Node Classification.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Node Classification\n",
    "\n",
    "In this notebook we consider the problem of node classification for the NDSSL data. We will predict the worker attribute (which is 0 for unemployed and 1 for employed) with 3 different models:\n",
    "- Multi-Layer Perceptron (MLP)\n",
    "- Graph Convolutional Network (GCN)\n",
    "- GraphSAGE\n",
    "\n",
    "The results of this notebook are used in Sec. 5.1 of the report."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from imports import *\n",
    "from nodepred import *\n",
    "from dataset import *\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "import matplotlib\n",
    "import matplotlib.style as style \n",
    "style.use('seaborn-paper')\n",
    "%matplotlib inline\n",
    "\n",
    "fontsize = 12\n",
    "plt.rcParams.update({\n",
    "    'font.size': fontsize, \n",
    "    'axes.labelsize': fontsize, \n",
    "    'legend.fontsize': fontsize,\n",
    "    'xtick.labelsize': fontsize,\n",
    "    'ytick.labelsize': fontsize,\n",
    "    'axes.titlesize': fontsize\n",
    "                    })"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<torch._C.Generator at 0x7f76bc042f70>"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "## set random seeds\n",
    "random.seed(123)\n",
    "np.random.seed(123)\n",
    "torch.manual_seed(123)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1.5.1\n",
      "device = cuda\n"
     ]
    }
   ],
   "source": [
    "print(torch.__version__)\n",
    "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
    "print('device =', device)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Load the Dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "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>person_id</th>\n",
       "      <th>household_id</th>\n",
       "      <th>age</th>\n",
       "      <th>gender</th>\n",
       "      <th>worker</th>\n",
       "      <th>relationship</th>\n",
       "      <th>household_income</th>\n",
       "      <th>household_size</th>\n",
       "      <th>zipcode</th>\n",
       "      <th>household_vehicles</th>\n",
       "      <th>household_workers</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>2509159</td>\n",
       "      <td>2201175</td>\n",
       "      <td>42</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>13</td>\n",
       "      <td>3</td>\n",
       "      <td>97201</td>\n",
       "      <td>3</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>2509160</td>\n",
       "      <td>2201175</td>\n",
       "      <td>43</td>\n",
       "      <td>2</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>13</td>\n",
       "      <td>3</td>\n",
       "      <td>97201</td>\n",
       "      <td>3</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>2509161</td>\n",
       "      <td>2201175</td>\n",
       "      <td>17</td>\n",
       "      <td>1</td>\n",
       "      <td>2</td>\n",
       "      <td>2</td>\n",
       "      <td>13</td>\n",
       "      <td>3</td>\n",
       "      <td>97201</td>\n",
       "      <td>3</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>2509162</td>\n",
       "      <td>2201176</td>\n",
       "      <td>41</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>11</td>\n",
       "      <td>2</td>\n",
       "      <td>97201</td>\n",
       "      <td>3</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>2509163</td>\n",
       "      <td>2201176</td>\n",
       "      <td>11</td>\n",
       "      <td>1</td>\n",
       "      <td>2</td>\n",
       "      <td>2</td>\n",
       "      <td>11</td>\n",
       "      <td>2</td>\n",
       "      <td>97201</td>\n",
       "      <td>3</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   person_id  household_id  age  gender  worker  relationship  \\\n",
       "0    2509159       2201175   42       1       1             1   \n",
       "1    2509160       2201175   43       2       1             1   \n",
       "2    2509161       2201175   17       1       2             2   \n",
       "3    2509162       2201176   41       1       1             1   \n",
       "4    2509163       2201176   11       1       2             2   \n",
       "\n",
       "   household_income  household_size  zipcode  household_vehicles  \\\n",
       "0                13               3    97201                   3   \n",
       "1                13               3    97201                   3   \n",
       "2                13               3    97201                   3   \n",
       "3                11               2    97201                   3   \n",
       "4                11               2    97201                   3   \n",
       "\n",
       "   household_workers  \n",
       "0                  2  \n",
       "1                  2  \n",
       "2                  2  \n",
       "3                  1  \n",
       "4                  1  "
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "node_attributes = pd.read_csv('data/NDSSL data/raw/node_attributes.csv')\n",
    "node_attributes.head(5)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Processing...\n",
      "{'x': None, 'edge_index': tensor([[      0,       0,       1,  ..., 1486224, 1378614, 1556530],\n",
      "        [      1,       2,       2,  ..., 1601329, 1601329, 1601329]]), 'edge_attr': None, 'y': None, 'pos': None, 'norm': None, 'face': None, 'edge_weight': tensor([10.9161, 12.7494, 12.5828,  ...,  0.0497,  0.1667,  0.1667])}\n",
      "Done!\n",
      "{'x': None, 'edge_index': tensor([[      0,       0,       1,  ..., 1486224, 1378614, 1556530],\n",
      "        [      1,       2,       2,  ..., 1601329, 1601329, 1601329]]), 'edge_attr': None, 'y': None, 'pos': None, 'norm': None, 'face': None, 'edge_weight': tensor([10.9161, 12.7494, 12.5828,  ...,  0.0497,  0.1667,  0.1667])}\n"
     ]
    }
   ],
   "source": [
    "## remove old processed files\n",
    "import os, shutil\n",
    "if os.path.exists('data/NDSSL data/processed'):\n",
    "    shutil.rmtree('data/NDSSL data/processed')\n",
    "\n",
    "dataset = load_dataset(dataset_name='NDSSL')\n",
    "data = dataset[0]\n",
    "dataset.process()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "torch.Size([1601330, 141])"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data_columns = ['age', 'gender', 'zipcode', 'relationship',  \n",
    "                'household_workers', 'household_income', 'household_size', 'household_vehicles']\n",
    "data.x = dataframe2onehot(node_attributes[data_columns], node_attributes)\n",
    "data.x.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([1, 2])"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.unique(node_attributes['worker'].values)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "torch.Size([1601330])"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data.y = torch.LongTensor(node_attributes['worker'].values - 1)\n",
    "data.y = data.y.type(torch.int64).reshape((len(node_attributes), 1))[:,0]\n",
