{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Implementing KL aneeling in Keras\n",
    "- [Generating Sentences from a Continuous Space](https://arxiv.org/pdf/1511.06349.pdf)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using Theano backend.\n",
      "Using gpu device 0: GeForce GTX TITAN X (CNMeM is disabled, cuDNN 5005)\n"
     ]
    }
   ],
   "source": [
    "from keras import backend as K"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "hp_lambda = K.variable(0)  # default values\n",
    "\n",
    "def original_loss(y_true, y_pred):\n",
    "    loss = K.categorical_crossentropy(y_true, y_pred)\n",
    "    KL = K.variable([100])\n",
    "    return loss + (hp_lambda) * KL"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "y_pred = K.placeholder(ndim=1)\n",
    "y_true = K.placeholder(ndim=1)\n",
    "f = K.function([y_true, y_pred], original_loss(y_true, y_pred))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "When lambda=0.0: [  1.19209304e-07]\n"
     ]
    }
   ],
   "source": [
    "K.set_value(hp_lambda, 0)\n",
    "print(\"When lambda={}: {}\".format(K.get_value(hp_lambda), f([[0, 1, 0], [0, 1, 0]])))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "When lambda=1.0: [ 100.]\n"
     ]
    }
   ],
   "source": [
    "K.set_value(hp_lambda, 1)\n",
    "print(\"When lambda={}: {}\".format(K.get_value(hp_lambda), f([[0, 1, 0], [0, 1, 0]])))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- K.set_valueをepochごとに呼び出せば良い\n",
    "- やり方としては\n",
    "    1. エポックごとにset_valueを実行\n",
    "    2. hp_lambdaをアップデートするコールバックを書く\n",
    "- の2通りありそう\n",
    "- 後者に関しては[url](https://github.com/fchollet/keras/blob/master/keras/callbacks.py)のLerningRateSchedulerが参考になりそう\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Callbacks"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from keras import callbacks\n",
    "\n",
    "\n",
    "class AneelingCallback(callbacks.Callback):\n",
    "    '''Aneeling theano shared variable.\n",
    "    # Arguments\n",
    "        schedule: a function that takes an epoch index as input\n",
    "            (integer, indexed from 0) and returns a new\n",
    "            learning rate as output (float).\n",
    "    '''\n",
    "    def __init__(self, schedule, variable):\n",
    "        super(AneelingCallback, self).__init__()\n",
    "        self.schedule = schedule\n",
    "        self.variable = variable\n",
    "        \n",
    "    def on_epoch_begin(self, epoch, logs={}):\n",
    "        assert hasattr(self.model.optimizer, 'lr'), \\\n",
    "            'Optimizer must have a \"lr\" attribute.'\n",
    "        value = self.schedule(epoch)\n",
    "        assert type(value) == float, 'The output of the \"schedule\" function should be float.'\n",
    "        K.set_value(self.variable, value)\n",
    "        \n",
    "        \n",
    "def schedule(epoch):\n",
    "    if epoch > 5:\n",
    "        return 10.0\n",
    "    return 0.0\n",
    "\n",
    "aneeling_callback = AneelingCallback(schedule, hp_lambda)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- テストはしてない"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 2",
   "language": "python",
   "name": "python2"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}