mirror of
https://github.com/saymrwulf/nvidia-AI.summerschool17-notebooks.git
synced 2026-11-22 10:03:18 +00:00
1898 lines
344 KiB
Text
1898 lines
344 KiB
Text
|
|
{
|
||
|
|
"metadata": {
|
||
|
|
"name": "102-nearest-neighbor-and-preproc"
|
||
|
|
},
|
||
|
|
"nbformat": 3,
|
||
|
|
"nbformat_minor": 0,
|
||
|
|
"worksheets": [
|
||
|
|
{
|
||
|
|
"cells": [
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Title: Nearest Neighbor Classification\n",
|
||
|
|
"Author: Thomas Breuel\n",
|
||
|
|
"Institution: UniKL"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Nearest Neighbor Classification on MNIST Data\n",
|
||
|
|
"============================================="
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"\n",
|
||
|
|
"from pylab import *\n",
|
||
|
|
"from scipy.spatial import distance\n",
|
||
|
|
"import tables\n",
|
||
|
|
"import random as pyrandom"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 2
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"We start off by loading some data. These are images of handwritten digits. This is a fairly typical pattern classification task.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"mnist = tables.openFile(\"mnist.h5\",\"r\")"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 3
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"The data is stored in an HDF5 file, an easy way of storing and accessing array data.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"mnist"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 4,
|
||
|
|
"text": [
|
||
|
|
"File(filename=mnist.h5, title='', mode='r', rootUEP='/', filters=Filters(complevel=0, shuffle=False, fletcher32=False))\n",
|
||
|
|
"/ (RootGroup) ''\n",
|
||
|
|
"/deskewed (EArray(60000, 28, 28), shuffle, zlib(9)) ''\n",
|
||
|
|
" atom := Float32Atom(shape=(), dflt=0.0)\n",
|
||
|
|
" maindim := 0\n",
|
||
|
|
" flavor := 'numpy'\n",
|
||
|
|
" byteorder := 'little'\n",
|
||
|
|
" chunkshape := (20, 28, 28)\n",
|
||
|
|
"/images (CArray(60000, 28, 28), shuffle, zlib(9)) ''\n",
|
||
|
|
" atom := Float32Atom(shape=(), dflt=0.0)\n",
|
||
|
|
" maindim := 0\n",
|
||
|
|
" flavor := 'numpy'\n",
|
||
|
|
" byteorder := 'little'\n",
|
||
|
|
" chunkshape := (83, 28, 28)\n",
|
||
|
|
"/labels (CArray(60000,), shuffle, zlib(9)) ''\n",
|
||
|
|
" atom := Int32Atom(shape=(), dflt=0)\n",
|
||
|
|
" maindim := 0\n",
|
||
|
|
" flavor := 'numpy'\n",
|
||
|
|
" byteorder := 'little'\n",
|
||
|
|
" chunkshape := (16384,)\n",
|
||
|
|
"/test_deskewed (EArray(10000, 28, 28), shuffle, zlib(9)) ''\n",
|
||
|
|
" atom := Float32Atom(shape=(), dflt=0.0)\n",
|
||
|
|
" maindim := 0\n",
|
||
|
|
" flavor := 'numpy'\n",
|
||
|
|
" byteorder := 'little'\n",
|
||
|
|
" chunkshape := (20, 28, 28)\n",
|
||
|
|
"/test_images (CArray(10000, 28, 28), shuffle, zlib(9)) ''\n",
|
||
|
|
" atom := Float32Atom(shape=(), dflt=0.0)\n",
|
||
|
|
" maindim := 0\n",
|
||
|
|
" flavor := 'numpy'\n",
|
||
|
|
" byteorder := 'little'\n",
|
||
|
|
" chunkshape := (41, 28, 28)\n",
|
||
|
|
"/test_labels (CArray(10000,), shuffle, zlib(9)) ''\n",
|
||
|
|
" atom := Int32Atom(shape=(), dflt=0)\n",
|
||
|
|
" maindim := 0\n",
|
||
|
|
" flavor := 'numpy'\n",
|
||
|
|
" byteorder := 'little'\n",
|
||
|
|
" chunkshape := (16384,)\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 4
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"We can access this on-disk data like arrays in memory.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"imshow(mnist.root.images[0],cmap=cm.gray,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 5,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x348e7d0>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAAD5CAYAAAADZljUAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\neuGFFyRJe/bsUW5uriorK/kY
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 5
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Access to on-disk data is slower, and it's easier to just copy everything into memory.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"images = array(mnist.root.images,'f')\n",
|
||
|
|
"labels = array(mnist.root.labels,'i')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 6
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"There is a separate _test set_.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"test_images = array(mnist.root.test_images,'f')\n",
|
||
|
|
"test_labels = array(mnist.root.test_labels,'i')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 7
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Let's look at some more samples.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"print labels[:10]\n",
|
||
|
|
"for i in range(25):\n",
|
||
|
|
" subplot(5,5,i+1)\n",
|
||
|
|
" gray()\n",
|
||
|
|
" imshow(images[i])"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"[5 0 4 1 9 2 1 3 1 4]\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAVsAAAD8CAYAAADZoQcPAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nLFxqs+3o6EBoaCiAWyM+Jvso
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 8
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Let's use the first 1000 samples as our training set.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"images1k = images[:1000].reshape(1000,28*28)\n",
|
||
|
|
"labels1k = labels[:1000]"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 9
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Now we define a character recognizer by nearest neighbor classification:\n",
|
||
|
|
"\n",
|
||
|
|
"- compute all the distances between the unknown image and the training samples\n",
|
||
|
|
"- return the class associated with the sample at the smallest distance.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def ocr(image):\n",
