mirror of
https://github.com/saymrwulf/nvidia-AI.summerschool17-notebooks.git
synced 2026-11-21 09:54:43 +00:00
497 lines
1 MiB
Text
497 lines
1 MiB
Text
|
|
{
|
||
|
|
"metadata": {
|
||
|
|
"name": "35-simple-pca"
|
||
|
|
},
|
||
|
|
"nbformat": 2,
|
||
|
|
"worksheets": [
|
||
|
|
{
|
||
|
|
"cells": [
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"from pylab import *",
|
||
|
|
"from scipy.spatial import distance",
|
||
|
|
"from scipy.ndimage import interpolation",
|
||
|
|
"import tables",
|
||
|
|
"import random as pyrandom"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 1
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def showgrid(images,rows=None,cols=10,size=1,cmap=cm.gray):",
|
||
|
|
" if rows is None: rows = (len(images)+cols-1)//cols",
|
||
|
|
" figsize(cols*size,rows*size)",
|
||
|
|
" for i in range(len(images)):",
|
||
|
|
" subplot(rows,cols,i+1)",
|
||
|
|
" imshow(images[i],cmap=cmap)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 2
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"with tables.openFile(\"mnist.h5\",\"r\") as mnist:",
|
||
|
|
" print mnist",
|
||
|
|
" images = array(mnist.root.deskewed,'f')",
|
||
|
|
" labels = array(mnist.root.labels,'i')",
|
||
|
|
" test_images = array(mnist.root.test_deskewed,'f')",
|
||
|
|
" test_labels = array(mnist.root.test_labels,'i')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"mnist.h5 (File) ''",
|
||
|
|
"Last modif.: 'Sun Jun 17 16:43:50 2012'",
|
||
|
|
"Object Tree: ",
|
||
|
|
"/ (RootGroup) ''",
|
||
|
|
"/deskewed (EArray(60000, 28, 28), shuffle, zlib(9)) ''",
|
||
|
|
"/images (CArray(60000, 28, 28), shuffle, zlib(9)) ''",
|
||
|
|
"/labels (CArray(60000,), shuffle, zlib(9)) ''",
|
||
|
|
"/test_deskewed (EArray(10000, 28, 28), shuffle, zlib(9)) ''",
|
||
|
|
"/test_images (CArray(10000, 28, 28), shuffle, zlib(9)) ''",
|
||
|
|
"/test_labels (CArray(10000,), shuffle, zlib(9)) ''",
|
||
|
|
""
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 3
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Principal Component Analysis",
|
||
|
|
"============================",
|
||
|
|
"",
|
||
|
|
"Recall that classifiers are functions $D(x)$ from feature vectors $x$ to classes $c$,",
|
||
|
|
"where $D(x)$ is determined by a training set $T = \\\\{(x_1,c_1),...,(x_N,c_N)\\\\}$.",
|
||
|
|
"",
|
||
|
|
"We have already considered a number of transformations of the input vectors: feature extraction,",
|
||
|
|
"canonicalization, kernels, etc. These are functions $f:x\\rightarrow\\xi$ that transform the input",
|
||
|
|
"feature vectors into some other space in which recognition is easier.",
|
||
|
|
"",
|
||
|
|
"There are also automatic ways of constructing such transformations.",
|
||
|
|
"One of the most important ones is _principal component analysis_ (PCA).",
|
||
|
|
"PCA is a linear transformation of the input space that results in the best",
|
||
|
|
"possible approximation to the input vectors with a given number of dimensions.",
|
||
|
|
"",
|
||
|
|
"That is, ",
|
||
|
|
"",
|
||
|
|
"- $f(x) = Mx+\\mu$",
|
||
|
|
"- $||f(x)-f(y)|| \\approx ||x-y||$ for $x,y \\in T$.",
|
||
|
|
"- $f(x)\\in R^m$ and $x\\in R^n$ where $m \\lt n$",
|
||
|
|
"",
|
||
|
|
"Note that we are not using the class labels in the construction.",
|
||
|
|
""
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"The construction of $f$ starts by computing the mean of the vectors in the training set",
|
||
|
|
"and subtracting it from the original data:",
|
||
|
|
"",
|
||
|
|
"$\\hat{\\mu} = \\frac{1}{N} \\sum_{i=1}^N x_i$",
|
||
|
|
"",
|
||
|
|
"(We use $\\hat{\\cdot}$ to indicate quantities that are estimated.)"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"mu = mean(images,axis=0)",
|
||
|
|
"gray()",
|
||
|
|
"imshow(mu)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 4,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x4469a90>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAAD5CAYAAAADZljUAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n1lkDdl4f3gIePU8f03n1LKiO
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 4
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Now we subtract the mean from all the data vectors.",
|
||
|
|
