mirror of
https://github.com/saymrwulf/nvidia-AI.summerschool17-notebooks.git
synced 2026-11-21 09:54:43 +00:00
2023 lines
866 KiB
Text
2023 lines
866 KiB
Text
|
|
{
|
||
|
|
"metadata": {
|
||
|
|
"name": "hmm-ocr"
|
||
|
|
},
|
||
|
|
"nbformat": 2,
|
||
|
|
"worksheets": [
|
||
|
|
{
|
||
|
|
"cells": [
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"from pylab import *",
|
||
|
|
"def ls(A): ",
|
||
|
|
" \"Normalize a matrix to be a stochastic matrix.\"",
|
||
|
|
" return A*1.0/maximum(1e-6,sum(A,0)[newaxis,:])",
|
||
|
|
"def vs(v): ",
|
||
|
|
" \"Normalize a vector to sum to 1.\"",
|
||
|
|
" return v*1.0/maximum(1e-6,sum(v))",
|
||
|
|
"def unary(i,n):",
|
||
|
|
" \"Generate a unary vector.\"",
|
||
|
|
" assert i<n",
|
||
|
|
" v = zeros(n)",
|
||
|
|
" v[i] = 1",
|
||
|
|
" return v"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1540
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Hidden Markov Models",
|
||
|
|
"====================",
|
||
|
|
"",
|
||
|
|
"- speech recognition",
|
||
|
|
"- online handwriting recognition",
|
||
|
|
"- OCR",
|
||
|
|
"- DNA sequence analysis"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Training Data",
|
||
|
|
"=============",
|
||
|
|
"",
|
||
|
|
"Let's build an HMM model of a character. To do this, we need to construct training data that we can use in the rest of this worksheet.",
|
||
|
|
"For that, we use a page of 30 rows of 26 upper case letters, arranged in a widely spaced grid."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"letter = 1-mean(imread(\"../scans/chars/letter.png\"),2)",
|
||
|
|
"gray()",
|
||
|
|
"imshow(letter[:300,:300])"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1625,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x99eb8050>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAXIAAAFsCAYAAADR+DLiAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nG6BUwO97+Nv94oyhLZ1HOVEm
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1625
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"h,w = letter.shape",
|
||
|
|
"ch,cw = h//30,w//26",
|
||
|
|
"lines = [letter[i*ch:(i+1)*ch,:] for i in range(30)]",
|
||
|
|
"chars = [[l[:,j*cw:(j+1)*cw] for j in range(26)] for l in lines]",
|
||
|
|
"chars = array(chars)",
|
||
|
|
"chars = array(chars>0.5*amax(chars),'B')",
|
||
|
|
"Es = chars[:,4,:,:]",
|
||
|
|
"gray()",
|
||
|
|
"imshow(Es[5])"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1626,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x9a46bc90>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAASkAAAFrCAYAAACNCX/lAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAGc9JREFUeJzt3dtvm/Udx/GPE6dJExzn0MRO6rRxQ5M2h6aB0koMrYMs\nRdoIlLVDwFQqQNxwM9A02P4Bmo5NWrvtCnUo2qSxXnUFARtVx+jGSreljGObkibUSW2XnBrnaCf5\n7QLiEdqkp8T+Bb9f0k9afHiebw68Zz99bDuMMUYAYKm0ZA8AAPMhUgCsRqQAWI1IAbAakQJgNSIF\nwGrXHanXX39d69at09q1a7V3796FnAkA/s9ch8nJSVNRUWE6OztNNBo19fX15qOPPpp1G0ksFot1\n1Wsu1/VI6sSJE7r55ptVXl6ujIwMPfjgg/rTn/50PZsCgHldV6R6enpUVlYW/9rn86mnp2fBhgKA\nGdcVKYfDsdBzAMBlXVekVq5cqUAgEP86EAjI5/Mt2FAAEHc9B85jsZhZs2aN6ezsNBMTExw4Z7FY\nN7zm4tR1cDqd+vWvf627775bU1NTevzxx7V+/frr2RQAzMvxxaOehd8wx60AXIO5UsQZ5wCsRqQA\nWI1IAbAakQJgNSIFwGpECoDViBQAqxEpAFYjUgCsRqQAWI1IAbAakQJgNSIFwGpECoDViBQAqxEp\nAFYjUgCsRqQAWI1IAbAakQJgNSIFwGpECoDViBQAqxEpAFYjUgCsRqQAWI1IAbAakQJgNSIFwGpE\nCoDViBQAqxEpAFYjUgCsRqQAWI1IAbAakQJgNSIFwGpECoDViBQAqxEpAFYjUgCsRqQAWI1IAbAa\nkQJgNSIFwGpECoDViBQAqxEpAFYjUgCsRqQAWI1IAbAakQJgNSIFwGpECoDViBQAqxEpAFYjUgCs\nRqQAWI1IAbAakQJgNSIFwGrzRioQCOjOO+9UTU2NamtrtX//fklSf3+/mpqaVFlZqW3btmlwcDAh\nwwJIPQ5jjJnrylAopFAopI0bN2p4eFi33nqrDh06pBdffFErVqzQM888o71792pgYEAtLS2zN+xw\nLPrwAL4+5kyRuQb33XefeeONN0xVVZUJhULGGGOCwaCpqqq65LaSWCwW66rXXOZ9JPVlXV1d2rp1\nqz744AOtWrVKAwMD0udbVkFBQfzrGTySAnAt5krRVR04Hx4e1o4dO7Rv3z65XK5Z1zkcDoIEYNFc\nMVKxWEw7duzQrl27tH37dkmSx+NRKBSSJAWDQRUXFy/ulABS1ryRMsbo8ccfV3V1tZ566qn45ffe\ne69aW1slSa2trfF4AcBCm/eY1N///nd985vf1IYNG+JP6fbs2aPNmzfrgQce0Llz51ReXq6DBw8q\nLy9v9oZ5CgjgGsyVoqs+cH6tiBSAa3FDB84BIFmIFACrESkAViNSAKxGpABYjUgBsBqRAmA1IgXA\nakQKgNWIFACrESkAViNSAKxGpABYjUgBsBqRAmA1IgXAakQKgNWIFACrESkAViNSAKxGpABYjUgB\nsBqRAmA1IgXAakQKgNWIFACrESkAViNSAKxGpABYjUgBsBqRAmA1IgXAakQKgNWIFACrESkAViNS\nAKxGpABYjUgBsBqRAmA1IgXAakQKgNWIFACrESkAViNSAKxGpABYjUgBsBqRAmA1IgXAakQKgNWI\nFACrESkAViNSAKxGpABYjUgBsBqRAmA1IgXAakQKgNWIFACrESkAViNSAKxGpABYjUgBsNpVRWpq\nakoNDQ1qbm6WJPX396upqUmVlZXatm2bBgcHF3VIAKnrqiK1b98+VVdXy+FwSJJaWlrU1NSk9vZ2\nNTY2qqWlZVGHBJDCzBUEAgHT2Nhojh49au655x5jjDFVVVUmFAoZY4wJBoOmqqrqkvtJYrFYrKte\nc3HqCp5++mk9//zzGhoail8WDofl8XgkSR6PR+Fw+EqbgSXS09OVkZERX/i/yclJxWIxxWIxTU5O\nJnscfGHeSL3yyisqLi5WQ0OD3nzzzcvexuFwxJ8Gwn4ul0slJSUqLS2N/x8NPnfhwgUFg0EFg0H1\n9/cnexx8Yd5Ivf322zp8+LBeffVVjY+Pa2hoSLt27ZLH41EoFJLX61UwGFRxcXGi5sUNcrlc8vv9\n2rBhg9avX5/scazS3t6u999/XyMjI0TKIo4vjh9d0d/+9jf9/Oc/18svv6xnnnlGhYWFevbZZ9XS\n0qLBwcFLDp7z6MpO1dXVuuuuu9TY2KhvfOMbyR7HKsePH9fRo0d19OhRvffee8keJ+XMlaIrHpP6\nspnw/OQnP9EDDzygAwcOqLy8XAcPHrzxCZEQTqdT2dnZcrvdKioqSvY4VsnLy1NOTo6czmv6zwKL\n7Kp/G1u3btXWrVslSQUFBTpy5MiiDQUAMzjjHIDVeFybYqanpxWNRjU2Nqbh4eFZ16WlpSk9PV1O\np1Pp6ekLsj9jjKampjQ5Oampqak5jzvM7H9m3wu1fyx9RCrFRCIRdXZ26sSJExoZGZl1XW5urjwe\nj0pKShbs9ISJiQmFw2EFg0GFw2FFo9E5b5ufny+v1yuv16sVK1YsyP6x9BGpFBOJRHT27FnFYjF1\nd3fPus7r9aq2tlYZGRkLFqloNKqenh69//77+uCDDzQ2NjbnbVevXq3a2lplZmYSKcQRqRQz80gq\nGAwqMzNz1nU333yznE6nSkpKFmx/M5H673//qzfffHPWKxe+qq6uTsuWLdPKlSsXbP9Y+ohUipl5\n2cdXj0dJ0rJly9TR0SGv16v8/Hy5XK74+mrQvrrNSCQSX9PT0/Hr+vr61N7ers7OTnV3dysSicy5\nHbfbrY6ODnk8HmVlZd3YN/ql7+nL3wfn7y09RApxw8PD6urqUmZmpiKRiPx+v/x+v9asWTNvpMbH\nx9Xd3a3Ozs74U8kvb/P06dMKBoNXfD3cxYsX1dHRIafTqQsXLizI95SXlxf/PnJycjggvwQRKcSN\njIyoq6srftxq8+bNSktLk8fjUX5+/pz3GxsbUyAQUFtbm06cOKGJiYn4ddFoVP39/err65sVr8sZ\nGhpSR0eHBgYGdOrUqQX5nrxer8bGxpSTk6Py8vIF2SYSi0ghbnR0VKOjo+ru7pbD4VBWVpZKS0tV\nW1s77/0mJiZ0/vx5ffjhh3rrrbc0Ojp6Xfufebr46aefXtf9L8fv9ys/P19+v3/W01AsHZzMCcBq\nRAqA1YgUAKsRKQBWI1IArEakAFiNSAGwGpECYDUiBcBqRAqA1XhZDL7WYrGYLl68qFAopLNnz877\ngajnz5/XwMDArNceIvmIFL7WxsbGdO7cObW1tWlsbGzed0Ho6OhQR0fHvO95hcQjUvham4nU+Pi4\nAoGA0tLmPsLR39+vCxcuECnLECl8rc28jczMOzvMxxgTX7AHkYI1cnNzVVhYqMLCwkvev2p4eFi9\nvb3q6+u7po9AJzpLH5GCNXJzc7VmzRqtW7dOfr9/1nWhUEinTp3SqVOnrilSWPqIFKyRm5uriooK\nbdmyRbfccsus686cOSNjjD777DN98sknSZoQyUCkYI3ly5erqKhI5eXlqqmpmXVdenq6wuGwenp6\ndP78eY2Pj2tsbEzj4+NXfO90LG1ECkuCy+WS3+/XyMiIcnJy1N3drUAgoEAgMO8n0GDpI1JYEmYi\nlZ2drbKyMp08eVIOh0O9vb1E6muOSGFJuOmmm5Sdna1Vq1ZpfHw8HqiF+lQZ2ItI4YZlZmaqpKRE\n1dXVikQiunDhgvr7+9Xf33/ZDyGd4XQ6VVBQoIKCAuXn56u2tlbl5eVyu92X3DYtLU1paWlyOp1y\nOBxatmxZ/H/j641I4YZlZWWprKxMsVhMbrdbZ86cUXt7uyYnJ+eNVEZGhkpLS1VZWanKykpVVFRo\nzZo1KigoSOD0sB2Rwg1bvny5fD6f3G63KioqlJ+fr8nJSQWDQZ0/f37O+2VkZKikpER1dXW6/fbb\n5fP55HK5lJubm8DpYTsihRuW
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1626
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"The individual letters aren't centered, so let's do that next."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"from scipy.ndimage import measurements,interpolation",
|
||
|
|
"",
|
||
|
|
"def centered(image,size=None):",
|
||
|
|
" center = array(measurements.center_of_mass(image))",
|
||
|
|
" tcenter = array(image.shape)/2",
|
||
|
|
" delta = tcenter-center",
|
||
|
|
" shifted = interpolation.shift(image,delta)",
|
||
|
|
" if size is None: return shifted",
|
||
|
|
" h,w = shifted.shape",
|
||
|
|
" th,tw = size",
|
||
|
|
" ch,cw = h//2-th//2,w//2-tw//2",
|
||
|
|
" return shifted[ch:ch+th,cw:cw+tw]"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1542
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"We now have a stack of letters \"E\" in sequence. We put them together into a long row of letters."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"Es = [centered(E,size=(35,30)) for E in Es]",
|
||
|
|
"signal = hstack(Es).T",
|
||
|
|
"figsize(12,8)",
|
||
|
|
"imshow(signal.T)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1637,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x7ebdd9d0>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAArsAAAA4CAYAAAD92DDdAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n9+8D0PT4EBFu3ryJpKQkREVF
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1637
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Next, we are going to transform vertical slices through this input into tokens, namely by clustering."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"from sklearn import cluster",
|
||
|
|
"km = cluster.KMeans(k=5)",