|
||
|
|
" ds = distance.cdist(array([image.ravel()]),images1k)\n",
|
||
|
|
" md = argmin(ds[0])\n",
|
||
|
|
" return labels1k[md]"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 10
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Mathematical Formulation)\n",
|
||
|
|
"\n",
|
||
|
|
"If the training set is:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ T = \\\\{(x_1,c_1) ... (x_N,c_N)\\\\} $$\n",
|
||
|
|
"\n",
|
||
|
|
"Mathematically, we write this as:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ D(x) = c_{\\arg\\min_i ||x-x_i||} $$"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Norms and Distances)\n",
|
||
|
|
"\n",
|
||
|
|
"Note that we can formulate this in terms of a distance function:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ d(x,y) = || x - y || $$\n",
|
||
|
|
"\n",
|
||
|
|
"The distance function is here defined in terms of a metric $|| \\cdot ||$.\n",
|
||
|
|
"\n",
|
||
|
|
"Usually, this is the Euclidean metric:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ ||x||^2 = \\sum x_i^2 $$"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Let's try it out.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"ocr(images1k[0])"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 10,
|
||
|
|
"text": [
|
||
|
|
"5"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 10
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Now let's compute error rates for the test set.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"predictions = [ocr(test_images[i]) for i in range(1000)]\n",
|
||
|
|
"print sum(array(predictions)!=array(test_labels[:1000]))/1000.0"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.172\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 11
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "heading",
|
||
|
|
"level": 1,
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Test and Training Set Size"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Before continuing, let's put our data in a standard form.\n",
|
||
|
|
"Usually, we represent data sets as a large 2D array, where the rows correspond to a sample.\n",
|
||
|
|
"\n",
|
||
|
|
"Since our original data is rank 3 (or maybe higher rank), we need to reshape.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def make2d(a):\n",
|
||
|
|
" return a.reshape(len(a),prod(a.shape[1:]))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 11
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"This wraps up nearest neighbor evaluation; all you do is give it a training set and a test set, and it returns the resulting error rate.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def nneval(training,ltraining,testing,ltesting,**kw):\n",
|
||
|
|
" ds = distance.cdist(make2d(testing),make2d(training),**kw)\n",
|
||
|
|
" closest = argmin(ds,axis=1)\n",
|
||
|
|
" predictions = [ltraining[i] for i in closest]\n",
|
||
|
|
" errors = sum(array(predictions)!=array(ltesting))\n",
|
||
|
|
" return errors*1.0/len(testing)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 12
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Here is the nearest neighbor experiment we just ran.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"nneval(images[:1000],labels[:1000],test_images[:1000],test_labels[:1000])"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 14,
|
||
|
|
"text": [
|
||
|
|
"0.17199999999999999"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 14
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Note that it is important that we use different test and training sets, otherwise we get incorrect estimates of performance. This is particularly pronounced with nearest neighbor classifiers (why?), but also happens with other classifiers.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"nneval(images[:1000],labels[:1000],images[:1000],labels[:1000])"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 15,
|
||
|
|
"text": [
|
||
|
|
"0.0"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 15
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "heading",
|
||
|
|
"level": 1,
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Margin of Error"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Let's try different randomly chosen test sets. Note that we don't get the same error estimate every time.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"errs = []\n",
|
||
|
|
"for i in range(10):\n",
|
||
|
|
" S = array(pyrandom.sample(xrange(len(test_images)),1000),'i')\n",
|
||
|
|
" e = nneval(images[:1000],labels[:1000],test_images[S],test_labels[S])\n",
|
||
|
|
" print e\n",
|
||
|
|
" errs.append(e)\n",
|
||
|
|
"print \"mean\",mean(errs)\n",
|
||
|
|
"print \"sd\",var(errs)**.5"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.147\n",
|
||
|
|
"0.143"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"\n",
|
||
|
|
"0.124"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"\n",
|
||
|
|
"0.154"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"\n",
|
||
|
|
"0.147"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"\n",
|
||
|
|
"0.128"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"\n",
|
||
|
|
"0.115"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"\n",
|
||
|
|
"0.135"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"\n",
|
||
|
|
"0.144"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"\n",
|
||
|
|
"0.129"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"\n",
|
||
|
|
"mean 0.1366\n",
|
||
|
|
"sd 0.0117234807118\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 16
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Binomial Distributions)\n",
|
||
|
|
"\n",
|
||
|
|
"Each of these estimates of the error rate is a random sample.\n",
|
