nvidia-AI.summerschool17-no.../teaching-ncso/28-simple-convnet-threshold.ipynb

525 lines
1.2 MiB
Text
Raw Permalink Normal View History

2017-09-22 17:38:09 +00:00
{
"metadata": {
"name": "51-simple-convnet-threshold"
},
"nbformat": 3,
"nbformat_minor": 0,
"worksheets": [
{
"cells": [
{
"cell_type": "code",
"collapsed": true,
"input": [
"from pylab import *\n",
"from scipy.ndimage import filters\n",
"def F(a): return array(a,'f')\n",
"figsize(12,6)"
],
"language": "python",
"metadata": {},
"outputs": [],
"prompt_number": 5
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Let's start with a simple image processing problem: we want to\n",
"recover a sharp image from a blurred, noisy image.\n",
"We first construct a sharp image."
]
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"temp = F(mean(imread(\"page.png\"),2))\n",
"roi = (slice(500,1000),slice(500,1000))\n",
"target = F(1-(temp[roi]>0.5))\n",
"imshow(target)"
],
"language": "python",
"metadata": {},
"outputs": [
{
"output_type": "pyout",
"prompt_number": 37,
"text": [
"<matplotlib.image.AxesImage at 0x4500250>"
]
},
{
"output_type": "display_data",
"png": "iVBORw0KGgoAAAANSUhEUgAAAXMAAAFsCAYAAAA+OlncAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nSjqhV9XnAqRddurUSfQLxb92
}
],
"prompt_number": 37
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Now we artificially blur and degrade the image."
]
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"image = filters.gaussian_filter(target+0.2*randn(*target.shape),3.0)\n",
"image -= amin(image); image /= amax(image)\n",
"imshow(image)"
],
"language": "python",
"metadata": {},
"outputs": [
{
"output_type": "pyout",
"prompt_number": 40,
"text": [
"<matplotlib.image.AxesImage at 0xccfdf10>"
]
},
{
"output_type": "display_data",
"png": "iVBORw0KGgoAAAANSUhEUgAAAXMAAAFsCAYAAAA+OlncAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\naPzN0pRU7GuMOFWvovMGAPCu
}
],
"prompt_number": 40
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"The simplest convolutional neural network has no hidden layer; it is the equivalent\n",
"of a sigmoid regression. The formula is:\n",
"\n",
"$$ Y = \\sigma(F*X+\\theta) $$\n",
"\n",
"Here\n",
"\n",
"- $F$ is the filter\n",
"- $X$ is the input image\n",
"- $Y$ is the cleaned up output\n",
"- $\\theta$ is a threshold"
]
},
{
"cell_type": "code",
"collapsed": true,
"input": [
"def sigmoid(x): return 1/(1+exp(-x))"
],
"language": "python",
"metadata": {},
"outputs": [],
"prompt_number": 54
},
{
"cell_type": "code",
"collapsed": true,
"input": [
"r=10\n",
"filter = F(0.01*randn(2*r+1,2*r+1))"
],
"language": "python",
"metadata": {},
"outputs": [],
"prompt_number": 55
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"When we initialize the filter with random weights, this is\n",
"the kind of processor we get."