|
||
|
|
"km.fit(signal)",
|
||
|
|
"centers = km.cluster_centers_",
|
||
|
|
"figsize(4,4)",
|
||
|
|
"imshow(centers.T,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1638,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x900a3910>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAD0AAAD6CAYAAAASj0/lAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nDzaPWZ2rfEZ2AAAAAElFTkSu
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1638
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"We can now represent the signal as a sequences of numbers, each indicating",
|
||
|
|
"which cluster center represents that particular slice best."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"outputs = km.predict(signal)",
|
||
|
|
"print outputs[:35]"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"[1 1 1 1 1 1 0 0 2 2 2 3 3 3 3 3 3 3 3 0 0 0 0 4 4 4 1 1 1 1 1 1 1 1 1]"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1639
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"We have a fairly reasonable number of examples for each slice type."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"from collections import Counter",
|
||
|
|
"Counter(outputs)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1640,
|
||
|
|
"text": [
|
||
|
|
"Counter({1: 278, 3: 230, 0: 194, 4: 102, 2: 96})"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1640
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"We can also reconstruct the original signal in terms of these slices."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"imshow(centers[array([1,1,1,0,2,2,2,3,3,3,3,3,0,0,4,4,1,1,1])].T,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1642,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x90313490>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAJYAAAD6CAYAAABDN77DAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\njsaP+O7fv4+ffvoJbrcbjY2N
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1642
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Manual Construction of a Model",
|
||
|
|
"================================"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"We'd like some kind of model that represents characters fairly well. How can we construct it?",
|
||
|
|
"",
|
||
|
|
"Let's start by constructing the \"average character\" and look at the sequence of slice codes it corresponds to."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"unaryEm = mean(array(Es),0)",
|
||
|
|
"figsize(4,4); imshow(Em)",
|
||
|
|
"s = km.predict(Em.T)",
|
||
|
|
"s",
|
||
|
|
"xlim(0,len(s)+1)",
|
||
|
|
"plot(2*s,color='r',linewidth=3)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1643,
|
||
|
|
"text": [
|
||
|
|
"[<matplotlib.lines.Line2D at 0x90484c50>]"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAMwAAAD9CAYAAADwBJj6AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAIABJREFUeJztfXtwG9d97rcAAeJNkCL4kKiYqi0lkkxLTOQok1oJPTad\n+LZRkjrjOBnbGtvqZNRpJ640qdzecSxPex0lM25GcTMdt1Uz8m3Hjf+oE8001rUzsRzXaavGtZo2\nSmI3Fm1ZIikRfOH93PsH+Ts8ODgLYBcLEiDON3MGC2Cxe3Zxvv09z+9ouq7rUFBQqAmOte6AgkIr\nQRFGQcEEFGEUFExAEUZBwQQUYRQUTEARRkHBBCwT5tixYxgaGsLo6ChGR0dx5swZO/uloNCU6LD6\nQ03TcPjwYRw+fNjO/igoNDXqUslUzFOh3WBZwgDAU089hWeeeQZ79uzBk08+iXA4XPK9pml1dU5B\nYS1QSRBUlDDj4+MYGRkpa6dPn8ahQ4dw8eJFnD9/HoODgzhy5IjhyVerPfbYY+p8LXq+Zrm2aqgo\nYV566aWqBwCAgwcP4lOf+lRN+yootDIs2zCTk5Ns+/nnn8fIyIgtHVJQaGZYtmGOHj2K8+fPQ9M0\nbNmyBU8//bSd/bKEsbExdb4WPV+rXJum16K4WYSmaTXphQoKzYJqY1ZF+hUUTEARRkHBBBRhFBRM\nQBFGQcEEFGEUFExAEUZBwQQUYRQUTEARRkHBBBRhFBRMQBFGQcEEFGEUFExAEUZBwQQUYRQUTEAR\nRkHBBBRhFBRMQBFGQcEEFGEUFExAEUZBwQQUYRQUTKCuQn5nzpzBww8/jEKhgIMHD+Lo0aNl+3zx\ni1+s5xQKCrbj93//9/HRj37U2o91i8jn8/r111+vX7x4Uc9ms/quXbv0CxculOwDQDXVmq49++yz\nhuMaqEwJyyrZuXPncMMNN2B4eBgulwv33HMPvv/971s9XMPwJIA05HfuAoD3rV3XmgKbAPwM9Y/C\nPIC/X+W+rwUsq2SXL1/G5s2b2fuhoSH827/9my2dsgu3AKi0tsB2AEcAfHl1utOU+D0AdpRgdAL4\nIoBvAWiuUVAZZ8+exdmzZ2vev67lLpodj9Wwz1ijO9HkuMnm492A1iLM2NhYSVG/xx9/vOL+llWy\nTZs24dKlS+z9pUuXMDQ0ZPVwtuMWALcvb+ex9Edqy80PILf83U0ANqx675oHO7ntEazcIzPtG9wx\nhhve47WFZcLs2bMHb731FiYmJpDNZvHd734X+/fvt7NvdYGXLv8XwK+590kA57j3H1uVHjUffAC2\nLG/nAbxp8TgT3PZwHf1pBVgmTEdHB/7iL/4Cn/jEJ7Bjxw58/vOfx/bt2+3sm2WI0uX/SPY5y22P\nNbg/zQr+33oLQNbicSa47WGrnWkR1BWHufPOO3HnnXfa1RfbUEm6EM4C+N/L22MN7k+zglfHLtRx\nnAlue7iO47QC1l2kvxbpAgA/gbJjdnDbP6/jOO9w2+/DOhxUHNbdtdUiXQBlxwClEqYewiQBXF3e\ndgMYrONYzY51RZhapQvhLLc91oD+NDvsUsmA9lHL1hVhapUuhLPc9pjdnWly2OUhI0xw28N1HquZ\nUZfRXwvqfXLVCgeA9y9v1yJdgBU7xoUVOybakN41H+zykBEmuO3hOo9F8AL4KwAfsul4XwLwap3H\naDhh1sLRXIt0AVbsmN9cfv8xAM83qlNNBjvVMaAxhHkAwL02HQtYCljXi3WlkgHAFIBjJvY/y22P\n2dmRJoddHjLCBLc9bMPxAPski51YdxLmbZhTL86iPeMxdnnICBPc9rANxwNK+3g/gH+v83iXqu9S\nFQ0nzC8bfYI60a52jN0qmSwWU6zzmLwUfBHAdJ3HswPrTiUzi3aMx9jtIQPsj8VsBhBc3o6iOcgC\nKMIAaD87xm4PGWGC2x6u81h2q4x2oeEq2VpA0zTW+Pey7wDgJ4UCkFtKlLlV0+Dp7DQ8rmy7FlRa\n+70SZOep5dz8+cT9P5jLAdklmvzK6YTf46l4DKPrpu/p9VImgw8XCgCArR0d+HeHg31PjX8v6yuB\nV8dWKzRRC9YdYTRNg8PhYK9i4z+nP/+/dR25aBQuACO6jvf5fJh1lAtfnmwy4vEQB1O9hDF65bcr\nnYv/3e6FBUaYCZ8P3V1dhv2m34gPHQAoFots8BeLRUwtLgLJJADgho4O+DweFIvFskb7i/3k3ysJ\ns4rQNA1Op1PaHA5HyTbhP+Nx7MlkAAC3Op34f/5yrz1POP5VhmpP1Vqvw0gyygaweD6j4+xcWGDf\nvRcOIywQRjyG7Jpp0FMrFAq4WigwwmxxOOD1elEoFKRNRhY6LqAIs2qgP9XpdKKjo6Niczqd7Hdv\nJBLYc+0aAGBfsYifBINlxzaSWDLwhBHJY+Zaam1GBJXtu/Wtt9g5ZiIRdC8/HIz6K7tuXdfLSDBX\nKADL93AYgNfrRT6fL2k8uYvFUj8af3+USraKINJ0dHTA5XIZto6Olcv/795e9mf/r4UFbJKpNcvH\nBj8AJedPdHTgueFhvOv3l6gt9RCGyClu89KDV494wvD7ewoFbFyWpAVNQ2xwEN3LDw7xtzxheOlM\nhBHJsMBd31A+D6/Xi1wuh1wuV/Jw4ckiuy/N6iED1glh+Kc8TxS3281aZ2dn2Wcul4v9bsLnQ/6X\nv0SHrqM7n8fY7GxdfboxFsPhO+5AHqUD0ex1yewuGXkAlNkHMsL8Bndd06EQQpEIey/aJLQtU22L\nxSLy+TxyuRx7TXPOg8FcDiG/H+lcDtlsFtlslpGN+lsoFFAsFqFpGlPtAGAnd5+aSR0DWpwwsj/S\n7XbD7/fD7/fD5/Ox7WoSBgBef/dd7OXUlXqwMRbD/mQSr2/fbgthjCQLb1PImrj/jfPz7Phzg4Ml\nhUvE39IAFqWLKGEKhQLbjnm9CKZScOk6PrRpE666XMhms8hxxKH36XQa6XQamUymZHsknweWz/0L\nG/4LO1E3YYaHhxEKheB0OuFyuXDu3LnqP7IJRBBecni9XoTDYYRCIYTDYXR1daGrq4sRhLdfeBsG\nAF594AFcvHgRnem09HwOGngOx8q2YMO878IF7PzJTwAAd77+OqLj4yguD7BG2zBA+aAX97vh9dfZ\n8dPXX19SW87ISSHzOspsmEKhgHhvL4LL1YRGw2G8vXEj8vl8CWnoNR6PY3FxEbFYDIuLi1hcXISu\n67hx2TUNABeW1V6rXka7UTdhNE3D2bNn0dPTY0d/TJ2XSOr1euHxeODxeBAKhbBhwwZs2LABvb29\n6O3txYYNG+ByuSp6yRi2bkWh/FMAQJF7ovPbPC7ddhu2/ed/wpVIoPvqVfzmpUu49PGP1+VWNvKM\nWfGSbZybY587RkZKCGN0DJmkE71k5CkrDA0By4R5v8eDjm3bmMpGjUgzNzeHmZkZRKNROBwORqwd\nqRTrzy8cDiZtmgG2qGRrxX6SMB6Ph6le3d3d6OvrQ39/PwYGBjAwMID+/n64XC7DOEytqOTm5RG9\n/34M/OVfAgB2/OM/wvHFL0IXpJmZc8pexc+qBQJp3/Dly+yzwM03l0kY2bbsmmXqm67r6Ni6FfiX\nfwEAvK9QgGPrVhQKhRLCULt69So8Hg+cTify+TxSqRSSiQQ+wEmYXzoc0CxI50bBFglz++23w+l0\n4ktf+hJ+93d/145+1QSRMKFQCN3d3YhEIti4cSOGhoZYIwNfNvBqhWywSvc7ehT63/89tMVFeN55\nB1tffx3Zu+82fb6q55F8V2lgackk3MuE0Z1O9N1yCyDJaqhEONqWkUvXdXSOrBSejSST8G7dimKx\nWOIcoNdgMMgcCKlUCgsLC/A7nQgs/34WwDWHAygYyfz6sWqlYgmvvfYaBgcHce3aNYyPj+MDH/gA\n9u3bV+9hawKvlhFxfD4fAoEAs2FINeM9Yg1HOAwcPgwcOwYA8P/5n8P/0ENAxxr7WP7nf9imtnUr\nuvr67D/HjpUIimdyEp7eXuZR
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1643
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Now we don't know how many of these slices we actually need.",
|
||
|
|
"",
|
||
|
|
"Well, we do approximately, but for the sake of argument, let's construct a generic model",
|
||
|
|
"that just represents the sequences of slices."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"s0 = s[s!=roll(s,-1)]",
|
||
|
|
"print s0",
|
||
|
|
"figsize(4,4)",
|
||
|
|
"subplot(121); imshow(tile(centers[s0],(10,1)).T,interpolation='nearest')",
|
||
|
|
"from itertools import chain",
|
||
|
|
"sr = array(list(chain(*[[x]*4 for x in s0])),'i')",
|
||
|
|
"subplot(122); imshow(tile(centers[sr],(5,1)).T,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"[1 0 2 3 0 4]"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1648,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x9108fd10>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAABYCAYAAADRJpqfAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n+XxwOp0wGAzIzs4OW7UEQm96
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1648
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Durational Models",
|
||
|
|
"=================="
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"In the figure above, for the first case, we have 1 state per output. This yields a very",
|
||
|
|
"compressed image. We need a _durational model_ (thinking of the $x$ axis as time).",
|
||
|
|
"",
|
||
|
|
