CompressionEval.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 162,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import sys, os\n",
    "sys.path.append(os.path.join(sys.path[0], 'tfGCN'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 163,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "%matplotlib inline\n",
    "\n",
    "import numpy as np\n",
    "import matplotlib\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "import time\n",
    "from tqdm import tqdm\n",
    "import tensorflow as tf\n",
    "\n",
    "from models import GCN, HashedGCN\n",
    "from utils import *"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 165,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Set random seed\n",
    "seed = 123"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Settings\n",
    "flags = tf.app.flags \n",
    "FLAGS = tf.app.flags.FLAGS\n",
    "flags.DEFINE_string('dataset', 'cora', 'Dataset string.')  # 'cora', 'citeseer', 'pubmed'\n",
    "flags.DEFINE_string('model', 'gcn', 'Model string.')  # 'gcn', 'gcn_cheby', 'hashed_gcn', 'dense'\n",
    "flags.DEFINE_float('learning_rate', 0.01, 'Initial learning rate.')\n",
    "flags.DEFINE_integer('epochs', 200, 'Number of epochs to train.')\n",
    "flags.DEFINE_integer('hidden1', 16, 'Number of units in hidden layer 1.')\n",
    "flags.DEFINE_float('dropout', 0.5, 'Dropout rate (1 - keep probability).')\n",
    "flags.DEFINE_float('weight_decay', 5e-4, 'Weight for L2 loss on embedding matrix.')\n",
    "flags.DEFINE_integer('early_stopping', 10, 'Tolerance for early stopping (# of epochs).')\n",
    "flags.DEFINE_integer('max_degree', 3, 'Maximum Chebyshev polynomial degree.')\n",
    "flags.DEFINE_integer('compression_factor1', 1, 'Hashed GCN weight compression factor (layer 1).')\n",
    "flags.DEFINE_integer('compression_factor2', 1, 'Hashed GCN weight compression factor (layer 2).')\n",
    "tf.app.flags.DEFINE_string('f', '', 'kernel')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Load data\n",
    "adj, features, y_train, y_val, y_test, train_mask, val_mask, test_mask = load_data(FLAGS.dataset)\n",
    "\n",
    "# Some preprocessing\n",
    "features = preprocess_features(features)\n",
    "\n",
    "support = [preprocess_adj(adj)]\n",
    "num_supports = 1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# Define placeholders\n",
    "placeholders = {\n",
    "    'support': [tf.sparse_placeholder(tf.float32) for _ in range(num_supports)],\n",
    "    'features': tf.sparse_placeholder(tf.float32, shape=tf.constant(features[2], dtype=tf.int64)),\n",
    "    'labels': tf.placeholder(tf.float32, shape=(None, y_train.shape[1])),\n",
    "    'labels_mask': tf.placeholder(tf.int32),\n",
    "    'dropout': tf.placeholder_with_default(0., shape=()),\n",
    "    'num_features_nonzero': tf.placeholder(tf.int32)  # helper variable for sparse dropout\n",
    "}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# Define model evaluation function\n",
    "def evaluate(features, support, labels, mask, placeholders):\n",
    "    t_test = time.time()\n",
    "    feed_dict_val = construct_feed_dict(features, support, labels, mask, placeholders)\n",
    "    outs_val = sess.run([model.loss, model.accuracy], feed_dict=feed_dict_val)\n",
    "    return outs_val[0], outs_val[1], (time.time() - t_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 78,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "model = GCN(placeholders, input_dim=features[2][1], logging=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 89,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Reseed\n",
    "np.random.seed(seed)\n",
    "tf.set_random_seed(seed)\n",
    "\n",
    "# Initialize session\n",
    "sess = tf.Session()\n",
    "sess.run(tf.global_variables_initializer())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 90,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Baseline accuracy: 0.80899906\n"
     ]
    }
   ],
   "source": [
    "cost_val = []\n",
    "\n",
    "# Train model\n",
    "for epoch in range(FLAGS.epochs):\n",
