From 529c8c1f847a204aa5ba234ea45c8edd4ae3d7bd Mon Sep 17 00:00:00 2001 From: William de Vazelhes Date: Fri, 8 Mar 2019 17:10:44 +0100 Subject: [PATCH 1/9] Export notebook to gallery --- examples/metric_plotting.ipynb | 708 ------------------------------- examples/plot_metric_examples.py | 398 +++++++++++++++++ 2 files changed, 398 insertions(+), 708 deletions(-) delete mode 100644 examples/metric_plotting.ipynb create mode 100644 examples/plot_metric_examples.py diff --git a/examples/metric_plotting.ipynb b/examples/metric_plotting.ipynb deleted file mode 100644 index f8661181..00000000 --- a/examples/metric_plotting.ipynb +++ /dev/null @@ -1,708 +0,0 @@ -{ - "cells": [ - { - "cell_type": "markdown", - "metadata": { - "collapsed": false - }, - "source": [ - "### Metric Learning and Plotting\n", - "\n", - "This is a small walkthrough which illustrates all the Metric Learning algorithms implemented in metric_learn, and also does a quick visualisation which can help understand which algorithm might be best suited for you.\n", - "\n", - "Of course, depending on the data set and the constraints your results will look very different; you can just follow this and change your data and constraints accordingly. " - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "#### Imports " - ] - }, - { - "cell_type": "code", - "execution_count": 1, - "metadata": { - "collapsed": false - }, - "outputs": [], - "source": [ - "%matplotlib inline\n", - "\n", - "import metric_learn\n", - "import numpy as np\n", - "from sklearn.datasets import load_iris\n", - "\n", - "# visualisation imports\n", - "import matplotlib.pyplot as plt\n", - "from mpl_toolkits.mplot3d import Axes3D" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "### Loading our data-set and setting up plotting\n", - "\n", - "We will be using the IRIS data-set to illustrate the plotting. You can read more about the IRIS data-set here: [link](https://en.wikipedia.org/wiki/Iris_flower_data_set). \n", - "\n", - "We would like to point out that only two features - Sepal Width and Sepal Length are being plotted. This is because it is tough to visualise more features than this. The purpose of the plotting is to understand how each of the new learned metrics transform the input space. " - ] - }, - { - "cell_type": "code", - "execution_count": 2, - "metadata": { - "collapsed": false - }, - "outputs": [], - "source": [ - "# loading our dataset\n", - "\n", - "iris_data = load_iris()\n", - "# this is our data\n", - "X = iris_data['data']\n", - "# these are our constraints\n", - "Y = iris_data['target']\n", - "\n", - "# function to plot the results\n", - "def plot(X, Y):\n", - " x_min, x_max = X[:, 0].min() - .5, X[:, 0].max() + .5\n", - " y_min, y_max = X[:, 1].min() - .5, X[:, 1].max() + .5\n", - " plt.figure(2, figsize=(8, 6))\n", - "\n", - " # clean the figure\n", - " plt.clf()\n", - "\n", - " plt.scatter(X[:, 0], X[:, 1], c=Y, cmap=plt.cm.Paired)\n", - " plt.xlabel('Sepal length')\n", - " plt.ylabel('Sepal width')\n", - "\n", - " plt.xlim(x_min, x_max)\n", - " plt.ylim(y_min, y_max)\n", - " plt.xticks(())\n", - " plt.yticks(())\n", - "\n", - " plt.show()" - ] - }, - { - "cell_type": "code", - "execution_count": 3, - "metadata": { - "collapsed": false - }, - "outputs": [ - { - "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAAdsAAAFsCAYAAACEtRP5AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAOAZZQsAgGf/AckQihPvOEbCAAAAAElFTkSuQmCC\n", - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - } - ], - "source": [ - "# plotting the dataset as is.