]
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"pred = sigmoid(filters.convolve(image,filter))\n",
"subplot(121); imshow(pred)\n",
"subplot(122); imshow(pred,vmin=0,vmax=1)\n",
"amin(pred),amax(pred)"
],
"language": "python",
"metadata": {},
"outputs": [
{
"output_type": "pyout",
"prompt_number": 56,
"text": [
"(0.43267395210529414, 0.50682288189825297)"
]
},
{
"output_type": "display_data",
"png": "iVBORw0KGgoAAAANSUhEUgAAA9MAAAFsCAYAAADPDurMAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nU7+5MUf8y/GhqS3VBZv43tY/
}
],
"prompt_number": 56
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Now, we can view the convolution basically as a large number of independent\n",
"training problems. Consider the image $I_\\hat{p}$, which is the image\n",
"$I$ shifted by $-p$. If we keep the filter fixed and shift the image,\n",
"then the output $C$ at pixel $p$ is given by:\n",
"\n",
"$$Y(p) = F \\cdot X_{\\hat{p}} + \\theta$$\n",
"\n",
"Generally, the filter $F$ has a small footprint, meaning that\n",
"it is zero outside a small region around the origin.\n",
"\n",
"In essence, this problem is just like training with lots of separate\n",
"training instances, except that we're trying to use a convolution\n",
"operation to implement this.\n",
"We need to keep track of the coordinates in the right way.\n",
"\n",
"$$\n",
"\\begin{eqnarray}\n",
"\\frac{\\partial}{\\partial F_{ij}} \\sum_p (~T(p)-Y(p)~)^2 &=& \\sum 2(T(p)-Y(p)) ~~ \\sigma'(F\\cdot I_{\\hat{p}}) ~~ I_{\\hat{p}}\\\\\\\n",
"&=&\\sum 2(T(p)-Y(p)) ~~ Y(p)(1-Y(p)) ~~ I_{\\hat{p},i,j}\n",
"\\end{eqnarray}\n",
"$$\n",
"\n",
"Here, we define $\\delta(p)$ as before:\n",
"\n",
"$$ \\delta(p) = 2(T(p)-Y(p)) ~~ Y(p)(1-Y(p)) $$"
]
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"delta0 = (target-pred)\n",
"subplot(121); imshow(delta0)\n",
"delta = (target-pred)*pred*(1-pred)\n",
"subplot(122); imshow(delta)"
],
"language": "python",
"metadata": {},
"outputs": [
{
"output_type": "pyout",
"prompt_number": 44,
"text": [
"<matplotlib.image.AxesImage at 0xcd7aa10>"
]
},
{
"output_type": "display_data",
"png": "iVBORw0KGgoAAAANSUhEUgAAAsEAAAFOCAYAAABnv8DkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nvdi3b59WRxXuiXcyxXFShY+B
}
],
"prompt_number": 44
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"This implements the update rule, the change to each individual weight."
]
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"dw = array([[sum(delta*roll(roll(image,i,0),j,1)) for j in range(-r,r+1)] for i in range(-r,r+1)])\n",
"dw /= prod(image.shape)\n",
"imshow(dw)\n",
"amin(dw),amax(dw),dw.dtype"
],
"language": "python",
"metadata": {},
"outputs": [
{
"output_type": "pyout",
"prompt_number": 45,
"text": [
"(-0.020560804507159271, -0.0030023010284572597, dtype('float64'))"
]
},
{
"output_type": "display_data",
"png": "iVBORw0KGgoAAAANSUhEUgAAAWwAAAFpCAYAAAC4UhIKAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nhvfh5ZpU8eAxC1Jv/TWAe/GQ
}
],
"prompt_number": 45
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Now we perform the updates multiple times.\n",
"Note that this is like a batch gradient update over all pixels in the image,\n",
"so each \"iter\" is really more like an \"epoch\"."