    "    t = time.time()\n",
    "    # Construct feed dictionary\n",
    "    feed_dict = construct_feed_dict(features, support, y_train, train_mask, placeholders)\n",
    "    feed_dict.update({placeholders['dropout']: FLAGS.dropout})\n",
    "\n",
    "    # Training step\n",
    "    outs = sess.run([model.opt_op, model.loss, model.accuracy], feed_dict=feed_dict)\n",
    "\n",
    "    # Validation\n",
    "    cost, acc, duration = evaluate(features, support, y_val, val_mask, placeholders)\n",
    "    cost_val.append(cost)\n",
    "\n",
    "#     # Print results\n",
    "#     print(\"Epoch:\", '%04d' % (epoch + 1), \"train_loss=\", \"{:.5f}\".format(outs[1]),\n",
    "#           \"train_acc=\", \"{:.5f}\".format(outs[2]), \"val_loss=\", \"{:.5f}\".format(cost),\n",
    "#           \"val_acc=\", \"{:.5f}\".format(acc), \"time=\", \"{:.5f}\".format(time.time() - t))\n",
    "\n",
    "    if epoch > FLAGS.early_stopping and cost_val[-1] > np.mean(cost_val[-(FLAGS.early_stopping+1):-1]):\n",
    "        print(\"Early stopping...\")\n",
    "        break\n",
    "        \n",
    "_, baseline_acc, _ = evaluate(features, support, y_test, test_mask, placeholders)\n",
    "print(\"Baseline accuracy:\", baseline_acc)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 148,
   "metadata": {
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n",
      "\n",
      "  0%|                                                    | 0/8 [00:00<?, ?it/s]\n",
      "\n",
      " 12%|█████▍                                     | 1/8 [06:18<44:05, 377.94s/it]\n",
      "\n",
      " 25%|██████████▊                                | 2/8 [12:24<37:27, 374.66s/it]\n",
      "\n",
      " 38%|████████████████▏                          | 3/8 [19:01<31:46, 381.33s/it]\n",
      "\n",
      " 50%|█████████████████████▌                     | 4/8 [25:37<25:42, 385.59s/it]\n",
      "\n",
      " 62%|██████████████████████████▉                | 5/8 [32:07<19:20, 386.97s/it]\n",
      "\n",
      " 75%|████████████████████████████████▎          | 6/8 [38:57<13:07, 393.87s/it]\n",
      "\n",
      " 88%|█████████████████████████████████████▋     | 7/8 [45:46<06:38, 398.30s/it]\n",
      "\n",
      "100%|███████████████████████████████████████████| 8/8 [52:30<00:00, 400.15s/it]\n",
      "\n"
     ]
    }
   ],
   "source": [
    "compression_factors = [1, 2, 4, 8, 16, 32, 64, 128]\n",
    "compression_accuracies = []\n",
    "num_runs = 5\n",
    "\n",
    "for cf in tqdm(compression_factors):\n",
    "    FLAGS.compression_factor1 = cf\n",
    "    FLAGS.compression_factor2 = 1\n",
    "    np.random.seed(seed)\n",
    "    tf.set_random_seed(seed)\n",
    "    all_acc = []\n",
    "    \n",
    "    for _ in range(num_runs):\n",
    "        model = HashedGCN(placeholders, input_dim=features[2][1], logging=True)\n",
    "        sess = tf.Session()\n",
    "        sess.run(tf.global_variables_initializer())\n",
    "\n",
    "        cost_val = []\n",
    "        # Train model\n",
    "        for epoch in range(FLAGS.epochs):\n",
    "            t = time.time()\n",
    "\n",
    "            feed_dict = construct_feed_dict(features, support, y_train, train_mask, placeholders)\n",
    "            feed_dict.update({placeholders['dropout']: FLAGS.dropout})\n",
    "\n",
    "            # Training step\n",
    "            outs = sess.run([model.opt_op, model.loss, model.accuracy], feed_dict=feed_dict)\n",
    "\n",
    "            # Validation\n",
    "            cost, acc, duration = evaluate(features, support, y_val, val_mask, placeholders)\n",
    "            cost_val.append(cost)\n",
    "\n",
    "            if epoch > FLAGS.early_stopping and cost_val[-1] > np.mean(cost_val[-(FLAGS.early_stopping+1):-1]):\n",
    "                break\n",
    "        _, compressed_acc, _ = evaluate(features, support, y_test, test_mask, placeholders)\n",
    "        all_acc.append(compressed_acc)\n",
    "    compression_accuracies.append(all_acc)\n",