    "data.y.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 1, 1])"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data.y[0:20]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "2"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "num_classes = len(torch.unique(data.y))\n",
    "num_classes"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Make the train/val/test split. We could pick random subsets of the nodes, or make the split by household."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "number of distinct households: 632626\n",
      "number of training households: 506100\n",
      "number of validation households: 126526\n",
      "number of testing households: 0\n",
      "size of training set: 1280901\n",
      "size of validation set: 320429\n",
      "size of testing set: 0\n"
     ]
    }
   ],
   "source": [
    "split_by_household = True\n",
    "\n",
    "if not split_by_household:\n",
    "    ## create a train/val/test split\n",
    "    node_ids = list(set(list(node_attributes['person_id'])))\n",
    "    np.random.shuffle(node_ids)\n",
    "\n",
    "    ntrain = int(0.8 * len(node_ids))\n",
    "    nval = len(node_ids) - ntrain #int(0.2 * len(household_ids))\n",
    "    ntest = len(node_ids) - (ntrain + nval)\n",
    "\n",
    "    node_ids_train = node_ids[:ntrain]\n",
    "    node_ids_val = node_ids[ntrain:ntrain + nval]\n",
    "    node_ids_test = node_ids[ntrain + nval:]\n",
    "\n",
    "    train_mask = torch.IntTensor(node_attributes['person_id'].isin(node_ids_train)).type(torch.int64)\n",
    "    val_mask = torch.IntTensor(node_attributes['person_id'].isin(node_ids_val)).type(torch.int64)\n",
    "    test_mask = torch.IntTensor(node_attributes['person_id'].isin(node_ids_test)).type(torch.int64)\n",
    "\n",
    "    train_idx = np.arange(len(node_attributes))[train_mask == 1]\n",
    "    val_idx = np.arange(len(node_attributes))[val_mask == 1]\n",
    "    test_idx = np.arange(len(node_attributes))[test_mask == 1]\n",
    "\n",
    "    print('size of training set: %i' %torch.sum(train_mask).item())\n",
    "    print('size of validation set: %i' %torch.sum(val_mask).item())\n",
    "    print('size of testing set: %i' %torch.sum(test_mask).item())\n",
    "else:\n",
    "    ## create a train/val/test split by household\n",
    "    household_ids = list(set(list(node_attributes['household_id'])))\n",
    "    np.random.shuffle(household_ids)\n",
    "\n",
    "    ntrain_households = int(0.8 * len(household_ids))\n",
    "    nval_households = len(household_ids) - ntrain_households #int(0.2 * len(household_ids))\n",
    "    ntest_households = len(household_ids) - (ntrain_households + nval_households)\n",
    "\n",
    "    households_train = household_ids[:ntrain_households]\n",
    "    households_val = household_ids[ntrain_households:ntrain_households + nval_households]\n",
    "    households_test = household_ids[ntrain_households + nval_households:]\n",
    "\n",
    "    print('number of distinct households: %i' %len(household_ids))\n",
    "    print('number of training households: %i' %len(households_train))\n",
    "    print('number of validation households: %i' %len(households_val))\n",
    "    print('number of testing households: %i' %len(households_test))\n",
    "\n",
    "    train_mask = torch.IntTensor(node_attributes['household_id'].isin(households_train)).type(torch.bool)\n",
    "    val_mask = torch.IntTensor(node_attributes['household_id'].isin(households_val)).type(torch.bool)\n",
    "    test_mask = torch.IntTensor(node_attributes['household_id'].isin(households_test)).type(torch.bool)\n",
    "\n",
    "    train_idx = np.arange(len(node_attributes))[train_mask == 1]\n",
    "    val_idx = np.arange(len(node_attributes))[val_mask == 1]\n",
    "    test_idx = np.arange(len(node_attributes))[test_mask == 1]\n",
    "\n",
    "    print('size of training set: %i' %torch.sum(train_mask).item())\n",
    "    print('size of validation set: %i' %torch.sum(val_mask).item())\n",
    "    print('size of testing set: %i' %torch.sum(test_mask).item())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "data.train_mask = train_mask\n",
    "data.val_mask = val_mask\n",
    "data.test_mask = test_mask"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Break down the number of edges by train/test:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "check: True\n",
      "fraction of train-train edges: 0.65\n",
      "fraction of train-test edges: 0.31\n",
      "fraction of test-test edges: 0.05\n"
     ]
    }
   ],
   "source": [
    "edge_tuple_list = tuple(map(tuple, np.transpose(data.edge_index.numpy())))\n",
    "edge_tuple_array = np.asarray(edge_tuple_list)\n",
    "\n",
    "edge_idx_train0 = np.isin(edge_tuple_array[:,0], train_idx)\n",
    "edge_idx_train1 = np.isin(edge_tuple_array[:,1], train_idx)\n",
    "\n",
    "Num_edges_train_train = np.sum(edge_idx_train0 & edge_idx_train1)\n",
    "Num_edges_test_test = np.sum(~edge_idx_train0 & ~edge_idx_train1)\n",
    "Num_edges_train_test = np.sum(~edge_idx_train0 & edge_idx_train1) \\\n",
    "                        + np.sum(edge_idx_train0 & ~edge_idx_train1)\n",
    "\n",
    "print('check:', edge_idx_train1.shape[0] \\\n",
    "      == (Num_edges_train_train + Num_edges_test_test + Num_edges_train_test))\n",
    "\n",
    "print('fraction of train-train edges: %.3f' % (Num_edges_train_train/data.edge_index.shape[1]))\n",
    "print('fraction of train-test edges: %.3f' % (Num_edges_train_test/data.edge_index.shape[1]))\n",
    "print('fraction of test-test edges: %.3f' % (Num_edges_test_test/data.edge_index.shape[1]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "batch_size = 10000\n",
    "n_epochs = 50"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Node classification: MLP "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "full_set = torch.utils.data.TensorDataset(data.x, data.y)\n",
    "train_set = torch.utils.data.Subset(full_set, train_idx)\n",
    "val_set = torch.utils.data.Subset(full_set, val_idx)\n",
    "#test_set = torch.utils.data.Subset(full_set, test_idx)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True)\n",
    "val_loader = torch.utils.data.DataLoader(val_set, batch_size=batch_size, shuffle=True)\n",
    "#test_loader = torch.utils.data.DataLoader(test_set, batch_size=batch_size, shuffle=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "number of hidden units = 256\n",
      "number of trainable parameters 168450\n"
     ]
    }
   ],
   "source": [
    "h = 256\n",
    "print('number of hidden units = %i' %h)\n",
    "\n",
    "model = mlp(data.x.shape[1], h, num_classes).to(device)\n",
    "print('number of trainable parameters %i' %model_params(model))\n",
    "\n",
    "optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "def train():\n",
    "    model.train()\n",
    "\n",
    "    total_loss = total_examples = 0\n",
    "    for x_batch, y_batch in train_loader:\n",
    "        x_batch = x_batch.to(device)\n",
    "        y_batch = y_batch.to(device)\n",
    "        optimizer.zero_grad()\n",
    "        out = model(x_batch)\n",
    "        loss = F.nll_loss(out, y_batch)\n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "        total_loss += loss.item() * x_batch.shape[0]\n",
    "        total_examples += x_batch.shape[0]\n",