]
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"theta = 0.0\n",
"for iter in range(1000):\n",
" pred = sigmoid(filters.convolve(image,filter)+theta)\n",
" err = sum((pred-target)**2)\n",
" delta = (pred-target)*pred*(1-pred)\n",
" delta /= prod(image.shape)\n",
" dw = array([[sum(delta*roll(roll(image,i,0),j,1)) for j in range(-r,r+1)] for i in range(-r,r+1)])\n",
" if iter%50==0:\n",
" print iter,err,\":\",(amin(pred),amax(pred)),sum(abs(delta)),(amin(dw),amax(dw)),(amin(filter),amax(filter)),theta\n",
" filter -= dw\n",
" theta -= sum(delta)"
],
"language": "python",
"metadata": {},
"outputs": [
{
"output_type": "stream",
"stream": "stdout",
"text": [
"0 24760.208943 : (0.0035678377385413352, 0.9978953848980896) 0.0458853929868 (0.0053400223104518819, 0.0072862025192962669) (-0.12178462, 0.29784608) 0.0\n",
"50"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 20176.1372269 : (0.00066173566353980708, 0.99803984359858355) 0.0355319773236 (-0.0012748398445276302, 0.00045858820214109207) (-0.14058399, 0.36322036) -0.260421732761\n",
"100"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 19031.330067 : (0.00048385370845032913, 0.99854400091214424) 0.0326746901224 (-0.0010332582710076623, 0.00040143795443587651) (-0.1455725, 0.41569439) -0.454887283628\n",
"150"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 18198.192199 : (0.0004155409322442416, 0.99872334330467938) 0.0308582819579 (-0.00088536597466849765, 0.00036149074890512781) (-0.15812871, 0.45914328) -0.622232347642\n",
"200"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 17539.3167994 : (0.00038990543483953337, 0.99878464293543867) 0.0295672706387 (-0.00078260360495212408, 0.0003293243177035594) (-0.17461969, 0.50059199) -0.772475198537\n",
"250"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 16994.1057221 : (0.00038608381544868965, 0.99878856138959937) 0.0285879240528 (-0.00070586597014078928, 0.00030194455200845989) (-0.18923596, 0.53775758) -0.910363441817\n",
"300"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 16529.5810543 : (0.00037515339279886427, 0.99875869969945386) 0.0278121315518 (-0.00064585058238400637, 0.00027882636954881774) (-0.20231509, 0.57152367) -1.03855836321\n",
"350"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 16125.6248064 : (0.00037309558600197144, 0.99870606309593635) 0.0271780527159 (-0.00059739241228377329, 0.00025924393283338615) (-0.21412554, 0.60258859) -1.15873135473\n",
"400"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 15769.0215644 : (0.00037939348845828587, 0.99863653733030056) 0.026647204231 (-0.00055734803392893924, 0.00024253697679875588) (-0.22487944, 0.63144696) -1.27202312692\n",
"450"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 15450.6061906 : (0.00039192129822601023, 0.9985536631379115) 0.026194124504 (-0.00052366992786349181, 0.00022816327581808566) (-0.23474531, 0.6584661) -1.37926288474\n",
"500"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 15163.7442414 : (0.0004093551676040092, 0.99849197991033856) 0.0258011740743 (-0.00049495441071344083, 0.00021569034133553074) (-0.24385791, 0.68392771) -1.48108354287\n",
"550"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 14903.4588846 : (0.00043080094880172627, 0.99849865740375987) 0.0254556803338 (-0.00047019424500519458, 0.00020477904207624254) (-0.2523264, 0.70805389) -1.57798726124\n",
"600"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 14665.9044865 : (0.00045560509780877642, 0.99849411870271354) 0.0251482815005 (-0.00044864629512610902, 0.0001951601962971246) (-0.26023978, 0.73102361) -1.67038494714\n",
"650"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 14448.0330459 : (0.00048325261579012488, 0.99848106919570145) 0.0248718997265 (-0.00042974212412478607, 0.00018662198306495541) (-0.26767153, 0.7529828) -1.75862145479\n",
"700"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 14247.3764536 : (0.00051330978740367025, 0.99846150259590782) 0.0246210815827 (-0.0004130403155082855, 0.00017935258926101555) (-0.2746824, 0.77405226) -1.84299226766\n",
"750"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 14061.9001809 : (0.00054539353479446818, 0.99843693083452123) 0.0243915641533 (-0.00039819017585347391, 0.00017361989512953811) (-0.28132361, 0.79433364) -1.92375494443\n",
"800"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 13889.9005715 : (0.00057915236730348807, 0.99840853151815401) 0.0241799662986 (-0.00038490986354262734, 0.0001684039036092902) (-0.28763828, 0.81391168) -2.00113722712\n",
"850"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 13729.9336051 : (0.00061425805257500328, 0.99837723930106792) 0.023983579772 (-0.00037296916839079324, 0.00016363017738825521) (-0.29366297, 0.83285952) -2.07534289738\n",
"900"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 13580.7621414 : (0.00062288689370889759, 0.99834380726690386) 0.0238002149688 (-0.00036217905485174084, 0.0001592367759027625) (-0.29942873, 0.85123914) -2.14655607795\n",
"950"
]
},
{
"output_type": "stream",
"stream": "stdout",
"text": [
" 13441.3162875 : (0.00062606354808537336, 0.99830885078316589) 0.0236280854244 (-0.00035238058175589759, 0.00015517494449664518) (-0.30496234, 0.86910409) -2.21494452426"
]
}
],
"prompt_number": 48
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"As you can see, the nonlinear filter that we computed restores some of the crispness in the text.\n",
"However, it is far from a good filter; for a good filter, we really need multiple units.\n",
"In addition, we have to be very careful about scaling the inputs and the outputs, learning rates, etc."