||
|
|
"This is similar to presidential polling.\n",
|
||
|
|
"\n",
|
||
|
|
"We're taking $n$ samples, and each sample is randomly true or false with the probability of error $p$.\n",
|
||
|
|
"It is determined by a binomial distribution.\n",
|
||
|
|
"\n",
|
||
|
|
"The variance of a binomial distribution with probability $p$ and sample size $n$ is:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ v = n p (1-p) $$\n",
|
||
|
|
"\n",
|
||
|
|
"We want to know the variance of the estimate of $p$, which is:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ \\hat{p} = \\frac{k}{n} $$\n",
|
||
|
|
"\n",
|
||
|
|
"$$ \\sigma_{\\hat{p}} = \\sqrt{\\frac{v}{n^2}} = \\sqrt{\\frac{p(1-p)}{n}} $$"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Margin of Error)\n",
|
||
|
|
"\n",
|
||
|
|
"We want to have a high confidence that our estimate of the probability of error is close to the true error rate. If we want that confidence to be 95 percent, then we need to choose a margin of error of $\\pm 2\\sigma$.\n",
|
||
|
|
"\n",
|
||
|
|
"Note that the standard deviation we calculate from the error rate is pretty close to the one we actually observe.\n",
|
||
|
|
"\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"sd = ((0.1*(1-0.1))/1000.0)**.5\n",
|
||
|
|
"print sd"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.00948683298051\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 66
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Comparing Classifiers)\n",
|
||
|
|
"\n",
|
||
|
|
"Often we want to be able to say: \"classifer 1 is better than classifier 2 because it gets a lower error rate\".\n",
|
||
|
|
"\n",
|
||
|
|
"When making such comparisons, it is essential to take into account the _margin of error_.\n",
|
||
|
|
"\n",
|
||
|
|
"If the error rate estimates of two classifiers are within the margin of error within each other, we cannot reliably determine which classifier is \"better\".\n",
|
||
|
|
"\n",
|
||
|
|
"In that case, we need more training/test data."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Publications)\n",
|
||
|
|
"\n",
|
||
|
|
"You *must* give the margin of error for any classifier error rate estimates in your reports / publications.\n",
|
||
|
|
"\n",
|
||
|
|
"In simple cases, it is sufficient to give the number of samples used for evaluation.\n",
|
||
|
|
"\n",
|
||
|
|
"In complex cases, you may want to use other kinds of _significance tests_."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "heading",
|
||
|
|
"level": 1,
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Preprocessing and Feature Extraction"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Classifier performance can often be improved with preprocessing and feature extraction.\n",
|
||
|
|
"This can be as simple as blurring.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"from scipy.ndimage import filters\n",
|
||
|
|
"def f(x):\n",
|
||
|
|
" return filters.gaussian_filter(x,1.0)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 16
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"gray(); imshow(f(images[0]))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 20,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x62ace50>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAAD5CAYAAAADZljUAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nJt+C3XPpy5z3Mho74LW8ixRZ
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 20
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Note the improved performance relative to the original classifier.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"nneval(amap(f,images[:1000]),labels[:1000],amap(f,test_images),test_labels)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 21,
|
||
|
|
"text": [
|
||
|
|
"0.114"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 21
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"It's useful to try different parameters for feature extractors. Note that we evaluate using the saem test set; this generally helps reduce the variance of our error estimates greatly.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"for sigma in [0.0,0.5,1.0,2.0,3.0]:\n",
|
||
|
|
" def f(x):\n",
|
||
|
|
" return filters.gaussian_filter(x,sigma)\n",
|
||
|
|
" print sigma,nneval(amap(f,images[:1000]),labels[:1000],amap(f,test_images),test_labels)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.131\n",
|
||
|
|
"0.5 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1234\n",
|
||
|
|
"1.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.114\n",
|
||
|
|
"2.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1506\n",
|
||
|
|
"3.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.2256\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 22
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"You can experiment with other preprocessing methods.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"imshow(filters.gaussian_laplace(images[0],2.0))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 23,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x62b57d0>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAAD5CAYAAAADZljUAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n34O+Rbu3anaV3hr+3e9+t8zP
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 23
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Again, you need to evaluate your preprocessing.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"for sigma in [0.5,1.0,2.0,3.0,4.0]:\n",
|
||
|
|
" def f(x):\n",
|
||
|
|
" return filters.gaussian_laplace(x,sigma)\n",
|
||
|
|
" print sigma,nneval(amap(f,images[:1000]),labels[:1000],amap(f,test_images),test_labels)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.5 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.256\n",