    "    \n",
    "    return total_loss/total_examples\n",
    "\n",
    "@torch.no_grad()\n",
    "def evaluator(data_loader):\n",
    "    model.eval()\n",
    "    \n",
    "    y_target = torch.empty(0)\n",
    "    y_pred = torch.empty(0)\n",
    "    loss = total_examples = 0.0\n",
    "    for x_batch, y_batch in data_loader:\n",
    "        x_batch,  y_batch = x_batch.to(device), y_batch.to(device)\n",
    "        total_examples += x_batch.shape[0]\n",
    "        \n",
    "        z = model(x_batch)\n",
    "        pred = 1.0*torch.argmax(nn.Softmax(dim=1)(z), axis=1).cpu()\n",
    "        loss += F.nll_loss(z, y_batch).item() * x_batch.shape[0]\n",
    "\n",
    "        y_target = torch.cat((y_target, 1.0 * y_batch.cpu()), 0)\n",
    "        y_pred = torch.cat((y_pred, pred), 0)\n",
    "        \n",
    "    y_target = y_target.numpy()\n",
    "    y_pred = y_pred.numpy()\n",
    "    \n",
    "    return loss/total_examples, y_target, y_pred"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train batch length: 1280901, val batch length: 320429\n",
      "epoch: 0||train loss: 0.6900||val loss: 0.6902||train acc: 0.5035||val acc: 0.5025||train auc: 0.5000||val auc: 0.5000\n",
      "epoch: 1||train loss: 0.1913||val loss: 0.1910||train acc: 0.9263||val acc: 0.9259||train auc: 0.9261||val auc: 0.9258\n",
      "epoch: 2||train loss: 0.1584||val loss: 0.1586||train acc: 0.9340||val acc: 0.9331||train auc: 0.9339||val auc: 0.9331\n",
      "epoch: 3||train loss: 0.1475||val loss: 0.1480||train acc: 0.9361||val acc: 0.9353||train auc: 0.9360||val auc: 0.9352\n",
      "epoch: 4||train loss: 0.1434||val loss: 0.1442||train acc: 0.9374||val acc: 0.9365||train auc: 0.9373||val auc: 0.9365\n",
      "epoch: 5||train loss: 0.1380||val loss: 0.1395||train acc: 0.9391||val acc: 0.9381||train auc: 0.9391||val auc: 0.9380\n",
      "epoch: 6||train loss: 0.1364||val loss: 0.1379||train acc: 0.9407||val acc: 0.9394||train auc: 0.9406||val auc: 0.9393\n",
      "epoch: 7||train loss: 0.1329||val loss: 0.1347||train acc: 0.9422||val acc: 0.9411||train auc: 0.9421||val auc: 0.9410\n",
      "epoch: 8||train loss: 0.1325||val loss: 0.1350||train acc: 0.9422||val acc: 0.9409||train auc: 0.9420||val auc: 0.9407\n",
      "epoch: 9||train loss: 0.1280||val loss: 0.1308||train acc: 0.9448||val acc: 0.9435||train auc: 0.9447||val auc: 0.9434\n",
      "epoch: 10||train loss: 0.1252||val loss: 0.1281||train acc: 0.9458||val acc: 0.9440||train auc: 0.9457||val auc: 0.9440\n",
      "epoch: 11||train loss: 0.1225||val loss: 0.1258||train acc: 0.9474||val acc: 0.9461||train auc: 0.9473||val auc: 0.9461\n",
      "epoch: 12||train loss: 0.1222||val loss: 0.1258||train acc: 0.9477||val acc: 0.9464||train auc: 0.9475||val auc: 0.9463\n",
      "epoch: 13||train loss: 0.1186||val loss: 0.1222||train acc: 0.9492||val acc: 0.9478||train auc: 0.9491||val auc: 0.9478\n",
      "epoch: 14||train loss: 0.1162||val loss: 0.1198||train acc: 0.9512||val acc: 0.9493||train auc: 0.9511||val auc: 0.9492\n",
      "epoch: 15||train loss: 0.1133||val loss: 0.1173||train acc: 0.9513||val acc: 0.9495||train auc: 0.9512||val auc: 0.9494\n",
      "epoch: 16||train loss: 0.1119||val loss: 0.1157||train acc: 0.9525||val acc: 0.9505||train auc: 0.9525||val auc: 0.9505\n",
      "epoch: 17||train loss: 0.1131||val loss: 0.1169||train acc: 0.9522||val acc: 0.9504||train auc: 0.9521||val auc: 0.9503\n",
      "epoch: 18||train loss: 0.1067||val loss: 0.1109||train acc: 0.9545||val acc: 0.9526||train auc: 0.9544||val auc: 0.9525\n",
      "epoch: 19||train loss: 0.1075||val loss: 0.1117||train acc: 0.9544||val acc: 0.9524||train auc: 0.9543||val auc: 0.9523\n",
      "epoch: 20||train loss: 0.1038||val loss: 0.1082||train acc: 0.9557||val acc: 0.9535||train auc: 0.9556||val auc: 0.9534\n",
      "epoch: 21||train loss: 0.1032||val loss: 0.1075||train acc: 0.9559||val acc: 0.9540||train auc: 0.9559||val auc: 0.9539\n",
      "epoch: 22||train loss: 0.1010||val loss: 0.1057||train acc: 0.9563||val acc: 0.9543||train auc: 0.9562||val auc: 0.9543\n",
      "epoch: 23||train loss: 0.0999||val loss: 0.1046||train acc: 0.9575||val acc: 0.9555||train auc: 0.9574||val auc: 0.9554\n",
      "epoch: 24||train loss: 0.0974||val loss: 0.1024||train acc: 0.9583||val acc: 0.9562||train auc: 0.9582||val auc: 0.9561\n",
      "epoch: 25||train loss: 0.0965||val loss: 0.1016||train acc: 0.9589||val acc: 0.9566||train auc: 0.9589||val auc: 0.9566\n",
      "epoch: 26||train loss: 0.0957||val loss: 0.1006||train acc: 0.9590||val acc: 0.9565||train auc: 0.9589||val auc: 0.9564\n",
      "epoch: 27||train loss: 0.0937||val loss: 0.0987||train acc: 0.9598||val acc: 0.9573||train auc: 0.9597||val auc: 0.9573\n",
      "epoch: 28||train loss: 0.0928||val loss: 0.0978||train acc: 0.9607||val acc: 0.9583||train auc: 0.9607||val auc: 0.9583\n",
      "epoch: 29||train loss: 0.0907||val loss: 0.0958||train acc: 0.9617||val acc: 0.9593||train auc: 0.9617||val auc: 0.9592\n",
      "epoch: 30||train loss: 0.0905||val loss: 0.0954||train acc: 0.9617||val acc: 0.9592||train auc: 0.9617||val auc: 0.9592\n",
      "epoch: 31||train loss: 0.0883||val loss: 0.0936||train acc: 0.9624||val acc: 0.9600||train auc: 0.9624||val auc: 0.9600\n",
      "epoch: 32||train loss: 0.0874||val loss: 0.0928||train acc: 0.9629||val acc: 0.9605||train auc: 0.9629||val auc: 0.9604\n",
      "epoch: 33||train loss: 0.0874||val loss: 0.0927||train acc: 0.9629||val acc: 0.9607||train auc: 0.9628||val auc: 0.9606\n",
      "epoch: 34||train loss: 0.0863||val loss: 0.0917||train acc: 0.9629||val acc: 0.9606||train auc: 0.9628||val auc: 0.9605\n",
      "epoch: 35||train loss: 0.0850||val loss: 0.0907||train acc: 0.9632||val acc: 0.9607||train auc: 0.9632||val auc: 0.9606\n",
      "epoch: 36||train loss: 0.0833||val loss: 0.0888||train acc: 0.9650||val acc: 0.9621||train auc: 0.9649||val auc: 0.9620\n",
      "epoch: 37||train loss: 0.0832||val loss: 0.0885||train acc: 0.9652||val acc: 0.9621||train auc: 0.9651||val auc: 0.9621\n",
      "epoch: 38||train loss: 0.0821||val loss: 0.0875||train acc: 0.9649||val acc: 0.9621||train auc: 0.9648||val auc: 0.9621\n",
      "epoch: 39||train loss: 0.0820||val loss: 0.0878||train acc: 0.9653||val acc: 0.9625||train auc: 0.9652||val auc: 0.9624\n",
      "epoch: 40||train loss: 0.0803||val loss: 0.0858||train acc: 0.9660||val acc: 0.9634||train auc: 0.9660||val auc: 0.9634\n",
      "epoch: 41||train loss: 0.0787||val loss: 0.0848||train acc: 0.9666||val acc: 0.9634||train auc: 0.9666||val auc: 0.9634\n",
      "epoch: 42||train loss: 0.0783||val loss: 0.0839||train acc: 0.9671||val acc: 0.9642||train auc: 0.9670||val auc: 0.9641\n",
      "epoch: 43||train loss: 0.0779||val loss: 0.0835||train acc: 0.9679||val acc: 0.9650||train auc: 0.9679||val auc: 0.9650\n",
      "epoch: 44||train loss: 0.0769||val loss: 0.0829||train acc: 0.9676||val acc: 0.9648||train auc: 0.9675||val auc: 0.9648\n",
      "epoch: 45||train loss: 0.0767||val loss: 0.0825||train acc: 0.9676||val acc: 0.9647||train auc: 0.9676||val auc: 0.9646\n",
      "epoch: 46||train loss: 0.0764||val loss: 0.0822||train acc: 0.9680||val acc: 0.9649||train auc: 0.9679||val auc: 0.9649\n",
