mnist digit recognition demo

digit.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import numpy\n",
    "import theano\n",
    "import theano.tensor as T\n",
    "import time"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "data=numpy.loadtxt(\"train.csv\", skiprows=1, delimiter=\",\", dtype=numpy.float32)\n",
    "test_data=numpy.loadtxt(\"test.csv\", skiprows=1, delimiter=\",\", dtype=numpy.float32)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "labels=numpy.zeros((data.shape[0],10), dtype=numpy.float32)\n",
    "for i in range(data.shape[0]):\n",
    "    labels[i,int(data[i,0])]=1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "h=100\n",
    "d=28*28\n",
    "c=10\n",
    "n=data.shape[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "data_shared = theano.shared(data[:,1:])\n",
    "labels_shared = theano.shared(labels)\n",
    "M1 = theano.shared(numpy.random.rand(d,h).astype(numpy.float32)-0.5)\n",
    "b1 = theano.shared(numpy.random.rand(h).astype(numpy.float32)-0.5)\n",
    "M2 = theano.shared(numpy.random.rand(h,h).astype(numpy.float32)-0.5)\n",
    "b2 = theano.shared(numpy.random.rand(h).astype(numpy.float32)-0.5)\n",
    "M3 = theano.shared(numpy.random.rand(h,c).astype(numpy.float32)-0.5)\n",
    "b3 = theano.shared(numpy.random.rand(c).astype(numpy.float32)-0.5)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "hidden_layer = T.nnet.sigmoid(data_shared.dot(M1)+b1[None,:])\n",
    "hidden_layer2 = T.nnet.sigmoid(hidden_layer.dot(M2)+b2[None,:])\n",
    "prediction = T.nnet.softmax(hidden_layer2.dot(M3)+b3[None,:])\n",
    "\n",
    "error = T.sqr(prediction-labels_shared).sum(1).mean()\n",
    "\n",
    "grad_M1 = T.grad(error,M1)\n",
    "grad_b1 = T.grad(error,b1)\n",
    "grad_M2 = T.grad(error,M2)\n",
    "grad_b2 = T.grad(error,b2)\n",
    "grad_M3 = T.grad(error,M3)\n",
    "grad_b3 = T.grad(error,b3)\n",
    "\n",
    "epsilon = T.scalar()\n",
    "f = theano.function([epsilon],T.sqrt(error),updates=[(M1,M1-epsilon*grad_M1),\n",
    "                                                    (M2,M2-epsilon*grad_M2),\n",
    "                                                    (M3,M3-epsilon*grad_M3),\n",
    "                                                    (b1,b1-epsilon*grad_b1), \n",
    "                                                    (b2,b2-epsilon*grad_b2),\n",
    "                                                    (b3,b3-epsilon*grad_b3)])\n",
    "\n",
    "input_data = T.matrix()\n",
    "test_hidden_layer = T.nnet.sigmoid(input_data.dot(M1)+b1[None,:])\n",
    "test_hidden_layer2 = T.nnet.sigmoid(test_hidden_layer.dot(M2)+b2[None,:])\n",
    "test_prediction = T.nnet.softmax(test_hidden_layer2.dot(M3)+b3[None,:])\n",
    "\n",
    "predict = theano.function([input_data],test_prediction)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "f(0.1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "eps = 0.05\n",
    "start = time.perf_counter()\n",
    "for i in range(10):\n",
    "    for j in range(10):\n",
    "        f(eps)\n",
    "    print(f(0.0))\n",
    "# print(time.perf_counter()-start)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "print(predict(data[:,1:]).argmax(1))\n",
    "print(data[:,0].astype(int))\n",
    "(predict(data[:,1:]).argmax(1)==data[:,0]).sum()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "numpy.savetxt(\"test_predictions.csv\",\n",
    "              numpy.array([range(1,test_data.shape[0]+1),predict(test_data).argmax(1)]).transpose(),\n",
    "              fmt='%d,%d',header='ImageId,Label', comments=\"\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.4.4"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}