"The second case uses exactly three states for each output, giving rise to perfectly uniformly",
|
||
|
|
"spaced outputs.",
|
||
|
|
"",
|
||
|
|
"For our first attempt, we are going to use self transitions; these allow variable durations",
|
||
|
|
"in each state. However, the probability distribution for how long we stay in a state ",
|
||
|
|
"has an exponential distribution."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(18,8)",
|
||
|
|
"subplot(221); xticks([]); yticks([]); imshow(imread(\"Figures/states-1.png\")); xlabel(\"1 state per output\")",
|
||
|
|
"subplot(222); xticks([]); yticks([]); imshow(imread(\"Figures/states-2.png\")); xlabel(\"self-transitions\")",
|
||
|
|
"subplot(223); xticks([]); yticks([]); imshow(imread(\"Figures/states-3.png\")); xlabel(\"3 states per output\")",
|
||
|
|
"subplot(224); xticks([]); yticks([]); imshow(imread(\"Figures/states-4.png\")); xlabel(\"durational model\")"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1657,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.text.Text at 0xa4e67bd0>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAA7gAAAHZCAYAAACsDpeNAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nFVbcsGHDEBoaisjISAQFBdXo
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1657
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"However, by using combinations of states that output the same symbol,",
|
||
|
|
"we can approximate arbitrary distributions.",
|
||
|
|
"",
|
||
|
|
"Particularly simple to approximate is a normal distribution for a durational",
|
||
|
|
"model, because of the central limit theorem about sums of random variables.",
|
||
|
|
"",
|
||
|
|
"The duration in the individual state is exponentially distributed."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(6,6)",
|
||
|
|
"p0 = 0.9",
|
||
|
|
"decay = vs(array([p0**n for n in range(100)]))",
|
||
|
|
"plot(decay)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1668,
|
||
|
|
"text": [
|
||
|
|
"[<matplotlib.lines.Line2D at 0x9bd31390>]"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAX4AAAFtCAYAAAAAmAHJAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nHwCGJuqmbOj1xRfS5ZdLXV3S
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1668
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"But if we add up the durations of multiple states, we get a peaked distribution",
|
||
|
|
"that approximates a normal distribution."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(6,6)",
|
||
|
|
"result = vs(decay)",
|
||
|
|
"for i in range(10):",
|
||
|
|
" result = vs(convolve(result,decay))",
|
||
|
|
"plot(result[:400])"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1667,
|
||
|
|
"text": [
|
||
|
|
"[<matplotlib.lines.Line2D at 0x9d6f7c50>]"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAYUAAAFtCAYAAADoNHanAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nbGyEy+VCWVkZVq1a5f/ZsmXL
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1667
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Transition Matrix for Manual Model",
|
||
|
|
"=================================="
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"So for now, let's just build a model similar to the model below."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(6,6)",
|
||
|
|
"xticks([]); yticks([]); imshow(imread(\"Figures/states-2.png\")); xlabel(\"self-transitions\")"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1679,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.text.Text at 0x8a709e10>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAV0AAADsCAYAAADXaXXTAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nkZiYiKioKKxbtw7Z2dkYM2YM
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1679
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"This is not going to match the \"durations\" of our character models very well,",
|
||
|
|
"but we can still have the learning algorithms optimize it."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"ns = len(s0) # number of states",
|
||
|
|
"no = len(centers) # number of centers"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1672
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Here, we implement the state transition matrix.",
|
||
|
|
"Note that we set a \"probability floor\", since we are going to use",
|
||
|
|
"the matrix for estimation later, and zeros (=something that never happens)",
|
||
|
|
"don't work so well.",
|
||
|
|
"",
|
||
|
|
"This transition matrix can be thought of as a linear function that",
|
||
|
|
"maps old states into new states.",
|
||
|
|
"",
|
||
|
|
"Note that a lot of literature on Markov models uses the opposite",
|
||
|
|
"convention for subscripts. Our transition matrices are multiplied",
|
||
|
|
"from the left with a column vector of old states and yield a new state."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"A = zeros((ns,ns))",
|
||
|
|
"for i in range(ns):",
|
||
|
|
" A[i,i] = 0.6",
|
||
|
|
" A[(i+1)%ns,i] = 0.4",
|
||
|
|
"A = ls(maximum(1e-2,A))",
|
||
|
|
"figsize(4,4)",
|
||
|
|
"imshow(A,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1699,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x5dcc3550>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAPYAAAD5CAYAAAAURMgdAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nu3cVj8cX/p527NhR1o/LKuok
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1699
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"We still need to write down how the individual states correspond to output symbols.",
|
||
|
|
"In fact, in our case, that's quite simple, since each state corresponds to a single output",
|
||
|
|
"symbol (but not the other way around).",
|
||
|
|
"",
|
||
|
|
"The `B` matrix is what makes Hidden Markov Models different from Markov chains.",
|
||
|
|
"In Markov chains, we can observe what state the chain is in.",
|
||
|
|
"In a Hidden Markov Model, we have imperfect information about the state of",
|
||
|
|
"the system.",
|
||
|
|
"`B` is a kind of \"confusion matrix\" or \"error matrix\" for observing the state of the system.",
|
||
|
|
"It's really the simplest such system; there may be much more general kinds of observations",
|
||
|
|
"and errors we can make.",
|
||
|
|
"",
|
||
|
|
"In HMMs for speech recognition, the \"observations\" are actually usually the \"slices\"",
|
||
|
|
"(short time spectrum) at a given time, and instead of approximating that as an output",
|
||
|
|
"symbol via clustering (see above), the relationship between the observation vector $b$",
|
||
|
|
"and the state is modeled directly."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"B = zeros((no,ns))",
|
||
|
|
"for i in range(ns): B[s0[i],i] += 1.0",
|
||
|
|
"B = ls(maximum(1e-2,B))",
|
||
|
|
"figsize(4,4)",
|
||
|
|
"imshow(B,interpolation='nearest')",
|
||
|
|
"print s0"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"[1 0 2 3 0 4]"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAPYAAADUCAYAAACvdt71AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n"
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1700
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Sampling from the Distribution",
|
||
|
|
"=============================="
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"First, let's define a simple function to sample from discrete distributions in general."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def rsample(dist):",
|
||
|
|
" v = add.accumulate(dist)",
|
||
|
|
" assert abs(v[-1]-1)<1e-3",
|
||
|
|
" val = rand()",
|
||
|
|
" return searchsorted(v,val)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1688
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Sampling from Hidden Markov Models happens in two stages.",
|
||
|
|
"",
|
||
|
|
"First, we sample the state sequence.",
|
||
|
|
"",
|
||
|
|
"Then, given the state sequence, we sample the observations."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"def chain_sample(A,n,state=0):",
|
||
|
|
" states = zeros(n,'i')",
|
||
|
|
" for i in range(n):",
|
||
|
|
" p = A[:,state]",
|
||
|
|
" state = rsample(A[:,state])",
|
||
|
|
" states[i] = state",
|
||
|
|
" return states",
|
||
|
|
"figsize(20,2)",
|
||
|
|
"plot(chain_sample(A,200))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1689,
|
||
|
|
"text": [
|
||
|
|