      "epoch: 47||train loss: 0.0751||val loss: 0.0807||train acc: 0.9685||val acc: 0.9655||train auc: 0.9685||val auc: 0.9655\n",
      "epoch: 48||train loss: 0.0742||val loss: 0.0800||train acc: 0.9690||val acc: 0.9659||train auc: 0.9690||val auc: 0.9659\n",
      "epoch: 49||train loss: 0.0744||val loss: 0.0802||train acc: 0.9688||val acc: 0.9655||train auc: 0.9688||val auc: 0.9655\n",
      "epoch: 50||train loss: 0.0736||val loss: 0.0794||train acc: 0.9694||val acc: 0.9662||train auc: 0.9693||val auc: 0.9662\n"
     ]
    }
   ],
   "source": [
    "loss_train_mlp = []\n",
    "acc_train_mlp = []\n",
    "auc_train_mlp = []\n",
    "ap_train_mlp = []\n",
    "loss_val_mlp = []\n",
    "acc_val_mlp = []\n",
    "auc_val_mlp = []\n",
    "ap_val_mlp = []\n",
    "\n",
    "for epoch in range(n_epochs+1):\n",
    "\n",
    "    # evaluate on both train and val sets\n",
    "    loss_train, y_target_train, y_pred_train = evaluator(train_loader)    \n",
    "    loss_val, y_target_val, y_pred_val = evaluator(val_loader)\n",
    "    \n",
    "    if epoch == 0:\n",
    "        print('train batch length: %i, val batch length: %i' %(len(y_pred_train), len(y_pred_val)))\n",
    "    \n",
    "    # compute the accuracy, average precision, AUC\n",
    "    acc_train = np.mean(y_pred_train == y_target_train)\n",
    "    auc_train = roc_auc_score(y_target_train, y_pred_train)\n",
    "    ap_train = average_precision_score(y_target_train, y_pred_train)\n",
    "    roc_train = roc_curve(y_target_train, y_pred_train)\n",
    "\n",
    "    acc_val = np.mean(y_pred_val == y_target_val)\n",
    "    auc_val = roc_auc_score(y_target_val, y_pred_val)\n",
    "    ap_val = average_precision_score(y_target_val, y_pred_val)\n",
    "    roc_val = roc_curve(y_target_val, y_pred_val)\n",
    "\n",
    "    # record the results\n",
    "    loss_train_mlp.append(loss_train)\n",
    "    acc_train_mlp.append(acc_train)\n",
    "    auc_train_mlp.append(auc_train)\n",
    "    ap_train_mlp.append(ap_train)\n",
    "    \n",
    "    loss_val_mlp.append(loss_val)\n",
    "    acc_val_mlp.append(acc_val)\n",
    "    auc_val_mlp.append(auc_val)\n",
    "    ap_val_mlp.append(ap_val)\n",
    "\n",
    "    # print results\n",
    "    print('epoch: %i||train loss: %.4f||val loss: %.4f||train acc: %.4f||val acc: %.4f||train auc: %.4f||val auc: %.4f' \\\n",
    "          % (epoch, loss_train, loss_val, acc_train, acc_val, auc_train, auc_val))\n",
    "    #print('epoch: %i||train loss: %.4f||val loss: %.4f||train acc: %.4f||val acc: %.4f' \\\n",
    "    #      % (epoch, loss_train, loss_val, acc_train, acc_val))\n",
    "    \n",
    "    # train\n",
    "    if epoch != n_epochs:\n",
    "        loss_train = train()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "fig, ax = plt.subplots(1, 2, figsize=(12,4))\n",
    "fig.subplots_adjust(wspace=0.3)\n",
    "\n",
    "ax[0].plot(np.arange(1, len(loss_train_mlp)+1), loss_train_mlp)\n",
    "ax[0].plot(np.arange(1, len(loss_val_mlp)+1), loss_val_mlp)\n",
    "ax[0].set_xlabel('epochs')\n",
    "ax[0].set_ylabel('BCE loss')\n",
    "#ax[0].set_xscale('log')\n",
    "#ax[0].set_yscale('log')\n",
    "#ax[0].set_ylim([np.min(loss_mlp), np.max(loss_mlp)])\n",
    "\n",
    "ax[1].plot(acc_train_mlp)\n",
    "ax[1].plot(acc_val_mlp)\n",
    "#ax[1].plot(ap_mlp, label='AP')\n",
    "ax[1].set_xlabel('epochs')\n",
    "ax[1].set_ylabel('accuracy')\n",
    "#ax[1].set_xscale('log')\n",
    "#ax[1].legend()\n",
    "\n",
    "#plt.savefig(\"figures/node_prediction_mlp.pdf\", bbox_inches='tight')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Node classification: Graph Neural Networks"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Compute GraphSAINT normalization: : 165000372it [01:43, 1588527.13it/s]                             \n"
     ]
    }
   ],
   "source": [
    "# set the number of steps to match what was used for the MLP\n",
    "walk_length = 100\n",
    "num_steps = 10\n",
    "\n",
    "loader = GraphSAINTRandomWalkSampler(data, batch_size=batch_size, walk_length=walk_length,\n",
    "                                     num_steps=num_steps, sample_coverage=100, save_dir=None)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Adjust the number of epochs so that in both the MLP and GNN approaches roughly the same number of samples are visited during training."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [],
   "source": [
    "mlp_epoch_size = 0\n",
    "for batch in train_loader:\n",
    "    mlp_epoch_size += batch[1].size()[0]\n",
    "    \n",
    "gnn_epoch_size = 0\n",
    "for batch in loader:\n",
    "    gnn_epoch_size += batch.y[batch.train_mask].size()[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "67"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "n_epochs = int((mlp_epoch_size/gnn_epoch_size)*n_epochs)\n",
    "n_epochs"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### GCN\n",
    "\n",
    "documentation: https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html?highlight=GCNConv#torch_geometric.nn.conv.GCNConv"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "number of hidden units = 256\n",
      "number of trainable parameters 168450\n"
     ]
    }
   ],
   "source": [
    "h = 256\n",
    "print('number of hidden units = %i' %h)\n",
    "\n",
    "model = GCN(data.x.shape[1], h, num_classes).to(device)\n",
    "model.set_aggr('mean')\n",
    "print('number of trainable parameters %i' %model_params(model))\n",
    "\n",
    "optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "def train():\n",
    "    model.train()\n",
    "\n",
    "    total_loss = total_examples = 0\n",
    "    for batch in loader:\n",
    "        batch = batch.to(device)\n",
    "        optimizer.zero_grad()\n",
    "        out = model(batch.x, batch.edge_index)\n",
    "        batch_size = out[batch.train_mask].shape[0]\n",
    "        loss = F.nll_loss(out[batch.train_mask], batch.y[batch.train_mask])\n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "        total_loss += loss.item() * batch_size\n",
    "        total_examples += batch_size\n",
    "    return total_loss / total_examples\n",
    "\n",
    "@torch.no_grad()\n",
    "def evaluator_train():\n",
    "    model.eval()\n",
    "    \n",
    "    y_target = torch.empty(0)\n",
    "    y_pred = torch.empty(0)\n",
    "    total_loss = total_examples = 0\n",
    "    for batch in loader:\n",
    "        batch = batch.to(device)\n",
    "        out = model(batch.x, batch.edge_index)\n",
    "        batch_size = out[batch.train_mask].shape[0]\n",
    "        total_examples += batch_size\n",
    "        total_loss += F.nll_loss(out[batch.train_mask], batch.y[batch.train_mask]).item() * batch_size\n",
    "        \n",
    "        y_target = torch.cat((y_target, 1.0 * batch.y[batch.train_mask].cpu()), 0)\n",
    "        y_pred = torch.cat((y_pred, 1.0 * torch.argmax(nn.Softmax()(out[batch.train_mask]), axis=1).cpu()), 0)\n",
    "        \n",
    "    y_target = y_target.numpy()\n",
    "    y_pred = y_pred.numpy()\n",
    "    \n",
    "    return total_loss/total_examples, y_target, y_pred\n",
    "\n",
