cothec / neural4coref

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "from keras.models import Model\n",
    "from keras.layers import Input, Dense, Dropout, Activation, BatchNormalization\n",
    "from keras.optimizers import SGD, Adam\n",
    "import numpy as np"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "# Data preparation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "filename = 'input_data.csv'\n",
    "raw_data = open(filename, 'rt')\n",
    "data = np.loadtxt(raw_data, delimiter= '\\t')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "print data.shape"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "Our dataset consists of ~466K examples (pairs of mentions), each example described by 1126 features. Labels say whether a pair belongs to the same cluster (1) or not (0)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "size_of_dataset = len(data)\n",
    "number_of_features = 1126\n",
    "\n",
    "X = data[:,0:1126]\n",
    "Y = data[:,1126] #last column consists of labels"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "Now let's split data into trainig and test set (90/10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "np.random.seed(999) #seed fixed for reproducibility\n",
    "mask = np.random.rand(size_of_dataset) < 0.9  #array of boolean variables\n",
    "\n",
    "training_set = X[mask]\n",
    "training_labels = Y[mask]\n",
    "\n",
    "test_set = X[~mask]\n",
    "test_labels = Y[~mask]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "# Neural network configuration"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "inputs = Input(shape=(number_of_features,))\n",
    "output_from_1st_layer = Dense(1000, activation='relu')(inputs)\n",
    "output_from_1st_layer = Dropout(0.5)(output_from_1st_layer)\n",
    "output_from_1st_layer = BatchNormalization()(output_from_1st_layer)\n",
    "output_from_2nd_layer = Dense(500, activation='relu')(output_from_1st_layer)\n",
    "output_from_2nd_layer = Dropout(0.5)(output_from_2nd_layer)\n",
    "output_from_2nd_layer = BatchNormalization()(output_from_2nd_layer)\n",
    "output = Dense(1, activation='sigmoid')(output_from_2nd_layer)\n",
    "\n",
    "model = Model(inputs, output)\n",
    "model.compile(optimizer='Adam',loss='binary_crossentropy',metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "model.fit(training_set, training_labels, batch_size=256, nb_epoch=25)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "source": [
    "# Evaluation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true,
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "scores = model.evaluate(test_set, test_labels)\n",
    "print(\"%s: %.2f%%\" % (model.metrics_names[1], scores[1]*100))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Playing with the model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "You can save the weights of the model to a file and later recreate the model without training by model.load_weights(\"my_weights.h5\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "model.save_weights(\"my_weights.h5\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "To have predictions for a test set we do"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "predictions = model.predict(test_set)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "and for a single example"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "single_example = test_set[4:5,:] #example number 5 from the test set\n",
    "prediction = model.predict(single_example)\n",
    "print '%.8f' % prediction[0]"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 2",
   "language": "python",
   "name": "python2"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2.0
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}