    "compression_accuracies = np.asarray(compression_accuracies)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 157,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "compression_accuracies_median = (1 - np.median(compression_accuracies, axis=1)) * 100\n",
    "compression_accuracies_std = np.std(compression_accuracies, axis=1) * 100"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 160,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYIAAAEKCAYAAAAfGVI8AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nuf8Hk7koLv7YnnUAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x1769eb314a8>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Plot compression vs accuracy\n",
    "x = np.arange(8)\n",
    "plt.xlabel(\"Compression factor (Layer 1 only)\")\n",
    "plt.ylabel(\"Test error (%)\")\n",
    "plt.errorbar(x, compression_accuracies_median, compression_accuracies_std, \\\n",
    "             label=\"HashNet\", linewidth=2.0, marker='v', capsize=3)\n",
    "fig = plt.plot(x, (1 - np.repeat(baseline_acc, 8)) * 100, label=\"Uncompressed\", linewidth=2.0, linestyle='--', color='black')\n",
    "plt.xticks(x, [\"1\", \"1/2\", \"1/4\", \"1/8\", \"1/16\", \"1/32\", \"1/64\", \"1/128\"])\n",
    "leg = plt.legend()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 161,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "with open('compression_layer1.png','wb') as f:\n",
    "    fig[0].figure.savefig(f, format='png', dpi=144)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 166,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n",
      "\n",
      "  0%|                                                    | 0/4 [00:00<?, ?it/s]\n",
      "\n",
      " 25%|██████████▊                                | 1/4 [07:15<21:46, 435.58s/it]\n",
      "\n",
      " 50%|█████████████████████▌                     | 2/4 [13:47<14:04, 422.38s/it]\n",
      "\n",
      " 75%|████████████████████████████████▎          | 3/4 [19:57<06:46, 406.67s/it]\n",
      "\n",
      "100%|███████████████████████████████████████████| 4/4 [26:38<00:00, 405.00s/it]\n",
      "\n",
      "\n",
      "\n",
      "  0%|                                                    | 0/6 [00:00<?, ?it/s]\n",
      "\n",
      " 17%|███████▏                                   | 1/6 [07:49<39:06, 469.26s/it]\n",
      "\n",
      " 33%|██████████████▎                            | 2/6 [15:47<31:28, 472.09s/it]\n",
      "\n",
      " 50%|█████████████████████▌                     | 3/6 [23:45<23:40, 473.62s/it]\n",
      "\n",
      " 67%|████████████████████████████▋              | 4/6 [31:51<15:54, 477.29s/it]\n",
      "\n",
      " 83%|███████████████████████████████████▊       | 5/6 [40:07<08:03, 483.00s/it]\n",
      "\n",
      "100%|███████████████████████████████████████████| 6/6 [48:46<00:00, 493.75s/it]\n",
      "\n"
     ]
    }
   ],
   "source": [
    "compression_factors_base = [1, 2, 4, 8]\n",
    "compression_factors_hashed = [1, 2, 4, 8, 16, 32]\n",
    "num_runs = 5\n",
    "\n",
    "base_compression_accuracies = []\n",
    "both_layer_compression_accuracies = []\n",
    "for cf in tqdm(compression_factors_base):\n",
    "    FLAGS.hidden1 = 16 // cf\n",
    "    np.random.seed(seed)\n",
    "    tf.set_random_seed(seed)\n",
    "    all_acc = []\n",
    "    for _ in range(num_runs):\n",
    "        model = GCN(placeholders, input_dim=features[2][1], logging=True)\n",
    "        sess = tf.Session()\n",
    "        sess.run(tf.global_variables_initializer())\n",
    "\n",
    "        cost_val = []\n",
    "        # Train model\n",
    "        for epoch in range(FLAGS.epochs):\n",
    "            t = time.time()\n",
    "\n",
    "            feed_dict = construct_feed_dict(features, support, y_train, train_mask, placeholders)\n",
    "            feed_dict.update({placeholders['dropout']: FLAGS.dropout})\n",
    "\n",
    "            # Training step\n",
    "            outs = sess.run([model.opt_op, model.loss, model.accuracy], feed_dict=feed_dict)\n",
    "\n",
    "            # Validation\n",
    "            cost, acc, duration = evaluate(features, support, y_val, val_mask, placeholders)\n",