|
||
|
|
"1.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.2013\n",
|
||
|
|
"2.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1123\n",
|
||
|
|
"3.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1235\n",
|
||
|
|
"4.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.171\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 24
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"A common feature extraction method is gradient filtering. This is motivated by biological vision systems, but also has some theoretical justifications.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"from scipy.ndimage import interpolation\n",
|
||
|
|
"def gf(x,sigma=1.0,power=1,zoom=1):\n",
|
||
|
|
" d0 = filters.gaussian_filter(x,sigma,order=(1,0))\n",
|
||
|
|
" d1 = filters.gaussian_filter(x,sigma,order=(0,1))\n",
|
||
|
|
" if power!=1:\n",
|
||
|
|
" d0 = d0**power\n",
|
||
|
|
" d1 = d1**power\n",
|
||
|
|
" result = hstack([d0,d1])\n",
|
||
|
|
" if zoom!=1:\n",
|
||
|
|
" result = interpolation.zoom(result,zoom)\n",
|
||
|
|
" return result"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 25
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"imshow(gf(images[0]))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 26,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x5eb3fd0>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAWwAAADCCAYAAABt/viTAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n99vNN99sd9xxR+v3ffv29XL7
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 26
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Note that we are getting slightly better performance than with blurring.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"for sigma in [0.5,1.0,1.5,2.0,2.5,3.0,4.0]:\n",
|
||
|
|
" def f(x):\n",
|
||
|
|
" return gf(x,sigma=sigma)\n",
|
||
|
|
" print sigma,nneval(amap(f,images[:1000]),labels[:1000],amap(f,test_images),test_labels)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.5 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1817\n",
|
||
|
|
"1.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1349\n",
|
||
|
|
"1.5 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1078\n",
|
||
|
|
"2.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1088\n",
|
||
|
|
"2.5 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1241\n",
|
||
|
|
"3.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1496\n",
|
||
|
|
"4.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.216\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 27
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Normalization and Metrics\n",
|
||
|
|
"=========================="
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Nearest Neighbor Classification)\n",
|
||
|
|
"\n",
|
||
|
|
"If the training set is:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ T = \\\\{(x_1,c_1) ... (x_N,c_N)\\\\} $$\n",
|
||
|
|
"\n",
|
||
|
|
"Mathematically, we write this as:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ D(x) = c_\\{\\arg\\min_i ||x-x_i||\\} $$\n",
|
||
|
|
"\n",
|
||
|
|
"Here, we used a _norm_, written as the function $||\\cdot ||$."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Norms)\n",
|
||
|
|
"\n",
|
||
|
|
"There are many different kinds of norms:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ || x ||_2 = \\left( \\sum x_i^2 \\right) ^ {1/2} $$\n",
|
||
|
|
"\n",
|
||
|
|
"$$ || x ||_1 = \\left( \\sum |x_i| \\right) $$\n",
|
||
|
|
"\n",
|
||
|
|
"$$ || x ||_d = \\left( \\sum |x_i|^d \\right)^ {1/d} $$\n",
|
||
|
|
"\n",
|
||
|
|
"$$ || x ||_\\infty = \\max |x_i| = \\lim_{d\\rightarrow\\infty} ||x||_d $$"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Norm Axioms)\n",
|
||
|
|
"\n",
|
||
|
|
"A norm captures the idea of a _length_ of a vector. It must satisfy:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ ||\\cdot|| : {\\mathbb R}^n \\rightarrow {\\mathbb R} $$\n",
|
||
|
|
"$$ ||x||>0 \\hbox{ if } x\\neq 0 $$\n",
|
||
|
|
"$$ ||\\alpha x|| = |\\alpha|\\,||x|| $$\n",
|
||
|
|
"\n",
|
||
|
|
"Triangle inequality:\n",
|
||
|
|
"$$ ||x + y|| \\leq ||x|| + ||y|| $$"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Metrics)\n",
|
||
|
|
"\n",
|
||
|
|
"A _metric_ is a way of measuring _distance_ in a space. In vector spaces, it is closely related to norms.\n",
|
||
|
|
"\n",
|
||
|
|
"$$ d(x,y) = ||x - y|| $$\n",
|
||
|
|
"\n",
|
||
|
|
"Abstractly, norms are required to have the following properties:\n",
|
||
|
|
"\n",
|
||
|
|
"$$d(x,y)\\geq 0$$\n",
|
||
|
|
"$$d(x,y)=0 \\hbox{ iff } x=y$$\n",
|
||
|
|
"$$d(x,y) = d(y,x) \\hbox{ (symmetry)}$$\n",
|
||
|
|
"$$d(x,z) \\leq d(x,y) + d(y,z) \\hbox{ (triangle inequality)}$$"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Similarity / Dissimilarity Functions)\n",
|
||
|
|
"\n",
|
||
|
|
"There is no strict requirement that we use norms or metrics for \"nearest neighbor\" classification.\n",
|
||
|
|
"\n",
|
||
|
|
"Often, we can use dissimilarity functions. Generally, we prefer dissimilarity functions that satisfy:\n",
|
||
|
|
"\n",
|
||
|
|
"$$d(x,y)\\geq 0$$\n",
|
||
|
|
"$$d(x,y)=0 \\hbox{ iff } x=y$$\n",
|
||
|
|
"$$d(x,y) = d(y,x) \\hbox{ (symmetry)}$$\n",
|
||
|
|
"\n",
|
||
|
|
"However, none of those are necessary for good performance of a nearest neighbor classifier.\n",
|
||
|
|
"\n",
|
||
|
|
"For a _dissimilarity function_, generally, the value of the function should increase with increasing dissimilarity."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"for p in [1,2,3,4]:\n",
|
||
|
|
" def f(x):\n",
|
||
|
|
" v = filters.gaussian_filter(x,sigma=1.5)\n",
|
||
|
|
" return v\n",
|
||
|
|
" print p,nneval(amap(f,images[:1000]),labels[:1000],amap(f,test_images[:1000]),test_labels[:1000],metric='minkowski',p=p)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