    "@torch.no_grad()\n",
    "def evaluator_val():\n",
    "    model.eval()\n",
    "    \n",
    "    y_target = torch.empty(0)\n",
    "    y_pred = torch.empty(0)\n",
    "    total_loss = total_examples = 0\n",
    "    for batch in loader:\n",
    "        batch = batch.to(device)\n",
    "        out = model(batch.x, batch.edge_index)\n",
    "        batch_size = out[batch.val_mask].shape[0]\n",
    "        total_examples += batch_size\n",
    "        total_loss += F.nll_loss(out[batch.val_mask], batch.y[batch.val_mask]).item() * batch_size\n",
    "        \n",
    "        y_target = torch.cat((y_target, 1.0 * batch.y[batch.val_mask].cpu()), 0)\n",
    "        y_pred = torch.cat((y_pred, 1.0 * torch.argmax(nn.Softmax()(out[batch.val_mask]), axis=1).cpu()), 0)\n",
    "\n",
    "    y_target = y_target.numpy()\n",
    "    y_pred = y_pred.numpy()\n",
    "    \n",
    "    return total_loss/total_examples, y_target, y_pred"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "number of training examples: 942650, number of validation examples: 236441\n",
      "epoch: 0||train loss: 0.6930||val loss: 0.6930||train acc: 0.5039||val acc: 0.5032||train auc: 0.5011||val auc: 0.5003\n",
      "epoch: 1||train loss: 0.6900||val loss: 0.6900||train acc: 0.5448||val acc: 0.5479||train auc: 0.5431||val auc: 0.5463\n",
      "epoch: 2||train loss: 0.6807||val loss: 0.6809||train acc: 0.5505||val acc: 0.5550||train auc: 0.5501||val auc: 0.5545\n",
      "epoch: 3||train loss: 0.6680||val loss: 0.6696||train acc: 0.5650||val acc: 0.5663||train auc: 0.5640||val auc: 0.5650\n",
      "epoch: 4||train loss: 0.6571||val loss: 0.6577||train acc: 0.5926||val acc: 0.5957||train auc: 0.5919||val auc: 0.5945\n",
      "epoch: 5||train loss: 0.6507||val loss: 0.6518||train acc: 0.6072||val acc: 0.6090||train auc: 0.6064||val auc: 0.6076\n",
      "epoch: 6||train loss: 0.6444||val loss: 0.6453||train acc: 0.6119||val acc: 0.6144||train auc: 0.6115||val auc: 0.6140\n",
      "epoch: 7||train loss: 0.6389||val loss: 0.6404||train acc: 0.6142||val acc: 0.6153||train auc: 0.6135||val auc: 0.6144\n",
      "epoch: 8||train loss: 0.6344||val loss: 0.6355||train acc: 0.6179||val acc: 0.6193||train auc: 0.6175||val auc: 0.6185\n",
      "epoch: 9||train loss: 0.6299||val loss: 0.6325||train acc: 0.6218||val acc: 0.6204||train auc: 0.6215||val auc: 0.6200\n",
      "epoch: 10||train loss: 0.6284||val loss: 0.6303||train acc: 0.6231||val acc: 0.6213||train auc: 0.6232||val auc: 0.6215\n",
      "epoch: 11||train loss: 0.6254||val loss: 0.6265||train acc: 0.6256||val acc: 0.6244||train auc: 0.6252||val auc: 0.6240\n",
      "epoch: 12||train loss: 0.6236||val loss: 0.6252||train acc: 0.6263||val acc: 0.6242||train auc: 0.6256||val auc: 0.6234\n",
      "epoch: 13||train loss: 0.6225||val loss: 0.6240||train acc: 0.6269||val acc: 0.6263||train auc: 0.6268||val auc: 0.6263\n",
      "epoch: 14||train loss: 0.6210||val loss: 0.6224||train acc: 0.6275||val acc: 0.6281||train auc: 0.6271||val auc: 0.6276\n",
      "epoch: 15||train loss: 0.6201||val loss: 0.6216||train acc: 0.6273||val acc: 0.6268||train auc: 0.6269||val auc: 0.6263\n",
      "epoch: 16||train loss: 0.6187||val loss: 0.6204||train acc: 0.6283||val acc: 0.6274||train auc: 0.6281||val auc: 0.6271\n",
      "epoch: 17||train loss: 0.6175||val loss: 0.6191||train acc: 0.6298||val acc: 0.6296||train auc: 0.6295||val auc: 0.6290\n",
      "epoch: 18||train loss: 0.6160||val loss: 0.6169||train acc: 0.6305||val acc: 0.6308||train auc: 0.6302||val auc: 0.6306\n",
      "epoch: 19||train loss: 0.6149||val loss: 0.6161||train acc: 0.6312||val acc: 0.6318||train auc: 0.6310||val auc: 0.6316\n",
      "epoch: 20||train loss: 0.6143||val loss: 0.6161||train acc: 0.6318||val acc: 0.6295||train auc: 0.6316||val auc: 0.6293\n",
      "epoch: 21||train loss: 0.6124||val loss: 0.6151||train acc: 0.6330||val acc: 0.6296||train auc: 0.6325||val auc: 0.6291\n",
      "epoch: 22||train loss: 0.6114||val loss: 0.6131||train acc: 0.6334||val acc: 0.6338||train auc: 0.6330||val auc: 0.6331\n",
      "epoch: 23||train loss: 0.6110||val loss: 0.6130||train acc: 0.6340||val acc: 0.6326||train auc: 0.6335||val auc: 0.6320\n",
      "epoch: 24||train loss: 0.6097||val loss: 0.6119||train acc: 0.6349||val acc: 0.6330||train auc: 0.6344||val auc: 0.6325\n",
      "epoch: 25||train loss: 0.6079||val loss: 0.6108||train acc: 0.6364||val acc: 0.6344||train auc: 0.6360||val auc: 0.6339\n",
      "epoch: 26||train loss: 0.6076||val loss: 0.6098||train acc: 0.6357||val acc: 0.6349||train auc: 0.6352||val auc: 0.6343\n",
      "epoch: 27||train loss: 0.6062||val loss: 0.6091||train acc: 0.6369||val acc: 0.6348||train auc: 0.6364||val auc: 0.6342\n",
      "epoch: 28||train loss: 0.6056||val loss: 0.6077||train acc: 0.6375||val acc: 0.6365||train auc: 0.6369||val auc: 0.6358\n",
      "epoch: 29||train loss: 0.6040||val loss: 0.6078||train acc: 0.6394||val acc: 0.6354||train auc: 0.6392||val auc: 0.6351\n",
      "epoch: 30||train loss: 0.6034||val loss: 0.6076||train acc: 0.6394||val acc: 0.6363||train auc: 0.6394||val auc: 0.6363\n",
      "epoch: 31||train loss: 0.6024||val loss: 0.6051||train acc: 0.6396||val acc: 0.6379||train auc: 0.6393||val auc: 0.6375\n",
      "epoch: 32||train loss: 0.6009||val loss: 0.6036||train acc: 0.6409||val acc: 0.6391||train auc: 0.6407||val auc: 0.6387\n",
      "epoch: 33||train loss: 0.6001||val loss: 0.6022||train acc: 0.6415||val acc: 0.6401||train auc: 0.6415||val auc: 0.6401\n",
      "epoch: 34||train loss: 0.5981||val loss: 0.6006||train acc: 0.6432||val acc: 0.6415||train auc: 0.6428||val auc: 0.6411\n",
      "epoch: 35||train loss: 0.5987||val loss: 0.6011||train acc: 0.6425||val acc: 0.6411||train auc: 0.6427||val auc: 0.6413\n",
      "epoch: 36||train loss: 0.5973||val loss: 0.5996||train acc: 0.6427||val acc: 0.6420||train auc: 0.6419||val auc: 0.6410\n",
      "epoch: 37||train loss: 0.5965||val loss: 0.5988||train acc: 0.6435||val acc: 0.6430||train auc: 0.6427||val auc: 0.6419\n",
      "epoch: 38||train loss: 0.5939||val loss: 0.5971||train acc: 0.6460||val acc: 0.6445||train auc: 0.6455||val auc: 0.6441\n",
      "epoch: 39||train loss: 0.5933||val loss: 0.5963||train acc: 0.6461||val acc: 0.6454||train auc: 0.6460||val auc: 0.6453\n",
      "epoch: 40||train loss: 0.5924||val loss: 0.5949||train acc: 0.6470||val acc: 0.6463||train auc: 0.6466||val auc: 0.6458\n",
      "epoch: 41||train loss: 0.5908||val loss: 0.5945||train acc: 0.6488||val acc: 0.6465||train auc: 0.6484||val auc: 0.6461\n",
      "epoch: 42||train loss: 0.5909||val loss: 0.5933||train acc: 0.6489||val acc: 0.6483||train auc: 0.6485||val auc: 0.6478\n",
      "epoch: 43||train loss: 0.5893||val loss: 0.5922||train acc: 0.6493||val acc: 0.6489||train auc: 0.6492||val auc: 0.6487\n",
      "epoch: 44||train loss: 0.5879||val loss: 0.5908||train acc: 0.6498||val acc: 0.6493||train auc: 0.6499||val auc: 0.6494\n",
      "epoch: 45||train loss: 0.5866||val loss: 0.5904||train acc: 0.6519||val acc: 0.6504||train auc: 0.6518||val auc: 0.6503\n",
      "epoch: 46||train loss: 0.5855||val loss: 0.5898||train acc: 0.6528||val acc: 0.6490||train auc: 0.6525||val auc: 0.6486\n",