"[<matplotlib.lines.Line2D at 0x805de290>]"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAABHwAAACOCAYAAABDnR8uAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\njz6q75pxrv+hQ8DJJ+u7rmze
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1689
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Here is code for sampling both the state sequence and the sequence of observations."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def hmm_sample2(A,B,n,state=0):",
|
||
|
|
" states = chain_sample(A,n,state=state)",
|
||
|
|
" outputs = array([rsample(B[:,s]) for s in states])",
|
||
|
|
" return array(outputs,'i'),states",
|
||
|
|
"def hmm_sample(A,B,n,state=0):",
|
||
|
|
" return hmm_sample2(A,B,n,state=state)[0]"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1690
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"hmm_sample(A,B,20,state=s[0])"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1691,
|
||
|
|
"text": [
|
||
|
|
"array([0, 2, 2, 2, 2, 2, 2, 3, 3, 3, 0, 0, 2, 2, 3, 3, 3, 3, 0, 4], dtype=int32)"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1691
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"We can now generate new, random images from this."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"imshow(centers[hmm_sample(A,B,200,state=s[0])].T)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1692,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x757b9ad0>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAApsAAACMCAYAAAAz3mJuAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n6HQ6GI1GWF5exq1bt/D4449r
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1692
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"It's a little hard to tell whether this sometimes looks wrong because of the observations",
|
||
|
|
"or because of the state sequence, so here is another sampling function that outputs the",
|
||
|
|
"\"best\" observation for each state."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def best_sample(A,B,n,state=0):",
|
||
|
|
" states = chain_sample(A,n,state=state)",
|
||
|
|
" outputs = array([argmax(B[:,s]) for s in states])",
|
||
|
|
" return array(outputs,'i')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1695
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"imshow(centers[best_sample(A,B,200,state=s[0])].T)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1694,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x673e8350>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAApsAAACMCAYAAAAz3mJuAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nIrAUbN53331429vehtdffx0z
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1694
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Since that looks kind of noisy, you might be wondering whether the model is doing anything at all.",
|
||
|
|
"Here is the output from a completely random HMM, but using the same output symbols."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"imshow(centers[hmm_sample(ls(rand(5,5)),ls(rand(no,5)),200,state=s[0])].T)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1697,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x61686050>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAApsAAACMCAYAAAAz3mJuAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nnQ7rtXSklATIZrPY3NxEo9Hg
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1697
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Forward Algorithm",
|
||
|
|
"========================"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"To get started, let's compute what's known as the _forward algorithm_.",
|
||
|
|
"",
|
||
|
|
"This algorithm computes $P(s_t | o_t ... o_1)$, that is, it updates our current",
|
||
|
|
"estimate of what state the system is in based on the current and all previous observations.",
|
||
|
|
"",
|
||
|
|
"To do this, it starts of with an initial state distribution (e.g., uniform), and",
|
||
|
|
"then computes the posterior of the distribution given the observation:",
|
||
|
|
"",
|
||
|
|
"$$P(s_1|o_1) = \\frac{P(o_1|s_1) P(s_1)}{P(o_1)}$$",
|
||
|
|
"",
|
||
|
|
"Here:",
|
||
|
|
"",
|
||
|
|
"- $P(s_1)$ is our prior",
|
||
|
|
"- $P(o_1)$ is the first observation",
|
||
|
|
"- $P(o_1|s_1)$ is given by the observation matrix $B$",
|
||
|
|
"",
|
||
|
|
"As usual, we don't bother computing $P(o_1)$ explicitly and",
|
||
|
|
"instead just normalize.",
|
||
|
|
"",
|
||
|
|
"Now, to do the same thing for $P(s_2|o_1,o_2)$, we reduce this to the above",
|
||
|
|
"problem, by observing that all the information about $o_1$ is already contained",
|
||
|
|
"in $P(s_1|o_1)$. ",
|
||
|
|
"",
|
||
|
|
"However, $P(s_1|o_1)$ is the state distribution prior",
|
||
|
|
"to the state transition, so in order to get a \"$P(s_2)$ prior\", we",
|
||
|
|
"need to multiply $P(s_2|s_1)P(s_1|o_1)$.",
|
||
|
|
"",
|
||
|
|
"This then gives us the complete forward algorithm:",
|
||
|
|
"",
|
||
|
|
"- obtain a new prior by computing $P(s_t|s_{t-1})P(s_{t-1}|o_{t-1})$ using the $A$ matrix",
|
||
|
|
"- update the new state using Bayes rule and the observation $o_t$",
|
||
|
|
"- repeat"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def hmm_forward(A,B,observations,p=None):",
|
||
|
|
" if p is None: p = dot(A,ones(len(A))/len(A))",
|
||
|
|
" fps = []",
|
||
|
|
" for i,o in enumerate(observations):",
|
||
|
|
" # update P(state|ovservation) using Bayes formula",
|
||
|
|
" p = B[o,:]*p",
|
||
|
|
" p /= sum(p)",
|
||
|
|
" fps.append(p)",
|
||
|
|
" # now compute the probabilities in the next state",
|
||
|
|
" p = dot(A,p)",
|
||
|
|
" return array(fps)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1701
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(6,6)",
|
||
|
|
"fps = hmm_forward(A,B,outputs)",
|
||
|
|
"figsize(5,5)",
|
||
|
|
"imshow(r_[signal[:30].T,fps[:30].T/amax(fps)],interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1703,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x4c22ed50>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAOoAAAExCAYAAAByGj4WAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nzkzP/K9sL5UBrF5g3z0DIH84
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1703
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Note that there are several things that this is _not_:",
|
||
|
|
"",
|
||
|
|
"- the sequence of most probable states at time $t$ does _not_ necessarily belong to any sequence of states that can acctually even occur (\"path\")",
|
||
|
|
"- the most probable state at time $t$ is _not_ necessarily the \"best\" state given the observations",
|
||
|
|
"",
|
||
|
|
"C.f.",
|
||
|
|
"",
|
||
|
|
"- \"The old man the boat.\"",
|
||
|
|
"- \"The horse raced past the barn fell.\"",
|
||
|
|
"",
|
||
|
|
"**Homework** Construct examples for these two casees."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Viterbi Decoding",
|
||
|
|
"==================="
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Let's now look at the problem of actually finding the best path.",
|
||
|
|
"",
|
||
|
|
"The algorithm for this is quite similar to what we have already seen for string edit distance and dynamic time warping."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"print ns,no",
|
||
|
|
"print len(outputs),amax(outputs)+1"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"6 5",
|
||
|
|
"900 5"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1564
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"What we want to find is the sequence of states $s$ such that the likelihood of that",
|
||
|
|
"state sequence given the observation is maximized:",
|
||
|
|
"",
|
||
|
|
"$$\\hat{s} = \\arg\\max_s \\prod_t P(o_t|s_t) P(s_t|s_{t-1})$$",
|
||
|
|
"",
|
||
|
|
"The key insight is that once we know the best path to each of the states at time $t-1$,",
|
||
|
|
"the subsequent search doesn't depend on it.",
|
||
|
|
"",
|
||
|
|
"So, we maintain the accumulated probabilities in an array $p_t(s)$ (called `probs` in the code).",
|
||
|
|
"",
|
||
|
|
"In addition, we maintain information about which state at times $t-1$ actually",
|
||
|
|
"gave rise to the best way of coming to each of the states at time $t$ (called `pred` in the code).",
|
||
|
|
"By tracing backwards, we can reconstruct the entire best path."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"probs = zeros((len(outputs),ns))",
|
||
|
|