" 1 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.159\n",
|
||
|
|
"2 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.147\n",
|
||
|
|
"3 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.143\n",
|
||
|
|
"4 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.152\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 18
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Normalization)\n",
|
||
|
|
"\n",
|
||
|
|
"A second important concept is that of _normalization_.\n",
|
||
|
|
"Often, an input vector is known only up to a scale factor.\n",
|
||
|
|
"\n",
|
||
|
|
"That is, $x$ and $\\alpha x$ really represent the same sample.\n",
|
||
|
|
"\n",
|
||
|
|
"We can remove this influence by transforming the data as in\n",
|
||
|
|
"\n",
|
||
|
|
"$$ x \\rightarrow \\frac{x}{||x||} $$\n",
|
||
|
|
"\n",
|
||
|
|
"Note that any norm will do, but they will yield different results for classification."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"for sigma in [0.5,1.0,2.0,3.0,4.0]:\n",
|
||
|
|
" def f(x):\n",
|
||
|
|
" v = gf(x,sigma=sigma)\n",
|
||
|
|
" v = v*1.0/norm(v.ravel())\n",
|
||
|
|
" return v\n",
|
||
|
|
" print sigma,nneval(amap(f,images[:1000]),labels[:1000],amap(f,test_images),test_labels)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.5 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1517\n",
|
||
|
|
"1.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1187\n",
|
||
|
|
"2.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.0921\n",
|
||
|
|
"3.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1276\n",
|
||
|
|
"4.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1734\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 29
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"An analogous way of computing this is to use the _consine dissimilarity_:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ d(x,y) = 1 - \\frac{x \\cdot y}{||x|| ~ ||y||} $$"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"# cosine dissimilarity\n",
|
||
|
|
"for sigma in [0.5,1.0,2.0,3.0,4.0]:\n",
|
||
|
|
" def f(x):\n",
|
||
|
|
" v = gf(x,sigma=sigma)\n",
|
||
|
|
" v = v*1.0/norm(v.ravel())\n",
|
||
|
|
" return v\n",
|
||
|
|
" print sigma,nneval(amap(f,images[:1000]),labels[:1000],\n",
|
||
|
|
" amap(f,test_images),test_labels,\n",
|
||
|
|
" metric='cosine')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.5 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1517\n",
|
||
|
|
"1.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1187\n",
|
||
|
|
"2.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.0921\n",
|
||
|
|
"3.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1276\n",
|
||
|
|
"4.0 "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1734\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 47
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Note)\n",
|
||
|
|
"\n",
|
||
|
|
"Note that a combination of normalization and feature extraction gives us the best recognition performance so far.\n",
|
||
|
|
"\n",
|
||
|
|
"You need to experiment to find good combinations.\n",
|
||
|
|
"\n",
|
||
|
|
"The best preprocessing for one classifier isn't necessarily the best preprocessing for another classifier."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Deskewing and Canonicalization\n",
|
||
|
|
"=============================="
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"There are other kinds of preprocessing that we can perform that greatly improve recognition.\n",
|
||
|
|
"\n",
|
||
|
|
"Generally, these are aimed at making recognition _invariant_ under transformation.\n",
|
||
|
|
"\n",
|
||
|
|
"We've already seen a couple of examples of this:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ x\\rightarrow \\alpha x ~~\\hbox{ canonicalize by }~~ x \\rightarrow \\frac{x}{|| x ||} $$\n",
|
||
|
|
"\n",
|
||
|
|
"$$ x\\rightarrow x + c ~~\\hbox{ canonicalize by }~~ x \\rightarrow x - \\bar{x} $$"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Canonical Elements)\n",
|
||
|
|
"\n",
|
||
|
|
"Formally, we say that the set of all feature vectors is divided into subsets, where each subset contains vectors that are equivalent from the point of view of recognition.\n",
|
||
|
|
"\n",
|
||
|
|
"For each subset, we pick a _canonical element_ via a canonicalization function.\n",
|
||
|
|
"\n",
|
||
|
|
"So, assume that some $x_r$ has class $c_r$.\n",
|
||
|
|
"\n",
|
||
|
|
"Then $S_r = \\\\{x | x = x_r + \\alpha \\\\}$ all is in the same class as $c_r$.\n",
|
||
|
|
"\n",
|
||
|
|
"The canonical element is found by the function:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ f(x) = x - \\bar{x} $$\n",
|
||
|
|
"\n",
|
||
|
|
"$$ S_r = \\\\{ x | f(x) = f(x_r) \\\\} $$\n",
|
||
|
|
"\n",
|
||
|
|
"Note that this does not work perfectly in the presence of errors."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Invariants)\n",
|
||
|
|
"\n",
|
||
|
|
"Classification is about decision functions $D(x)$\n",
|
||
|
|
"\n",
|
||
|
|
"A decision function is _invariant_ under a transformation $f$ if $D(f(x)) = D(x)$ for all $x$.\n",
|
||
|
|
"\n",
|
||
|
|
"A decision function is _invariant_ under a group transformation $G$ if $D(g(x)) = D(x)$ for all $x$ and for all $g\\in G$.\n",
|
||
|
|
"\n",
|
||
|
|
"The set of $ \\\\{ g\\cdot x | g \\in G \\\\} $ is often a set of equivalent samples and we can pick a canonical element by computing an appropriate transformation $g(x)$ and then using $g(x)\\cdot x$ as a canonical element."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"(Translations)\n",
|
||
|
|
"\n",
|
||
|
|
"Assume that classes are invariant under _shifts_ or _translations_ of the input image.\n",