]
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"figsize(20,6)\n",
"subplot(131); imshow(target)\n",
"subplot(132); imshow(image)\n",
"subplot(133); imshow(sigmoid(filters.convolve(image,filter)+theta))"
],
"language": "python",
"metadata": {},
"outputs": [
{
"output_type": "pyout",
"prompt_number": 51,
"text": [
"<matplotlib.image.AxesImage at 0xcd0df50>"
]
},
{
"output_type": "display_data",
"png": "iVBORw0KGgoAAAANSUhEUgAABIAAAAFmCAYAAAAVjeFiAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAIABJREFUeJzsfXl4nGW5/j37klkyk2SyJ23apjvdgBZb2rK0LAoHRdTj\nckQ84vFSz0FFPcpPAREVlUvwqEe53HBjPYhFy1agrZTSlkKha5o0S9Osk8wkM5NMJslkfn/kut8+\n82bSlhq2+t3XlSuZycz3vd+7Ps/93s/zmjKZTAYGDBgwYMCAAQMGDBgwYMCAAQMGzliY3+oCGDBg\nwIABAwYMGDBgwIABAwYMGHhjYRBABgwYMGDAgAEDBgwYMGDAgAEDZzgMAsiAAQMGDBgwYMCAAQMG\nDBgwYOAMh0EAGTBgwIABAwYMGDBgwIABAwYMnOEwCCADBgwYMGDAgAEDBgwYMGDAgIEzHAYBZMCA\nAQMGDBgwYMCAAQMGDBgwcIbjDSGAnnjiCcyZMwezZs3CHXfc8UbcwoABAwYMvINhrBMGDBgwYGAy\nGGuEAQMGDLwxMGUymcxUXjCdTmP27NnYtGkTysvLcc455+C+++7D3Llzp/I2BgwYMGDgHQpjnTBg\nwIABA5PBWCMMGDBg4I3DlCuAdu7ciZkzZ2LatGmw2Wz40Ic+hL/85S9TfRsDBgwYMPAOhbFOGDBg\nwICByWCsEQYMGDDwxsE61Rdsa2tDZWWlel1RUYEdO3ao1yaTaapvacCAAQNnDKZYlPm2hLFOGDBg\nwMDp40xfJ062RgDGOmHAgAEDJ8KJ1okpJ4Deygl5w4YNuOKKK3DxxRfjmWeemfRzx44dw44dO/DB\nD34Qo6Ojb2IJDbxTIQfRRRddhGefffYtLI0BA+9snMo64fP5YLFYsn74vUwmg7GxMYyOjqqfTCYD\nk8kEi8UCs9msPmexWOByuVBQUIDq6mpMmzYNoVAINpsNAwMDiEQi6Ovrw/DwMMxmM5xOJ1wuFywW\nC5LJJHp6etDZ2Yne3l7E43EMDQ1hZGQEAGC1WmGz2WCz2WA2m5HJZJBOpzE2NgaTyQSz2azeZ3nT\n6bSaT0wmk/pheWUdpVIp+P1+9ezpdBqpVArDw8Pqmc1mM6xWK6xWq6onYmxsDGNjY1nXZZlk+fg9\nln94eBgjIyNZdWq1WtWz8AeA+q7dbofNZsPY2BhGRkaQTCaRSqVUXfGefE5+n++z3XLVRSKRgMvl\nUs9iMpnUM8tn5/dGR0fVs8t6GB0dxcjIiGoD3kv2Rz6rfN50Oq3aTraPzWZDIBBAVVUVFixYgPnz\n5yMUCmFoaAhNTU3Yv38/6uvrEQ6HkUwms9pCr0f5vOxDspwjIyNwu90T2phgH+M12a/4Op1