      "epoch: 47||train loss: 0.5844||val loss: 0.5876||train acc: 0.6530||val acc: 0.6504||train auc: 0.6526||val auc: 0.6499\n",
      "epoch: 48||train loss: 0.5833||val loss: 0.5856||train acc: 0.6540||val acc: 0.6534||train auc: 0.6536||val auc: 0.6531\n",
      "epoch: 49||train loss: 0.5820||val loss: 0.5858||train acc: 0.6543||val acc: 0.6526||train auc: 0.6542||val auc: 0.6525\n",
      "epoch: 50||train loss: 0.5813||val loss: 0.5851||train acc: 0.6555||val acc: 0.6528||train auc: 0.6554||val auc: 0.6525\n",
      "epoch: 51||train loss: 0.5806||val loss: 0.5843||train acc: 0.6559||val acc: 0.6523||train auc: 0.6556||val auc: 0.6518\n",
      "epoch: 52||train loss: 0.5802||val loss: 0.5846||train acc: 0.6556||val acc: 0.6526||train auc: 0.6558||val auc: 0.6527\n",
      "epoch: 53||train loss: 0.5792||val loss: 0.5826||train acc: 0.6548||val acc: 0.6543||train auc: 0.6551||val auc: 0.6546\n",
      "epoch: 54||train loss: 0.5773||val loss: 0.5821||train acc: 0.6586||val acc: 0.6545||train auc: 0.6583||val auc: 0.6541\n",
      "epoch: 55||train loss: 0.5765||val loss: 0.5808||train acc: 0.6577||val acc: 0.6552||train auc: 0.6578||val auc: 0.6554\n",
      "epoch: 56||train loss: 0.5761||val loss: 0.5800||train acc: 0.6580||val acc: 0.6554||train auc: 0.6578||val auc: 0.6552\n",
      "epoch: 57||train loss: 0.5751||val loss: 0.5798||train acc: 0.6588||val acc: 0.6561||train auc: 0.6586||val auc: 0.6558\n",
      "epoch: 58||train loss: 0.5746||val loss: 0.5794||train acc: 0.6588||val acc: 0.6544||train auc: 0.6584||val auc: 0.6538\n",
      "epoch: 59||train loss: 0.5730||val loss: 0.5779||train acc: 0.6599||val acc: 0.6573||train auc: 0.6598||val auc: 0.6572\n",
      "epoch: 60||train loss: 0.5722||val loss: 0.5768||train acc: 0.6606||val acc: 0.6573||train auc: 0.6606||val auc: 0.6573\n",
      "epoch: 61||train loss: 0.5716||val loss: 0.5752||train acc: 0.6602||val acc: 0.6586||train auc: 0.6602||val auc: 0.6586\n",
      "epoch: 62||train loss: 0.5706||val loss: 0.5758||train acc: 0.6614||val acc: 0.6594||train auc: 0.6615||val auc: 0.6595\n",
      "epoch: 63||train loss: 0.5709||val loss: 0.5755||train acc: 0.6606||val acc: 0.6588||train auc: 0.6607||val auc: 0.6589\n",
      "epoch: 64||train loss: 0.5693||val loss: 0.5761||train acc: 0.6621||val acc: 0.6588||train auc: 0.6619||val auc: 0.6586\n",
      "epoch: 65||train loss: 0.5699||val loss: 0.5743||train acc: 0.6618||val acc: 0.6585||train auc: 0.6614||val auc: 0.6580\n",
      "epoch: 66||train loss: 0.5684||val loss: 0.5736||train acc: 0.6623||val acc: 0.6587||train auc: 0.6619||val auc: 0.6582\n",
      "epoch: 67||train loss: 0.5682||val loss: 0.5742||train acc: 0.6626||val acc: 0.6585||train auc: 0.6629||val auc: 0.6588\n"
     ]
    }
   ],
   "source": [
    "loss_train_gcn = []\n",
    "acc_train_gcn = []\n",
    "auc_train_gcn = []\n",
    "ap_train_gcn = []\n",
    "loss_val_gcn = []\n",
    "acc_val_gcn = []\n",
    "auc_val_gcn = []\n",
    "ap_val_gcn = []\n",
    "\n",
    "for epoch in range(n_epochs+1):\n",
    "\n",
    "    # evaluate on both train and val sets\n",
    "    loss_train, y_target_train, y_pred_train = evaluator_train()\n",
    "    loss_val, y_target_val, y_pred_val = evaluator_val()\n",
    "    \n",
    "    if epoch == 0:\n",
    "        print('number of training examples: %i, number of validation examples: %i' %(len(y_pred_train), len(y_pred_val)))\n",
    "    \n",
    "    # compute the accuracy, average precision, AUC\n",
    "    acc_train = np.mean(y_pred_train == y_target_train)\n",
    "    auc_train = roc_auc_score(y_target_train, y_pred_train)\n",
    "    ap_train = average_precision_score(y_target_train, y_pred_train)\n",
    "    roc_train = roc_curve(y_target_train, y_pred_train)\n",
    "\n",
    "    acc_val = np.mean(y_pred_val == y_target_val)\n",
    "    auc_val = roc_auc_score(y_target_val, y_pred_val)\n",
    "    ap_val = average_precision_score(y_target_val, y_pred_val)\n",
    "    roc_val = roc_curve(y_target_val, y_pred_val)\n",
    "\n",
    "    # record the results\n",
    "    loss_train_gcn.append(loss_train)\n",
    "    acc_train_gcn.append(acc_train)\n",
    "    auc_train_gcn.append(auc_train)\n",
    "    ap_train_gcn.append(ap_train)\n",
    "    \n",
    "    loss_val_gcn.append(loss_val)\n",
    "    acc_val_gcn.append(acc_val)\n",
    "    auc_val_gcn.append(auc_val)\n",
    "    ap_val_gcn.append(ap_val)\n",
    "\n",
    "    # print results\n",
    "    print('epoch: %i||train loss: %.4f||val loss: %.4f||train acc: %.4f||val acc: %.4f||train auc: %.4f||val auc: %.4f' \\\n",
    "          % (epoch, loss_train, loss_val, acc_train, acc_val, auc_train, auc_val))\n",
    "    \n",
    "    # train\n",
    "    if epoch != n_epochs:\n",
    "        loss_train = train()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "fig, ax = plt.subplots(1, 2, figsize=(12,4))\n",
    "fig.subplots_adjust(wspace=0.3)\n",
    "\n",
    "ax[0].plot(np.arange(1, len(loss_train_gcn)+1), loss_train_gcn)\n",
    "ax[0].plot(np.arange(1, len(loss_val_gcn)+1), loss_val_gcn)\n",
    "ax[0].set_xlabel('epochs')\n",
    "ax[0].set_ylabel('BCE loss')\n",
    "#ax[0].set_xscale('log')\n",
    "#ax[0].set_yscale('log')\n",
    "#ax[0].set_ylim([np.min(loss_history), np.max(loss_history)])\n",
    "\n",
    "ax[1].plot(acc_train_gcn)\n",
    "ax[1].plot(acc_val_gcn)\n",
    "#ax[1].plot(ap_history, label='AP')\n",
    "ax[1].set_xlabel('epochs')\n",
    "ax[1].set_ylabel('accuracy')\n",
    "#ax[1].set_xscale('log')\n",
    "#ax[1].legend()\n",
    "\n",
    "#plt.savefig(\"figures/node_prediction_gcn.pdf\", bbox_inches='tight')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### GraphSAGE"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "number of hidden units = 256\n",
      "number of trainable parameters 335618\n"
     ]
    }
   ],
   "source": [
    "h = 256\n",
    "print('number of hidden units = %i' %h)\n",
    "\n",
    "model = SAGE(data.x.shape[1], h, num_classes).to(device)\n",
    "model.set_aggr('mean')\n",
    "print('number of trainable parameters %i' %model_params(model))\n",
    "\n",
    "optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [],
   "source": [
    "def train():\n",
    "    model.train()\n",
    "\n",
    "    total_loss = total_examples = 0\n",
    "    for batch in loader:\n",
    "        batch = batch.to(device)\n",
    "        optimizer.zero_grad()\n",
    "        out = model(batch.x, batch.edge_index)\n",
    "        batch_size = out[batch.train_mask].shape[0]\n",
    "        loss = F.nll_loss(out[batch.train_mask], batch.y[batch.train_mask])\n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "        total_loss += loss.item() * batch_size\n",
    "        total_examples += batch_size\n",
    "    return total_loss / total_examples\n",
    "\n",
    "@torch.no_grad()\n",
    "def evaluator_train():\n",
    "    model.eval()\n",
    "    \n",
    "    y_target = torch.empty(0)\n",
    "    y_pred = torch.empty(0)\n",
    "    total_loss = total_examples = 0\n",
    "    for batch in loader:\n",
    "        batch = batch.to(device)\n",
    "        out = model(batch.x, batch.edge_index)\n",
