
        "## Training\n",
        "\n",
        "- We train the model with the full graph, so batch size = 1\n",
        "- For large graphs its also possible to train in batches, by using the NeighborSampler\n",
        "- This is a transductive setup, as all nodes are used during training\n",
        "- Inductive training can be achieved using GraphSAGE\n"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "VpPB6t0A8EwC",
        "outputId": "fc670849-2ff4-447b-a0b3-96a0b24c90df"
      },
      "source": [
        "# Use GPU for training\n",
        "device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
        "model = GNN()\n",
        "model = model.to(device)\n",
        "graph = graph.to(device)\n",
        "\n",
        "# Loss function\n",
        "loss_fn = torch.nn.BCEWithLogitsLoss()\n",
        "optimizer = torch.optim.Adam(model.parameters(), lr=0.00001)  \n",
        "\n",
        "\n",
        "def train():\n",
        "      model.train()\n",
        "      optimizer.zero_grad() \n",
        "      out = model(graph.x, graph.edge_index)  \n",
        "      preds = out[graph.train_mask]\n",
        "      targets = torch.unsqueeze(graph.y[graph.train_mask], 1)\n",
        "      loss = loss_fn(preds.float(), targets.float())  \n",
        "      loss.backward() \n",
        "      optimizer.step()\n",
        "      return loss\n",
        "\n",
        "def test():\n",
        "      model.eval()\n",
        "      optimizer.zero_grad() \n",
        "      out = model(graph.x, graph.edge_index)  \n",
        "      preds = out[graph.test_mask]\n",
        "      targets = torch.unsqueeze(graph.y[graph.test_mask], 1)\n",
        "      loss = loss_fn(preds.float(), targets.float())  \n",
        "      return loss\n",
        "\n",
        "for epoch in range(0, 800):\n",
        "    tr_loss = train()\n",
        "    if epoch % 100 == 0:\n",
        "      loss = test()\n",
        "      print(f'Epoch: {epoch:03d}, Test loss: {loss:.4f} | Train loss: {tr_loss:.4f}')"
      ],
      "execution_count": 10,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Epoch: 000, Test loss: 0.6880 | Train loss: 0.6885\n",
            "Epoch: 100, Test loss: 0.6843 | Train loss: 0.6822\n",
            "Epoch: 200, Test loss: 0.6824 | Train loss: 0.6774\n",
            "Epoch: 300, Test loss: 0.6809 | Train loss: 0.6737\n",
            "Epoch: 400, Test loss: 0.6796 | Train loss: 0.6700\n",
            "Epoch: 500, Test loss: 0.6785 | Train loss: 0.6658\n",
            "Epoch: 600, Test loss: 0.6778 | Train loss: 0.6614\n",
            "Epoch: 700, Test loss: 0.6776 | Train loss: 0.6568\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 376
        },
        "id": "n8mkEZFBHy2U",
        "outputId": "a1e45400-528e-4568-ce23-51000277654c"
      },
      "source": [
        "from sklearn.metrics import roc_auc_score\n",
        "\n",
        "df = pd.DataFrame()\n",
        "# Model predictions'\n",
        "out = torch.sigmoid(model(graph.x, graph.edge_index))\n",
        "df[\"preds\"] = out[graph.test_mask].round().int().cpu().detach().numpy().squeeze()\n",
        "df[\"prob\"] = out[graph.test_mask].cpu().detach().numpy().squeeze().round(2)\n",
        "\n",
        "# Groundtruth\n",
        "df[\"gt\"] = graph.y[graph.test_mask].cpu().detach().numpy().squeeze()\n",
        "\n",
        "print(\"Test ROC: \", roc_auc_score(df[\"gt\"], df[\"preds\"]))\n",
        "df.head(10)"
      ],
      "execution_count": 11,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Test ROC:  0.5687391641761471\n"
          ]
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/html": [
              "<div>\n",
              "<style scoped>\n",
              "    .dataframe tbody tr th:only-of-type {\n",
              "        vertical-align: middle;\n",
              "    }\n",
              "\n",
              "    .dataframe tbody tr th {\n",
              "        vertical-align: top;\n",
              "    }\n",
              "\n",
              "    .dataframe thead th {\n",
              "        text-align: right;\n",
              "    }\n",
              "</style>\n",
              "<table border=\"1\" class=\"dataframe\">\n",
              "  <thead>\n",
              "    <tr style=\"text-align: right;\">\n",
              "      <th></th>\n",
              "      <th>preds</th>\n",
              "      <th>prob</th>\n",
              "      <th>gt</th>\n",
              "    </tr>\n",
              "  </thead>\n",
              "  <tbody>\n",
              "    <tr>\n",
              "      <th>0</th>\n",
              "      <td>0</td>\n",
              "      <td>0.49</td>\n",
              "      <td>0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>1</th>\n",
              "      <td>1</td>\n",
              "      <td>0.63</td>\n",
              "      <td>1</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>2</th>\n",
              "      <td>1</td>\n",
              "      <td>0.54</td>\n",
              "      <td>1</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>3</th>\n",
              "      <td>0</td>\n",
              "      <td>0.45</td>\n",
              "      <td>0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>4</th>\n",
              "      <td>1</td>\n",
              "      <td>0.57</td>\n",
              "      <td>1</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>5</th>\n",
              "      <td>0</td>\n",
              "      <td>0.49</td>\n",
              "      <td>0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>6</th>\n",
              "      <td>1</td>\n",
              "      <td>0.52</td>\n",
              "      <td>1</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>7</th>\n",
              "      <td>0</td>\n",
              "      <td>0.48</td>\n",
              "      <td>0</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>8</th>\n",
              "      <td>1</td>\n",
              "      <td>0.69</td>\n",
              "      <td>1</td>\n",
              "    </tr>\n",
              "    <tr>\n",
              "      <th>9</th>\n",
              "      <td>1</td>\n",
              "      <td>0.57</td>\n",
              "      <td>1</td>\n",
              "    </tr>\n",
              "  </tbody>\n",
              "</table>\n",
              "</div>"
            ],
            "text/plain": [
              "   preds  prob  gt\n",
              "0      0  0.49   0\n",
              "1      1  0.63   1\n",
              "2      1  0.54   1\n",
              "3      0  0.45   0\n",
              "4      1  0.57   1\n",
              "5      0  0.49   0\n",
              "6      1  0.52   1\n",
              "7      0  0.48   0\n",
              "8      1  0.69   1\n",
              "9      1  0.57   1"
            ]
          },
          "metadata": {},