"pred = zeros((len(outputs),ns))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1565
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"probs[0] = vs(ones(ns))*B[outputs[0]]",
|
||
|
|
"probs[0]"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1566,
|
||
|
|
"text": [
|
||
|
|
"array([ 0.1646382 , 0.00163059, 0.0013433 , 0.00129166, 0.00028223,",
|
||
|
|
" 0.00154323])"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1566
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"for t in range(1,len(outputs)):",
|
||
|
|
" for j in range(ns):",
|
||
|
|
" for k in range(ns):",
|
||
|
|
" c = probs[t-1,k]*A[j,k]*B[outputs[t],j]",
|
||
|
|
" if c>probs[t,k]:",
|
||
|
|
" probs[t,k] = c",
|
||
|
|
" pred[t,k] = j"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1567
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Let's look at the probabilities."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"xlim([0,60]); figsize(12,12)",
|
||
|
|
"imshow(probs[:60].T,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1706,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x7421b510>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAdsAAABHCAYAAABPqe24AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAACdZJREFUeJzt3V9Mm9Ufx/FPByRT0WXMUchKAtlfyljbWSUxWYKrbIkT\nHGEXW3QmuE2jWSK70Wi8mDcbbkYzzYyJiWZyMXfnn2VrxAATXdzigKBCwlSadEjJqlZtZWGU87tY\nftUOKLSzwNr362pPn56H7z4JfNPT55zHYowxAgAAabNovgsAACDT0WwBAEgzmi0AAGlGswUAIM1o\ntgAApFnurQy2WCz/VR0AANw2kl3IM2Oz9Xq9ampqUjQa1d69e/Xiiy8m/IHvvPPOtNc6d+5cwp/1\n448/Tnvu6tWrCceGw+Fpz0UikYRjx8fHpz03MTGRcCwAIDsEg0EtW7YspQ+aCaeRo9Go9u/fL6/X\nq76+Pp08eVL9/f0pFwoAQDZK2GwvXryoVatWqbS0VHl5edq5c6c++eSTuaoNAICMkHAaeWhoSCUl\nJbFjm82mCxcuxL3n4MGDsX9XV1f/p8UBALBQHDlyRHfccUdKYxM229nMS/+72UpSX19fSoUAALCQ\nvfDCC1q2bJleffXVpMcmnEZesWKF/H5/7Njv98tmsyVfIQAAWSxhs3W73bp8+bJ8Pp/GxsZ06tQp\n1dXVzVVtAABkhITTyLm5uVq5cqVWrVqlnJwcHTx4UOXl5XHvycnJiTtmqQwAAPEsMz1ir7OzU/n5\n+XryySf13XffxQ+2WLRoUfyHY5otACAT/XudbbKbWsy4XeOmTZu0dOnSlIsDACDb3dJ2jVLyW1YB\nAHA7upWlPzNOI0uSz+dTbW0t08gAgKyV1mlkAABwa2i2AACk2YzNdteuXXrwwQc1MDCgkpISffDB\nB3NRFwAAGSPhDVJ+v1+BQEBLly5VQUGBnn76aTU2Ns5VbQAAZISEN0gFAgEFAgE5nU6Fw2Hdd999\n+vjjj2MbW3CDFAAgW6TtBqmioiI5nU5JUn5+vsrLy/XLL7+kXikAAFlo1utsfT6furu7VVVVFfc6\n62wBANkg7etsw+Gwqqur9corr2j79u3/DGYaGQCQJdK6zvb69etqaGjQE088EddoAQDA7CRstsYY\n7dmzR3a7XU1NTXNVEwAAGSVhs21ra1NLS4veffddLV68WFarVV6vd65qAwAgIyRsth6PR5FIRKOj\nowqHwyorK1N+fv5c1QYAQEaY8TvbO++8U5I0NjamaDSqgoKCtBcFAEAmmXHpz8TEhDZu3KiffvpJ\nzz77rOx2e9x5lv4AALJB2pf+SNIff/yhrVu3qrm5WdXV1TcGs/QHAJAl5uQRe0uWLNG2bdv07bff\nJl0gAADZLGGzDQaDCoVCkqTR0VG1trbK5XLNSWEAAGSKhM12eHhYmzdvlsPhUEFBgYLBoDwez1zV\nBgBARkh4g1RlZaW6urr0xhtv6NKlS/rrr78mvYfvaAEASGzG72yvXLmiM2fOaO/evdx5DABACmZs\ntgcOHNDRo0cn3XUMAABmJ+E08unTp1VYWCiXy6WOjo45KgkAgIUnbetsX375ZbW0tCg3N1fXrl3T\nn3/+qYaGBn344Yc3BlssqVUMAMBt5lbW2c56U4tz587p9ddf12efffbPYJotACBLzMmmFhLNFQCA\nVMy4N3Jpaanuuece5eTkKC8vb9J5tmsEACCxGZutxWJRR0cHT/sBACBFs5pGZn0tAACpm9Un24cf\nflg5OTl65plntG/fvrjzNGIAQDZI6yP2hoeHVVxcrKtXr6qmpkZvv/22Nm3adGMwj9gDAGSJtN6N\nXFxcLElavny56uvrdfHixdSqBAAgSyVstn///Xfs4QORSESff/65Kisr56QwAAAyRcLvbEdGRlRf\nXy9JGh8f1+OPP64tW7bEvYdpYwBANrj33ntTHjvrHaQAAEBqeJQPAABpRrMFACDNaLYAAKRZys3W\n6/Vq3bp1Wr16tV577bX/sqbb3lNPPSWr1Rp35/Zvv/2mmpoarVmzRlu2bFEoFJrHChcGv9+vhx56\nSBUVFVq/fr3eeustSWQ1lWvXrqmqqkpOp1N2u10vvfSSJLKaTjQalcvlUm1trSRymkppaak2bNgg\nl8ulBx54QBI5TScUCmnHjh0qLy+X3W7XhQsXks4qpWYbjUa1f/9+eb1e9fX16eTJk+rv70/pP5GJ\nGhsb5fV6415rbm5WTU2NBgYG5PF41NzcPE/VLRx5eXl688039cMPP+ibb77R8ePH1d/fT1ZTWLx4\nsdrb29XT06Pe3l61t7frq6++IqtpHDt2THa7PfakMnKa7P/73nd3d8f2TyCnqT3//PN65JFH1N/f\nr97eXq1bty75rEwKzp8/b7Zu3Ro7Pnz4sDl8+HAql8pYg4ODZv369bHjtWvXmkAgYIwxZnh42Kxd\nu3a+SluwHnvsMdPa2kpWM4hEIsbtdpvvv/+erKbg9/uNx+MxbW1t5tFHHzXG8Ps3ldLSUhMMBuNe\nI6fJQqGQKSsrm/R6slml9Ml2aGhIJSUlsWObzaahoaFULpU1RkZGZLVaJUlWq1UjIyPzXNHC4vP5\n1N3draqqKrKaxsTEhJxOp6xWa2z6nawmO3DggI4ePRq3lSw5Tfb/fe/dbrfee+89SeQ0lcHBQS1f\nvlyNjY3auHGj9u3bp0gkknRWKTVbHiJ/aywWCxn+SzgcVkNDg44dO6a777477hxZ/WPRokXq6enR\nlStX9OWXX6q9vT3uPFlJp0+fVmFhoVwu17R715LTDV9//bW6u7t19uxZHT9+XJ2dnXHnyemG8fFx\ndXV16bnnnlNXV5fuuuuuSVPGs8kqpWa7YsUK+f3+2LHf75fNZkvlUlnDarUqEAhIuvFwh8LCwnmu\naGG4fv26GhoatHv3bm3fvl0SWc1kyZIl2rZtmy5dukRWNzl//rw+/fRTlZWVadeuXWpra9Pu3bvJ\naQpT7XtPTpPZbDbZbDbdf//9kqQdO3aoq6tLRUVFSWWVUrN1u926fPmyfD6fxsbGdOrUKdXV1aVy\nqaxRV1enEydOSJJOnDgRayzZzBijPXv2yG63q6mpKfY6WU0WDAZjdzuOjo6qtbVVLpeLrG5y6NAh\n+f1+DQ4O6qOPPtLmzZvV0tJCTjeZbt97cpqsqKhIJSUlGhgYkCR98cUXqqioUG1tbXJZpfql8Zkz\nZ8yaNWvMypUrzaFDh1K9TEbauXOnKS4uNnl5ecZms5n333/f/Prrr8bj8ZjVq1ebmpoa8/vvv893\nmfOus7PTWCwW43A4jNPpNE6n05w9e5asptDb22tcLpdxOBymsrLSHDlyxBhjyCqBjo4OU1tba4wh\np5v9/PPPxuFwGIfDYSoqKmJ/w8lpaj09PcbtdpsNGzaY+vp6EwqFks6KvZEBAEgzdpACACDNaLYA\nAKQZzRYAgDSj2QIAkGY0WwAA0oxmCwBAmv0PPnKvletW5wgAAAAASUVORK5CYII=\n"
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1706
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"That's not very good; the probabilities become very small quickly",
|
||
|
|
"because we are multiplying together.",
|
||
|
|
"At the end, there is no information left."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"probs[-1]"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1707,
|
||
|
|
"text": [
|
||
|
|
"array([ 0., 0., 0., 0., 0., 0.])"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1707
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"We need to rewrite the algorithm to use logarithms. With that, we obtain:"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"costs = 999999*ones((len(outputs),ns))",
|
||
|
|
"pred = zeros((len(outputs),ns))",
|
||
|
|
"costs[0] = -log(vs(ones(ns))*B[outputs[0]])",
|
||
|
|
"print costs[0]"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"[ 1.80400496 6.41881359 6.61262349 6.65182564 8.1727776 6.47388061]"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1570
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"for t in range(1,len(costs)):",
|
||
|
|
" for j in range(ns):",
|
||
|
|
" for k in range(ns):",
|
||
|
|
" c = costs[t-1,j]-log(A[k,j]*B[outputs[t],j])",
|
||
|
|
" if c<costs[t,k]:",
|
||
|
|
" costs[t,k] = c",
|
||
|
|
" pred[t,k] = j"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1571
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Now we can actually look at the costs in a sensible way."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"subplot(211); plot(costs)",
|
||
|
|
"subplot(212); imshow(costs[:60].T,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1572,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x8aed7210>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAXwAAAExCAYAAAB/IRl4AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n4LL8GM6nfQYjVQb6uOdjT8ZS
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1572
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"We can also trace the predecessors back."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"t = len(costs)-1",
|
||
|
|
"state = argmin(costs[t])",
|
||
|
|
"states = []",
|
||
|
|
"while t>0:",
|
||
|
|
" state = pred[t,state]",
|
||
|
|
" states.append(state)",
|
||
|
|
" t -= 1",
|
||
|
|
"states.append(0)",
|
||
|
|
"states = array(states)[::-1]"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1573
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(18,8)",
|
||
|
|
"xlim(0,60)",
|
||
|
|
"imshow(costs[:60].T,interpolation='nearest')",
|
||
|
|
"plot(states[:60],'ro')",
|
||
|
|
"plot(argmin(costs[:60],1),'b*')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1711,
|
||
|
|
"text": [
|
||
|
|
"[<matplotlib.lines.Line2D at 0xd3955d0>]"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAABBEAAACTCAYAAADYzIGrAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nZOj1+n6vQy2PaX19vewMNfQL
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1711
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Note that the state estimates above almost agree, but not quite: the minimum cost state is not always on the best path."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Plotting the state transitions and preferred output labels on top of the data is also useful."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"bout = argmax(B[:,states[:60]],0)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1729
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(10,8); xlim((0,60)); ylim((35,-1))",
|
||
|
|
"imshow(signal[:60].T,interpolation='nearest')",
|
||
|
|
"plot(2*states[:60],'ro')",
|
||
|
|