    "        batch_size = out[batch.train_mask].shape[0]\n",
    "        total_examples += batch_size\n",
    "        total_loss += F.nll_loss(out[batch.train_mask], batch.y[batch.train_mask]).item() * batch_size\n",
    "        \n",
    "        y_target = torch.cat((y_target, 1.0 * batch.y[batch.train_mask].cpu()), 0)\n",
    "        y_pred = torch.cat((y_pred, 1.0 * torch.argmax(nn.Softmax()(out[batch.train_mask]), axis=1).cpu()), 0)\n",
    "        \n",
    "    y_target = y_target.numpy()\n",
    "    y_pred = y_pred.numpy()\n",
    "    \n",
    "    return total_loss/total_examples, y_target, y_pred\n",
    "\n",
    "@torch.no_grad()\n",
    "def evaluator_val():\n",
    "    model.eval()\n",
    "    \n",
    "    y_target = torch.empty(0)\n",
    "    y_pred = torch.empty(0)\n",
    "    total_loss = total_examples = 0\n",
    "    for batch in loader:\n",
    "        batch = batch.to(device)\n",
    "        out = model(batch.x, batch.edge_index)\n",
    "        batch_size = out[batch.val_mask].shape[0]\n",
    "        total_examples += batch_size\n",
    "        total_loss += F.nll_loss(out[batch.val_mask], batch.y[batch.val_mask]).item() * batch_size\n",
    "        \n",
    "        y_target = torch.cat((y_target, 1.0 * batch.y[batch.val_mask].cpu()), 0)\n",
    "        y_pred = torch.cat((y_pred, 1.0 * torch.argmax(nn.Softmax()(out[batch.val_mask]), axis=1).cpu()), 0)\n",
    "\n",
    "    y_target = y_target.numpy()\n",
    "    y_pred = y_pred.numpy()\n",
    "    \n",
    "    return total_loss/total_examples, y_target, y_pred"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "number of training examples: 943971, number of validation examples: 237165\n",
      "epoch: 0||train loss: 0.6929||val loss: 0.6931||train acc: 0.5024||val acc: 0.5020||train auc: 0.5046||val auc: 0.5041\n",
      "epoch: 1||train loss: 0.6555||val loss: 0.6557||train acc: 0.6516||val acc: 0.6517||train auc: 0.6493||val auc: 0.6488\n",
      "epoch: 2||train loss: 0.5403||val loss: 0.5403||train acc: 0.7553||val acc: 0.7555||train auc: 0.7542||val auc: 0.7539\n",
      "epoch: 3||train loss: 0.4598||val loss: 0.4582||train acc: 0.7858||val acc: 0.7878||train auc: 0.7852||val auc: 0.7869\n",
      "epoch: 4||train loss: 0.4028||val loss: 0.4025||train acc: 0.8290||val acc: 0.8302||train auc: 0.8284||val auc: 0.8295\n",
      "epoch: 5||train loss: 0.3491||val loss: 0.3477||train acc: 0.8480||val acc: 0.8483||train auc: 0.8476||val auc: 0.8479\n",
      "epoch: 6||train loss: 0.3052||val loss: 0.3036||train acc: 0.8654||val acc: 0.8661||train auc: 0.8648||val auc: 0.8654\n",
      "epoch: 7||train loss: 0.2694||val loss: 0.2670||train acc: 0.8814||val acc: 0.8827||train auc: 0.8810||val auc: 0.8823\n",
      "epoch: 8||train loss: 0.2425||val loss: 0.2406||train acc: 0.8988||val acc: 0.8998||train auc: 0.8987||val auc: 0.8997\n",
      "epoch: 9||train loss: 0.2305||val loss: 0.2289||train acc: 0.9032||val acc: 0.9036||train auc: 0.9030||val auc: 0.9032\n",
      "epoch: 10||train loss: 0.2191||val loss: 0.2166||train acc: 0.9097||val acc: 0.9106||train auc: 0.9095||val auc: 0.9103\n",
      "epoch: 11||train loss: 0.2119||val loss: 0.2125||train acc: 0.9134||val acc: 0.9123||train auc: 0.9133||val auc: 0.9121\n",
      "epoch: 12||train loss: 0.2064||val loss: 0.2063||train acc: 0.9165||val acc: 0.9160||train auc: 0.9163||val auc: 0.9158\n",
      "epoch: 13||train loss: 0.1970||val loss: 0.1995||train acc: 0.9212||val acc: 0.9191||train auc: 0.9210||val auc: 0.9189\n",
      "epoch: 14||train loss: 0.1918||val loss: 0.1920||train acc: 0.9241||val acc: 0.9226||train auc: 0.9241||val auc: 0.9225\n",
      "epoch: 15||train loss: 0.1882||val loss: 0.1894||train acc: 0.9255||val acc: 0.9240||train auc: 0.9254||val auc: 0.9238\n",
      "epoch: 16||train loss: 0.1805||val loss: 0.1844||train acc: 0.9280||val acc: 0.9254||train auc: 0.9278||val auc: 0.9252\n",
      "epoch: 17||train loss: 0.1759||val loss: 0.1806||train acc: 0.9298||val acc: 0.9270||train auc: 0.9297||val auc: 0.9268\n",
      "epoch: 18||train loss: 0.1718||val loss: 0.1768||train acc: 0.9310||val acc: 0.9278||train auc: 0.9309||val auc: 0.9275\n",
      "epoch: 19||train loss: 0.1693||val loss: 0.1736||train acc: 0.9318||val acc: 0.9290||train auc: 0.9317||val auc: 0.9289\n",
      "epoch: 20||train loss: 0.1654||val loss: 0.1701||train acc: 0.9327||val acc: 0.9302||train auc: 0.9325||val auc: 0.9300\n",
      "epoch: 21||train loss: 0.1633||val loss: 0.1683||train acc: 0.9331||val acc: 0.9311||train auc: 0.9330||val auc: 0.9310\n",
      "epoch: 22||train loss: 0.1620||val loss: 0.1680||train acc: 0.9331||val acc: 0.9305||train auc: 0.9330||val auc: 0.9303\n",
      "epoch: 23||train loss: 0.1596||val loss: 0.1664||train acc: 0.9340||val acc: 0.9309||train auc: 0.9339||val auc: 0.9308\n",
      "epoch: 24||train loss: 0.1569||val loss: 0.1634||train acc: 0.9344||val acc: 0.9321||train auc: 0.9343||val auc: 0.9320\n",
      "epoch: 25||train loss: 0.1549||val loss: 0.1635||train acc: 0.9347||val acc: 0.9313||train auc: 0.9346||val auc: 0.9311\n",
      "epoch: 26||train loss: 0.1536||val loss: 0.1599||train acc: 0.9351||val acc: 0.9330||train auc: 0.9350||val auc: 0.9329\n",
      "epoch: 27||train loss: 0.1519||val loss: 0.1609||train acc: 0.9355||val acc: 0.9323||train auc: 0.9353||val auc: 0.9322\n",
      "epoch: 28||train loss: 0.1505||val loss: 0.1589||train acc: 0.9358||val acc: 0.9325||train auc: 0.9357||val auc: 0.9324\n",
      "epoch: 29||train loss: 0.1497||val loss: 0.1590||train acc: 0.9360||val acc: 0.9320||train auc: 0.9358||val auc: 0.9318\n",
      "epoch: 30||train loss: 0.1474||val loss: 0.1557||train acc: 0.9368||val acc: 0.9338||train auc: 0.9367||val auc: 0.9337\n",
      "epoch: 31||train loss: 0.1472||val loss: 0.1566||train acc: 0.9366||val acc: 0.9334||train auc: 0.9365||val auc: 0.9332\n",
      "epoch: 32||train loss: 0.1451||val loss: 0.1554||train acc: 0.9373||val acc: 0.9331||train auc: 0.9372||val auc: 0.9330\n",
      "epoch: 33||train loss: 0.1446||val loss: 0.1556||train acc: 0.9374||val acc: 0.9329||train auc: 0.9373||val auc: 0.9328\n",
      "epoch: 34||train loss: 0.1434||val loss: 0.1535||train acc: 0.9378||val acc: 0.9338||train auc: 0.9376||val auc: 0.9337\n",
      "epoch: 35||train loss: 0.1424||val loss: 0.1538||train acc: 0.9379||val acc: 0.9338||train auc: 0.9378||val auc: 0.9336\n",
      "epoch: 36||train loss: 0.1414||val loss: 0.1536||train acc: 0.9384||val acc: 0.9328||train auc: 0.9383||val auc: 0.9327\n",
      "epoch: 37||train loss: 0.1409||val loss: 0.1521||train acc: 0.9382||val acc: 0.9339||train auc: 0.9381||val auc: 0.9338\n",
      "epoch: 38||train loss: 0.1398||val loss: 0.1521||train acc: 0.9390||val acc: 0.9336||train auc: 0.9389||val auc: 0.9336\n",
      "epoch: 39||train loss: 0.1406||val loss: 0.1513||train acc: 0.9385||val acc: 0.9345||train auc: 0.9383||val auc: 0.9344\n",
      "epoch: 40||train loss: 0.1389||val loss: 0.1525||train acc: 0.9389||val acc: 0.9334||train auc: 0.9388||val auc: 0.9332\n",
      "epoch: 41||train loss: 0.1376||val loss: 0.1519||train acc: 0.9395||val acc: 0.9340||train auc: 0.9394||val auc: 0.9339\n",