          "execution_count": 11
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "ggcoaQiDhpdU"
      },
      "source": [
        "## Explaining the predictions"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "7WPJWBHfgomc",
        "outputId": "d611104c-4432-4a33-f053-3cd78c921966"
      },
      "source": [
        "from torch_geometric.nn import GNNExplainer\n",
        "# Initialize explainer\n",
        "explainer = GNNExplainer(model, epochs=200, return_type='log_prob')\n",
        "\n",
        "# Explain node\n",
        "node_idx = 7\n",
        "node_feat_mask, edge_mask = explainer.explain_node(node_idx, graph.x, graph.edge_index)\n",
        "print(\"Size of explanation: \", sum(edge_mask > 0))"
      ],
      "execution_count": 12,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stderr",
          "text": [
            "Explain node 7: 100%|██████████| 200/200 [00:02<00:00, 82.58it/s] \n"
          ]
        },
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Size of explanation:  tensor(25, device='cuda:0')\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "ttnxQWNRiSda",
        "outputId": "e2de96e0-2f76-4fcf-c3fa-15e6c3b23bdf"
      },
      "source": [
        "print(\"Node features: \", graph.x[node_idx])\n",
        "print(\"Node label: \",  df[\"gt\"][node_idx])\n",
        "print(\"Node prediction: \",  df[\"preds\"][node_idx])"
      ],
      "execution_count": 13,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Node features:  tensor([-0.2367, -0.1935,  3.7567, -0.2128, -0.2124,  5.4706, -0.5386, -0.4115,\n",
            "        -0.1622,  0.2116, -0.1529, -0.3459, -0.1957, -0.2217, -0.4750, -0.2006,\n",
            "        -0.4221, -0.2727, -0.4964, -0.2263, -0.2821, -0.2192, -0.2268, -0.2027,\n",
            "         1.0373, -0.0218, -0.2800, -0.2361, -0.2226, -0.2057, -0.3231, -0.1938,\n",
            "         3.7738, -0.2022, -0.5585,  4.2744, -0.0594, -0.3480, -0.3727, -0.4935,\n",
            "        -0.2020, -0.2183,  1.8572, -0.0779, -0.2263, -0.1528,  2.0653, -0.4668,\n",
            "        -0.2193, -0.1163,  6.3437, -0.2100, -0.6597, -0.2962, -0.2531,  1.8709,\n",
            "        -0.2680, -0.2187,  0.7300, -0.2951, -0.2064, -0.1886, -0.2112, -0.2915,\n",
            "        -0.2250,  3.5988, -0.3386,  4.3039, -0.2275,  0.1154, -0.2643,  2.2388,\n",
            "         0.1718, -0.2528,  2.2578, -0.2046,  3.3064, -0.4842,  3.8121, -0.2549,\n",
            "        -0.3419, -0.1695, -0.3224, -0.2822, -0.3064, -0.2185, -0.2841, -0.4582,\n",
            "        -0.2062, -0.2791, -0.2172, -0.5902, -0.6998, -0.1862, -0.1884, -0.1220,\n",
            "        -0.3154, -0.2393, -0.3001, -0.1737,  6.4014, -0.0901, -0.1913, -0.2501,\n",
            "         2.6771, -0.3410, -0.2074, -0.1444, -0.2569, -0.2624, -0.4364, -0.1613,\n",
            "        -0.2820, -0.1735, -0.2014,  3.4549, -0.2631,  1.9961, -0.1275, -0.2393,\n",
            "        -0.2141,  3.3722, -0.1754, -0.2139, -0.3749, -0.6348, -0.2558, -0.1742],\n",
            "       device='cuda:0')\n",
            "Node label:  0\n",
            "Node prediction:  0\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "vyYKjY2JiQrX"
      },
      "source": [
        "## Create visualizations"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "WdFHVeS-iu36",
        "outputId": "7b78ea7d-1c42-42cb-ebe2-206fb136881d"
      },
      "source": [
        "# Show shape of masks\n",
        "print(node_feat_mask.shape)\n",
        "print(edge_mask.shape)"
      ],
      "execution_count": 14,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "torch.Size([128])\n",
            "torch.Size([77774])\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "oU3uqczhklnV"
      },
      "source": [
        "### The size of the following plot depends on the depth of the GNN!\n",
        "- Colors are edge labels\n",
        "- Greyed-out edges have an impact, but very low (< 0.1)"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 575
        },
        "id": "DVknzGYRhry8",
        "outputId": "7a432cfd-1a68-49a2-864f-70e47a6adc63"
      },
      "source": [
        "import matplotlib.pyplot as plt\n",
        "from pylab import rcParams\n",
        "rcParams['figure.figsize'] = 10, 10\n",
        "\n",
        "# Visualize result\n",
        "ax, G = explainer.visualize_subgraph(node_idx, graph.edge_index, edge_mask, y=graph.y)\n",
        "plt.show()"
      ],
      "execution_count": 15,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "iVBORw0KGgoAAAANSUhEUgAAAjwAAAIuCAYAAAC7EdIKAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADh0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uMy4yLjIsIGh0dHA6Ly9tYXRwbG90bGliLm9yZy+WH4yJAAAgAElEQVR4nOzdd5hU1cHH8e/M9t6XpohYKDaKJhbAQu9dmoCggmJAiaLR5NWQvBGNih3UaFABGyWAgDRRIZqoYHwVBUKiVJVdlm0zszs75b5/XHbpbLszuzP7+/jsw+zMveecWXlmf5xqMwwDERERkXBmr+sGiIiIiASaAo+IiIiEPQUeERERCXsKPCIiIhL2FHhEREQk7CnwiIiISNiLPNOLmZmZRosWLYLUFBEREZGa27p16yHDMLJO9doZA0+LFi3YsmVLYFolIiIiYiGbzbbndK9pSEtERETCngKPiIiIhD0FHhEREQl7CjwiIiIS9hR4REREJOwp8IiIiEjYU+ARERGRsKfAIyIiImHvjBsPioSyHOCfwGfA50AxYAOygU7A5cAvgfi6aqCIiASNAo+EFQP4EHj8yJ8xgAPwn3DdGiAW8AHjgLuAVsFrpoiIBJmGtCRs7MXsuRmIGWjcQBEnhx2AsiOvOYFXgPbAVKAkKC0VEZFgU+CRsDAfaIs5dOWo5r0ezKDzKnAh8LW1TRMRkXpAgUdC3lPA7Zi9Nd5alFMC7AeuwZz3IyIi4UOBR0LaG8BvAZeFZTqA7sAOC8sUEZG6pcAjIWs3cAeBmXfjAIZSux4jERGpPxR4JCT5gZGYE5MDwcAMVI8GqHwREQkuBR4JSeuAbzGXlQeKC3gEKAxgHSIiEhwKPBKS/kz1V2PVhA1znpCIiIQ2BR4JOfuBf5R/U1oKv/gFXHYZXHQRPPyw+fwPP8Avfwnnnw8jRkBZmfn8nj3QtStceilcdx3s33+04F69IDUV+vWreMoFPBn4tyQiIgGmwCMh5xMgqvybmBjYuBH+7//gq69gzRr45z/h/vth+nT4z38gLQ1efdW8/t57Ydw4+PpreOgheOCBowXPmAHz559U30/A4QC/JxERCSwFHgk5/+CY4SybDRITzccej/lls5khaNgw8/nx42HZMvPxd9/BDTeYj6+/HpYvP1pw166QlHRSfbHAlwF4HyIiEjwKPBJyPsNcRVXB54N27SA7G7p3h/POM4emIo8cFXfWWXDggPn4sstg6VLz8d/+BsXFkJd3xvpK0e7LIiKhToFHQk7xiU9ERJjDWfv3w+efw44zbBn4xBPw8cfQvr35Z7Nm5v1nUHaqOkVEJKTotHQJObbTvZCaag5T/eMfUFAAXq/Zy7N/vxlsAJo2PdrD43DAkiXmfZXUp38ZiIiENn2OS8jJOPab3Fwz3ACUlMD69dCmjRl8Fi82n3/9dRg40Hx86BD4j5yfPmsWTJxYaX0xQJpVjRcRkTqhwCMh5xqgYhDqp5/McHPppXDFFeYcnn794LHHYPZsc1l6Xh7ccot5/UcfQatWcOGFcPAg/Pa3Rwvu3BmGD4cPPjDn/axdC0A00D54b09ERALAZhjGaV+8/PLLjS1btgSxOSKVWwGMBYqCVF8UkA8kBKk+ERGpGZvNttUwjMtP9Zp6eCTkdMacSBwsbVHYEREJdQo8EnLSgAEE5y9vEnB/EOoREZHAUuCRkDQDc0PAQLMDQ4JQj4iIBJYCj4Sky4H+mCuoAiUBmBPgOkREJDgUeCRkvUjg5tZEAZ2AUQEqX0REgkuBR0JWKrAciLe43AigETCfM2xyKCIiIUWBR0JaJ+BvWBd6ooAmwKdAlkVliohI3VPgkZDXA/gIaAbE1aKcBKALsBU4u/bNEhGRekSBR8LCFcAu4FbM1VvVWcGViLn8fA6wHsi2vHUiIlLXFHgkbMQBzwI/YO6dk+73E5efTzLH/0WPAlIwj4xoAzwPHATGoTk7IiLhSqelS9hpDPwe+I3Hw3a/n++BbUAh5oTkpKIirkpOpgMnHEQqIiJhS4FHwpanrIzzoqJoDww98pzX6+XbH37g3HPPJTk5uS6bJyIiQaQhLQlbZWVlREdHH/dcSUkJcXFxFBYW4vP56qhlIiISbAo8ErY8Hs9JgcflcpGcnFwRekREpGFQ4JGw5PV6sdvt2O1H/4q73W7sdjsxMTHExcXh9Xpxu9112EoREQkWzeGRsFRWVkZUVNRxzxmGQVJSEmVlZRiGQUZGBjab1mWJiDQECjwSlk41fyc21tydx+v14vf7iYiIqIumiYhIHdCQloSlUwWecna7Hb/fH+QWiYhIXVLgkbBjGAY+n4/IyFN3YNrtdq3QEhFpYBR4JOx4vV4iIyNPOz/HbrdjGEaQWyUiInVJgUfCjsfjOWnC8rE0pCUi0vAo8EjYqSzw2Gw29fCIiDQwCjwSdqoSeNTDIyLSsCjwSNipypCWenhERBoWBR4JK16vl4iIiEo3FNSwlohIw6LAI2Glst6dcpq4LCLSsCjwSFjxeDyn3X/nWOrhERFpWBR4JKyoh0dERE5FZ2kBucBW4N9ACeYPpRHQEbgQ0IlLoaOqgUc9PCIiDUuDDTwHgZeBF4B8IA4oBXyADYgFDMAD9AJmAFcfeU3qp/LjIqpyKKiWpouINCwNLvA4gHuAN458X3rkz7ITrvMc83gFsAE4C1iI2fMj9U9Ve3dAS9NFRBqaBjWHZxNwAWbYKeVo2KmMATiBnUBn4D6OD0RSP1Qn8KiHR0SkYWkwgecdzKGpn6l60DmVEsxhsN61LEespx4eERE5nQYReFYAEzDDihVcwKfAQMBrUZlSe+WnpFeFJi2LiDQsYR949gCjsC7slCsB/g48anG5UnM+n0+BR0RETimsA48BjObkCclWcQGPAN8GqHypOq/Xi91ur/RIiXIKPCIiDUtYB54lwP8R2GGnUuCWAJYvVVOd3h0REWl4wjrwzMJcXRVIBvA1sD3A9ciZVWf+joiINDxhG3i+5ZgQMnEiZGfDxRcfveDwYejeHS64wPwzP998Pj8fBg+GSy+FX/wCtm07es+pysFcov50AN+LVK66gUdDWiIiDUvYBp73MXdNBuDmm2HNmuMvePRR6NoVdu0y/3z0yPTjRx6Bdu3g66/hjTfgrruO3nOqcjCHzFZY/g6kOtTDIyIiZxK2gedjjpms3KULpKcff8Hy5TB+vPl4/HhYtsx8/N13cMMN5uPWrWH3bjh48PTlHJEHFFrXfKmmmgQe9fCIiDQcYRt4vqzsgoMHoUkT83HjxkdDzWWXwdKl5uPPP4c9e2D//krriwf+VdPGSq0YhoHf76/SGVrlqrqaS0REwkPYBp6i6lxss5lfAL/5DRQUmMNazz0H7dtDFX6RGpi9PBJ8Xq+3WmFHREQanrCd9FDpKUmNGsFPP5m9PD/9ZE5GBkhOhnnzzMeGAeeeCy1bVlqfwTFzhiSoajKcpUnLIiINS9j28MRUdsGAAfD66+bj11+HgQPNxwUFUHZk9s8rr5jzdpKTK63PBiTUtLFSK5qwLCIilQnbwNPq2G9GjYKrroKdO+Gss+DVV82hq/XrzWXpGzaY3wNs324uO2/VCt5/H5555szlHOEBjl+sLsGiwCMiIpUJ298SXYDPOTK09dZbp77ogw9Ofu6qq+Df/z719acrBzM5Nq9eE8UiNdllWUNaIiINS9j28FxHcIeYfoE5rCXBpx4eERGpTNgGnh5UYR6PRZKAXwepLjmeYRgYhoHdHrZ/lUVExAJh+1siApgGxAahrligdxDqkZP5fL4aLUnXkJaISMMStoEHYCrmhoCBlAA8gRmwJPhqGngMw9DmgyIiDUhYT3xIBeYDwwFXAMqPAn7p9zMtPZ3ZLVrQsWNHzj33XM4991wGDRpEQoIWqgeaAo+IiFRFWPfwAPQBRmF9T48NSAYW2O107dqVH374gR9//BGfz8fmzZvZt2+fxTXKqfh8Ps3fERGRSjWI3xQvYU5itir02DF7jzYDTYDnnnuO1NRUMjMzWbZsGbfccgutW7e2qDY5k+qeoVVOPTwiIg1Lgwg8EcBiYDy1Dz0xa9aQPHMmnwNtjjzXtGlTJk+eTFxcHDNnzmTUqFE88sgjeDyeWtYmldGQloiIVEWDCDxghp45wEqgEdXfoycSiAMmX301Ld97jy/fffe41++66y62b99Oz5492bx5M1u2bKFHjx58//33VjRfTqOmgUdERBqWBhN4yl0PfA88CbTEDD5xp7k2AnOeTjxwC/Av4JnkZOa/8QYPPvgg27dvr7g2ISGBzZs3ExMTQ5MmTViyZAkjRoygW7du/O1vfwvkW2rQ1MMjIiJVYTvTXiSXX365sWXLliA2J7gMzBDzT8z5ON8AJZirrxoD1wJXYO7anHjCve+88w5PPPEEGzduJCkp6bR1fPXVV9x8881cf/31zJo1i9jYYOwM1DAYhsHPP/9MkyZNqn2v0+nE6/WSkpISgJaJiEhdsNlsWw3DuPyUrzXkwFNbv/71r8nLy+O11147Y29BUVER06ZNY9euXbz++uucf/75QWxl+PJ6veTl5dGoUaNq3+twOPD5fAo89YUB7AC2AP8A/gOUYXa/XgL8EvP8lrPrqoEiEgrOFHga3JCWlWbNmsW+ffuYM2fOGa9LTk5m3rx53HLLLfTo0YN3T5j/IzVT0xVa5TSkVQ8UAy8A5wKXA1OAucB64GNgDebOnhOBC49cswjQegARqSYFnlqIiYnhtdde45lnnuGTTz4547U2m42JEyeyfPlyHn30UaZNm0ZJSUmQWhqeajNhWXN46oFFwFnA/cAezN1BHae4zgCKgFJgK2b4OQ/4LDjNFJHwoMBTS82bN+fFF19k4sSJ7N27t9LrL7nkEjZt2oTL5eKGG25g586dQWhleKpt4JE6Ugz0AyZgBhlnNe93APswVyDcC/gtbZ2IhCkFHgvccMMN3H333YwePRqns/JP78TERP7yl79wxx130Lt3b958880gtDL81HZJunp46kABcCWwgeoHnROVYA5/3Qj4almWiIQ9BR6L3H777bRr147bb78dv7/yf3LabDbGjRvHypUreeqpp7jjjjtwuQJx4lf48vv9NT5WQkNadcCN2SvznyOPreAC3gdutag8EQlbCjwWsdlsPPnkkxw8eJA///nPVb6vbdu2fPzxxwBcd9117NixI1BNDDsKPCHmf4B/Y66+spILeBfQdlcicgYKPBaKiYlhwYIFLFiwgBUrVlT5vvj4eObOncu0adPo27cvq1atCmArw4cCTwjZCjyPGU4CwYU5J+hQgMoXkZCnwGOx7OxsFi5cyK9//Wu2bdtWrXtvuukmFi1axL333stjjz1WpaGxhszv99c4tNQmLEkN3Ik55yaQSoBHAlyHiIQsfeIHwGWXXcaf//xnRo8eTW5ubrXu7dChAx9++CEbNmxg3LhxVZoE3VAZhlHj0KLAE0Q7ga+DUE8Z8Arm8nURkRPoEz9AhgwZwo033sj48eMpK6vepIXGjRuzcuVK0tLS6NatG7t37w5MI0NcbUJLbXqHpJqeB7xHv/Xhoz3t6Uc/AMYwhla04mIuZiIT8Zywq+AXfEEkkSxmccVzr/M6Fxz573VeP3qxASwN4HsRkZClwBNADz74IGlpadx3333V3vclJiaGZ599lokTJ9K9e3c++uijwDQyRNW2h6Y2vUNSTWs4bmfkZ3iGNrSp+H4MY9jBDr7hG0oo4RVeqXjNh4/7uZ8e9Kh47jCHmclMPuMzPudzZjKTfPLNFx1H6hMROYE+8QPIbrfz0ksv8dlnn/Hqq69W+36bzcZtt93GvHnzmDRpEnPmzNGGeUfUtodGQ1pB4sbcRfmI/exnFau49Zh15H3og+3If7/gF+xnf8Vrz/EcQxlKNtkVz61lLd3pTjrppJFGd7qz5tiU849AviERCVX6xA+wxMRE3nrrLR5//HE2bdpUozI6derEBx98wMKFC5kyZQqlpZqkoPk7IeLfQOzRb+/mbv7Mn7Gf4qPHg4f5zKcXvQA4wAH+xt+4gzuOu+4ABzj7mFNEz+IsDnDg6AU/YA5tiYgcQ5/6QdCiRQteeeUVbr31Vv7zn//UqIyzzz6b9evXU1JSQp8+ffjxxx8tbmVo0fydEFFExafMSlaSTTYd6XjKS6cwhS50oTOdATMcPcZjpwxHlbJqY0MRCRsKPEHSuXNnHnroIYYNG0ZOTk6NyoiPj2fevHn069ePG264gc8//9ziVoaO2u7Box6eIDkmV37CJ6xgBS1owUhGspGN3MRNAMxkJrnkMpvZFddvYQsjGUkLWrCYxUxhCstYRjOasY99FdftZz/NaHa0IgN9sonISWxnmhNy+eWXG1u2bAlic8LfY489xvvvv8/KlStJTEyscTlr165lypQpPPLII4wYMcLCFoYGp9OJx+MhNTW12ve63W4cDgcZGRkBaJkcZzvwC046Bf0jPuIJnmAlK3mFV/grf+UDPiCOuFMWczM3049+DGMYhzlMRzryJV8C0IEObGUr6aSbF0dh9vCoE0+kwbHZbFsNw7j8VK/p30FBdt9993HZZZdx88034/F4Kr/hNHr27MmqVav44x//yOzZsxvcZGbN4QkRF1Dp8NLt3M5BDnIVV9GOdvyBP5zx+nTS+R/+hyuO/PcQDx0NOwDno7AjIidRD08d8Hq9jBkzhszMTJ5//vlazSf56aefGD58OFdccQWPP/44kZGRFra0/ioqKsJut9eol6w2vUNSA20xe3qC5VbgL0GsT0TqDfXw1DORkZHMmzePHTt28MgjtdsLv0mTJqxZs4bdu3dz0003NZgT1zWHJ3QY/Q180b7gVJYE9A1OVSISWvSpX0fi4+N5++23Wbp0KfPmzatVWYmJibz77rukp6fTt2/fah9nEYp0jlb95/f7KS4u5uDwgzhtQToiJQqObOAsInIcferXoaysLBYtWlQxkbk2oqKieOGFF+jRowfdunWr8fL3UKLAUz95PB4KCgrIycnB5/OR2S6T5C7Jgf+0iQOmAg1jVFdEqkmf+nWsZcuWvPnmm0ydOpXazpey2Ww88MAD3HPPPfTu3Tusl60bhqHAU8+43W7y8vI4fPgwERERZGdnk5qaas4rex6ICXADUoAZAa5DREKWPvXrgQ4dOjBnzhzGjBljSc/MuHHjmDNnDqNGjeK9996zoIX1T20Cj8/nIyIiwuIWNVwlJSXk5uZSVFREXFwc2dnZJCUlHR8qLwT+ACQEqBHxwFsBLF9EQp4CTz3Ro0cPfvvb3zJ8+PAab0x4rO7du7NkyRJmzJjByy+/bEELw4cCT+0ZhoHL5SInJwen00lSUhJZWVnEx8efNoh6p3rxXunlNFvt1Fw8cDdwncXlikhY0Wh3PTJu3Dh+/PFHRowYwXvvvVerjQkB2rVrx7p16xgyZAj79u1j5syZYTOUU9MeHsMwtEqrFsqDjsPhIDIykpSUFGJiTj1WZRgGbrcbt9uNy+Xi0KFDZL2SReOJjeGfQIkFDYoHJgH/a0FZIhLW9Klfz9x///20a9eOUaNGUVJS+98IzZs3Z/369XzxxRfcdttttdrssD6paeDR/J2aqVhxdfAgbreb9PR0MjIyTht2wNzvyOl04vf78fl8xMfHk3lWJqwFJlC7np4IzLDzGPAU2mhQRCqlT/56xmaz8cQTT9CkSRNuuukm3O7an4KYlpbGsmXLKC4uZvz48ZaUWddqurO0hrOqx+/3U1RUdHTFVWYm6enpREVFVXpvYmIiCQkJuN1uEhMTSUlJMScwRwEvAOuBszD3zqkqO+Y8nQ7AN8Cvqv+eRKRhUuCphyIiIpgzZw5JSUlMmDDBkl6Z2NhYFixYQGRkJCNHjrSk96iu1aSHR4Gnanw+H4WFheTk5GAYBllZWUdXXFWR0+mksLCQ9PR0fD4fCQknzCi+BtgNLDzyOApIBqJPKCjumOcHAxuAz4CWNXprItJA6WiJeqysrIzx48cTFxfHyy+/bMmxEV6vlzvvvJO9e/fy7rvvkpRUnX9e1x8///wz2dnZ1R6ecjgc+P1+kpOTA9Sy0Ob3+3E4HLhcLuLj40lMTKzREGBhYSFut5uMjAwiIiLIy8sjPT39zCG1EPgXsBX4ASjF7M1pDXQELgViq/+eRKThONPREgo89Zzb7WbUqFFkZ2czZ84cS+af+P1+7rnnHv7v//6PJUuWkJaWZkFLg+unn36iUaNG1f55FBYWEhERUesJ4eHG7/dXzLmJi4s7eVl5NcrJz88HzKFUzZcSkWDSWVohLCYmhoULF7Jv3z7uueceS05Ft9vtzJ49m6uuuop+/fqF7FEUGtKqPcMwcDgcFXN0srKySElJqVFQ8Xq95ObmEhUVRUZGhsKOiNQr+kQKAXFxcbzzzjts27aNBx54wJLQY7PZ+N///V/69u1Lnz59+PHHHy1oafDUZpWWAo/583M6neTk5ODxeMjMzCQ1NbXGP5vS0lIOHTpEcnKyhgtFpF5S4AkRiYmJLF68mH/+85/84Q9/sCz0PPjgg4wZM4bevXuzZ88eC1oaeLV57+rhMXdGzsnJobS0lPT0dNLS0mo1P6y4uJjCwkIyMjKIi7N6V0EREWto48EQkpKSwtKlS+nXrx9xcXHcd999lpR79913Ex8fT+/evVmxYgXnn3++JeUGks7Rqr7S0lKKioqw2+2kpaURHX3icqjqMQyD/Px8/H4/WVlZDfbnKiKhQYEnxKSnp7Ns2TL69etHbGws06ZNs6TcSZMmERcXR9++fVm6dCkXXXSRJeUGQk2Hs3w+H3a7vcZhKVSVlZVRVFSEYRgkJycTG1v7pU4+n4+8vDyio6NJS0trcD9TEQk9CjwhKDs7m+XLl9OnTx9iY2OZNGmSJeWOHTuW+Ph4Bg0axLvvvkv79u0tKbe+KA88DYXP56O4uBi3201SUhLx8fGWlOt2u8nPzycpKenkvXVEROopBZ4Q1aRJE5YvX07//v3xer1MmTLFknKHDh1KbGwsw4cPZ+nSpVx66aWWlGslm81Wo3k8DWX+TvmEZIfDQXx8PNnZ2Zb1wDidToqLi0lLSzvjsRIiIvWNAk8Ia968OatXr2bgwIE4nU7uvfdeS36x9e3bF5/Px7Bhw1i+fDlt2rSxoLV1z+v1WrJ5Y31WUlJCUVERUVFRZGVlWRbwDMOgoKAAr9drabkiIsES3p/+DUCzZs1YvXo1gwYNwuVy8dBDD1kSegYMGEBpaSmDBw9m5cqVITGRuTJerzdseyU8Hg+FhYUYhkFqaqql79Pj8ZCfn090dDSZmZmaryMiIUmBJwxkZ2ezcuVKhg4disvlYtasWZbMVbnxxhspLS1lwIABvP/++5xzzjkWtLb2ajOkFW49POWHe5aWlgZkTk35EFZKSoqWnItISGs4MzjDXHp6OsuXL+err77irrvuwufzWVLuuHHjmD59Ov379w+5zQlPFE5DWsfukGy328nOzrY07JQvOXe5XGRmZirsiEjIU+AJI8nJySxdupS9e/cyefJkS05ZB7jtttu49dZb6d+/PwcPHrSkzGDz+/0AYbFKq6ysjNzcXNxuN5mZmSQnJ1v6vjweD7m5udjtdjIzM8MmJIpIwxb6n/5ynISEBN555x2Ki4u5+eabcbvdlpQ7bdo0RowYwcCBA8nLy7OkzNqo7rBWOPTu+P1+CgoKyM/PJzk5mYyMDMvfk9PpJC8vj6SkJFJSUjRfR0TChgJPGIqNjWX+/PlERUUxevRoSkpKLCl3xowZ9O7dm4EDB1JQUGBJmcES6oGnpKSE3NxcbDYbWVlZlmweeKzyU841hCUi4UqBJ0xFR0fzyiuvkJWVxbBhw3A4HLUu02az8dBDD9GpUyeGDh1qSZm10RB6eMp3NC7f+6amJ5mficfj4dChQxrCEpGwpsATxiIjI5kzZw6tWrVi0KBBlvTK2Gw2Zs2axSWXXMLw4cNxuVwWtLRm7agOr9cbcnvHOBwOcnNziYmJISsrq9ZnX52uDg1hiUhDoMAT5ux2O08++SRXXnkl/fr14+eff651mTabjdmzZ9OiRQtGjRpFaWmpBS0NrFDq4SmfNFw+KTkxMdHyIOLz+Th06BClpaVkZWVpCEtEwp4CTwNgs9n44x//yNChQ+nRowc7duyodZl2u53nn3+e1NRUbrvtNsuWwVdVdScth8IePIZhUFhYSF5eHomJiQGZlAzgcrnIzc0lNjaWzMzMkOv5EhGpCQWeBsJmszF9+nR+97vfMWDAAD755JNalxkREcHLL7/M4cOHeeCBB2q0GWAwhMIp6W63m5ycHAzDIDs7OyA9Ln6/n8OHD+N0OsnIyCAxMdHyOkRE6isFngbmxhtv5C9/+Qs333wzS5curXV5MTExvPnmm2zevJlnn33WghZarz4PZ5X36hQUFJCamkpqampA9goqLS0lJyeHyMhIMjMziYqKsrwOEZH6rH7+FpCAuvbaa1m+fDk33ngj+/fvZ+rUqbXq/UhJSWHJkiV069aNpk2bMnz4cAtbe2rVGdKqr4GnrKyMgoKCioM+AxF0/H4/hYWFeDwe0tPTAzLxWUQkFNS/3wISFG3btmXdunUMGzaM/fv3M2vWrFrN5WjatCmLFy+mf//+ZGdnc+2111rY2pPZ7faK3ZMrU98Cj2EYFBcX43K5SE1NtXxPnXJut5uCggJiY2PJysqq10N6IiKBpiGtBqxp06asWbOGnTt3Mn78+FpvUNi2bVtef/11Jk6cyLZt2yxq5alVp4fH4/HUmyGc8j1vvF4v2dnZAQk7Jw6Tabm5iIgCT4OXnJzMokWLiI+PZ8CAARw6dKhW5XXq1IknnniC4cOHs3//fotaeTKbzVblHh6Px1MvenjK97xJSEggPT09IENYbreb3Nxc/H4/WVlZxMTEWF6HiEgoUuARoqOjeemll+jcuTM9e/bkhx9+qFV5gwcPZurUqQwZMiRgR1DY7fYq9fB4vV7sdnudHhp64p438fHxltdRfjREQUEBycnJpKWlhcVBqSIiVtEnogBHj4341a9+Ra9evdi6dWutypsyZQrdu3cP2MaEVR3SquvhrGDseeNyucjJySEiIiJgw2QiIqFOgaR/0G8AACAASURBVEeOM2HCBJ599llGjBjBe++9V6uy/vjHP9K4cWMmTZpU5eGnqqrqpOW6CjzH7nlTvluy1bxeL3l5eRX76iQnJ2uujojIaSjwyEl69uzJ4sWLeeCBB5g1a1aNw4rdbufFF18kLy/P8o0J63MPT1lZGbm5uURERATkMM7yVV6HDh2qOGervkzKFhGprxR45JTatWvHxo0b+fjjjxk7dmyNT0Yv35hw06ZNzJkzx7L21dceHofDweHDh0lJSQnI6qjyMOXxeMjKytJuySIiVaTAI6eVnZ3NihUryM7Oplu3bnz//fc1KiclJYVFixbx7LPPsmHDBkvaVpUenvLzvYJxVlT5EFb5xGSr59GUbyCYn59PUlIS6enpOgNLRKQaFHjkjKKjo3nqqaeYPHkyvXr1YuPGjTUq56yzzuL1119n8uTJ7Nq1q9btqsqy9GD17pT3ukRGRpKRkWF5ECkpKSE3NxdAJ5uLiNSQAo9UyYQJE3jttdeYMmUKzz//fI3m41x55ZU8/PDDjBw5ksLCwlq1pyrL0oMReI4dwrJ60nBZWRmHDh3C4XCQlpZGSkqKlpqLiNSQPj2lyq6++mo2bNjAokWLuP3222u03HzcuHF069aNCRMmVAw51URd9/AEcgjL5/ORn59Pfn4+8fHxZGVl6QwsEZFaUuCRajnrrLNYs2YNXq+X3r178+OPP1a7jD/96U/4fD4efvjhGrejLnt4AjWEVb76qnyFV3Z2dkA2KRQRaYgUeKTa4uLieOWVVxg0aBBdu3bls88+q9b9kZGRvPbaa6xatYq33nqrRm2obNKy3+/H7/dbviQ8UENYJSUl5OTkVKy+0p46IiLWUuCRGrHZbNx1110899xz3HTTTbz++uvVuj8tLY233nqL3/3ud3zxxRfVrr+yZelW9+4YhkF+fj4lJSWWDmGVHybqcDhITU3V6isRkQBR4JFa6datG++//z5z5szhzjvvxOVyVfne1q1b88ILLzB27NgaDY2dKfRYGXjKz8ICLDsewu/3U1BQQF5eXsU8HR30KSISOAo8Umvnn38+GzduxO/3c8MNN7Bjx44q39urVy8mT57M6NGjKSkpqVa9wQg85Sul4uLiSEtLq/UwU/k8nZycHOx2u+bpiIgEiQKPWCIhIYG5c+cybdo0+vXrx4IFC6q8dP3uu+/mvPPOY+rUqVW6pzzk2O12fD4fJSUlJwWfsrKyWq9scrlcFfN1arujsWEYOJ1ODh48iNfrJTMzk+TkZC0zFxEJEn3aiqVGjx7NqlWreOGFF5g8eTJOp7PSe2w2G88//zy7du3imWeeqfT64uJiiouLiYiIwOfznbSnj8/nwzCMGk9YNgyDwsJCHA4HmZmZtZ6vU36audvtJiMjg7S0NMsnU4uIyJkp8IjlWrVqxcaNG4mJieG6667j22+/rfSeuLg43nrrLV588UXWrVt3xmsTExNxOp34/X6cTifR0dHH9ZR4PJ4a9+6U769T3gtTm2BSvvKqpKSEtLQ00tPTdciniEgdUeCRgIiLi+O5555jxowZDBw4kHnz5lU6XNW0aVPmzZvHlClT2Lt372mvi4iIIDExkeLiYpxO50lzYMrKymoULDweD7m5uURFRZGRkVHj4abS0lJyc3NxOBykpKSQkZGhjQNFROqYAo8E1I033sj777/Pq6++yq233kpxcfEZr7/qqquYPn06Y8eOPeNOzuVzajwez0lDTjWZv1NaWkpeXh7JyckkJydX695j6z106BBFRUUkJSVp5ZWISD2iwCMBd8EFF7BhwwZSUlK47rrr+Prrr894/ZQpUzjnnHO4//77T3+RDxrnNybrkyx4DXgDWAnGAaNKQ1pOp7PiaIvi4mIKCwvJyMio0cGcHo+HvLy8iqMgsrOzLT8tXUREakeBR4IiNjaW2bNn89vf/pahQ4fyyiuvnHaIy2az8cILL/D3v/+dN9988+gLHmAJcBWQALFdYkmdlgq/AqYAY8B/np+oVlHYRtrgNPsZ+v3+ip6mgoICSktLyczMrPYwmNvtJi8vj8OHDxMbG6sl5iIi9ZjtTPMqLr/8cmPLli1BbI40BN9//z0TJ06kadOmPPvss2RmZrJlyxZiYmK45JJLKq777rvv6NevHyuWr+Dizy6G+wAvcOZRMZMdiAPOAeYBvzj6ktPprBgus9ls1d5fp7S0FIfDgd/vJzExkbi4OB0DISJSD9hstq2GYVx+qtfUwyNB17JlS9auXUurVq3o1KkT69atIycnh9tvv/24E9Tbtm3LozMeZcLVE+BuIJ+qhR0AP+AEvgOuA+4B3OZLRUVFuFwuDMMgOjq60nlF5cpXXRUXF5OQkFDRo6OwIyJS/6mHR+rUp59+yu23307Xrl359ttvGTt2LGPHjjVf3AZcC98WfstFvotqV1EccBm4lrvY9dMukpKSSExMJCYmhtjY2NNOLjYMA5fLhcPhICIigqSkJE1EFhGpp87Uw6PAI3Vq9+7d5OfnM3fuXDZv3kxpaSnbtm0j4UAC/BIoBKq2YXPlYsFziQfXey4SMxPPeCZW+c7ITqeTqKgoEhMTtbRcRKSeO1Pg0XavUqcWLlzIkiVLOHDgALGxsRw4cIBpd07j1U2vWht2AEohalsUKQ+nwIunvqR8M0On00lMTIw2CxQRCRPq4ZF6wev1sn//fr788kuyF2XTaUUnqPrB69UTB6wFOh99yuPxVExmjo2NJTExUcc/iIiEGPXwSL0XGRlJixYtaFHWAm4CqndwevWUAGOA3VBaVorT6cTr9VbsoaMDPUVEwo8Cj9QvT2MuPQ+wssNlFCwuwHa9jcTERGJjY7XaSkQkjOmfslJ/uIDXAQ/sYx/Xcz1tactFXMQzmKeoj2AE7Y7814IWtKMdAB48jGc8l3AJbWjDLGYBsJOdFde3ox3JJPM0T2N32kl9MZWsrCztoyMi0gCoh0fqj81U/I2MJJIneZIOdKCYYjrSke505x3eqbj8Hu4hhRQAFrEIN26+4RtcuGhLW0Yxila04iu+AsCHj2Y0YzCDiSQS/g6UAVp8JSIS9tTDI/XH51RMVG5CEzrQAYAkkmhDGw5woOJSA4N3eZdRjALAhg0nTrx4KaGEaKJJ5vhDQD/gA87jPM7hHPOJWMy9fkREJOwp8Ej9sYlTzt/ZzW7+xb/4Jb+seG4zm2lEIy7gAgCGMYwEEmhCE5rTnHu5l3TSjyvnbd6uCEiAuRvzlwF4HyIiUu9oSEssV1ZWhsvlqtjPpvzL5XJRWlqK1+vF6/Xi8/mO+9O+3c4oRhHH0RPLHTgYylCe5unjemze4q3jwsvnfE4EEfzIj+STT2c6041utKSl2SbKWMGKirk9gLlaKzfgPw4REakHFHjkJF6vl7y8PA4dOlTxlZubW/Fnfn7+cYHmxHADkJCQQEJCAvHx8cc9jouLIyIigsjIyIqv8u/jSuPw4KkIPB48DGUoYxjDEIYcbR9elrKUrWyteO5N3qQXvYgiimyyuYZr2MKWisDzPu/TgQ40otHRN+rHPIFdRETCngJPA1JUVMS+ffvYs2cPBw4cIDc3tyLIHBtsioqKSE9PJzMzs+IrKyuLrKwsLrvsMtLS0ioCTGJi4kmhpsZHMHwM5JkPDQxu4Rba0IZf8+vjLtvABlrTmrM4q+K55jRnIxsZy1icOPkn/+Ru7q54/cQeIQCigPiaNVVEREKLAk8YKSoqYu/evezZs4e9e/dWfJV/7/F4OOecc2jevDnNmjUjOzubtm3bkpWVdVywSU1NPeM5UwHTjoo5NZ/wCfOZzyVcUrH0/BEeoQ99Tp6LA9zJnUxgAhdxEQYGE5jApVwKgBMn61nPS7x0fH1xQOsAvycREakXdLREiCksLGT79u3s3LmTnTt3smfPnuMCTYsWLTj77LNp3rw5zZs3rwg4zZs3Jz09vX7vN/NXYBrgDFJ9McAPQJMg1SciIgGloyVCUEFBATt27Kj42r59Ozt27KC4uJhWrVrRpk0bLrzwQq688sqKQJOWlla/A80xysrKcDqdx83hifhFBBH+IPYsJaGwIyLSQCjw1LGioiK+/fZbdu7cyfbt2yt6b4qLi2ndujVt2rShVatWdO3aldatW9OsWbOwOOspMjKS2NhYSktLKSgooLCwELvdTtvz2gZnb5xY4M4g1CMiIvWCAk8QlZaW8s0337B161a+/PJLvvzyS3788UfatGlTEWy6d+9eEWxCpbemurxeLyUlJZSWluLz+YiOjiYpKYmmTZvCb4DbAUeAG2EcqUdERBoEBZ4A8Xq97Ny5ky1btlSEm127dnHhhRfSoUMHOnXqxF133UWrVq2IjAz//w1lZWWUlpZSWlqKYRjExsaSkpKC3++noKCApk2bEhUVBcOB3xLYwBMHjAUaB7AOERGpV8L/N20QGIbB7t272bp1a0Xvzddff02zZs3o0KEDHTt25KabbuLiiy8mLi6u8gLDgN/vx+12U1paitvtJiIigtjYWFJTUyuWrRuGQU5ODhkZGWbYAfNcq0XAtZgbAwZCCrT+sDX5jfLp0KED55xzDueccw79+/fn4osvDlClIiJSl+o08Pg5enZjKM1KMQyDH374gc2bN7N582b+/ve/A3D55ZfTsWNHHnzwQdq3b09ycnIlJYUXj8dTEXI8Hg8xMTHExsaSnJx8ymXuNpuNRo0anVzQFcB04BmsX7EVD7wL9+y4h3vuuYd9+/bRpUsX/H4/hw4dsrgyERGpL4K6LH0v8BawEXO7lUNABOADUjC3YbkOGEn92h6lvAfn2IBjGAadO3emU6dOdO7cmXPPPTds59ycjmEYuN3uipADEBsbS2xsLNHR0bX7eRjARMzeHqtCTxzwBjDMbHv37t05//zzWbt2Lb/61a+46667GsTwoohIuDrTsvSgBJ5PgYeAvx/53n2GayMxN8C9GPg90KfWtVefYRjs2bOnIuBs3rwZv99P586dK0JOy5YtG1zAAbMXp6ysrCLoREVFVYQcy8OCAdwHRc8XkVSahI2