\n", - "plot(X, Y)" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "### Metric Learning\n", - "\n", - "Why is Metric Learning useful? We can, with prior knowledge of which points are supposed to be closer, figure out a better way to understand distances between points. Especially in higher dimensions when Euclidean distances are a poor way to measure distance, this becomes very useful.\n", - "\n", - "Basically, we learn this distance: $D(x,y)=\\sqrt{(x-y)\\,M^{-1}(x-y)}$.\n", - "And we learn this distance by learning a Matrix $M$, based on certain constraints.\n", - "\n", - "Some good reading material for the same can be found [here](https://arxiv.org/pdf/1306.6709.pdf). It serves as a good literature review of Metric Learning. \n", - "\n", - "We will briefly explain the metric-learning algorithms implemented by metric-learn, before providing some examples for it's usage, and also discuss how to go about doing manual constraints.\n", - "\n", - "Metric-learn can be easily integrated with your other machine learning pipelines, and follows (for the most part) scikit-learn conventions." - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "### Large Margin Nearest Neighbour\n", - "\n", - "LMNN is a metric learning algorithm primarily designed for k-nearest neighbor classification. The algorithm is based on semidefinite programming, a sub-class of convex programming (as most Metric Learning algorithms are).\n", - "\n", - "The main intuition behind LMNN is to learn a pseudometric under which all data instances in the training set are surrounded by at least k instances that share the same class label. If this is achieved, the leave-one-out error (a special case of cross validation) is minimized. \n", - "\n", - "You can find the paper [here](http://jmlr.csail.mit.edu/papers/volume10/weinberger09a/weinberger09a.pdf)." - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "#### Fit and then transform!" - ] - }, - { - "cell_type": "code", - "execution_count": 4, - "metadata": { - "collapsed": false - }, - "outputs": [], - "source": [ - "# setting up LMNN\n", - "lmnn = metric_learn.LMNN(k=5, learn_rate=1e-6)\n", - "\n", - "# fit the data!\n", - "lmnn.fit(X, Y)\n", - "\n", - "# transform our input space\n", - "X_lmnn = lmnn.transform()" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "So what have we learned? The matrix $M$ we talked about before.\n", - "Let's see what it looks like." - ] - }, - { - "cell_type": "code", - "execution_count": 5, - "metadata": { - "collapsed": false - }, - "outputs": [ - { - "data": { - "text/plain": [ - "array([[ 2.49193844, 0.35638993, -0.39984418, -0.77608969],\n", - " [ 0.35638993, 1.68815388, -0.90376817, -0.07406329],\n", - " [-0.39984418, -0.90376817, 2.37468946, 2.18784107],\n", - " [-0.77608969, -0.07406329, 2.18784107, 2.94523937]])" - ] - }, - "execution_count": 5, - "metadata": {}, - "output_type": "execute_result" - } - ], - "source": [ - "lmnn.metric()" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "Now let us plot the transformed space - this tells us what the original space looks like after being transformed with the new learned metric." - ] - }, - { - "cell_type": "code", - "execution_count": 6, - "metadata": { - "collapsed": false - }, - "outputs": [ - { - "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAAdsAAAFsCAYAAACEtRP5AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAAAAAElFTkSuQmCC\n", - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - } - ], - "source": [ - "plot(X_lmnn, Y)" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "Pretty neat, huh?\n", - "\n", - "The rest of this notebook will briefly explain the other Metric Learning algorithms before plottting them.\n", - "Also, while we have first run `fit` and then `transform` to see our data transformed, we can also use `fit_transform` if you are using the bleeding edge version of the code. The rest of the examples and illustrations will use `fit_transform`." - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "### Information Theoretic Metric Learning \n", - "\n", - "ITML uses a regularizer that automatically enforces a Semi-Definite Positive Matrix condition - the LogDet divergence. It uses soft must-link or cannot like constraints, and a simple algorithm based on Bregman projections. \n", - "\n", - "Link to paper: [ITML](http://www.cs.utexas.edu/users/pjain/pubs/metriclearning_icml.pdf). " - ] - }, - { - "cell_type": "code", - "execution_count": 7, - "metadata": { - "collapsed": false - }, - "outputs": [], - "source": [ - "itml = metric_learn.ITML_Supervised(num_constraints=200)\n", - "X_itml = itml.fit_transform(X, Y)" - ] - }, - { - "cell_type": "code", - "execution_count": 8, - "metadata": { - "collapsed": false - }, - "outputs": [ - { - "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAAdsAAAFsCAYAAACEtRP5AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nRNxMxVZERMTNVGxFRETc7P8AXsVpWPQcdcQAAAAASUVORK5CYII=\n", - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - } - ], - "source": [ - "plot(X_itml, Y)" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "### Sparse Determinant Metric Learning\n", - "\n", - "Implements an efficient sparse metric learning algorithm in high dimensional space via an $l_1$-penalised log-determinant regularization. Compare to the most existing distance metric learning algorithms, the algorithm exploits the sparsity nature underlying the intrinsic high dimensional feature space.\n", - "\n", - "Link to paper here: [SDML](http://lms.comp.nus.edu.sg/sites/default/files/publication-attachments/icml09-guojun.pdf). \n", - "\n", - "One feature which we'd like to show off here is the use of random seeds.