Keras - User Specific Loss Function Using Backend

myloss.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Keras - User Specific Loss Function Using Backend\n",
    "- IPython notebook version of the example in http://stackoverflow.com/questions/38684768/customize-keras-loss-function-in-a-way-that-the-y-true-will-depend-on-y-pred"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "%matplotlib inline\n",
    "import numpy as np"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using gpu device 0: GeForce GTX 1070 (CNMeM is disabled, cuDNN 5005)\n",
      "Using Theano backend.\n"
     ]
    }
   ],
   "source": [
    "import theano\n",
    "from keras import backend as K\n",
    "from keras.layers import Dense\n",
    "from keras.models import Sequential"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def customized_loss(y_true, y_pred):\n",
    "    #loss = K.switch(K.equal(y_true, -1), 0, K.square(y_true-y_pred))\n",
    "    loss = np.square(y_true-y_pred)\n",
    "    return K.sum(loss)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "model = Sequential([ Dense(3, input_shape=(4,)) ])\n",
    "model.compile(loss=customized_loss, optimizer='sgd')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([[ 0.52077309,  0.16107648,  0.28060555,  0.68680049]]),\n",
       " array([[ 1, -1,  0]]))"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "x = np.random.random((1, 4))\n",
    "y = np.array([[1,-1,0]])\n",
    "x, y"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([[ 0.29694095, -0.39778247,  0.13337782],\n",
       "        [-0.82495385, -0.65100169,  0.43441007],\n",
       "        [-0.49311569, -0.50998557,  0.22923476],\n",
       "        [-0.20695877, -0.85389435,  0.4840104 ]], dtype=float32),\n",
       " array([ 0.,  0.,  0.], dtype=float32))"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "W, b = model.get_weights()\n",
    "W, b"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[-0.25875217 -1.04157531  0.53617591]] [[-0.25875219 -1.04157531  0.53617595]] [[  2.01450578e-08  -1.82999171e-09  -4.01655094e-08]]\n"
     ]
    }
   ],
   "source": [
    "output = model.predict(x)\n",
    "output_calc = np.dot( x, W) + b\n",
    "err = output - output_calc\n",
    "print( output, output_calc, err)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1/1 [==============================] - 0s\n",
      "1.87367010117 1.87367022163\n"
     ]
    }
   ],
   "source": [
    "loss_calc = np.sum(np.square(y-output_calc))\n",
    "print( model.evaluate(x, y), loss_calc ) # keras's loss"
   ]
  }
 ],
 "metadata": {
  "anaconda-cloud": {},
  "kernelspec": {
   "display_name": "Python [conda root]",
   "language": "python",
   "name": "conda-root-py"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.5.2"
  },
  "toc": {
   "nav_menu": {
    "height": "45px",
    "width": "252px"
   },
   "navigate_menu": true,
   "number_sections": true,
   "sideBar": true,
   "threshold": 4,
   "toc_cell": false,
   "toc_section_display": "block",
   "toc_window_display": false
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}