a/bz/G/FfXrC/wBOLn8A+4Ghf4pdffsnIkSN57733mDFjBg6Hg5deeokLLrjAojcgIiLBVGf78DiBezD/UV3V6RjeI19fADcC3YBXgYxANPAYhw8fZsOGDWzcuJFNmzbh8/kqem/uv/9+BRy3m7KyMux2O9HR0cTFxZGamhrYJfI2cDzsoKxLGbbJNigEXNUsIwHO7ng23xjf8Oz3zx533ESHDh0YOnQoOTk5LFu2jFdffZUePXpw3333MXny5LBY/i8iIqaA9fD8AHTBPBqpNnNPYzCnXawHOtainBMZhsF3333H2rVrWbNmDd999x1dunShW7dudOnShfPOO08B55iAExMTQ3R0dFCPnPB6vRw6dIjMzEwiyyLN5Ps4kI/ZTXi6gz/jMHuHOmAuc+8HB348wPXXX89f//pXOnXqdNo6v//+e26//XYiIyOZO3cu55xzjrVvSkREAiboQ1rfA7/A/L3kr/bdp5aIOffnilqUUVJSwscff8zatWtZu3YtkZGR9OrVi549e3LNNdcQGxtrUWtDR30KOCc6dOgQcXFxJCQkHH3SgLn3zuWKkiu4/L+Xw7eYvT52zIlgV2Am7R7A+ceX9+GHH3LHHXfw0Ucf0bjx6dek+3w+XnjhBZ566ikee+wxbrzxRqvfmoiIBEBQA48Dc8LxT1gXdsqlYP5+a3bC8yUlJXg8nlOuitq/f39FL86nn35Ku3bt6NmzJz179uTCCy9sUL04Pp+vIuB4PB48Hk+9CjjHcjgcuN1uMjKOH8xct24d/fr1Y9KkSTz//PPVLvexxx7j448/ZsWKFZXOOfrmm2+YMGECHTt25MknnyQxMbHa9YmISPCcKfBYPknhbsxhLKvDDphzgm7CHK0oV1RURN++fVm8eDFg7v/y+eef8/vf/56rr76aLl268PnnnzN69Gi+++47Vq1axbRp02jVqlVYh53yfXCKi4s5fPgwBw8eJDc3F6fTic1mIyEhgezsbLKzs0lNTSUuLq7ehB2v14vD4SA1NfW45//xj39wxx13EB8fT3FxcY3KnjFjBnFxcfzhD3+o9NpLLrmETZs2ER0dTefOnfnqq69qVKeIiNQ9S3t4NgM9Cdx+cQAJwFzMjXKLi4sZPHgwl112GUOGDGHFihWsWLGC5ORk+vbtS69evejYsWO9+UUeKIZhnNRz4/P5iIqKIjo6uuLPUPk5nGooa9u2bQwdOpTBgwfz3//+l927d/PFF1/UqPy8vDyuvfZaHnvsMfr27Vule5YuXcq9997L9OnTufPOOzWhWUSkHgrakFYXzNATaM2BrwoK6NG9O4ZhUFhYSHZ2NoMGDWLAgAG0atUqCK0IPsMw8Hq9eL1ePB5PxWOfz0dkZORx4SZU95NxOp2UlJSQmZl53PPz588nLS2NhQsXMnjwYKZPn86uXbuIj4+vUT1bt25lxIgRrF+/nnPPPbdK9+zdu5cJEyaQkpLCiy++SHZ2do3qFhGRwAjKkNb3mEvJg+EwMPbee/nXv/5FTk4OrVq1okuXLlx00UVhEXbKe2xKSkooKiri8OHD5OTk8PPPP5Ofn09paSk2m424uDjS09Np3LgxWVlZpKSkEB8fH7Jhx+fzUVxcfNJQFsDYsWPp2rUrn3zyCT169GDgwIE4HDU/cKtjx47cf//9jB07tmLTxMo0b96cNWvW0L59ezp16sTGjRtrXL+IiASXZT08DwOzOLJSeN8+GDcODh4Emw0mTYK77oIRI2DnTvOGggJITYWvvoKFC+Hxx48W9vXX8OWX0K4dvPMO/OlP4PNBv37w2GPYMPfoedXpZPfu3Xz//fd8//33pKamMn78+Jr9JILMMAx8Pl9FD03542N7bMq/oqKiiIyMJCIiIqznHR0+fJioqCiSkpJO+fratWt5/vnnee+99yypzzAMJk+ejGEYvPzyy9X62W7atIlJkyYxfPhw/ud//qfifDAREak7QRnS6szRnZT56Sfzq0MHKC6Gjh1h2TJo2/boDffcAykp8NBDxxf0zTcwaBD897+Qlwft28PWrZCVBePHm0Gqa1fOxjyqor46VaA59nvDMIiIiCAiIqIizBwbchqa0tJSioqKyMrKOm3wmD59Oueffz533nmnpfX27t2bgQMHcvfdd1d+wzHy8vKYMmUKP//8M/PmzaNly5aWtUtERKov4DstG8DXxz7RpIn5BZCUBG3awIEDRwOPYcC778KphgTeegtGjjQff/89XHCBGXYAunWDJUuga1d+wly1lXByCQFjGAZ+v/+kL5/Pd9zj0wWaqKioiu816fWo8nlYaWlppw07Xq+X1atXs3btWkvrjo2N5c033+T666+ndevW9OrVq8r3ZmRk8Pbbb/Pyyy/TtWtX7dkjIlKPWRJ4XJxhx//du+Ff/4Jf/vLoc5s3Q6NGZpg50TvvwPLl5uPzzzeHwHbvhrPOMnuJysoAczPd3cBF1Win3+/HMIxTfp342olBpvzLbrcTERGB3W4/7is6OrrisQJN9RQVFVXsA3Q6n376Kc2aNaNFixaW19+kSRPmz5/PqFGjWLVqVbXmgdlsNiZPnszVV1/NhAkT2LBhA7Nnz9aePSIi9YwlgceNeeq598QXHA4YOhSefhqO3RTwrbdg1KiTC/rsM4iPh4svNr9PS4O5c825P3Y7XH21OdR1pOyf3W4aY/YQlCt/fKpQY7PZsNls2O32isfHfh37fGRkJDExMScFG7FWWVkZJSUlla54WrZsGQMHDgxYO6644gr++Mc/MmrUKD744APS0tKqdX/5nj33338/nTp1Yt68ebRv3z5ArRURkeqyJPBEcYqNBj0eM+yMGQNDhhx93uuFpUvNeTknevvtk4NQ//7mF8DLL0P5XjKxsaRGRZEIxw2DlD8+VaCR+qV8KCslJeWMYdLr9bJq1SrLh7NONGrUKLZt28aECRNYvHhxtedSxcfH89xzz7F06VKGDh2qPXtEROoRSz6JEzB7eCoYBtxyizl359e/Pv7iDRugdWtziOpYfr85r6d8/k65nBzzz/x8mDMHbr0VAHdkJC1iYiqGQsq/oqKijlvVVN5rI/WP0+nEbrcTFxd3xusCOZx1opkzZxIREcHvfve7GpcxZMgQPvroI5YtW8bw4cPJz8+3sIUiIlITlgQeO+b5WRU++QTmzzcnJbdrZ36tXm2+dqpeHIBNm+Dss+HElS533WVOdr7mGvjNb+DCCwFIBjJOLkVChM/nO+XxEaeyfPnygA5nHSsyMpK//vWvbNiwgQULFtS4nPI9e1q1asW1117Ltm3bLGyliIhUl2XL0u8CnicwZ2idSg8gsAMcEkh5eXnExMRUOrnX6/Vy0UUXsXbt2qD08JTbtWsXvXv3ZsGCBVx55ZW1KmvRokXcd999PP744wwbNsyiFoqIyImCstPyGMyVU8GQBNwSpLrEeiUlJfj9/uPOyjqdYA5nHeuCCy5g7ty53HzzzRw4cKBWZQ0fPpwVK1Ywc+ZMHnzwQbzek6b3i4hIgFkWeK4AmllVWCXswKAg1SXWMgyDoqIiUlJSqjS3KpjDWSfq3r07d955J6NHj6akpHZH4pav4vruu+8YPHgwhw4dsqiVIiJSFZYFHhvm8RKB3ggwHpgOaCP/0ORwOCommFfG6/WycuXKOgs8AL/61a9o06YNd955J2ca/q2KtLQ0lixZQseOHbn22mv56quvLGqliIhUxtL1sqOA9li01v00GgMPBLB8CRyfz4fT6ST52D2ZzuDTTz+ladOmQR/OOpbNZuPpp59mz549zJ49u9blRURE8Pvf/55HHnmEwYMH8+abb1rQShERqYylgccGvAnEWlnoMeKARah3J1QVFRWRkJBARERE5RdjDmcNGlT3g5exsbEsXLiQV155hffff9+SMgcOHMjq1at5/PHHuffeeyk7soO4iIgEhuU7op0NrMAcerJSHPAS0MHiciU4ysrKKCsrq/KRC/VhOOtYjRs3ZsGCBUydOpUdO3ZYUmabNm346KOP2Lt3L/379+fgwYOWlCsiIicLyBaw12OGngSLKogDXgTGWlCW1I3CwkKSk5OrvAnkpk2bOOuss+p0OOtEHTt25E9/+hMjR47k8OHDlpSZkpLC22+/zXXXXcd1113HF198YUm5IiJyvIDted8V+BK4lJpPZI4HzgE+BsZZ1C4JPpfLhc1mq3RH5WMtXLiQUafaoLKOjRgxgoEDBzJmzBhKS0stKdNut/PAAw8we/ZsRowYwbx58ywpV0REjgroIT8XAluA/8XcFTmpivclHrl2OrATc8m7hCbDMCguLiYlJaXK9xQUFPDBBx8wdOjQALas5h5++GGaNGnC5MmT8fl8lpXbu3dv1q1bx5w5c5g6dSput9uyskVEGrqAn2oYAdwN/AzMB3pihp9oIOWYr9gjf3YG5gI5mEEpJtANlIAqLi4mJiaGqKioKt+zZMkSbrjhhmqfWB4sdruduXPnkpeXx4MPPljr5erHOv/88/nwww/Jz8+nd+/etd70UERETEE7xjkSGAisAQ4B+4F1wHuYR0T8B8gHNgE3EbiVXhI8Xq8Xl8tV5WXo5d58803GjBkToFZZIyYmhoULF7J582aee+45S8tOTExk/vz59O3bl+uvv56qHu8iIiKnF8gtc84o68iXhK+ioiISExOx26ueq7dv387PP//MddddF7iGWSQlJYUlS5bQvXt3GjVqxIgRIywr22azcc8993DRRRcxfPhwnn32Wfr3729Z+SIiDU2dBR4Jb263G4/HU+1hqYULFzJy5Mgq79VT15o0acLixYvp378/2dnZXH/99ZaW36tXL/72t78xYsQI9u7dy5QpU6q80k1ERI4K2pCWNCzVOS+rnMfjYdGiRYwePTqALbNe69ateeONN7jtttv4+uuvLS+/Xbt2bNiwgTfeeIMZM2ZYOlFaRKShUOARy5UftBkbW72ZWOvXr6dly5acd955gWhWQF111VUVy8r37t1reflnn30269at49///jejR4/G5XJZXoeISDhT4BHLFRcXV3uiMoTGZOUzGTBgANOnT2fo0KGWbUx4rJSUFBYvXkx6ejq9evXi559/trwOEZFwpcAjlnK5XERERBATU70NBXJzc/n73/9eL87Oqo1JkybRr18/Ro4cWdHTZaXo6GjmzJlD37596dq1K9u3b7e8DhGRcKTAI5Yp32QwKamqW0wetWjRIvr06VPls7bqs4ceeoiWLVsyceJEvF6v5eXbbDbuv/9+HnroIfr27ctHH31keR0iIuFGgUcs43K5iIqKIjq6eufZG4bBwoULQ26y8unYbDaeffZZ3G43M2bMsHRjwmONGDGCN954g1tuuYWFCxcGpA4RkXChwCOWqE3vzpdffklJSQlXX311AFpWN6Kjo3njjTfYunUrTzzxRMDq6dSpE6tXr+bRRx/lkUceCVi4EhEJdQo8YgmHw1HtIyTKvfzyy0ycOLFaGxSGgsTERBYtWsSCBQsC2gPTqlUrPvjgA9avX8/kyZN1BpeIyCmE128YqRN+vx+n01mj3p3c3FzWrVsX0quzzqRRo0YsXryYmTNnsm7duoDVk52dzerVq3E6nQwZMoSCgoKA1SUiEooUeKTWHA4HsbGxREZWf+Pu1157jYEDB9bbg0KtcMEFF7Bw4UKmTJnC1q1bA1ZPXFwcb7zxBpdeeindunVjz549AatLRCTUKPBIrfj9flwuV416dzweD/PmzWPSpEkBaFn9csUVV/DCCy8watQovvnmm4DVExERwaxZs7jtttvo3r17QAOWiEgoUeCRWikuLiY+Pr5GZ1+tXLmS8847j7Zt2wagZfVPz549eeKJJxg2bFjA98+ZPHkyTz/9NMOGDePDDz8MaF0iIqFAgUdqzOfzUVJSUuO9c15++WVuu+02+eNIcwAAIABJREFUi1tVvw0YMIA//elPDB48mF27dgW0rj59+rBgwQJuueUWVq9eHdC6RETqO52WLjXmcDiIj4+v0eqqr7/+mn379tGnT58AtKx+GzZsGGVlZQwcOJBVq1Zx7rnnBqyua665hiVLljB8+HCcTifDhw8PWF0iIvWZAo/UiN/vp6SkhOzs7Brd/5e//IWJEyfWaKJzOBg9ejRlZWUMGDCA1atXc/bZZwesrvbt27NixQqGDBmCw+FgwoQJAatLRKS+api/baTWHA4HcXFxNerdOXz4MCtXruSLL74IQMtCx80334zb7aZ///6sXr2apk2bBqyutm3bsnr1agYMGEBxcTHTpk0LWF0iIvWRAo9UW/nKrKysrBrdP3/+fPr06UNmZqbFLQs9kydPrujpWblyJY0bNw5YXS1btmTt2rUVoefBBx/EZrMFrD4RkfpEk5al2lwuF7GxsTVameX1enn11Vcb3GTlM5k6dSojR45k0KBBHDp0KKB1NWvWjDVr1rB69Wp+85vf4Pf7A1qfiEh9ocAj1WIYBg6Ho8Yrs9asWUPjxo1p166dxS0Lbffeey/9+/dn4MCB5OfnB7SurKwsVq1axdatW5k6dSo+ny+g9YmI1AcKPFItLpeL6OjoGk82njt3LpMnT7a4VeHhwQcfpGvXrgwePJjCwsKA1pWamsry5cvZt28fEydOpKysLKD1iYjUNQUeqbLa9u589tln/PjjjwwcONDiloUHm83GzJkzufLKKxk2bBgOhyOg9SUkJPDuu+/idrsZPXo0JSUlAa1PRKQuKfBIlZWUlBAZGUl0dHSN7n/qqaeY9v/t3Xd4VVX69vHvTnLSCSEUB2HEgoCACFgoIygISCgplFAChF4FAWmDCAKKUiQoSpXee0dAmtiAF0XAAZ1B9AeOohAChFSSnPePQ3UoycneJ+3+eHERkr2f9eSaEe+svfZa/frl21fRM8IwDN555x0qVqxIy5YtSUhIsHQ8b29vFi1aRMGCBWnevLnlIUtEJLso8EiGZWV25/vvv+fIkSO0bdvW5K7yHsMweO+993jkkUdo3bo1SUlJlo5ns9mYNWsWjz/+OE2bNrV8DZGISHZQ4JEMSUpKwjAMvLy8nLp/ypQp9OrVy+n78xs3NzemTp3KAw88QLt27UhOTrZ0PHd3d6ZMmULNmjUJDg7mjz/+sHQ8ERFXU+CRDImLi3PqRHSAU6dOsXfvXu3wm0nu7u5Mnz4dX19foqKiLA89hmHw1ltvER4ezssvv8yZM2csHU9ExJUUeOS+kpOTsdvteHt7O3X/1KlT6dy5s9OBKT/z8PDg448/xsPDwyULiw3DYOjQofTo0YOXX36Z//u//7N0PBERV1HgkfuKj493eu3O2bNnWb9+Pd27dze5q/zD09OT+fPnU7RoUZo1a8bly5ctH7NXr17079+fJk2a8Ouvv1o+noiI1RR45J7S0tJISUnBx8fHqfs/+ugj2rRpo2MkssjDw4Np06ZRoUIFQkJCiImJsXzM7t270717d5o0acLZs2ctH09ExEoKPHJP8fHx+Pr6OnXmUmxsLEuWLKFPnz4WdJb/uLm5MXHiROrUqUPjxo1dEkL69u1LZGQkTZs25dy5c5aPJyJiFQUeuSu73U5CQgJ+fn5O3T9r1iwaN25MiRIlTO4s/zIMg1GjRhEREUFwcDCnT5+2fMzBgwcTFhZGSEgIFy5csHw8EREraAc4uavExEQ8PT2dOiQ0Pj6eOXPmsHXrVgs6k4EDB+Lv709wcDDr1q2jTJkylo43fPhwkpOTCQ0NZfPmzRQsWNDS8UREzKYZHrmr+Ph4p2d3Fi5cSM2aNSldurTJXcl13bt3Z8SIETRt2pSjR49aOtb1Yy9q1qxJWFgYcXFxlo4nImI2BR65o5SUFOx2u1MbBSYkJPDBBx8wcOBACzqTW7Vp04aJEyfSrFkzDhw4YOlYhmHw7rvvUrlyZZo3b058fLyl44mImEmBR+4oK7M7s2fPpnr16lSqVMnkruROQkJCmDFjBm3btmXv3r2WjnX92IvSpUsTERGhA0dFJNdQ4JH/kZaWRnJyMr6+vpm+99KlS3z44YcMHz7cgs7kburVq8fixYvp2rWr5eumrh97Ubx4cdq0aWP5WV8iImZQ4JH/ER8fj4+Pj1Ovon/44Yc0bNiQxx9/3ILO5F5q1KjBqlWr6N+/P6tWrbJ0LHd3d2bMmEHBggVp3749KSkplo4nIpJVCjxym6y8in7u3DnmzJnDkCFDLOhMMqJKlSps2LCBkSNHMn/+fEvHuvXYi06dOnH16lVLxxMRyQoFHrlNYmIiNpsND4/M71gQHR1NREQEf//73y3oTDLqiSeeYMuWLUyePJmpU6daOpbNZmP+/PkkJSXRo0cP0tLSLB1PRMRZCjxyG2cXK//6668sX75cb2blEI8++ijbtm1j4cKFjBs3DrvdbtlYXl5eLF68mJiYGHr37k16erplY4mIOEuBR264evUq6enpTp2KPmHCBDp16kSxYsUs6Eyc8eCDD7J161a2bdtG//79LX3k5OPjw7Jlyzh9+jSvvvqqQo+I5DgKPHJDYmKiU4eEnjx5kk8++YS+ffta0JVkRdGiRdm6dSu//fYbrVu3tnTDQF9fX1atWsWJEyd4/fXXLRtHRMQZCjwC3Fys7Myr6O+88w69e/cmMDDQgs4kq/z9/Vm2bBkPPfQQDRs25LfffrN0rNWrV7Nr1y7L1w+JiGSGAo8AkJycjIeHR6YXKx87dowvv/ySHj16WNSZmMHDw4PJkycTERFBvXr1OHbsmGVjBQYGsnbtWqZNm8bq1astG0dEJDN0eKgAOD278/bbbzNw4ECn7hXXMgyDV199lYceeoiwsDBmzJhB/fr1LRmrZMmSrF69mqZNm1K0aFFeeOEFS8YREckozfAI6enppKSkZHr9zv79+zl+/DhRUVEWdSZWCA8PZ+nSpfTp08fSvXoqVKjAggUL6NSpk6UzSiIiGaHAIyQkJODt7Z2pnZXT09MZPnw4b7zxhlMHjEr2qlatGtu2bWPq1Km8+eablr1VVatWLSZNmkTLli05c+aMJWOIiGSEAo+QkJCQ6dmdZcuWYbPZaNGihUVdidUeffRRduzYwf79++nSpYtlZ2I1a9aMfv36ER4eTmxsrCVjiIjcjwJPPnf16lXsdnumZmni4uJ46623ePfdd506b0tyjsKFC7N+/XoMwyA0NJSYmBhLxunduzcvv/wyrVq10gnrIpItFHjyOWcWK0+ePJk6depQpUoVi7oSV/L29ubjjz+mZs2a1K9fn59++smSccaOHUvJkiXp0qWLjqAQEZdT4MnH7HY7iYmJmQo8P//8M4sWLWLkyJEWdiau5ubmxqhRo+jXrx/BwcHs37/fkjGmT59OXFwcQ4YMsfS4CxGRv1LgyceSkpKw2Wy4u7tn+J5Ro0bRu3dv/va3v1nYmWSXjh07Mm3aNCIjI1m7dq3p9b28vFi6dCn79+9n8uTJptcXEbkb7cOTj2X2cda+ffs4evQos2fPtrAryW716tVjw4YNRERE3Dgby8y1WgUKFGD16tXUq1eP4sWL07ZtW9Nqi4jcjWZ48qnre+9k9KDQ1NRUhg8fztixY/Uaej5QsWJFdu7cyapVqxgwYAApKSmm1i9evDhr167ljTfeYOfOnabWFhG5EwWefCopKSlTe+8sWrSIQoUK0aRJE4s7k5ziwQcfZPv27Zw9e5aQkBDOnj1rav2yZcuyZMkSunXrxnfffWdqbRGRv1LgyacSExMzPLtz8eJF3n33XcaNG6fX0PMZf39/li5dSp06dahTpw4HDx40tX716tWZOnUqERER/PLLL6bWFhG5lQJPPpTZx1kTJ06kUaNGPPnkkxZ3JjmRm5sbQ4cOJTo6mrZt2zJv3jxT37Bq0qQJQ4YMITw8nPPnz5tWV0TkVgo8+VBmHmf95z//YeXKlQwfPtwFnUlO1rBhQ7Zv387MmTPp16+fqTszd+3albCwMCIjI01fLyQiAgo8+VJGH2fZ7XYGDRrEwIEDKVq0qAs6k5zuscceY9euXVy+fJlGjRrx3//+17Tab7zxBkFBQQwaNEh79IiI6RR48pnMPM5avnw5ly9fplu3bi7oTHILPz8/5s+fT0hICHXr1uXLL780pa6bmxuzZ8/mwIED2vpAREynwJPPZPRxVkxMDG+++SZTpkzBw0PbNcntDMOgf//+TJ8+nY4dOzJjxgxTZmX8/f1ZsWIFEyZM4LPPPjOhUxERBwWefCajj7PeeOMNWrRowVNPPeWCriS3qlu3Ljt37mTx4sX06NHDlINBH374YebMmUPnzp315paImEaBJx/J6OOszz77jC+++EILlSVDSpUqxaeffgpAgwYNOH36dJZrvvDCCwwdOpRWrVpx5cqVLNcTEVHgyUeSkpLw8vK65+OspKQkXnvtNSZOnIifn58Lu5PczMfHh5kzZ9K2bVvq1avHnj17slyzW7duVK9ena5du5Kenm5ClyKSnynw5COJiYn4+Pjc85r33nuPihUr8vLLL7uoK8krDMOgV69ezJ07l549e/L+++9naV2PYRhMnDiRixcv8vbbb5vYqYjkRwo8+URGHmedOHGC+fPnM378eBd2JnnN888/z+7du1m/fj2dOnUiPj7e6Vqenp4sWrSI5cuXs2bNGhO7FJH8RoEnn7jf46z09HQGDBjA8OHDeeCBB1zcneQ1JUqU4JNPPsHPz4+XXnqJf//7307XKlq0KMuWLWPQoEEcOXLExC5FJD9R4Mknrr+OfjcLFiwAICoqylUtSR7n7e3Nhx9+SK9evQgODmbx4sVOP+KqVKkS0dHRtGnThj///NPkTkUkP1DgycPsdjsXL17EbreTnJx818Dzxx9/MG7cOKKjo3Fz0/8lxDyGYRAVFcWWLVv46KOP6NKlC5cvX3aqVlhYGO3atSMyMpLk5GSTOxWRvE7/dcvjEhMTSU5OxmazkZqaSkJCwv9cM2zYMKKionjiiSeyoUPJD8qVK8eePXsICgri+eefd/rU9WHDhlGsWDEGDhyo4ydEJFMUePIwwzDw8PAgLi4Od3d3Lly4gLu7+23XbN68mWPHjjFo0KBs6lLyC29vbyZNmsQ777xDZGQk7733HmlpaZmq4ebmxqxZs/j222+ZOXOmRZ2KSF6kwJPH2Ww2Ll26REJCAoGBgXh5ed34WkxMDIMHD+ajjz7K0O7LImZo3Lgxe/fuZdeuXYSFhfH7779n6n4/Pz+WL1/OpEmTTNnvR0TyBwWePM4wDOLj4ylUqND/hJqhQ4fSvHlzqlWrlk3dSX5VokQJNm3aRK1atXjhhRfYtm1bpu4vVaoU8+bNo2vXrpw6dcqiLkUkL1Hgye3OAcuAV4GngYeBvwNlgRbgM8eHh2IfwtfH97bbNm7cyNGjRxkxYoSLGxZxcHd3Z8iQISxcuJBBgwYxbNgwkpKSMnx/rVq1+Oc//0nr1q2ztNePiOQPxr0W/j3zzDP2Q4cOubAdybADwARgK2ADrgB3+p/SG3AHigNDgXZw/sp5nn/+eRYtWsSzzz7rqo5F7urixYv069ePU6dOMXfuXMqUKZOh++x2O7169SI1NZXZs2ff89gUEcn7DMP4xm63P3Onr2mGJ7e5BEQCdYF1QBIQx53DDte+Hg+cBPoDj8PgjoNp1aqVwo7kGIGBgSxYsICuXbsSHBzMokWLMvQWlmEYREdHc/z4cebMmeOCTkUkt1LgyU0OAI8Ba4EE7h5y7iYe0n5Nw2ebD8Pjh2f+fhELGYZBx44d2bJlC9OmTaNz585cunTpvvf5+PiwaNEixo0bx7fffuuCTkUkN1LgyS0+B14CYnDM2jjJHXempU3Da7oXdEGhR3Kc63v2FC5cmFq1anHgwIH73vPYY48xZcoU2rdvz4ULF1zQpYjkNgo8ucG/gGAcj6bMkgCsAIaZWFPEJLfu2dOuXTsmTZp03z17QkJCCA0NpVu3bqSnp9/9wqvA98A+4DPgWyDRvN5FJGdS4MnpUoEWOAKK2RKAqcDXFtQWMcH1PXv27t1LcHAwP/300z2vHz16NFeuXGHSpEm3f+E0jnBfDvADagIhQChQBwjA8YZjV+Coyd+EiOQICjw53Ts4/rK26tFTIhCBfsKVHKtEiRJs3LiRZs2aUb9+fWbNmnVjBicuLo6YmJgb19psNubPn8/HH3/s2JTwJNAAxzYN0cCPOGZ44nC8AHAJuIzjB4v/A+YDNYDKwJeu+g5FxBUUeHKyOByBx4rZnVvFAkssHkMkC9zc3OjZsyfbt29n5cqVhIWFcebMGfbt20ebNm1ue6OrePHifDzrY7pHdCf2yVjYhWPdW0oGBkrD8e/bEaA+0Afr//0TEZdQ4MnJFuOa/4XigfFoAbPkeI8//jjbtm2jTp06vPjii8TExJCSksLq1atvXpQMtSfUZuaVmdiSbHCP5Tz3lAjMA6ri2OBTRHI1BZ6cbAoQD0kk8RzP8RRPUYEKjGLUbZf1ox/++N/482lOU4c6VKEKlajEVrYCcJWrRBHFkzzJEzzBO7xzs8jvwGEXfE8iWeTh4cGAAQPYtGkTs2fPxsPDgxEjRpCQkOB4XNUU2At1k+re9u+FUxKBU0B1QC9/ieRqCjw5VTyOv2gBL7zYzW6OcITv+I5tbGM/+wE4xCFiib3t1rd4iwgiOMxhlrOc3vQGYBWrSCaZYxzjG75hJjP5hV8cN6WhNQuSa0yePJm2bdsSEBDAlStX+Pnnn3nllVccC5O/xNw1aVeBX4FwNAsqkot5ZHcDchdHAF/gMhgYN35SvXrtHwODNNIYzGCWspR1rLtxq4HBZS4DcIlLPMiDNz4fTzyppJJIIp54EkCA46YkHK/p9nXdtyjirL59+xISEsKZM2c4ffo0Bw8exOOcByzHmgX4KcA3wFwc+1eJSK6jwJNTHcfx5sg1aaTxNE9zkpP0oQ/VqMb7vE8IIRSn+G23vsmbNKABU5lKPPHsZCcALWjBBjZQnOIkkEA00QQRdPPGIy74vkRMYLPZKF26NKVLlwYgqn0UlMbatw3jcRzSGwYUtnAcEbGEHmnlVPE4HjNd44473/Edv/IrBznIPvaxilX0vcOUzDKW0ZGO/MqvbGUr7WlPOukc5CDuuPMbv/EzP/Me73Hq+nMz0KvpknvtxDULi9MBHdklkisp8ORU7sAdDn4OJJA61GEPezjJSUpTmod5mAQSKI3jp905zCGCCABqUIMkkjjPeZaylIY0xIaNYhTjH/yDQxy6fUyR3GgCcCXjl//Ij1S+5Z8AApjClPvfmAhMxvk3v0Qk2yjw5FRFAJvjw3Oc4yIXAUgkkU/5lKd5mrOc5Zdr//jiy0lOAvAQD7GLXQCc4ARJJFGUojzEQ+xmNwDxxLOf/ZSj3M0xNU0vuVEijvVnmVCWsnx37Z9v+AZffAknPGM3x6M3GkVyIa3hyamqcuOnyN/5nSiiSCONdNKJIIImNLnrre/xHt3oRjTRGBjMZz4GBn3oQyc6UYEK2LHTiU5UopLjJgN43vLvSsR8RwAfHG9TOWEXu3iMxyhFqYzdkI5jAfPTzo0nItlDgSenKs2NNTyVqMTh+/xIeeWW+fzylOfLO7xj7o8/q1h15wL+OLbUF8ltDuF02AFYznLa0CbjNyQAXwDdnR9TRFxPj7RyKjegHndcx2OJq8BLLhpLxExncHrBfQopbGQjLWmZuRt/dm48Eck+Cjw52WAcJztbzQCCgaIuGEvEbMnO3/oJn1CVqjzAAy4bU0SyhwJPTlYLKOaCcXyAIS4YR8QKvs7fuoxlmXucdZ2P82OKSPZQ4MnJDGAR1v7l6g2E4jgrSCQ3KoNTM6HxxPMpn9KMZpm70