|
||
|
|
"\n",
|
||
|
|
"We can use the _centroid_ as a canonicalizer.\n",
|
||
|
|
"\n",
|
||
|
|
"This is like considering the input image as a point cloud and computing the \"average point\"."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def centroid(image):\n",
|
||
|
|
" c0,c1 = mgrid[:image.shape[0],:image.shape[1]]\n",
|
||
|
|
" xm = sum(c0*image)/sum(image)\n",
|
||
|
|
" ym = sum(c1*image)/sum(image)\n",
|
||
|
|
" return xm,ym"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 20
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"y,x = centroid(images[0])\n",
|
||
|
|
"imshow(images[0])\n",
|
||
|
|
"plot([y],[x],'r+')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 23,
|
||
|
|
"text": [
|
||
|
|
"[<matplotlib.lines.Line2D at 0x4ee9dd0>]"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAQMAAAD9CAYAAAC4PhIWAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nxzEzM6NTf7kD0JyQtvulYiXm
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 23
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"The MNIST data has already been normalized so that the centroid is approximately at the center of the image (but it can't hurt to check).\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"for i in range(10):\n",
|
||
|
|
" print centroid(images[i])"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"(14.035783690482875, 13.645193370657632)\n",
|
||
|
|
"(13.553301030232202, 14.097568454157251)\n",
|
||
|
|
"(14.103585733921463, 13.543178344551675)\n",
|
||
|
|
"(14.487537587239133, 13.993461209772029)\n",
|
||
|
|
"(14.079522362948207, 13.71181319082735)\n",
|
||
|
|
"(14.074792275054866, 14.018848223932769)\n",
|
||
|
|
"(13.908193650058521, 14.408703656101897)\n",
|
||
|
|
"(13.945435392218119, 14.323666012968628)\n",
|
||
|
|
"(14.336858182082644, 14.11541252284602)\n",
|
||
|
|
"(13.822177247288515, 14.459001526253239)\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 25
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"The MNIST data hasn't been _skew corrected_. That is, characters may be slanted and/or rotated.\n",
|
||
|
|
"\n",
|
||
|
|
"To find this, we compute the major and minor axes of the covariance matrix (the second moments).\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def moments(image):\n",
|
||
|
|
" c0,c1 = mgrid[:image.shape[0],:image.shape[1]]\n",
|
||
|
|
" m0 = sum(c0*image)/sum(image)\n",
|
||
|
|
" m1 = sum(c1*image)/sum(image)\n",
|
||
|
|
" m00 = sum((c0-m0)**2*image)/sum(image)\n",
|
||
|
|
" m11 = sum((c1-m1)**2*image)/sum(image)\n",
|
||
|
|
" m01 = sum((c0-m0)*(c1-m1)*image)/sum(image)\n",
|
||
|
|
" return array([m0,m1]),array([[m00,m01],[m01,m11]])"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 26
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"c,v = moments(images[0])\n",
|
||
|
|
"print c"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"[ 14.03578369 13.64519337]\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 27
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"l,evs = eig(v)\n",
|
||
|
|
"print l\n",
|
||
|
|
"print evs"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"[ 45.88981162 12.77832657]\n",
|
||
|
|
"[[ 0.9490232 0.31520622]\n",
|
||
|
|
" [-0.31520622 0.9490232 ]]\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 29
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"y,x = c\n",
|
||
|
|
"((e00,e01),(e10,e11)) = evs\n",
|
||
|
|
"imshow(images[0])\n",
|
||
|
|
"plot([y,y+10*e10],[x,x+10*e11],'r')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 33,
|
||
|
|
"text": [
|
||
|
|
"[<matplotlib.lines.Line2D at 0x4d4ef90>]"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAQMAAAD9CAYAAAC4PhIWAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\naGw+g3Q6jdnZWZ36yx2A5oS0
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 33
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Now we want to rotate or deskew the character. For rotation, we make this vector vertical.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"\n",
|
||
|
|
"from scipy.ndimage import interpolation"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 35
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"subplot(121); imshow(images[0])\n",
|
||
|
|
"angle = arctan2(evs[0][1],evs[0][0])*180/pi\n",
|
||
|
|
"subplot(122); imshow(interpolation.rotate(images[0],angle))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 36,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x4d4f190>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAWwAAAC0CAYAAACwhRZPAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAIABJREFUeJztfW2MY+V59uWvGY+/7flcdldZSkB8ZL+aVciPkFI1G5Ai\nNiAioGpgVaCpkNoIgQLJjyS0P9jdSqiitD9QodGiSij8gazUgBIkQoBKRUpYNRL9EbUsWZad2Zmx\nx98e2+Pz/uC9nr3P48cez+x4xsd+LunR8cx4fI5n7nP5fu6P6/Y5juPAwsLCwmLg4d/pC7CwsLCw\n6A2WsC0sLCw8AkvYFhYWFh6BJWwLCwsLj8AStoWFhYVHYAnbwsLCwiOwhG1hoeGNN97A9ddfj2uv\nvRanTp3a6cuxsFDYNGFbo7YYRqytreFv/uZv8MYbb+DDDz/Eyy+/jP/5n//Z6cuysPgMzibQbDad\na665xvnoo4+cer3uHDx40Pnwww9dzwFgl119X1uN//zP/3Ruu+029fWJEyecEydOWNu2a1tXJwSx\nCbz//vv4/Oc/j3379gEA7rvvPvzsZz/DDTfcsJmXs7AYGFy4cAF79+5VX+/Zswf/9V//1fa86elp\nlMtlRKPRLb8GL72ul67VK6+7uLjY8WebComYjPrChQubeSkLi4GCz+fr6XnlchmNRgPlchn1er3P\nV2UxzKjX6yiXy2p1w6Y87F6N2sLCa9i9ezfOnz+vvj5//jz27NnT9rxoNNo3b81itDA2NoaxsTH1\ndaVS6fjcTXnYvRq1hYXXcOTIEfz+97/HuXPnUK/X8dOf/hTHjh0zPjcUCvXlGrz0ul66Vi++ro5N\nEfZGjNrCwksIBoP453/+Z9x222248cYbce+993bMzUivaCvhpdf10rV68XV1+P5/1nvDeP311/Ho\no49ibW0NDz30EH7wgx+4X9iGTSy2AZs03yuCz+fD9PT0tp/XYjSwuLjY0a43TdjrwRK2xXbAErbF\nsKEbYdtORwsLCwuPwBK2hYWFhUdgCdvCwsLCI9hUHbaFhYXFdqBj8m1Ec2SWsC0sLAYO6yWT+fNR\nI24bErGwsBgomMjacZyO3x8lWMK2sLCw8AhsSMTCwmKgYApzdPreqHnYlrAtLCwGFpKou8WrR4W4\nLWFbWFgMJHw+nyLpXsjacRwjwQ8TmVvCtrCwGChIkjY