"plot(bout+0.5,'yo')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1735,
|
||
|
|
"text": [
|
||
|
|
"[<matplotlib.lines.Line2D at 0xa7442f90>]"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAlEAAAFpCAYAAAC8vX2DAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nFlhMAhbyQ7qIan2J6nFL1Jcg
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1735
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Here is everything wrapped up into a single function."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def hmm_viterbi(A,B,outputs):",
|
||
|
|
" ns,ns1 = A.shape",
|
||
|
|
" no,ns2 = B.shape",
|
||
|
|
" assert ns==ns1",
|
||
|
|
" assert ns==ns2",
|
||
|
|
" costs = 999999*ones((len(outputs),ns))",
|
||
|
|
" pred = zeros((len(outputs),ns))",
|
||
|
|
" costs[0] = -log(vs(ones(ns))*B[outputs[0]])",
|
||
|
|
" # propagate the costs forward with dynamic programming",
|
||
|
|
" for t in range(1,len(costs)):",
|
||
|
|
" for j in range(ns):",
|
||
|
|
" ck = costs[t-1,j]-log(A[:,j]*B[outputs[t],:])",
|
||
|
|
" pred[t] = where(ck<costs[t],j,pred[t])",
|
||
|
|
" costs[t] = minimum(ck,costs[t])",
|
||
|
|
" # trace the states backwards",
|
||
|
|
" t = len(costs)-1",
|
||
|
|
" state = argmin(costs[t])",
|
||
|
|
" states = []",
|
||
|
|
" while t>0:",
|
||
|
|
" state = pred[t,state]",
|
||
|
|
" states.append(state)",
|
||
|
|
" t -= 1",
|
||
|
|
" states = array(states,'i')[::-1]",
|
||
|
|
" return states"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1576
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Let's make sure it still works."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"imshow(signal[:60].T,interpolation='nearest')",
|
||
|
|
"states = hmm_viterbi(A,B,outputs)",
|
||
|
|
"figsize(6,6)",
|
||
|
|
"plot(2*states[:60],'ro')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1736,
|
||
|
|
"text": [
|
||
|
|
"[<matplotlib.lines.Line2D at 0xcadd750>]"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAlIAAAFdCAYAAADIYYM7AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nTzVr1gybPTlV/ZWOM+rszPCP
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1736
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Viterbi Training",
|
||
|
|
"================"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"The Viterbi algorithm gives us a simple way of \"learning\" the matrices `A` and `B`.",
|
||
|
|
"This is called _Viterbi training_. While commonly used in some applications, it ",
|
||
|
|
"differs from the Baum-Welch procedure originally used for training Hidden Markov Models.",
|
||
|
|
"",
|
||
|
|
"(You can think of Viterbi training as being analogous to $k$-means and Baum Welch ",
|
||
|
|
"being analogous to Gaussian mixture model; the latter uses \"soft assignment\".)"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"The idea behind Viterbi training is that, once we have aligned our observations with the data, we have an estimate",
|
||
|
|
"of the hidden, unobservable variable, namely the sequence of states.",
|
||
|
|
"Once we have that, we can directly update the `A` and `B` matrices simply by counting."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"states = hmm_viterbi(A,B,outputs)",
|
||
|
|
"A1 = zeros(A.shape)",
|
||
|
|
"for t in range(1,len(states)):",
|
||
|
|
" A1[states[t],states[t-1]] += 1",
|
||
|
|
"A1 = ls(A1)",
|
||
|
|
"B1 = zeros(B.shape)",
|
||
|
|
"for t in range(0,len(states)):",
|
||
|
|
" B1[outputs[t],states[t]] += 1",
|
||
|
|
"B1 = ls(B1)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1738
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"A1[:,1]"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1739,
|
||
|
|
"text": [
|
||
|
|
"array([ 0. , 0.61538462, 0.38461538, 0. , 0. , 0. ])"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1739
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Here is a comparison of the transition matrix before and after.",
|
||
|
|
"The structure hasn't changed much, but the probabilities have been adjusted."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(10,8)",
|
||
|
|
"subplot(121); imshow(A,interpolation='nearest')",
|
||
|
|
"subplot(122); imshow(A1,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1741,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0xca12d10>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAkUAAAEYCAYAAABShoJ1AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nVGBkZKTqCaVs3bq16gnTqtVq
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1741
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"As a consequence, the durational model has improved."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(12,6)",
|
||
|
|
"subplot(211); imshow(centers[hmm_sample(A,B,200,state=s[0])].T)",
|
||
|
|
"subplot(212); imshow(centers[hmm_sample(A1,B1,200,state=s[0])].T)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1743,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0xa76d1e10>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAArsAAAFICAYAAABQjTaoAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAIABJREFUeJzsfVuMZUd19trnfj+n7z3TPfYMvjAzYAfHBvIQC6I/TqJI\nIBARBCUBcUkiHhIhIwVeAJuHYB4QIuQFKVZkhGSFlyhIKFaEFJzIUcQLRCRB2MLMTHefS5/T536/\n/g/WKq9dp6r22vvs09N29ie1eqb33lWrVq1brbpZi8ViAQECBAgQIECAAAECvAkRutsEBAgQIECA\nAAECBAiwLgTBboAAAQIECBAgQIA3LYJgN0CAAAECBAgQIMCbFkGwGyBAgAABAgQIEOBNiyDYDRAg\nQIAAAQIECPCmRRDsBggQIECAAAECBHjTwnOw+8ILL8D169fhgQcegK997Wt+0hQgQIAAAQIECBAg\ngC+wvJyzO5vN4K1vfSv88Ic/hIODA3jnO98Jzz//PNy4ceO1Qi3Ld0IDBAgQIECAAAECBFDBFM56\nyuz++Mc/hvvvvx+uXr0K0WgU/vAP/xD+6Z/+yTOBAQIECBAgQIAAAQKsA56C3ZOTE7hy5Yr4/+Hh\nIZycnPhGVIAAAQIECBAgQIAAfsBTsBssUwgQIECAAAECBAjwRoCnYPfg4ACOjo7E/4+OjuDw8NA3\nogIECBAgQIAAAQIE8AOegt3HHnsMXnnlFbh16xaMx2P4h3/4B3j/+9/vN20BAgQIECBAgAABAqyE\niKePIhH427/9W/jd3/1dmM1m8KlPfUqcxBAgQIAAAQIECBAgwEWBp6PHHAsN1vQGCBAgQIAAAQIE\nOCf4fvRYgAABAgQIECBAgABvBATBboAAAQIECBAgQIA3LYJgN0CAAAECBAgQIMCbFkGwGyBAgAAB\nAgQIEOBNC0+nMXCQz+fFvy3LWtumtcViIX68YD6fw2w2Ez8U4XBY/AAATKdT8SO/F41GIRKJQCj0\n+vhB1WYuLzjtMtHuBrq6sE2RSERLs6kt8/kcptMpTCYTmE6nEIlElHzSlYtl6+ijbZ/P5yyehUIh\n0aZIRC3+q8qqF3k30a0rj7YXeY280L3HRTgcFjwKhUI2XpvopH1M66bfO8mqjl5KE+ok1quixUl+\n5DopfaFQSNQnP+PSztEfmWaZdl3ZTrqP+iCXT/tHrsMkt7Ruy7KE/qp469QOU9myTnNhkhmuzaFA\n2p3o9yJb8ru6MiKRCMRiMYhGo65ol/sKf+bzubDFSIcT7Sqdxr+r3uXID9Yv04B1mWyzX+D2L6Vd\n1jm0EeFweKl/5DY6le3FNkejUZsdVGEVX+TX2QXIh+l06jsv0J9Ho1Ho9XrGd9cmUTs7O4IY/PE7\n4KUC6MYwUkwmExgOhzAajWA8HtuexWIxSCQSkEgkYD6fQ7/fh36/vxTsRqNRSKVSkE6nIRaLAYDa\nGHB5MZ/PRbtMHY+0D4dDmEwmbpotgLyTHaZlWRCLxSCdTkMqlVIqlVN/TqdT6Pf70Ov1YDqdQjwe\nF3wyGTPKJ6RRxYvRaGRrP4dnoVAIkskkpFIpSKVSyndWkVMv8o5yjD/c8mjfjcdjwQsqn7KOcA1J\nOBwWPIpGozZem9qO36TTaeHk0cni97KeUZhkPxKJCHmMx+MAoHe66IiwTCcbMZ/PBX3D4dCm+4vF\nQrSfSzsAOOqPim7LsoTzlNtGna1J92ezGViWtcRDy7IgkUgImuSBuax3cr3IP+QtDoJk+4Ht0JVH\nIcsntmk0GrEH8CaZiUajor1omznQBTEm2nWYzWZa/THRHo/HIZvNQiaTEfLuBKpz8/nc1lfj8Rh6\nvR70+33BX5XtpwiFQjaZoQMZGViPzDNK03Q6Ff2LtgppCIfDjrbZL1A9c+pjHe3RaBQSiQTE43GI\nRqPim/l8butvnc31apupHUwkEsY2cnRQBkcuuKC2czgcOuoJx4dTRCIRIS93PdhFo+F2ZM0BCqAq\n28rFaDSCbrcLnU5nicHJZFIYm/l8DpZlwWQygcFgYHsPDerGxgYkk0ntaIpmik3OD0dBODo00d7p\ndFYKzizLEgZHdoxobAuFwlJwyhkxTiYTaDQaMJlMoNfrQSwWg2w2CxsbG0bHg/ISDodFJkDFi16v\nB91u11af06AHDWo+n7fNPrhplwlIt5OjpKBZWd3sgqw/lC84qOh0OgDwmlzI76FhddOOVCoFhUIB\nkskkdDodCIVCRqMcDochk8nAxsYGFAoFG43D4RC63S5YluU4W6GjFweVGxsbwvE66RnKt5ONmM1m\n0O12RRup7gOA0DM3tJv0h4IGECYbIfc38lNVnk6nE4kE5HI52NjYsMkTzWSb6gUAW3adPqOZRJr1\nwn4w8Qx/ut2u0GnuAB7tpUq+0TajHHPAmf1R0W6iT6c/JtrRXu7s7EA6nWbRLtNE29Hv920zFQDm\no5oAXg92c7kcFAoFITOqWQEqP/Q59WeTyUT4W2wzth8TETrb7CdoRpDriyeTCXQ6Heh2uzCfzyGR\nSEAmk4FsNmvzZ/P5HLrdLoTDYWUCA7Gqbd7Y2BD2ydRGp3hDRRPV6VXR6XRsvlwHnd0yAf15oVCA\nYrFofNdzsHv16lXI5XJimujHP/7x0nMkhk6H+gk6Te41s9nv96HRaEAsFhPBAiKfz8PGxgZsbGzA\nZDIBy7JgOBxCu922vZdIJGBjYwMuX74slFTliKkxcFIwbJMp2EXaG43GEu1cjEYjGAwGSwpnWRak\n02nY2tqCy5cv25SZO005GAwgFArBcDiEZrMJyWRSlGcauSOP0DDTpRAU7XYb6vU6RKNR6Ha7MBgM\nYDgcGnkWiUQgm83C3t4e7O3t2Z65mdoylY/0uw12VQG9Tn8oXyaTCbTbbSHH/X7fVgbyxc1oPRaL\nQT6fh/39fcjn80LO0GipEI1GYWdnB3Z3d2F3dxcWi4Xot16vB41GA+LxOLRaLW29GBgPBoOl/o7H\n41AoFODSpUuwsbEBAHo9Q76hAXWyEdPpFBqNBiQSCSEjqPuLxUK0n2ZwTLRPJhOt/sigcmeyEchL\nDBh0uk+zKFSnw+EwpNNp2NnZgcuXL9sCORrg6YLdyWQilmfgtDbyVRfscgNGLKfZbAo+00GbCdPp\nVKv7aJsvXboEuVyOVV4oFBLtM/U31/+gbDUaDYhEIrY+wQz9YDBYoh3t5eHhIRQKBRbtdKkCXToW\njUah1WpBIpEQtgl5ZgpAwuGwCLgvX74sZhxUwS6VH1WwO5lMYDweQ71eF3rW7/dF+6PRKORyOaVt\n9ht0eYtJPinto9HI1o+Y4MIkF2I+n9ve0/F3sViItru1zWgHt7a2tO+ZdFoHarNR31cFtZ2mmTHK\nC27wH41GIZ/Pw97eHvzv//6v8V3Pwa5lWfCjH/0INjc3lc/f9a53AcDrUxum0b1X4AjVKQNqQqfT\ngdPTUzg9PYVGo2F7trW1JRx3v98X60JqtZrtvVwuB/fccw/cuHED9vf3tQETnVIyBUKm9Zdc2rlo\nNptQq9Xg7OzMxsNwOAxbW1tw3333wc2bN5eCU05Q2G63IR6PQ7/fh1KpBBsbG3Dt2jW4efOmceRO\nZYb2scyLs7Mz0f5qtQpnZ2dQq9WMCpVMJmF/fx+uX78ODz744NLzVYNdL/JOp8pUwa6uPJoNrtfr\nghc0+FksFoIvbkbqmUwGDg4O4MaNG3Dp0iVR9unpqVYmo9EoXL58GS5dugSXLl0SU+mY2cLvz87O\ntPUOh0Mhj3I/ZjIZuHLlCty8eRMuX77M0jOkwclGTCYTWxsLhQLs7e3Bzs4OLBYL8fd6va4tYzAY\nCF53Oh3Y3NzU6g8FHTyabARtR6vVEjQ1m03be/V6Xegz7atIJAI7Oztw//33w82bN20BONanGqTJ\n601pMC6vTZQDd7qcRAXZhlerVdEueQZNh36/r5UZapu5AZQ86ODSrsN4PNbqT6/XEzIjB8zZbBYO\nDw/hbW97G5t2eU0sbUej0YBS