    "            cost_val.append(cost)\n",
    "\n",
    "            if epoch > FLAGS.early_stopping and cost_val[-1] > np.mean(cost_val[-(FLAGS.early_stopping+1):-1]):\n",
    "                break\n",
    "        _, compressed_acc, _ = evaluate(features, support, y_test, test_mask, placeholders)\n",
    "        all_acc.append(compressed_acc)\n",
    "    base_compression_accuracies.append(all_acc)\n",
    "\n",
    "FLAGS.hidden1 = 16\n",
    "    \n",
    "for cf in tqdm(compression_factors_hashed):\n",
    "    FLAGS.compression_factor1 = cf\n",
    "    FLAGS.compression_factor2 = cf\n",
    "    np.random.seed(seed)\n",
    "    tf.set_random_seed(seed)\n",
    "    all_acc = []\n",
    "    for _ in range(num_runs):\n",
    "        model = HashedGCN(placeholders, input_dim=features[2][1], logging=True)\n",
    "        sess = tf.Session()\n",
    "        sess.run(tf.global_variables_initializer())\n",
    "\n",
    "        cost_val = []\n",
    "        # Train model\n",
    "        for epoch in range(FLAGS.epochs):\n",
    "            t = time.time()\n",
    "\n",
    "            feed_dict = construct_feed_dict(features, support, y_train, train_mask, placeholders)\n",
    "            feed_dict.update({placeholders['dropout']: FLAGS.dropout})\n",
    "\n",
    "            # Training step\n",
    "            outs = sess.run([model.opt_op, model.loss, model.accuracy], feed_dict=feed_dict)\n",
    "\n",
    "            # Validation\n",
    "            cost, acc, duration = evaluate(features, support, y_val, val_mask, placeholders)\n",
    "            cost_val.append(cost)\n",
    "\n",
    "            if epoch > FLAGS.early_stopping and cost_val[-1] > np.mean(cost_val[-(FLAGS.early_stopping+1):-1]):\n",
    "                break\n",
    "        _, compressed_acc, _ = evaluate(features, support, y_test, test_mask, placeholders)\n",
    "        all_acc.append(compressed_acc)\n",
    "    both_layer_compression_accuracies.append(all_acc)\n",
    "    \n",
    "base_compression_accuracies = np.asarray(base_compression_accuracies)\n",
    "both_layer_compression_accuracies = np.asarray(both_layer_compression_accuracies)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 170,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "both_layer_compression_median = (1 - np.median(both_layer_compression_accuracies, axis=1)) * 100\n",
    "both_layer_compression_std = np.std(both_layer_compression_accuracies, axis=1) * 100\n",
    "base_compression_median = (1 - np.median(base_compression_accuracies, axis=1)) * 100\n",
    "base_compression_std = np.std(base_compression_accuracies, axis=1) * 100"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 171,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYIAAAEKCAYAAAAfGVI8AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\njDF+zhKBMcb4OUsExhjj5/4ff4OejgSS1UMAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x177731a8ac8>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Plot comparison between hashed and equiv size unhashed\n",
    "x = np.arange(len(compression_factors_hashed))\n",
    "plt.xlabel(\"Compression factor\")\n",
    "plt.ylabel(\"Test error (%)\")\n",
    "plt.errorbar(x, both_layer_compression_median, both_layer_compression_std, label=\"HashNet\", linewidth=2.0, marker='v', capsize=3)\n",
    "fig = plt.errorbar(x[:len(compression_factors_base)], base_compression_median, base_compression_std, \\\n",
    "                   label=\"Neural Network, Equiv. Size\", linewidth=2.0, marker='o', linestyle='--', color='black', capsize=3)\n",
    "plt.xticks(x, [\"1\", \"1/2\", \"1/4\", \"1/8\", \"1/16\", \"1/32\", \"1/64\", \"1/128\"])\n",
    "leg = plt.legend()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 172,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "with open('compression_equiv.png','wb') as f:\n",
    "    fig[0].figure.savefig(f, format='png', dpi=144)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python [default]",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.1"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}