9lmi1WuoxE5P684YpdGIJ28LCYqAgybXX\nypBRKfOzhG1hoWHfvn1IJBIIBAIIhUJ4//33d/qShgY6oXbzfFutlvKaZdhD/q7f71fedyeyNn0A\neJXYLWFbWGjw+Xz41a9+hUwms9OX4jl0I8ReKz34davVUqRtIm8A8Pv9CAQC6tjtXKYPB6/FuS1h\nW1gY4JUbeJCge7KdvNhuMWXpSZOo19bWsLa2puLVkrSDwaB6zV6ujddlSlB64X9uCdvCQoPP58PX\nvvY1BAIB/PVf/zX+6q/+aqcvyRNYj/Q6edKdvvb5fPD7/eoxj1x+vx/BYBChUAjBYBDBYFCRPIle\neuaS8G1IxGJd6CVKGzUan8+HQCCgVjAYRCAQML6eNGy51tbW0Gw20Wg00Gw20Ww2sba25jJ6Lnku\nufUE2m8uvi6P+mP59aB7Mu+99x527dqFxcVFHD16FNdffz1uueUW13PkdOtQKLRtI6IGHevFkeX/\n3nQ/yJCK3+9HKBSC3+93kTOH1tJmpZ02Gg1Uq1XUajXUajXU63XloZO0B42s6/U6Go1GT8+1hN1n\nrEeiG4Hf78f4+LhaY2NjGB8fdyVe5Hn0+J7f70e9XkelUkG1WlXHRqOBiYmJtiVvDD6mFwO4t6b1\neh2rq6vqRpGP5eKNM8jYtWsXAGB6ehp33XUX3n///TbCttPSP0Mvtc8y9ixL70x11pK0g8Ggy/7C\n4TAikQii0SgikQgmJiZctu33+1Eul1EoFNQCoMiQ3rY83yBgI1PTr4iwbTa9OzqRdS+ZbRMCgYAy\nWmm49LL1c9ALlx5zpVJRxkzy9fl8iEajiMfjSCQSiMfjiMfj6qYIh8PqSMPSb8BqtYpyuYxSqdR2\nLJVKyrtfXV3d8r/zVqJSqWBtbQ3xeBzlchm/+MUv8OMf/3inL2sgYQphyMemkAfDFbRR3gum1wkG\ngxgfH1cOBO0zlUohmUwiHo+32f3KygrC4TD8fr/a2QFo86wHkbh7wRURts2m9wbpPUgj5eNeEQgE\nMD4+jmg0ikQioVYoFGoja24n9VUsFrG0tKTIul6vw3EcRCIRJJNJTE5OIpPJIJPJIB6PIxqNIhaL\nqWM4HDZ6TKVSCSsrK8jn864jr41kPeg3yMLCAu666y4AQLPZxF/8xV/g61//+g5flbfQiaxNXnY3\n4gyFQgiHw4jFYojH40in05icnFQrlUq5Xh8AIpEIfD4fGo0GyuUyarUaWq0Wms3muufzAq44JDLo\n29udgmmrJ5fcxvUKacDJZFIRK+N80nsnuXO7xce5XM5F1uVyGc1mE9FoFKlUClNTU5idncXMzAzS\n6TQSiQSSyaQ6RiKRthKrVquFfD6P5eVlLC0tqePY2Bh8Ph9arRZWV1eVpz3IuPrqq3H27NmdvgxP\noFNNtanumTFm/p60VenEyJVOp5HJZFxORDKZRCqVQiqVQiwWU/FfLunF64lHeW55fV6pEAG2wMP2\nejbdFH/r9rUpVtzt+/JrmShkmGKjhE3DlYYsEzMyHEKylisYDKLZbKqYc7VaRSAQUDfH1NQUZmZm\nMDc3h3Q6jWQyaSRsfYVCIbRaLTQaDayurqJSqSAcDmN8fFwlhzbyXi0GG72QtXwubVImseU9Ytp5\nptNpzM7OYm5uDnNzc8hkMojFYmpNTEygUqmgUqmgXC4rwjaVA+qELc/tFbIGrpCwe8mmDzpMn/by\nZ/KxXjGhL/l9E5nL5J1M6PWKQCDgIk8eSYZ6SET/cAgGg6jX60gkEqjVasrAy+Wy8qqnpqaQyWSQ\nSqWQSCQQiUQU0QNQ20veDHwsbxy5qtUqVldX0Wg0XLoPFoOFbqTVyZHRPVVTUpGgA8F7QP/wlrtO\n3kPpdBozMzPYu3cv9u7di8nJSZVPmZiYQDAYRD6fBwBVHdKprI/wutNwRYTdSzZ9kNGpmqJTOINx\nYJYXdftaD1GwwiMcDqvFSoxe4ff7EY1G21anpKMpZt5oNJBMJlV5nd/vR7VaxfT0tFqZTAbpdBrR\naFR5ydzOylI9lgY2Gg1F2DppS8L2QoWIRTs6CSp1qwiRzwOgChNo+36/v62ZRXcu6GHv2bMHf/RH\nf4TJyUnXfcbXbjabqFQqrusxkfZGk/yDiE0T9rBk06XnLOua9ZAGCZfxYP2x/JoehL7C4bCq7OAx\nHA73fK2yrI9ESuPXP1yA9m2q4ziKcJk1DwaDWF1dVYkchltSqRTC4bC6OXTCbjQaqNfrarFMkKtU\nKqmywXq9rs5pMZhgHLcXQusW/qAtSuINhUKYmJhQVUcTExPKnuRr6CWorAThfcLdqKzrLxaLyOfz\nKBQKWFlZQaFQUMlGWYOte9pexaYJe1iy6SRTPa6sk3YgEHCVuHV7TI9UrxFlpQXL5uLxOCYmJnq+\nVt4Iuidi2qICcBmrbB7gjTk2NoaJiQnU63Wk02nXSqVSbclM4HJIhDFwLpN3XSqVVE229bAHHyYv\n2vT/0jU99F1dMBhsKwelg8LVraWc30smk4rc19bWUK1WXTu71dVVLC8vu1Y2m1WkLR2FkSfsYcmm\n09CkN2AKMZCwpXfMkIR8HI1GXQX9crG6Q67NNGD06gXJ7DkX3y+9nmg0irW1NVdMnMvv97clF6WH\nzeYYetd6/NoUErEYTHSyKT1OzaNM7ukhxbGxMUQiEZdjwtJQLhnS4FG3NUnYzWYT1WrV1fBVLpeR\nzWaRy+XUMZfLIZ/Po1wuY3V1VYX+Rp6wvQY9rktPwES+9Fr1MjnZBSg9an2x+1D3siORiMpuM5HX\nryQIKzakgXOx65CPScIkYHaL
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 36
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"# computing a rotation angle for each image\n",
|
||
|
|
"def rotation_angle(image):\n",
|
||
|
|
" c,v = moments(image)\n",
|
||
|
|
" l,evs = eig(v)\n",
|
||
|
|
" i = argmax(l)\n",
|
||
|
|
" angle = arctan2(evs[i][1],evs[i][0])*180/pi\n",
|
||
|
|
" return angle\n",
|
||
|
|
"def derotate(image):\n",
|
||
|
|
" return interpolation.rotate(image,rotation_angle(image),reshape=0,order=1)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 62
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Result of canonicalizing by rotation.\n",
|
||
|
|
"\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"print labels[:10]\n",
|
||
|
|
"for i in range(16):\n",
|
||
|
|
" subplot(8,4,2*i+1)\n",
|
||
|
|
" gray()\n",
|
||
|
|
" imshow(images[i])\n",
|
||
|
|
" subplot(8,4,2*i+2)\n",
|
||
|
|
" imshow(derotate(images[i]))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"[5 0 4 1 9 2 1 3 1 4]\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAT4AAAD9CAYAAAAlMxueAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n3kNxcTG2bduG06dP66pa/QJ9
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 63
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"It turns out that canonicalizing by rotation isn't the best thing to do. It's generally better to _deskew_.\n",
|
||
|
|
"Skewing is a coordinate tranformation of the form:\n",
|
||
|
|
"\n",
|
||
|
|
"$$ M = \\left(\\begin{array}{cc} 1 & 0 \\\\\\ \\alpha & 1 \\end{array} \\right) $$\n",
|
||
|
|
"\n",
|
||
|
|
"This is also called a _shear_ transformation. The further a pixel is removed from $y=0$, the further its $x$ coordinate is shifted."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"# deskew transformation\n",
|
||
|
|
"def deskew(image):\n",
|
||
|
|
" c,v = moments(image)\n",
|
||
|
|
" alpha = v[0,1]/v[0,0]\n",
|
||
|
|
" affine = array([[1,0],[alpha,1]])\n",
|
||
|
|
" ocenter = array(image.shape)/2.0\n",
|
||
|
|
" offset = c-dot(affine,ocenter)\n",
|
||
|
|
" return interpolation.affine_transform(image,affine,offset=offset,order=1)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 46
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"subplot(121); imshow(images[3])\n",
|
||
|
|