Oq/Kz\nruU44ff5Odk+rF/ZTnJcyjLw+3pdsY71Msprsf1l+WWbDg8PIx6Pw263q76vt+GpXJvXY5uyn/A6\n/B77gBwzrCfZPvL5ZTvI/mSz2WC1WtU4lvOExWJR/zebzaqvyjluZGRE9d3BwcEJ7X+m4VR9iaKi\nopxtxn7J+WxoaAijo6Nq7nC73XA4HBgdHUUikUAymcTo6CgcDgdmzZqFT33qU1i9ejVMJhN27NiB\ne++9F/v27UMqlYLb7casWbOwdu1a1NbWIhKJ4MUXX8SePXvQ09Oj5k7ZB+x2O5xOJ8xmsypPOp2G\n2WyG3W6H3W6HyWTC6OgoUqlU1tifbP1jv00mkygoKEBeXh4sFguGhoYQj8fVMwGYMN75Xf2aHFdm\nsxk2m03d22q1wuFwwOl0YnR0FAMDA4jH46qsvJasd318yH4OACMjI1n3tFqtcDqdaj3lGNHHM59D\n/s9isSAWi8Hr9aqxI9cuu90Oh8MBq9WKsbExDA0NqT7BsnHsyfWP/+PY57Vlf+OaJ9dMXkuuN06n\nE+9///vx3e9+F8XFxRgdHcULL7yAr371qzhw4ABGR0dhs9ng8/ng8/lUX9HXY85FnINoCwDA0NCQ\n+i7vPzw8PKEt5Bwlr8m+INtTzoOyr+Tqo7JM8h7y+rLv8rscAzabTX1GPhf7I/uOnLdtNpsaP6lU\nCkNDQ+jv71fjLdcaIMuv21Dse+l0WrWrXLf0NVuuV/I5c/1P9l9ZT7IOM5kMhoeHlf0k+zn7I+su\nl43G/0WjUZwIU04AlZeXo7W1Vb1ubW1FRUXFVN8mJ3bt2oVgMIhNmzbhiiuuUO9v3boVq1evBjDu\nvJeXl2PTpk0G+WPglLB06dKs1ytWrDAIIAMG/gGcyjrBxU06kvrieirgYj8yMoKBgQEkEgl4vV5l\nbAYCATgcDgwPDytjAABSqZT6fDKZxPDwsDI4uABLx1CWWy+jXPxZJmkY8Fn17wBQxgGACQaXNNZZ\nDmmEyO9JgyedTqv/SWNa/paOqcPhgMPhUAbeyMiIMs6kc2O1WjE6Ogqz2ZzlYJ9MrSDrQa+7yb4r\nnSDWkV6PuV7zeWUd83p6GeTfuQw7aTSOjY0p45MGdy5CUieA9Hpgm/K3JCckwaCTeLpxy/bRDX39\nufQ2kCSfXn7ZHhaLRfULGuUkaUhQSscQQNazTEbQELLf53qOXOSpbqDLz+qkGNtUJwKlIS4d71wE\n04mei9Dbhd+jc03H1G63T+hHqVRq0jY80/F6fAndOePcKMk2+X/OYXSidfI1kUigs7MTg4ODyM/P\nR21tLS655BL4/X5EIhH4/X7MmTMHVVVVAICenh50dXVhYGAgizjQ+6nsm5PNd5JEkQ6/7kDzuyw/\nNwdMJpMiDPR+ye9K0oVzvMlkyiIx5NorxwfLr28u0EF3Op0wmUyKYJHXMJvNcDgccLlcWSR9rnkx\n1+aFjsnm5lxrCJ15vpZ1mKutOPeSHCApoLefLDvHPYkMkrXSwQeOE1983+12KyJJkgP6ppG+Runt\nos+XkoyX8zafRZI0klDnteRYkfO0bE/+LcmHXBslLpcLTqcTADA4OIhkMqnWSLkRJZ+JbaDPyXJO\nZNuS3J1srtTHv7SH5LXYBvJ9+ayyvU9Gosk+LddVfp7PK8epJBNl/bB/yc0YuZHA8sq+eTJMOQF0\n9tlno76+Hs3NzSgrK8MDDzyA++67b6pvkxO33XYbnn76aWzfvh2PPfaYev/w4cOora1Vr9evX4+n\nn376TSmTgXc+vvKVr2S9vv322zFt2jRcf/31b1GJDBh4Z+NU1gm5YHKh1ndCJfQFmO8B4wbJ0NAQ\nEokEotGoMlIdDgdsNhvy8vJgt9uVkcjP9vX1oa+vL8uwl85qLgdxssVXd7D1MvJv3ZHVjQN+jveV\nBIDcuaVjScNekhO6AyHLJQ1Tm80Gj8cDn88Hj8cDq9WKkZERZcCNjo7CarXC4/HA7XbDbrdjeHgY\niURC1aVUg0jDST67bnjzs7Ke9F03/Rqy/nK9L/tVLudLGsG6Skoqveikk2Ch6om7jhaLRZGNrGu9\nPScjKaTRO9mzyx/2A+7M22w2ZDIZpcDSd351o5S/WW/SoJd1IY1a4Liiwev1wuv1KsOe42ZgYABD\nQ0M521Fvdx064SGvIQkavc11o5wGcq52l9/Vr8X3JLknnT32O94jl+pKVybxu1R8sM3y8vLg9Xrh\n8Xiy1CGJRAL9/f3KGchVh2c6Xo8vwbo1mcYVNHI3XKr3WIecm9iW0sFOp9Po6+vDa6+9htraWtTW\n1sJqtWLWrFkwm83o6+uDy+VCMBiE2WxGR0cH6uvr0dbWhkQiMUFRksspzNWW+rjMtZ7Iz9CB5Xwt\n53USQFL9I+/DPmuxWOB2uxEMBmG32xGLxdDX14dUKqWuzzJIZ1M6ndJB9vl8mDZtGtxuN1paWtDW\n1pY1Hu12O0pKSlBSUoLe3l60tbVlbcTra56uRJRjXT6TXrdyjMq1h+NJqotkXcs65j3Zl/Q2lOOc\nP3JTKT8/H8lkEu3t7Vn3ikQiaG9vh8/nw8jICDo6OhRxJ+2V4eFh1Z/Zf3XCKxcZpNej7NtUvrlc\nLoyMjCgFlz6nkWQBssltQrd5
}
],
"prompt_number": 51
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Looking at the filter itself is also interesting:"
]
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"imshow(filter)"
],
"language": "python",
"metadata": {},
"outputs": [
{
"output_type": "pyout",
"prompt_number": 53,
"text": [
"<matplotlib.image.AxesImage at 0xfe8c950>"
]
},
{
"output_type": "display_data",
"png": "iVBORw0KGgoAAAANSUhEUgAAAWwAAAFpCAYAAAC4UhIKAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nNKx7GWiOdPnZurLzFY0JKpvN
}
],
"prompt_number": 53
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Note that this filter combines a number of sources of information:\n",
"\n",
"- it performs a kind of deconvolution\n",
"- it also takes advantage of the statistics of the input image (hence the anisotropy)\n",
"- it depends to some degree on the absolute input values, since the sigmoid depends on that"
]
}
],
"metadata": {}
}
]
}