      "epoch: 42||train loss: 0.1376||val loss: 0.1510||train acc: 0.9395||val acc: 0.9343||train auc: 0.9394||val auc: 0.9341\n",
      "epoch: 43||train loss: 0.1364||val loss: 0.1516||train acc: 0.9400||val acc: 0.9338||train auc: 0.9399||val auc: 0.9337\n",
      "epoch: 44||train loss: 0.1357||val loss: 0.1512||train acc: 0.9405||val acc: 0.9340||train auc: 0.9404||val auc: 0.9338\n",
      "epoch: 45||train loss: 0.1352||val loss: 0.1516||train acc: 0.9403||val acc: 0.9342||train auc: 0.9402||val auc: 0.9342\n",
      "epoch: 46||train loss: 0.1347||val loss: 0.1505||train acc: 0.9405||val acc: 0.9340||train auc: 0.9404||val auc: 0.9339\n",
      "epoch: 47||train loss: 0.1341||val loss: 0.1502||train acc: 0.9412||val acc: 0.9339||train auc: 0.9411||val auc: 0.9338\n",
      "epoch: 48||train loss: 0.1334||val loss: 0.1497||train acc: 0.9414||val acc: 0.9344||train auc: 0.9414||val auc: 0.9343\n",
      "epoch: 49||train loss: 0.1324||val loss: 0.1492||train acc: 0.9417||val acc: 0.9350||train auc: 0.9416||val auc: 0.9348\n",
      "epoch: 50||train loss: 0.1318||val loss: 0.1487||train acc: 0.9419||val acc: 0.9357||train auc: 0.9418||val auc: 0.9356\n",
      "epoch: 51||train loss: 0.1321||val loss: 0.1479||train acc: 0.9418||val acc: 0.9350||train auc: 0.9417||val auc: 0.9349\n",
      "epoch: 52||train loss: 0.1317||val loss: 0.1506||train acc: 0.9417||val acc: 0.9345||train auc: 0.9416||val auc: 0.9343\n",
      "epoch: 53||train loss: 0.1308||val loss: 0.1490||train acc: 0.9424||val acc: 0.9349||train auc: 0.9423||val auc: 0.9347\n",
      "epoch: 54||train loss: 0.1297||val loss: 0.1496||train acc: 0.9430||val acc: 0.9341||train auc: 0.9428||val auc: 0.9340\n",
      "epoch: 55||train loss: 0.1298||val loss: 0.1493||train acc: 0.9424||val acc: 0.9346||train auc: 0.9423||val auc: 0.9345\n",
      "epoch: 56||train loss: 0.1295||val loss: 0.1476||train acc: 0.9429||val acc: 0.9357||train auc: 0.9428||val auc: 0.9356\n",
      "epoch: 57||train loss: 0.1294||val loss: 0.1480||train acc: 0.9427||val acc: 0.9353||train auc: 0.9426||val auc: 0.9351\n",
      "epoch: 58||train loss: 0.1285||val loss: 0.1499||train acc: 0.9435||val acc: 0.9344||train auc: 0.9433||val auc: 0.9342\n",
      "epoch: 59||train loss: 0.1276||val loss: 0.1489||train acc: 0.9437||val acc: 0.9353||train auc: 0.9436||val auc: 0.9352\n",
      "epoch: 60||train loss: 0.1271||val loss: 0.1477||train acc: 0.9443||val acc: 0.9350||train auc: 0.9442||val auc: 0.9350\n",
      "epoch: 61||train loss: 0.1262||val loss: 0.1479||train acc: 0.9446||val acc: 0.9358||train auc: 0.9445||val auc: 0.9357\n",
      "epoch: 62||train loss: 0.1265||val loss: 0.1487||train acc: 0.9446||val acc: 0.9347||train auc: 0.9445||val auc: 0.9346\n",
      "epoch: 63||train loss: 0.1259||val loss: 0.1481||train acc: 0.9453||val acc: 0.9354||train auc: 0.9453||val auc: 0.9353\n",
      "epoch: 64||train loss: 0.1256||val loss: 0.1489||train acc: 0.9445||val acc: 0.9358||train auc: 0.9444||val auc: 0.9356\n",
      "epoch: 65||train loss: 0.1237||val loss: 0.1486||train acc: 0.9458||val acc: 0.9348||train auc: 0.9457||val auc: 0.9347\n",
      "epoch: 66||train loss: 0.1233||val loss: 0.1474||train acc: 0.9461||val acc: 0.9355||train auc: 0.9460||val auc: 0.9354\n",
      "epoch: 67||train loss: 0.1225||val loss: 0.1479||train acc: 0.9465||val acc: 0.9358||train auc: 0.9463||val auc: 0.9357\n"
     ]
    }
   ],
   "source": [
    "loss_train_sage = []\n",
    "acc_train_sage = []\n",
    "auc_train_sage = []\n",
    "ap_train_sage = []\n",
    "loss_val_sage = []\n",
    "acc_val_sage = []\n",
    "auc_val_sage = []\n",
    "ap_val_sage = []\n",
    "\n",
    "for epoch in range(n_epochs+1):\n",
    "\n",
    "    # evaluate on both train and val sets\n",
    "    loss_train, y_target_train, y_pred_train = evaluator_train()\n",
    "    loss_val, y_target_val, y_pred_val = evaluator_val()\n",
    "    \n",
    "    if epoch == 0:\n",
    "        print('number of training examples: %i, number of validation examples: %i' %(len(y_pred_train), len(y_pred_val)))\n",
    "    \n",
    "    # compute the accuracy, average precision, AUC\n",
    "    acc_train = np.mean(y_pred_train == y_target_train)\n",
    "    auc_train = roc_auc_score(y_target_train, y_pred_train)\n",
    "    ap_train = average_precision_score(y_target_train, y_pred_train)\n",
    "    roc_train = roc_curve(y_target_train, y_pred_train)\n",
    "\n",
    "    acc_val = np.mean(y_pred_val == y_target_val)\n",
    "    auc_val = roc_auc_score(y_target_val, y_pred_val)\n",
    "    ap_val = average_precision_score(y_target_val, y_pred_val)\n",
    "    roc_val = roc_curve(y_target_val, y_pred_val)\n",
    "\n",
    "    # record the results\n",
    "    loss_train_sage.append(loss_train)\n",
    "    acc_train_sage.append(acc_train)\n",
    "    auc_train_sage.append(auc_train)\n",
    "    ap_train_sage.append(ap_train)\n",
    "    \n",
    "    loss_val_sage.append(loss_val)\n",
    "    acc_val_sage.append(acc_val)\n",
    "    auc_val_sage.append(auc_val)\n",
    "    ap_val_sage.append(ap_val)\n",
    "\n",
    "    # print results\n",
    "    print('epoch: %i||train loss: %.4f||val loss: %.4f||train acc: %.4f||val acc: %.4f||train auc: %.4f||val auc: %.4f' \\\n",
    "          % (epoch, loss_train, loss_val, acc_train, acc_val, auc_train, auc_val))\n",
    "    \n",
    "    # train\n",
    "    if epoch != n_epochs:\n",
    "        loss_train = train()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 71,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x360 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "import matplotlib.ticker as tck\n",
    "lw = 3\n",
    "\n",
    "fig, ax = plt.subplots(1, 2, figsize=(12, 5))\n",
    "fig.subplots_adjust(wspace=0.3)\n",
    "\n",
    "ax[0].text(-0.1, 1.05, '(a)', transform=ax[0].transAxes, size=15)\n",
    "ax[1].text(-0.1, 1.05, '(b)', transform=ax[1].transAxes, size=15)\n",
    "\n",
    "ax[0].plot(np.arange(1, len(loss_train_mlp)+1), loss_train_mlp, label='MLP', linewidth=lw)\n",
    "ax[0].plot(np.arange(1, len(loss_train_sage)+1), loss_train_sage, label='GraphSAGE', linewidth=lw)\n",
    "ax[0].plot(np.arange(1, len(loss_train_gcn)+1), loss_train_gcn, label='GCN', linewidth=lw)\n",
    "ax[0].set_xlabel('epochs')\n",
    "ax[0].set_ylabel('loss')\n",
    "ax[0].legend()\n",
    "ax[0].xaxis.set_minor_locator(tck.AutoMinorLocator())\n",
    "ax[0].yaxis.set_minor_locator(tck.AutoMinorLocator())\n",
    "\n",
    "ax[1].plot(acc_train_mlp, linewidth=lw, label='MLP')\n",
    "ax[1].plot(acc_train_sage, linewidth=lw, label='GraphSAGE')\n",
    "ax[1].plot(acc_train_gcn, linewidth=lw, label='GCN')\n",
    "ax[1].xaxis.set_minor_locator(tck.AutoMinorLocator())\n",
    "ax[1].yaxis.set_minor_locator(tck.AutoMinorLocator())\n",
    "ax[1].set_xlabel('epochs')\n",
    "ax[1].set_ylabel('accuracy')\n",
    "fig.tight_layout()  # otherwise the right y-label is slightly clipped\n",
    "\n",
    "plt.savefig(\"figures/node_prediction_results.pdf\", bbox_inches='tight')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python (pytorch geometric)",
   "language": "python",
   "name": "pygeo"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}