"subplot(122); imshow(deskew(images[3]))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 47,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x471efd0>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAWwAAACzCAYAAACtgCb3AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAIABJREFUeJztnVuIJHf1x789PX2/zs49Oxs3JFly310MKmK8kGyMoGtC\nRBI0Lsl6wRcRRcyDQnzRzUOQeHkIGsPCH4J5yeUhWaKIGgOyD+6CGNGAu2SzmftM3+8z9X9Yzm9P\n/fpX1T093b3TVecDxfRMd1dVz5z6zqnzO5eAZVkWBEEQhD3P2LU+AUEQBKE7RLAFQRBGBBFsQRCE\nEUEEWxAEYUQQwRYEQRgRRLAFQRBGhJ4F+8yZM7jllltw88034+mnn+7nOQnCNUVsW9izWD3QarWs\nG2+80bpw4YLVaDSsw4cPW++8847tNQBkk23gW78R25ZtL2xOjKMHzp49i5tuugkHDx4EADzyyCN4\n9dVXceutt/ayO0HYM3Rr29dffz0AIJ/PI5PJDOXc5Fijdaxej/fee+85PtdTSOTy5cs4cOCA+n5h\nYQGXL1/uZVeCsKfo1rbz+Tzy+TxqtRpqtdowT1HwGLVaTdlTPp93fW1PHnYgEOjpxARhr9OtbXOv\nKRqNDup09jSWoauFaMPOiUajNhtyE+2eBHv//v24dOmS+v7SpUtYWFjoZVeCsKfYqW1HIpFhnNae\nOZZJpJ2e71a898LnGpXj9RQSufvuu/Huu+/i4sWLaDQa+P3vf4/jx4/39cQE4VqwU9sepnd9LY9l\nWVZHsdbp9vVe/R0O4ng9edjj4+P41a9+hc9+9rPY2trCyZMnZcFR8ARi2zujk0dtWZaESfpIwNrp\nv81udyx/JGEIDMh8XQkEAipLxA+4/Y7pOafrXXRg57z33nuOv/OePGxBEPwDia5JnHci1Nfin6vX\nEMEWBKErdBE2ibKbUPPnRLx7QwRbEAQb3YY5uhFw037ptSLaO0cEWxAERScR5WLtJNxO+9VFWkR7\n54hgC4JgRM/wcBJrJ09bF2OTaAs7QwRbEAQFF1S3xUUuuvTYJNBuzzvtW3BGBFsQBBvdxqKdRNv0\nc9Mxul20FK4igi0Igism75oLM/9ef85pf2NjY8Y4uBTauCOCLQiCEacUPb5tb2+3Paavpn2RSFuW\nhbGxMSXcXKhFtJ0RwRYEQdFtxgcXZ9r07/V9kliPjY3ZnuOLkSLU7ohgC4IAoLvYdaeFQxJjEmfT\nRp4197CdMksEOyLYgiA44pTpMTY2psIaQLtAkxDr4kzv0/dp8tC5ly5cQQRbEARXeMyahze4dzw+\nPo5gMIjx8XHb42AwiLGxMQSDQfV4a2urbWu1Wuprq9Vqi2kLVxDBFgTBiFPmhx7OCAaDiEQiCIfD\nti0UCrUJ+NjYGBqNBprNJprNpnpMXymWzT1sKba5igi2IAjGIhk9bW97e9sWAiHPORQKIRKJIBaL\nqS0ajSrRpo0Em+Zg1mo11Ot11Go15X3TcVqtVtv5iWiLYAuC4IKbl01ecygUQjQaRTweRzKZRCKR\nQCKRaPO6Q6EQxsbGUC6XUalU1BYMBtU/CRJrnu4nXEUE26O4xRX1LRgMGvfBb1v5bay+MCSLQ6OP\nW8yYp9zxeDSJNRfsRCKBVCqFVCqFWCyGcDiMSCSitrGxMSSTSVSrVdtWLBZRKpVQLBYxPj6uYuOm\nlEE/I4LtUQKBgLqQaItEIuox3bbSravp/ZVKBaVSybZVKhW1MMQXiMQT8iY8XY9CIrpokyjHYjEl\n2JlMBvF4vM3uxsfH28Ih9XoduVwOuVwOoVAIgUAAW1tb2N7eti1MAhDB3s2bDx48iHQ6reJYZ8+e\n7dd5CbuEBDsWiyGZTLZtqVRKPY7FYsYc3EKhgI2NDayvr2NjY0N5PuRxA968gPxq106FK1y0dbEm\nL1sX7HQ6jUQiYYtrx2IxjI+PK/vhWzwet4l1vV5Hq9VSd3WAN21tp+xKsAOBAP785z9j3759/Tof\noU9wwU6lUpiYmEAmk8HExAQmJiaQzWbV42QyaXsfsb6+jqWlJSQSCYRCIVt8EYDygLwWa/SzXeud\n+ng4hIdEuFibBDuTySCZTCIej9u2cDiMZrNpE+NWq4VoNKq8+EajgUqlgnq9bgvTeNHWdsquQyJ+\n/uXtZQKBAMLhMGKxGNLpNCYmJjA1NYXp6em2LZPJGLMESKzHx8exvb2t4o3A1QuIvB+v4Xe7dpqG\nzr1sEm2nkEgqlVILkHwh0pSHTQ5BvV5HpVJBoVBApVIBYHcM/M6uPez77rsPwWAQ3/rWt/CNb3yj\nX+cl7ABTdVk4HEY8Hkc6nUY2m8Xk5CRmZmZs2/T0NGZmZpDNZtV++D4BoFwuI5/PY2NjQy0iNRoN\ntVjp1CZzlBG7dsapzJy8bRJvWifRPexoNNq2YG1ZFmq1GorFIgqFAgqFApLJJCqVii17hPK0Af/+\nQ92VYL/99tuYn5/H6uoqjh07hltuuQX33HNPv85N6BK+Yk9fY7EYJicnMTk5qTzrmZkZTE5OIp1O\nIx6PIxKJ2DJE3DIFTBeIVy+abuw6n8+rxyRQXob+1qYSclog5AJM7+GvoUpGUwtWKr6hNZdsNotm\ns6nWTUiw6/W6LZbtBRukhdduGOv8Emfm5+cBANPT03jooYd8sziz1yCPmow9k8lg3759Sqx5KGRy\nchKZTAaJRALhcNjW2wEwN6X3G93YdSaTUZvXxFq/azIV0Dhtpvarumjz19HxKLQSj8eRSqWQzWaR\nzWbV4mU0GlV53LqXP+pEo1GbPbnRs2BXKhUUi0UAV26b33zzTdx555297k7YBWNjYyqFjwR7YmLC\nJtok2Pv27TN62G6i7ScR95tdO/Wt1p93EmvuXTsNMtDT87hoE8FgsM3pyGazSKVSylZJsEm03T6H\nV+k5JLK8vIyHHnoIANBqtfCVr3wF999/f99OTOge3dhpkVH3smdmZlTGB1WfkYdtaiLvl4uA4ye7\n5mLs5qm6CXE3IREeDqESdN1L1m2YN4Oq1WqoVCpKsLe3tx3vALzgcbvRs2DfcMMNOH/+fD/PRegR\nPSSSzWaVWOsxbKo24wuUHCej94uA+8Wu3f6WTr1EegmJmOLY1JaVt2ilGHY8Hldx7q2tLdRqNZTL\nZRSLxbaQiOlzeF20pdLRA+j5seRB86Y71HjH1DQe8L6hC3bc8pmdStP581yEqeKVP6bv9TCIXvxC\nom56HY9xc/um19Nr+Pl53YZFsD0Av63UBdtJrAVhJ2O5TKKtC7TTpnvf/DHt15SbzUVbF+xAIKA8\ncd6G1euIYHsAk4fNvWuKG5omVfvByAVnTIt3TjbBwxCmdD3TpnvNFH/m/yyogtbkletNp0KhkC3F\nkO/PD4hgewC3kAgJNi9y8YtxC73Raa3CKSRCGy85N4U6dKcBQEcPm99B6rngfrJnEWyPYPKweftU\nPRziJyMXesOpPJ3HnQF0DImYUv/0/Tp55HpIJBQK2Tr5+c2ORbBHDF6CTpteCpxIJFTjHWqfShVj\nBMX9+MWnFzdsbGwgn8+jWCyiWq2qntj8IvRL9ohwFVMrBN0mnd5D6BkoZHe8/zq3Ny7mPA7uN0Sw\nRwh+W8hbXKbTaVVowNP5stmsqhIjwebGTl5Mq9VSjZ1qtZp6vLKygg8++ACLi4tYW1tDoVBQr2k0\nGmpBSfA2po59fAAGnyyjr5/wuzza9IVHchRIrOv1OqrVqm0qDR92oIdb/IQI9ohBt4W8+IUEmxfL\nUM+QZDJpFGy+WNRoNJDP55HP51EoFNTjtbU1rK2tYXV11SbY9XpdCbbfLhg/YUr94/2weWtVXbRN\nU41oM+Vrc1skp4GLNRdvU4jFL02hRLBHCPKw9Uky1M6SelyTYPMOadQcnnvV5KlUq1Xk83msrq6q\nbW1tDZubm0q8aatUKrZFJb/emvoN7mWbuvN1Emu+lqILrJuHzb/S1u1dnRdzs0WwRwgu2FQVRi1U\nTSERfiFxD5u8GbpAuGBfvnxZ
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 47
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"# effect of deskewing\n",
|
||
|
|
"print labels[:10]\n",
|
||
|
|
"for i in range(16):\n",
|
||
|
|
" subplot(8,4,2*i+1)\n",
|
||
|
|
" gray()\n",
|
||
|
|
" imshow(images[i])\n",
|
||
|
|
" subplot(8,4,2*i+2)\n",
|
||
|
|
" imshow(deskew(images[i]))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"[5 0 4 1 9 2 1 3 1 4]\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAT4AAAD9CAYAAAAlMxueAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\niooKiMVi8Hg8vPvuu5g7dy7m
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 49
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Let's now look at the error rates achieved with deskewing:\n"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"print \"derotated\",nneval(amap(derotate,images[:1000]),labels[:1000],amap(derotate,test_images),test_labels)\n",
|
||
|
|
"print \"deskewed\",nneval(amap(deskew,images[:1000]),labels[:1000],amap(deskew,test_images),test_labels)\n",
|
||
|
|
"print \"original\",nneval(images[:1000],labels[:1000],test_images,test_labels)\n",
|
||
|
|
"\n"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"derotated "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.1308\n",
|
||
|
|
"deskewed "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.0761\n",
|
||
|
|
"original "
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"0.131\n"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 65
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"metadata": {},
|
||
|
|
"source": [
|
||
|
|
"Not that deskewing results in substantial improvements, while derotating hardly gives an improvement at all.\n",
|
||
|
|
"\n",
|
||
|
|
"(We could analyze this further with confusion matrices, and there may be ways of fixing this.)"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [],
|
||
|
|
"language": "python",
|
||
|
|
"metadata": {},
|
||
|
|
"outputs": []
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"metadata": {}
|
||
|
|
}
|
||
|
|
]
|
||
|
|
}
|