YDra/1FJPdQ4MnpauLYyt6q0OMHTLeotogrPINTf5P54UcMMRSkYOZu1AJ/kVxJgSc3mIjjJ0pvk+v6AVsgs3/fi+Qo5cjSW1qZlgo858LxRMQUCjy5gTeOrfOfwpyZHgNH2NkKVDOhnkh28gBa47qdwktf+yUiuYoCT27hD+wFupK10OMLlAUOALWz3pZIjjAA8HTBOP5ogb9ILqXAk5t4Ax/gmO15FMdfvhndp8ePm29jHQUqWNGgSDaphGPhvdVbqQYALSweQ0QsoZ2Wc6OawEngcxzre3bimM53w7GWwX7tzx44doV9BMdPwO2AAtnQr4grLMKxnicTh4hmig/0rd6XL2p8Qd26dXn44YcpVaoUVapU4e9//7tFg4qIWRR4cisDxyOp2jjO9jkJfAucx3EkhS+OWZyncM3mhSLZrQTwIdAbR9A3ky/QBeq8WIfZbWeTnJxM3bp1OXjwIEeOHOGNN94weUARMZsCT17ghmMvkjLZ3YhINosCTgGTMC/0+AL1gWho5t6MZs2akZqayo4dOxg+fDhRUVEmDSQiVtIaHhHJW0YDI8nyG4127IzyGEVSaBKs5sZbYGPHjuXw4cMsXLiQhQsX0qpVK86ePZvFpkXEago8IpL3DAX2AH/HuaMnPMEoaPBn3T8ZWnQodnf7jS899thjtGnThoMHD7J9+3aefvppXnjhBdatW2dS8yJiBcNut9/1i88884z90KFDLmxHRMRESTjebJyM4xHXFRyL+u/m+qL+bsA/Id4nnoYNG9KzZ08iIyNvXGa320lNTcVmswFw+PBhevbsSZUqVZg0aRL+/v5WfDcich+GYXxjt9ufudPXNMMjInmXN46tGH4DVgJ9gMrc3LXcwLGSsQzQCZgFnAPeA4qAn58fc+fO5a233uLo0aM3yhqGcSPsAFSpUoW9e/fi6+vLiy++yJEjR6z/3kQkUzTDIyL5UxqO2Z4MvLqxceNGxo4dy6effkpgYOA9r92wYQNDhgxhwIAB9OjRA8PI6GZZIpJVmuEREfmr63tVZUBISAgNGzbklVdeIT09/Z7XhoaGsmPHDtasWUNkZCQxMTFZ71VEskyBR0QkA0aOHMmlS5f44IMP7nttqVKl2Lp1K2XLluXFF1/kyy+/dEGHInIvCjwiIhlgs9mYNWsWc+fOZe/evRm6ftSoUbz//vt069aN8ePHk5aWZn2jInJHCjwiIhlUvHhxZs6cSZ8+fTh16lSG7qlbty579uxh//79hIWF8dtvv1ncpYjciQKPiEgm1KhRg2HDhtGhQwcuX76coXseeOABVq9ezUsvvUTdunXZtm2bxV2KyF8p8IiIZFL79u2pVasWPXv2zPBjKnd3d/r378/ChQsZNmwYr7/+OsnJyRZ3KiLXKfCIiDhhzJgxJCcn89Zbb2Xqvueee449e/Zw5swZGjZsmOFHYyKSNQo8IiJOsNlsfPzxx2zZsoWVK1dm6t5ChQqxYMEC2rVrR8OGDVmzZo1FXYrIdTotXUTESYUKFWLRokWEh4fz2GOP8fTTT2f4XsMw6NKlC9WqVSMqKorvvvuOUaNG4eGhv5ZFrKAZHhGRLChbtixTpkyhc+fOTr2BVbFiRXbu3MmJEydo2bKlNioUsYgCj4hIFjVo0IAuXbrQsWNHEhMTM31/oUKFWLFiBVWqVKFevXocO3bMgi5F8jcFHhERE/Tt25fHHnuM/v37c68zCu/G3d2dkSNHMnLkSJo3b651PSImU+ARETGBYRhER0fzyy+/ZOj4ibsJDw9n3bp1jBs3jpEjR5KammpilyL5lwKPiIhJvL29mT9/PvPmzcvS5oIVKlRg586d/Otf/yIiIoILFy6Y2KVI/qTAIyJiouLFizNv3jwGDBjAiRMnnK5zfV1PpUqVqFevHt9//72JXYrkPwo8IiImq1KlCmPHjqVDhw5Zmp3x8PDgzTffZMSIETRr1oy1a9ea2KVI/qLAIyJigRYtWhAaGkpUVBRJSUlZqnU97Lz11lu8+eabWtcj4gQFHhERiwwfPpwHH3yQXr16ZfjMrbu5vl/P0aNHadWqFbGxsSZ1KZI/KPCIiFjEzc2NqVOncunSJV5//XWnXle/VVBQECtXrqRixYrUq1ePH374waRORfI+BR4REQt5enoyf/58Dhw4kKXX1a/z8PBg9OjRDB06lNDQUL744gsTuhTJ+xR4REQsFhAQwLJly1i4cCErVqwwpWZERARz5syhS5cumT68VCQ/0il1IiIu8Le//Y3ly5cTFhZG0aJFqVu3bpZrPv/882zcuJHWrVtz+vRpXnvtNQzDMKFbkbxHMzwiIi7y+OOPM2/ePPr06cORI0dMqVm2bFm2b9/Oli1bePXVV7l69aopdUXyGgUeEREXeu6555g8eTLt27fnl19+MaVmsWLF2Lx5M+fOnaNNmzbExcWZUlckL1HgERFxseDgYAYOHEjr1q05f/68KTX9/PxYtGgRjzzyCI0bN+a3334zpa5IXqHAIyKSDTp27EhoaCiRkZEkJCSYUtPDw4MJEybQokULXn75ZR1HIXILBR4RkWwybNgwypUrR7du3Uxbe2MYBv369WPMmDE0b96cPXv2mFJXJLdT4BERySaGYTBp0iTsdjtDhgzJ8saEtwoPD2fBggX06tWLJUuWmFZXJLdS4BERyUY2m43Zs2dz/PhxJkyYYGrt6tWrs3nzZiZPnsy4ceNMDVQiuY0Cj4hINvPz82PJkiWsW7eO2bNnm1q7dOnSbN++nT179tC7d29SUlJMrS+SWyjwiIjkAEWKFGH16tXMmDGDhQsXml5748aNxMXF0a5dOxITE02tL5IbKPCIiOQQJUuWZM2aNURHR7Ns2TJTa/v4+DB//nwKFy5MRESE9uqRfEeBR0QkB3n44YdZtWoV7777LmvXrjW1toeHBx999BFlypQhPDyc2NhYU+uL5GQKPCIiOUzp0qVZuXIlo0aNYvPmzabWdnNzY9KkSdSsWZOQkBDOnTtnan2RnEqHh4qI5EBly5Zl2bJltGrVCk9PTxo0aGBabcMwGD16NAVAxHZ5AAAeNklEQVQKFKBRo0asW7eOkiVLmlZfJCfSDI+ISA5VsWJFlixZQv/+/dm9e7eptQ3DYPDgwXTu3JkmTZrw888/m1pfJKdR4BERycEqV67MggUL6NOnD1988YXp9Xv16sXAgQNp0qQJJ06cML2+SE6hwCMiksM9++yzzJkzh+7du3PgwAHT63fo0IExY8YQHh7Od999Z3p9kZxAgUdEJBeoWbMmM2bMoFOnTnz77bem12/evDlTpkyhVatWfP3116bXF8luCjwiIrlE7dq1ef/992nXrh1Hjx41vX7Dhg2ZNWsWUVFRpq8ZEsluCjwiIrlI/fr1mTRpEm3btuX48eOm13/hhRdYvHgxvXr1YsuWLabXF8kuei1dRCSXadSoEVevXqV169asXr2aMmXKmFr/ueeeY9WqVbRq1Yr4+HgiIiJMrS+SHRR4RERyodDQUFJTU2nevDlLliyhUqVKptavVKkS69evp1mzZhiGQcuWLU2tL+JqCjwiIrlU8+bN8fHxoXXr1syfP5/nnnvO1Pply5ZlzZo1hIeH4+npSWhoqKn1RVxJgUdEJBdr1KgRPj4+REVFMXPmTGrXrm1q/XLlyrFq1SpatGiBp6cnwcHBptYXcRUtWhYRyeXq1KnDvHnz6NmzJ9u2bTO9fsWKFVm+fDn9+/dn586dptcXcQUFHhGRPKB69eosXbqUQYMGsW7dOtPrV65cmcWLF9OnTx/27dtnen0RqynwiIjkEZUrV2b16tW8+eabLF682PT6zz77LPPmzaNr16589dVXptcXsZICj4hIHlKuXDnWrVtHdHQ0M2fONL1+zZo1mT17Np06deLQoUOm1xexigKPiEge8+ijj7Jx40bmz5/Pe++9h91uN7X+Cy+8wEcffURkZCRHjhwxtbaIVRR4RETyoBIlSrBhwwY2bdrEmDFjTA899erVu3H21vfff29qbRErKPCIiORRxYoVY/369Xz99dcMGTKE9PR0U+sHBwczfvx4WrZsyY8//mhqbRGzKfCIiORhgYGBrF69mpMnT/LKK6+Qmppqav3Q0FBGjx5Ns2bN+Omnn0ytLWImBR4RkTzO39+fpUuXcvHiRbp27UpSUpKp9SMiIhg+fDjh4eH88ssvptYWMYsCj4hIPuDj48OCBQvw8vKiRYsWXLhwwdT6kZGR9O/fn/DwcH7//XdTa4uYQYFHRCSfsNlsTJ8+nRo1atC4cWN+/vlnU+t37tyZqKgoWrVqxeXLl02tLZJVCjwiIvmIm5sbr7/+Or169SIkJMT0vXReffVVqlevTvv27UlOTja1tkhWKPCIiORDHTp0IDo6mvbt27N582bT6hqGwTvvvENgYCB9+vQx/c0wEWcp8IiI5FP16tVjxYoVjBgxgpkzZ5q2V4+7uzszZ87k999/Z+TIkabUFMkqBR4RkXysUqVKbNmyhSVLljBixAjS0tJMqevt7c2SJUvYvXs306ZNM6WmSFYo8IiI5HMlSpRg06ZN/PDDD3Tu3JnExERT6gYGBrJq1SqmT5/OmjVrTKkp4iwFHhERoWDBgixfvpyAgADCw8M5f/68KXVLlCjBihUrGD58OPv27TOlpogzFHhERARwvLb+wQcfULduXRo1asTJkydNqVu+fHnmzp1Lt27ddO6WZBsFHhERucEwDIYMGcKAAQMIDQ1l//79ptT9xz/+wfjx42ndujWnT582paZIZijwiIjI/2jTpg3Tpk2jc+fObNiwwZSaYWFh9O3bl5YtW5q+07PI/SjwiIjIHb3wwgusWbOG0aNHM378eFP21OnRowfBwcFERkaafqaXyL0o8IiIyF098cQTbN++na+++ooOHTqYcmTEyJEjKVWqFF27djX99HaRu1HgERGReypatCirV6/moYceomHDhvznP//JUj03Nzc++OAD4uPjGT58uElditybAo+IiNyXzWZj3Lhx9O3bl9DQULZt25alep6enixcuJDPP/+cuXPnmtSlyN15ZHcDIiKSe7Rp04Zy5crRuXNnjh49yqBBg3Bzc+5n5wIFCrB06VIaNmxImTJleP75503uVuQmzfCIiEimVKlShR07dvDFF18QFRWVpXU9jzzyCLNnz6Zr16788ssv5jUp8hcKPCIikmlFixZlzZo1lChRguDg4Cyt66lduzaDBg0iMjKSuLg4E7sUuUmBR0REnGKz2Xj33Xfp06cPoaGhbN++3elaXbp0oVq1avTs2dOU199F/kqBR0REsqRt27YsXryYYcOGMXHiRKcCi2EYvPvuu1y+fJm3337bgi4lv1PgERGRLKtatSrbt29n3759dOzY0alHU56ensybN4+1a9fqdHUxnQKPiIiYolixYqxdu5bixYs7vV9PkSJFWLx4Mf/85z85fPiwBV1KfqXAIyIiprHZbIwfP54+ffoQEhLCqlWrMl2jQoUKREdH06FDB/744w8LupT8SIFHRERM17ZtW9asWcP7779P3759iY+Pz9T9jRs3Jioqivbt25OcnGxRl5KfKPCIiIglypcvz44dO/Dw8KB+/focO3YsU/e/9tprlCxZkv79+2O32y3qUvILBR4REbGMr68v0dHRDB48mIiICD7++OMMhxfDMPjoo4/44Ycf+OijjyzuVPI6BR4REbFceHg4W7duZeXKlXTs2JHY2NgM3efj48PixYuZNm0an376qcVdSl6mwCMiIi7xyCOPsGXLFh5++GHq1q3L/v37M3RfiRIlmD9/Pq+88gqnTp2yuEvJqxR4RETEZWw2G6NHj2bChAl07dqV6Oho0tLS7nvfc889x5AhQ+jYsSOJiYku6FTyGgUeERFxufr16/Ppp5+yb98+WrZsydmzZ+97T+fOnSlfvjyvvfaaFjFLpinwiIhItihevDirV6/m+eefp379+uzateue1xuGweTJkzly5AgLFy50UZeSVxj3SsnPPPOM/dChQy5sR0RE8qOvv/6a3r17Exoayuuvv47NZrvrtSdPnqRx48asWLGCypUr3/a1S8DnwMFrv58D0gF/4GmgJvAP4FFrvg3JZoZhfGO325+509c0wyMiItmuRo0a7N69m1OnTtG4ceN7HktRunRpJk6cSKdOnW687XUE6AD8DYgExgF7gX8BJ4D/B8wAegMVgWrAWiDVsu9IchoFHhERyREKFSrEggULaNeuHU2bNmXWrFmkp6fzww8/8Prrr992bUhICE2aNKFrz550SU+nJrAUSAIuA3dbBn0FSMQxAxQFVAaOW/YdSU6iwCMiIjmGYRh06NCBTz75hI0bN9K8eXM8PDzYtm0bX3311W3XNhk5kq3u7iz+4QcSuHvIuZsrOMLOM0C0Oe1LDqbAIyIiOc4jjzzChg0bqFu3Lk2bNuXFF19k1KhRpKenA/AV8JLNRtLSpaSUL+/0OHYcMz4jgOFmNC45lgKPiIjkSMeOHSMtLY0ePXqwZ88ejh49ypw5czgBvAxk7jjSe0sA3gcmm1hTchYFHhERyZG8vLy4cuUKx48fp3DhwqSnpzNmzBiaY27YuS4Bx0zPvyyoLdlPr6WLiEiuMdpuZ4JhkGBRfQMoBxwFPCwaQ6yj19JFRCTXuwyMtzDsgGNNzxlgnYVjSPZQ4BERkVxhIa75j9YVYLwLxhHXUuAREZFcIZpb1u507gzFikHFijcveOMNqFQJKleGBg3gt98cn9+w4ebnn3kGvvji5j1DhzpqVKwIK1bc+PRx4N8Wfz/iWgo8IiKS413C8ajpho4dYdu22y8aPBiOHoXvvoMmTWDMGMfnX3oJjhxxfH7uXOja1fH5LVvg228dnz9wACZNgsuXAcd/HL9A8hIFHhERyfEOAz63fqJ2bQgKuv2igICbH8fHg2E4Pvb3v/nxrZ8/ftxRx8MD/Pwcs0DXQlQ88KUF34dkHwUeERHJ8f4FpGTkwtdfh7//HZYsuTnDA7BuHZQrB40bO2Z5AJ56yhFwEhLg/HnYswfO3JxH+tbMb0CynQKPiIjkeFfI4EGfb7/tCC2RkfDhhzc/Hx4OP/wA69c71vqAY51Po0ZQsya0aQM1aoC7+21jSt6hwCMiIjlepv9jFRkJa9b87+dr14ZTpxwzOuCYEfruO/j0U7DboUwZ58eUHE3/e4qISI4XBHjd76L//Ofmxxs2OB5hAZw86Qgz4FiknJwMhQtDWhrExDg+f/So41eDBreNKXmHNpIUEZEcryp/+Qm9TRvYu9cxU1OyJIweDVu3wo8/gpsblCoFM2Y4rl2zBhYuBJsNfHwcr58bBly9CrVqOa4JCIDFix0LmK+NVct13564gI6WEBGRHC8F8Aeuumi8AGAu0NxF44k5dLSEiIjkap5ANReOlwK86MLxxHoKPCIikisMAQq4YBw3oAlQ2AVjieso8IiISK7QCPB1wTjewGAXjCOupcAjIiK5gjswD2tDjxfQNC2NqqkZ2vVHchG9pSUiIrlGMBACrAOSLajvB4w+d47jf/5JYGAg3t7eeHh44O3tja+vK+aXxCqa4RERkVxlJvAwYDO5ri+wCXi8WDEKFCiAm5sbNpsNX19f3G/ZgVlyJwUeERHJVQJwnGT+KBnYjDAj/vwTPxxhpybg5uZGkSJF8Pb2Ji0tjStXruDhoQciuZ0Cj4iI5DpFgINAGFlb0+MLeDVpwtgNG6h7y+f9/f1JS0sjICAAb29vzp07R2JiYlZalmymwCMiIrlSALAcWAkUJXOvrPvieBvrFeCzGTP4oH9//nPL0RSGYVCwYEHsdjv+/v4ULlyYuLg4Ll68yL027JWcS4FHRERytcbAb8AioAaOt3ECcIQaN8DAsXFhwLXfSwCjgf8C44FqVasycuRIIiMjiY+Pv1HX29sbLy/HQzObzUbRokWx2+2cP3+eVL3FlevoaAkREclTkoBjwLdADJCG4+2rijjO5Cpyh3vsdju9e/cmOTmZOXPmYBjGXevHx8cTFxdHwYIF8fHxMf8bEKfd62gJBR4REREgMTGR+vXr065dO3r27HnPa69evUpsbCxeXl4EBATcMyCJ6+gsLRERkfvw8fFh8eLFTJw4kS+//PKe115/xJWenq5HXLmEAo+IiMg1Dz/8MLNnz6Zjx46cPn36ntcahkGhQoXw9fXl/Pnzeosrh1PgERERuUXdunUZOHAgrVq14sqVK/e93s/P78ZbXJcuXdJbXDmUAo+IiMhf9OzZk6pVq9KtWzfS09Pve70eceV8CjwiIiJ/YRgGkydPJiYmhrfffjvD99z6iCs52YrTvsRZCjwiIiJ34OXlxZIlS1i+fDmrV6/O8H1+fn4EBQURGxt7274+kr0UeERERO6iaNGiLFu2jMGDB3P48OEM3+fp6UnRokVJSEjQ7sw5hAKPiIjIPVSqVIn333+ftm3bcvbs2Qzf5+7uTpEiRbDb7cTExJCWlmZhl3I/CjwiIiL3ERISQlRUFG3atCEpKSnD911f1+Pl5cX58+e5evWqhV3KvSjwiIiIZMDQoUN56KGH6Nu3b6YfURUoUICCBQsSExOj/XqyiQKPiIhIBhiGwYwZM/jhhx+YMmVKpu/39vamSJEiXL58mcuXL1vQodyLAo+IiEgG+fj4sGzZMmbNmpWpN7eu8/DwoGjRoly9epWYmJgM7fEj5lDgERERyYSSJUuyevVqhgwZwmeffZbp+93c3ChcuDAeHh6cP39ei5ldRIFHREQkkypUqMCCBQvo1KkTx44dc6pGwYIF8fPz02JmF1HgERERcUKtWrWYNGkSLVu25MyZM07V8PPzu7GYWTszW8sjuxsQERHJrZo1a8bZs2dp1qwZO3bsoFChQpmu4e3tTVBQEBcuXCAgIABfX18LOhXN8IiIiGRB7969adCgAa1atXL6lXNPT0+KFClCXFwccXFxJncooMAjIiKSZWPHjqVkyZJ069bN6UXI19/gSkpK4uLFiyZ3KAo8IiIiWeTm5sb06dO5dOkSQ4cOdfrsLDc3N4oUKUJaWhoXLlzQGVwmUuARERExwfXT1b/66iuio6OdrmMYBkFBQbi5uXH+/Hnt1WMSBR4RERGTBAQEsGbNGubMmcPSpUudrmMYBoGBgXh7e3Pu3DlSU1NN7DJ/0ltaIiIiJipevDjr1q2jSZMm+Pr6EhYW5nStAgUK4O7uTkxMDEFBQdhsNhM7zV8UeERERExWpkwZ1qxZQ1hYGF5eXgQHBztdy9fXF8MwiImJoXDhwgo9TtIjLREREQs8+eSTrFq1ij59+rBr164s1fLx8SEwMJCYmBhSUlJM6jB/UeARERGxSNWqVVm6dCldu3bl888/z1Itb29vChUqxIULF7QrsxMUeERERCxUvXp1FixYQIcOHThw4ECWanl5eREUFERsbCxJSUkmdZg/KPCIiIhYrHbt2syePZs2bdpw+PDhLNXy9PQkKCiIixcvOr2zc36kwCMiIuIC9erVY+rUqbRs2ZLvv/8+S7U8PT0pXLgwly5dUujJIL2lJSIi4iKNGzcmOTmZZs2asWnTJsqWLet0LZvNRpEiRYiJiSE9PR0/Pz8TO817FHhERERcqFmzZiQnJxMaGsrWrVt59NFHna7l4eFB4cKFiYmJwW634+/vb2KneYsCj4iIiIu1adOGpKQkmjRpwqZNm3jsscecruXh4XFjpgdQ6LkLBR4REZFs0KlTJwzDoFGjRqxfv54nnnjC6Vru7u4ULlyY8+fPYxiGHm/dgQKPiIhINunYsSO+vr40bdqUNWvW8NRTTzldy93dnSJFitwIPb6+viZ2mvsp8IiIiGSjiIgIvL29CQ8PZ/ny5Tz33HNO17o+0xMTE4NhGPj4+JjYae6m19JFRESyWUhICDNmzKBVq1ZZ3pH5+kLmS5cuaXPCWyjwiIiI5AANGjS4sSPzzp07s1Treui5ePGiQs81CjwiIiI5RO3atVm+fDndu3dn06ZNWapls9lu7Miss7e0hkdERCRHqVatGuvWraN58+YkJSXRsmVLp2tdP4biwoULBAUF4enpaWKnuYsCj4iISA7z1FNPsXHjRsLDw0lMTKRDhw5O1/L09Lxxynp+Dj0KPCIiIjlQ+fLl2bJlCyEhIVy+fJlXXnnF6VpeXl4EBgZy4cIFChcujM1mM7HT3EFreERERHKo0qVLs2PHDhYuXMiwYcNIT093upa3tzcFCxbkwoULpKWlmdhl7qDAIyIikoOVLFmSHTt2cOTIETp16pSlt658fHzw9/e/ceBofqLAIyIiksMFBgaybt06DMMgLCyM2NhYp2v5+fnh7e3NhQsXsNvtJnaZsynwiIiI5ALe3t7MnTuXqlWr0qBBA86cOeN0rYCAANzd3bMUnHIbBR4REZFcws3NjXHjxtGpUyfq16/P0aNHna4VGBiI3W7n0qVLJnaYcynwiIiI5DK9e/fm3XffJSwsjN27dztVwzAMgoKCSElJIS4uzuQOcx4FHhERkVwoLCyMxYsX061bN5YuXepUDcMwKFy4MAkJCSQkJJjcYc6ifXhERERyqZo1a7J161aaN2/Of//7XwYNGoRhGJmq4ebmRuHChTl//jzu7u54eXlZ1G320gyPiIhILla2bFl27tzJxo0b6d27t1PnZnl4eBAUFERsbCxXr161oMvsp8AjIiKSy/3tb39j27ZtXLlyhUaNGnH27NlM1/D09CQwMJCYmBhSU1Mt6DJ7KfCIiIjkAX5+fixYsIAGDRpQp04dvv3220zX8Pb2pkCBAnlyjx4FHhERkTzCzc2NoUOHMnHiRJo3b86KFSsyXcPPzw8vL688t0ePAo+IiEge06RJE7Zs2cLbb7/NiBEjMn12VkBAAHa7ncuXL1vUoesp8IiIiORB5cuXZ+/evRw5coSWLVtmaoNBwzAoVKgQiYmJJCYmWtil6yjwiIiI5FFBQUGsW7eO0qVLU6dOHf79739n+N7rr6tfunSJlJQUC7t0DQUeERGRPMzDw4MJEyYwYMAAGjZsyI4dOzJ1b2BgILGxsZl+LJbTKPCIiIjkA+3bt2fZsmX07duXSZMmkZ6enqH7vL298fPzy/VvbinwiIiI5BPVqlVj9+7dbN++nRYtWhATE5Oh+/z9/fHw8ODixYsWd2gdBR4REZF8pESJEmzdupWKFStSq1Yt9u/fn6H7AgMDSU1NzbUHjSrwiIiI5DM2m40xY8YQHR1Nu3btmDJlyn0fcV0/XT0hIYGkpCQXdWoeBR4REZF86uWXX2bv3r1s3ryZiIiI+z7icnd3JygoiIsXL+a64ycUeERERPKxkiVL8sknn/DEE09k6BGXzWYjICAg1y1i9sjuBkRERCR72Ww2xo4dS82aNYmMjKRfv3707dsXN7c7z4v4+vqSkpLCxYsXKVSo0P983Q78DHwD/Au4jGOG5QGgClAVCLLoe7kb417p7JlnnrEfOnTIhe2IiIhIdvr111+JiooiKCiImTNnEhR052hit9s5f/48vr6++Pn5AfBfYDowDUjGMasShyMAAdgAHyARKAMMBVoC3ib1bhjGN3a7/Zk7fU2PtEREROSGkiVLsm3bNsqUKcPzzz/PV199BcDWrVtZt27djeuuL2KOi4sjNiWFXkBpYBIQCyTgmNm5dVrl6rXPXcUx89Mbx6zP0r9cZwUFHhEREbmNzWbj7bffZvLkyURFRTFmzBgeeOABhgwZwpUrV25c5+7uzonAQMpdvMgCIAnHzE5GXcERgLoBwUDGdgVyjh5piYiIyF39+eef9O7dm3PnzlG8eHEqVqzIiBEjANgMtAIS7HYwjCyNYwMeBL669rsz9EhLREREnHL48GHCwsKoUaMG+/btIzo6mt9++43dQASOR1dZDTvgeMz1X6AmcCHL1f6XAo+IiIjcVVxcHHv37uXgwYMYhsGlS5fo1KMHzXEsPjZTKvA70MnkuqBHWiIiIpIJ58+fp4O7O7sKFSLFojF8gUVAs0zep0daIiIiYoofixRhn4VhBxyPybrhmPExiwKPiIiIZNgErq3bsdhVYJOJ9RR4REREJEP+BHZwy545Dz8MTz4JlSvDM9eeJLVq5fhz5cqOr1eu7Ph8Sgp06uS4/qmnYO/em4VXrIBKlaBCBRg6FHBsWDjexN51tISIiIhkyF4cr4/fdlb6nj1QpMjNP69YcfPj116DggUdH8+e7fj92DH4808IDob/9/8gNhYGD4ZvvoGiRSEqCnbtgpde4lsc+/p4mdC7ZnhEREQkQw7g2CwwQ+x2WLkS2rRx/Pn4cahb1/FxsWIQGAiHDsGpU/D4446wA1CvHqxZAziOnDhmUu8KPCIiIpIhX/KXIyAMAxo0gKefhlmzbr/488/hgQccYQYcj7E2boTUVPj5Z8eMzpkzULo0/Pgj/PKL42vr1zs+D6QBh03qXY+0REREJENi//qJL76AEiUcj6jq14dy5aB2bcfXli27ObsD0LkznDjhWOtTqhTUrAnu7lCoEEyf7lj74+bm+PxPPwGQcqcxnaTAIyIiIhnyPzv3lSjh+L1YMQgPh4MHHYEnNRXWrnXM4lzn4QHR0Tf/XLMmlCnj+LhpU8cvcMwUubvfuCzdpN71SEtEREQyxP/WP8THQ1zczY937ICKFR1/3rnTMdtTsuTN6xMSHNcBfPqpIwCVL+/4859/On6PjYVp06BrV8CxQLqASb1rhkdEREQy5FluWVPzxx+OWR1wzOi0bQsNGzr+vHz57Y+zwBFqXn7Z8diqRAlYtOjm1159FY4ccXw8cuSNmR8bUMmk3nW0hIiIiGTIAuAVMvGmVhbZcBwk6n+/C6/R0RIiIiKSZbUw97iH+3mEjIed+1HgERERkQx5FKjsorH8gddMrKfAIyIiIhk2FPNmXe4lHYg0sZ4Cj4iIiGRYCPA41gYIP2D0td/NosAjIiIiGeYGrMCc863uVv9xYIAFdUVEREQy7HFgCuBrQe0CwCrA/X4XZpICj4iIiGRad2Ak5oUeAwgAdgOlTap5KwUeERERccpQYCqO0JOVGRlf4CHga6CqCX3diQKPiIiIOK0z8D3wNJl/e8sG+AC9gB+B8ua2dhsFHhEREcmSR4D9wDqgPuCN4/HUnUKG17Wv+eMIOkeBSVi3CPo6naUlIiIiWWYA9a79+gNHADoIHMBxFIUb8ABQG8ds0LM4ZndcRYFHRERETPUAEHrtV06hR1oiIiKS5ynwiIiISJ6nwCMiIiJ5ngKPiIiI5HkKPCIiIpLnKfCIiIhInqfAIyIiInmeAo+IiIjkeQo8IiIikucZdrv97l80jHPA/7muHRERERGnlbLb7UXv9IV7Bh4RERGRvECPtERERCTPU+ARERGRPE+BR0RERPI8BR4RERHJ8xR4REREJM/7//9GnPX1qQ6TAAAAAElFTkSuQmCC\n",
            "text/plain": [
              "<Figure size 720x720 with 1 Axes>"
            ]
          },
          "metadata": {}
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "RHj6MNDblFvF"
      },
      "source": [
        "#### For further details: https://pytorch-geometric.readthedocs.io/en/latest/_modules/torch_geometric/nn/models/gnn_explainer.html"
      ]
    }
  ]
}