"This actually doesn't change the input too much."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"centered = images - mu[newaxis,:,:]",
|
||
|
|
"showgrid(centered[:50])"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAkwAAAEzCAYAAADD+FxKAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nt2NtbQ3T09N49dVXkc1mRcDM
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 5
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def make2d(a): return a.reshape(len(a),prod(a.shape[1:]))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 6
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Next, we compute the covariance matrix (we already have seen that when looking at Gaussian densities)."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"cov = dot(make2d(centered).T,make2d(centered))/len(centered)",
|
||
|
|
"print cov.shape,amin(cov),amax(cov)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"(784, 784) -0.114308 0.20846"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 7
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(8,8)",
|
||
|
|
"imshow(cov,cmap=cm.jet,vmin=-0.2,vmax=0.2)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 19,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x9e83ed0>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAeIAAAHcCAYAAAAKmal+AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nk6TQt2/6YgUsd7ofX7npXfx+
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 19
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"The matrix $M$ of the PCA transformation is now given by the eigenvectors",
|
||
|
|
"of the covariance matrix.",
|
||
|
|
"The top $m$ eigenvectors give us a projection matrix from $n$ to $m$ dimensions",
|
||
|
|
"that represents the best linear approximation to the input vectors.",
|
||
|
|
"(Here \"top\" refers to the eigenvectors in descending order of magnitude of",
|
||
|
|
"their corresponding eigenvectors.)"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"evals,evecs = linalg.eig(cov)",
|
||
|
|
"top = argsort(-evals)",
|
||
|
|
"evals = evals[top]",
|
||
|
|
"evecs = evecs.T[top] # NB: eigenvectors come out backwards",
|
||
|
|
"print evals.shape,evecs.shape"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"(784,) (784, 784)"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 9
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"We can look at these eigenvectors as images and see that they represent",
|
||
|
|
"meaningful variations of the input in some way;",
|
||
|
|
"the first eigenvector, for example, reprents the variation along",
|
||
|
|
"the skinny/rounded character axis."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(8,8)",
|
||
|
|
"for i in range(36):",
|
||
|
|
" subplot(6,6,i+1)",
|
||
|
|
" imshow(real(evecs[i]).reshape(28,28),vmin=-0.1,vmax=+0.1,cmap=cm.jet)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAdwAAAHbCAYAAABhpitNAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nq2InnyHls7zCTRXtgGKSiS2x
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 20
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"evecs.shape"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 11,
|
||
|
|
"text": [
|
||
|
|
"(784, 784)"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 11
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"To take a data set and transform mit, we now construct a `project` function.",
|
||
|
|
"It first subtracts the mean and then projects using the $M$ matrix we ",
|
||
|
|
"constructed above."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def project(data,k=50):",
|
||
|
|
" centered = data-mu.ravel()[newaxis,:]",
|
||
|
|
" return dot(evecs[:k],data.T).T",
|
||
|
|
" "
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 12
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"When we project the input vectors into this new space,",
|
||
|
|
"the vectors don't look all that meaningful",
|
||
|
|
"However, we can see that the most variation exists in the first",
|
||
|
|
"few components."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"projected = project(make2d(images))",
|
||
|
|
"print projected.shape",
|
||
|
|
"figsize(6,6)",
|
||
|
|
"imshow(projected[:60],interpolation='nearest')"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"(60000, 50)"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 13,