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1743
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Let's wrap this up as well."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def hmm_viterbi_training(A,B,outputs):",
|
||
|
|
" states = hmm_viterbi(A,B,outputs)",
|
||
|
|
" A1 = ones(A.shape)",
|
||
|
|
" for t in range(1,len(states)):",
|
||
|
|
" A1[states[t],states[t-1]] += 1",
|
||
|
|
" A1 = ls(A1)",
|
||
|
|
" B1 = ones(B.shape)",
|
||
|
|
" for t in range(0,len(states)):",
|
||
|
|
" B1[outputs[t],states[t]] += 1",
|
||
|
|
" B1 = ls(B1)",
|
||
|
|
" return A1,B1"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1582
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Forward-Backward Algorithm",
|
||
|
|
"============================="
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Above, we already saw how we can compute the _forward probabilities_",
|
||
|
|
"",
|
||
|
|
"$$P(s_t|o_{t-1}...o_1)$$",
|
||
|
|
"",
|
||
|
|
"What we might want to ask instead, however, is the probability that",
|
||
|
|
"the system is in state $s_t$ given all observations:",
|
||
|
|
"",
|
||
|
|
"$$P(s_t|o_N...o_1)$$",
|
||
|
|
"",
|
||
|
|
"This is what the _forward backward algorithm_ does for us.",
|
||
|
|
"",
|
||
|
|
"We observe that",
|
||
|
|
"",
|
||
|
|
"$$P(s_t=s) = \\frac{P(o_N...o_t|s_t) P(s_t|o_{t-1}...o_1)}{P(o_N...o_1)} = \\frac{P(o_N...o_t|s_t) P(s_t|o_{t-1}...o_1)}{P(o_N...o_1)}$$",
|
||
|
|
"",
|
||
|
|
"The second factor is the forward probabilities that we have already computed.",
|
||
|
|
"",
|
||
|
|
"The first factor is an accumulated likelihood over a path. It is similar to the forward computation",
|
||
|
|
"in the Viterbi algorithm, but instead of finding the best path, we add up the contributions from all paths",
|
||
|
|
"that arrive in a particular state (we couldn't do that in the Viterbi algorithm because then we wouldn't",
|
||
|
|
"have had a way of tracing back the best path). To propagate the probabilites backwards, we use the",
|
||
|
|
"transpose of the state matrix.",
|
||
|
|
"",
|
||
|
|
"In the algorithm below, we renormalize the first factor after each step to avoid underflow."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def hmm_forward_backward(A,B,observations):",
|
||
|
|
" p = dot(A,ones(len(A))/len(A))",
|
||
|
|
" fps = []",
|
||
|
|
" bps = []",
|
||
|
|
" # the forward step",
|
||
|
|
" for i,o in enumerate(observations):",
|
||
|
|
" # update P(state|ovservation) using Bayes formula",
|
||
|
|
" p = B[o,:]*p",
|
||
|
|
" p /= sum(p)",
|
||
|
|
" fps.append(p)",
|
||
|
|
" # now compute the probabilities in the next state",
|
||
|
|
" p = dot(A,p)",
|
||
|
|
" fps = array(fps)",
|
||
|
|
" # the backward step",
|
||
|
|
" p = ones(len(A))",
|
||
|
|
" bps = []",
|
||
|
|
" for i,o in enumerate(observations[::-1]):",
|
||
|
|
" # update P(state|ovservation) using Bayes formula",
|
||
|
|
" p = B[o,:]*p",
|
||
|
|
" p /= sum(p)",
|
||
|
|
" bps.append(p)",
|
||
|
|
" # now compute the probabilities in the previous",
|
||
|
|
" p = dot(A.T,p)",
|
||
|
|
" bps = array(bps)[::-1]",
|
||
|
|
" smoothed = array(fps)*array(bps)",
|
||
|
|
" smoothed /= sum(smoothed,1)[:,newaxis]",
|
||
|
|
" return smoothed,fps,bps"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1744
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Given our nice state sequence above, this would look completely boring, so let's run this with some noisy transition matrix."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(6,6)",
|
||
|
|
"A2,B2 = ls(A1+0.5*rand(*A1.shape)),ls(B1+0.5*rand(*B1.shape))",
|
||
|
|
"smoothed,fps,bps = hmm_forward_backward(A2,B2,outputs)",
|
||
|
|
"imshow(r_[signal[:30].T,smoothed[:30].T/amax(smoothed)],interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1748,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x57381e90>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAARIAAAFpCAYAAABQ0qchAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nx6urqz3Hb9265Tnut5dmIfY1
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1748
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Here you can see the contributions of the forward and backward directions."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"subplot(311); imshow(smoothed[:30].T,interpolation='nearest')",
|
||
|
|
"subplot(312); imshow(fps[:30].T,interpolation='nearest')",
|
||
|
|
"subplot(313); imshow(bps[:30].T,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1749,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x57390cd0>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAWYAAAFJCAYAAACo3If1AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nIoNhMBMRGQyDmYjIYBjMREQG
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1749
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Compared to the _forward algorithm_, this algorithm does take advantage of all the information in the label sequence.",
|
||
|
|
"",
|
||
|
|
"However, it is still not guaranteed that the states that maximize the posterior probability at each time $t$ actually",
|
||
|
|
"are all on the best path."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Baum-Welch Reestimation",
|
||
|
|
"======================="
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"The _forward backward algorithm_ now gives us everything we need for the second HMM training",
|
||
|
|
"algorithm, _Baum Welch reestimation_.",
|
||
|
|
"",
|
||
|
|
"The idea here is similar to Viterbi training, but with two important differences:",
|
||
|
|
"",
|
||
|
|
"- instead of the sequence of states on the best path, we use the most probable states at each time $t$",
|
||
|
|
"- instead of \"hard updates\" (0/1 indicator whether we are in a state), we update using the state probabilities computed by the forward backward algorithm."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"ns = 5",
|
||
|
|
"A0 = A.copy()",
|
||
|
|
"B0 = B.copy()"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1750
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"_,fps,bps = hmm_forward_backward(A0,B0,outputs)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1751
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"The re-estimate of the matrix $B$ is quite simple: we compute the probability of being in state $s$ at times $t$",
|
||
|
|
"and then update the \"soft counts\" in the new `B` matrix."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def re_B(A,B,outputs,fps,bps):",
|
||
|
|
" B1 = zeros(B.shape,'f')",
|
||
|
|
" for t in range(1,len(outputs)):",
|
||
|
|
" state = vs(fps[t]*bps[t])",
|
||
|
|
" B1[outputs[t]] += state",
|
||
|
|
" return ls(maximum(1e-3,B1))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1588
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"B1 = re_B(A0,B0,outputs,fps,bps)",
|
||
|
|
"subplot(121); imshow(B0,interpolation='nearest')",
|
||
|
|
"subplot(122); imshow(B1,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1589,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x9391f490>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAWYAAACZCAYAAAAclZBuAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAAAASUVORK5CYII=\n"
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1589
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"The re-estimation for the matrix $A$ is slightly trickier, since we are updating counts",
|
||
|
|
"for a transition.",
|
||
|
|
"",
|
||
|
|
"To do this, the correct procedure can be described as follows (I will not prove that here):",
|
||
|
|
"",
|
||
|
|
"- we compute the forward probabilities to get into states at time $t-1$",
|
||
|
|