|
||
|
|
"text": [
|
||
|
|
"<matplotlib.image.AxesImage at 0x661abd0>"
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAATQAAAFqCAYAAACUBaYYAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n6IEHHqBjX3755SCmivTKXsSs
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 13
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"We can also reconstruct the high dimensionsal (original) representation",
|
||
|
|
"by reversing this process."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"def reconstruct(projected):",
|
||
|
|
" return dot(projected,evecs[:projected.shape[1]])",
|
||
|
|
"reconstructed = reconstruct(projected)",
|
||
|
|
"print reconstructed.shape"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"(60000, 784)"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 14
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"Note that the reconstructed images are not exactly like the original images;",
|
||
|
|
"some detail has been distorted, more for some characters than for others.",
|
||
|
|
"",
|
||
|
|
"Because it allows us to represent large numbers of data components with fewer ones,",
|
||
|
|
"PCA is commonly used in data compression.",
|
||
|
|
"Implicitly, it forms an important part of JPEG and MPEG image and video compression."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"figsize(8,8)",
|
||
|
|
"showgrid(reconstructed[:50].reshape(50,28,28))"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "display_data",
|
||
|
|
"png": "iVBORw0KGgoAAAANSUhEUgAAAkwAAAEzCAYAAADD+FxKAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n19cRi8WQSCSwu7sL4JYnn+/N
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 21
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"PCA as Preprocessing for Classifiers",
|
||
|
|
"====================================",
|
||
|
|
"",
|
||
|
|
"We started out motivating the use of PCA as preprocessing for pattern recognition.",
|
||
|
|
"Let's try this out.",
|
||
|
|
"",
|
||
|
|
"For this to work, we need to transform both the training and the test set vectors.",
|
||
|
|
"The labels, of course, stay the same.",
|
||
|
|
"",
|
||
|
|
"Here, we use nearest neighbor classification, but PCA can be used as a preprocessing",
|
||
|
|
"step to many kinds of pattern recognition algorithms."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"test_projected = project(make2d(test_images))",
|
||
|
|
"print test_projected.shape"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "stream",
|
||
|
|
"stream": "stdout",
|
||
|
|
"text": [
|
||
|
|
"(10000, 50)"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 16
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [
|
||
|
|
"def nneval(training,ltraining,testing,ltesting,metric='euclidean',p=2):",
|
||
|
|
" ds = distance.cdist(make2d(testing),make2d(training),metric=metric,p=p)",
|
||
|
|
" closest = argmin(ds,axis=1)",
|
||
|
|
" predictions = [ltraining[i] for i in closest]",
|
||
|
|
" errors = sum(array(predictions)!=array(ltesting))",
|
||
|
|
" return errors*1.0/len(testing)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [],
|
||
|
|
"prompt_number": 17
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": false,
|
||
|
|
"input": [
|
||
|
|
"nneval(projected[:1000],labels[:1000],test_projected,test_labels)"
|
||
|
|
],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": [
|
||
|
|
{
|
||
|
|
"output_type": "pyout",
|
||
|
|
"prompt_number": 18,
|
||
|
|
"text": [
|
||
|
|
"0.071199999999999999"
|
||
|
|
]
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"prompt_number": 18
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "markdown",
|
||
|
|
"source": [
|
||
|
|
"There are some conditions where PCA doesn't work well.",
|
||
|
|
"For example, if we know that the information about class labels is primarily contained",
|
||
|
|
"in just a few input dimensions, PCA tends to mix this information up with other,",
|
||
|
|
"irrelevant dimensions, and some classifiers then work less well."
|
||
|
|
]
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"cell_type": "code",
|
||
|
|
"collapsed": true,
|
||
|
|
"input": [],
|
||
|
|
"language": "python",
|
||
|
|
"outputs": []
|
||
|
|
}
|
||
|
|
]
|
||
|
|
}
|
||
|
|
]
|
||
|
|
}
|