"- we compute the backwards probabilities to get into states at time $t$ and combine that with the observation at time $o_t$",
|
||
|
|
"",
|
||
|
|
"We might now just want to update by taking the outer product of the state vector at times $t-1$ and $t$ and add",
|
||
|
|
"that as soft counts to our new transition matrix. If all the values were 0/1 (as in Viterbi training), that would be correct. ",
|
||
|
|
"However, for soft updates, we also need to take into account the existing weight assigned to the transition by the existing",
|
||
|
|
"transition matrix $A$ (we counted that in neither the forward nor the backward computations).",
|
||
|
|
"",
|
||
|
|
"Therefore, the full reestimation for the matrix $A$ looks like this:"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def re_A(A,B,outputs,fps,bps):",
|
||
|
|
" A1 = zeros(A.shape)",
|
||
|
|
" for t in range(1,len(outputs)):",
|
||
|
|
" state0 = vs(fps[t-1])",
|
||
|
|
" state1 = vs(bps[t]*B[outputs[t],:])",
|
||
|
|
" A1 += vs(outer(state1,state0)*A)",
|
||
|
|
" return ls(maximum(1e-3,A1))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1590
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Running one reestimation step has an effect similar to what the Viterbi update had: some of the weights get updated,",
|
||
|
|
"but the overall structure of the transition matrix isn't changing much."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"A1 = re_A(A0,B0,outputs,fps,bps)",
|
||
|
|
"subplot(121); imshow(A0,interpolation='nearest')",
|
||
|
|
"subplot(122); imshow(A1,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1591,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x93edb190>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAWYAAACyCAYAAABx/mWbAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n3Zx7enp09OhRbd682ftjab7P
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1591
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Let's do some more updates and see what we get."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def reestimate(A,B,outputs):",
|
||
|
|
" smoothed,fps,bps = hmm_forward_backward(A,B,outputs)",
|
||
|
|
" return re_A(A,B,outputs,fps,bps),re_B(A,B,outputs,fps,bps)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1592
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"for i in range(50):",
|
||
|
|
" A1,B1 = reestimate(A1,B1,outputs)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1593
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(20,6)",
|
||
|
|
"subplot(211); imshow(centers[hmm_sample(A0,B0,400)].T)",
|
||
|
|
"subplot(212); imshow(centers[hmm_sample(A1,B1,400)].T)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1594,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x94172e90>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAABHkAAAE1CAYAAABzzjxoAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\noqyE6SCVSsHq6iqsrq7aDKlo
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1594
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"As before, the reestimated model gives much nicer outputs than the model we started with.",
|
||
|
|
"",
|
||
|
|
"The odd variation in character size is a consequence of the fact that the width of each part of the letter varies randomly with an exponential distribution.",
|
||
|
|
"Given the small number of states our model has, this is pretty close to what can be achieved."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Constrained Models and Parameter Tying",
|
||
|
|
"================"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"A simple way of dealing with this would seem to be to adopt a more complicated model with more states and transitions,",
|
||
|
|
"as suggested above in the section on durational models.",
|
||
|
|
"",
|
||
|
|
"The problem with that is that we have limited training data, and increasing, say, to 100 states requires then that we estimate",
|
||
|
|
"10000 transition probabilities; we don't have enough training data to do that.",
|
||
|
|
"",
|
||
|
|
"The way out of this dilemma is to constrain the structure of the transition matrices (i.e., force some of their elements",
|
||
|
|
"to be small or zero) and/or to \"tie\" parameters together.",
|
||
|
|
"",
|
||
|
|
"Here, I will illustrate a simple constraint on the state transition matrix, a restricted for of Bakis model.",
|
||
|
|
"",
|
||
|
|
"To do this, we generate a mask of those values in the final transition matrix that are allowed to be non-zero.",
|
||
|
|
"This matrix allows self loops and circular progression throuh a sequence of staes."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"ns = 15",
|
||
|
|
"from scipy import linalg",
|
||
|
|
"v = zeros(ns)",
|
||
|
|
"v[:2] = 1.0",
|
||
|
|
"bakis = linalg.circulant(v)",
|
||
|
|
"imshow(bakis,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1754,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x69891c90>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAWwAAAFpCAYAAAC4UhIKAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n+V5BQYEGDXryr82ZM0eNjY3e
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1754
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Now we start with a completely random set of weights, albeit constrained by the form of the parameters above."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"A0 = ls((1.0+rand(ns,ns))*bakis)",
|
||
|
|
"B0 = ls(1.0+rand(no,ns))",
|
||
|
|
"subplot(121); imshow(A0,interpolation='nearest')",
|
||
|
|
"subplot(122); imshow(B0,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1755,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x5798cf10>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAWwAAACyCAYAAABm3PVSAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nWK3WQJdBNCZeMQkAX1aAk3uz
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1755
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"To estimate the parameters of this constrained model, we estimate the parameters of a general",
|
||
|
|
"model and then gradually restrict the model into its desire form.",
|
||
|
|
"That's not the best way of doing this computation (for that, you should modify the",
|
||
|
|
"forward backward and reestimation procedures), but it will do here."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"A1,B1 = A0,B0",
|
||
|
|
"for i in range(2,200):",
|
||
|
|
" A1,B1 = reestimate(A1,B1,outputs)",
|
||
|
|
" A1 = ls(A1*maximum(1.0/i,bakis))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1756
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"These are the final transition matrices."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(10,10)",
|
||
|
|
"subplot(121); imshow(A1,interpolation='nearest')",
|
||
|
|
"subplot(122); imshow(B1,interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1759,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x6b1a6a50>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAksAAAEYCAYAAABMT7LGAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n61//qpMnT9odCYCDxVWWzpw5
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1759
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"And here are randomly generated samples from the final output."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(6,6)",
|
||
|
|
"imshow(centers[hmm_sample(A1,B1,400)].T)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 1760,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x9bff5e90>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAWwAAAA7CAYAAABfY7BCAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n3+9njXU4HQBAQw579uzZmD17
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 1760
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"This looks a lot nicer than the other models, and spacing and size are much more controlled.",
|
||
|
|
"Keep in mind that we started with random weights!",
|
||
|
|
"",
|
||
|
|
"However, this model still cannot represent relationships between different parts of a character:",
|
||
|
|
"if the first half of a character is \"wide\", then the second half should be as well, but that",
|
||
|
|
"relationship is not represented here.",
|
||
|
|
"To represent that, we would need more complex models (e.g., with multiple parallel paths)."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": []
|
||
|
|
}
|
||
|
|
]
|
||
|
|
}
|
||
|
|
]
|
||
|
|
}
|