\n", - "Some of the algorithms feature randomised algorithms for selecting constraints - to fix these, and get constant results for each run, we pass a numpy random seed as shown in the example below." - ] - }, - { - "cell_type": "code", - "execution_count": 9, - "metadata": { - "collapsed": false - }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "/Users/bhargavvader/Open_Source/metric-learn/venv/lib/python2.7/site-packages/sklearn/covariance/graph_lasso_.py:252: ConvergenceWarning: graph_lasso: did not converge after 100 iteration: dual gap: 2.377e-04\n", - " ConvergenceWarning)\n" - ] - } - ], - "source": [ - "sdml = metric_learn.SDML_Supervised(num_constraints=200)\n", - "X_sdml = sdml.fit_transform(X, Y, random_state = np.random.RandomState(1234))" - ] - }, - { - "cell_type": "code", - "execution_count": 10, - "metadata": { - "collapsed": false - }, - "outputs": [ - { - "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAAdsAAAFsCAYAAACEtRP5AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nbEVERBymYisiIuKw/wO7E19JBWZHrgAAAABJRU5ErkJggg==\n", - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - } - ], - "source": [ - "plot(X_sdml, Y)" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "### Least Squares Metric Learning\n", - "\n", - "LSML is a simple, yet effective, algorithm that learns a Mahalanobis metric from a given set of relative comparisons. This is done by formulating and minimizing a convex loss function that corresponds to the sum of squared hinge loss of violated constraints. \n", - "\n", - "Link to paper: [LSML](http://web.cs.ucla.edu/~weiwang/paper/ICDM12.pdf)" - ] - }, - { - "cell_type": "code", - "execution_count": 11, - "metadata": { - "collapsed": false - }, - "outputs": [], - "source": [ - "lsml = metric_learn.LSML_Supervised(num_constraints=200)\n", - "X_lsml = lsml.fit_transform(X, Y)" - ] - }, - { - "cell_type": "code", - "execution_count": 12, - "metadata": { - "collapsed": false - }, - "outputs": [ - { - "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAAdsAAAFsCAYAAACEtRP5AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAIABJREFUeJzs3Xd4jXcbwPHvk5O9JJGIRIQYiVnE3mIXtXftvWmrrdVB\naUurqLYULVW1N7U3sRWJGUSQkEhEZM/zvH9E9U1PgkZOTsT9uS7X65z3Pr/nfnpF7vP8pqKqKkII\nIYTQHyNDJyCEEELkd1JshRBCCD2TYiuEEELomRRbIYQQQs+k2AohhBB6JsVWCCGE0DNjfTWsKIqs\nKRJCCPHGUVVV+fd7eiu2Ty+oz+aFEEKIPEVRdOosIN3IQgghhN5JsRVCCCH0TIqtEEIIoWdSbIUQ\nQgg9k2IrhBBC6JkUWyGEEELPpNgKIYQQeibFVgghhNAzKbZCCCGEnkmxFUIIIfRMiq0QQgihZ1Js\nhRBCCD2TYiuEEELomRRbIYQQQs+k2AohhBB6JsVWCCGE0DMptkIIIYSeSbEVQggh9EyKrRBCCKFn\nUmyFEEIIPZNiK8RLOnHiBD179aJLtx7s2bPH0OkIIV4jUmyFeAknT56kdZt3MHMrg32Zqrzbuy/b\nt283dFpCiNeEoqqqfhpWFFVfbQuR23r37YuRc0la9hgAwMk927i8fzP79uw2cGZCiLxEURRUVVX+\n/b482QrxEtLS0jA2MXn2WmNiQlpamgEzEkK8TowNnYAQr4PBAwfSuVt3LKysMTE1Z/XcL5jz7SxD\npyWEeE1IN7IQL2n37t18N3ceaWlpDB44gG7duhk6JSFEHpNVN7IUWyGEECKHyJitEEIIYSBSbIUQ\nQgg9k2IrhBBC6JkUWyGEEELPpNgKIYQQeibFVgghhNAzKbZCCCGEnkmxFUIIIfRMiq0QQgihZ1Js\nhRBCCD2TYiuEEELomRRbIYQQQs+k2AohhBB6JsVWCCGE0DMptkIIIYSeSbEVQggh9EyKrRBCCKFn\nUmyFEEIIPZNiK/KlpKQkBg8Zip29Ay6uRViwcKGhUxJCvMGMDZ2AEPrw8YSJnL92iy/X7CXmcSTT\nxg/CvWhRWrdubejUhBBvIHmyFfnSzl276DTiQ+ydnHH3LEuTrn3ZsWuXodMSQryhpNiKfMne3p7Q\nO4HPXofdCcTRwcGAGQkh3mSKqqr6aVhRVH21LcSLHD16lHbtO1CzRVtiHz8i+Polzpw+haOjo6FT\nE0LkY4qioKqqovO+FFuRX127do3t27djaWlJjx49sLe3N3RKQoh8ToqtEEIIoWdZFVsZsxVCCCH0\nTIqtEEIIoWdSbIUQQgg9k2IrhBBC6JkUWyGEEELPpNgKIYQQeibFVuhdXFwcly9fJjIy0tCpCCGE\nQUixFXp17NgxipcoQau27SnuUYJFixcbOiUhhMh1sqmF0JvU1FRc3dwYMOUbKtX1IfTubaYP6sQJ\n32N4enoaOj0hhMhxsqmFyHUPHz4kLU1Lpbo+ABR296BUhcpcuXLFwJkJIUTukmIr9MbR0RFtWioB\nF88B8Dg8jMAr/pQqVcrAmQkhRO6Sw+OF3piamvLHihX06t2HIsVLEhx0i48/+pAKFSoYOjUhhMhV\nMmYr9O7hw4dcvXqVokWLUqJECUOnI4QQeiOn/gghhBB6JhOkhMglvr6+vN26DQ18GrNg4ULkS6cQ\nQoqtEDno/PnzvNO2He7VfajVoS+zvpvH3HnzDJ2WEMLApBtZiBz0wfjx3I2DjkPGAXDD7y9WfzOF\nK5f8DJyZECI3SDeyELlAo9GQmpz07HVKchIajfwzE+JNJ0t/hMhBAwcMoHbdulhY22Dr4MiWJfP4\n6ouphk5LCGFg0o0sRA67dOkS3343h7i4eHp060LHjh0NnZIQIpfI0h8hhBBCz2TMVgghhDAQGbMV\nQgiRZyUlJXH69GkURaFGjRqYmpoaOqVskWIrhBAiT3r06BGN6tUhJfoRWlXFqmBhDh71xc7OztCp\n/WfSjSyEECJHJCcn88G4MZQs5oZ3xXLs3LnzldqbPOFjiitRzGzgyDcNnSicEs6nUyblULa5S4qt\nEEKIHDH+/XEc3rSSceVNaeMQQ6/uXTh79my227tx/SpVnExRFAVFUajsZMqNa9dyMOPcI8VWCCFE\njtiwbh2D3ypAMTszqhWxpom7JVu3bsl2e1Wq1eBwSCKpWpWUNJUj95OoWqNmDmace6TYCiGEyBGW\nlhZEJqQ+ex2VDNbWNtlub9r0GZgXq8CgHfcYvOMedqUqM+XTz3Ii1Vwn62zFa+fOnTusX7+ewoUL\n06NHD4yM5DujEHnBqlWrGDt8CC2LWRCepHI5xphzF/xwcnLKdpuqqhIcHIyiKBQpUgRF0VnCmqfI\nphYiX1i/fj29+vTFpVgJHj8MpaCDPdevXsHYWCbWC5EXHDx4kO3btmJboADDh4+gUKFChk4pV0mx\nFfmCXUFHuo2eRKN2XUlOSuTTPm1p08yHH3/80dCpCSGE7CAl8of4uFiq1G8MgKmZOZXq+nDtNZ2d\nKIR4c0ixFa+VAgXsObR5NQAxUY85uWcbtWrVMnBWQgjxfNKNLF4rx44do2Wr1miMTUiIi6Vqtar4\nHj0qk6SEEHmCjNmKfCMxMRFfX19cXV0pW7asodMRQohnpNgKIYQQeiYTpIQQQggDkWIrhBBC6JkU\nWyFe0r59+2j9TltatmrNhg0bDJ2OEOI1IsVWiJdw+PBhuvXoiXuNxng1bMOI0WNZt26dodMSQrwm\nZIKUEC/h3d69MStalmZd+gBw5sBOzu9cy8F9ew2cmRAiL5EJUkK8AkVR0KZpn73WarUY5fEN0YUQ\neYfs3i6yJS0tjU6dOuHn50fZsmXZunUrGo3G0GnpzYhhw2jTth0aYw0mpmZsWPANvyz62dBpCSFe\nE9KNLLLFxdUN1diEGk1ace7QHpLiYwi7H5KvC+6xY8f4fv6PpKalMnjgAN5++21DpySEyGNkUwuR\nYzZt2kSPd3vx464zWNrYkpgQz6iW1Zk7+1sGDx5s6PSEEMJgZMxW5JgbN25gXcAOSxtbAMwtLLG1\nL0hQUJBhExNCiDxKiq34z/r06UPskyh2/LGEJ5ER7Fv/O48fhjJgwABDpyaEEHmSdCOLbFm2bBmj\nxo4jOTERUzMzvpn5NcOHDzd0WkIIYVAyZiuEEELomYzZCiGEEAYixVYIIfKItLQ0Q6cg9ESKrRBC\nGNiuXbso7FQQUxMTvN+qQGBgoKFTEjlMiq3IMyIjIzl69CgBAQFZxsTExNC1ew9sbAvgVtSdVatW\n5WKGQuS8oKAgenbrwthK1qzv6kllk0e0adkcmfOSv0ixFXnC8ePH8SxThmFj36d2vfp8MP7DTOMG\nDx1KWFwys7ccY+iMHxk99j1OnjyZy9kKkXNOnz5NeWdryheyRGOk0M6zAMEhIURGRho6NZGDpNiK\nbHnw4AEdO3fBs2w52nXoSHBw8Cu116Pnu/SbNJMpv2xi5voDrN2wkYMHD+rE7dmzh26jJ2FjZ0+p\nilWo27oT+/fvf6VrC2FIzs7O3ItKJPnpQRcPYlNI06rY2toaODORk6TYiv8sJSWFps1boC3gzMDP\n52HqXByfJk1JTEzMdnvB9+5SpX4TAKxsClDGu2am3cn29g48uJM+nqWqKmF3b+Hg4JD9mxHCwBo0\naECdRk2ZeCScBRei+PToQ+Z+/z0mJiaGTk3kIDn1R/xn165dIzo2jq4jP0ZRFNw9y3L+8B78/f2p\nXr36f27PxMSEkqU9Ob5zM/VadyQq4iGXTh3j8/dH6sTOmf0t/QcOolbztoSH3CUxKpw+ffrkxG0J\nYRCKovDHmrXs2LGD4OBgplerRtWqVQ2dlshhUmzFf2ZhYUFCXBwpyUmYmpmTmpJMXEw0FhYW2W5z\n3ZrVtGrdhu1L5xMZ8ZCJEyZQt25dnbi2bduyf+8e9u3bh51PLXr06IGVldWr3I4QBqcoCq1btzZ0\nGkKPZAcp8Z+pqkrX7j0IuBNMlYYt8Du2nyKOdmzdvAnlFQ5UT0xM5NatWzg5OVGoUKEczFgIIXKH\nbNcoclRqaioLFy7E/9JlypUtw4gRI2SMSQjxxpNiK4QQQuiZ7I0shBBCGIgUWyGEEELPpNgKIYQQ\neibFVgghhNAzKbZCCCGEnsmmFiIDrVbLhg0bCAwMpEqVKjRv3tzQKQkhxGtPlv6IZ1RV5d3efTh3\n8RKeVWpy/sgeBvbry9TPPzN0akII8VqQdbbihc6dO8c7HTrx1Zq9mJpb8CQygg/a1Sf43l3s7e0N\nnZ4QQuR5ss42j7h48SKDBg+hT79+ee5ouMjISAq5FMHUPH2P4wIOjljb2hIVFWXgzIQQ4vUmxTYX\n+fn54dO4CXGWjhi7lKZbj3fZtm2bodN6xtvbm5CgmxzfvYW4mCds/20htjY2FC1a1NCpCSHEa026\nkXPR0GHDiTa1o23/9KPjzhzYyZmtf3D08CHDJvZ/zp49S78BA7kdGEilypX54/fleHh4GDotIYR4\nLWTVjSyzkXNRckoKZgUsn702s7AkNTXVgBnpqlatGpf8Lho6DSGEyFek2Oaivr170alrNwoUdMLC\nypo/Zn/OlAkfGTotIYQQeibdyLlsx44dzPp2NikpKfTr24dBAwe+0hmwImvh4eEcOXIES0tLmjRp\ngqmpqaFTEkLkc7L0R7xRLl++TOMmTSlWpgIxUZHYmJty6MB+LC0tX/xhIYTIJim2wmAePXpEQEAA\nbm5uuTazuXHTZnjU8KFplz6oqsoPE4bTtnE9JkyYkCvXF0K8mWSdrchRWq2W0aNHU7duPYYNG5bl\nRK+9e/dS2suLAcNGUvGtSnw3Z06u5HfvXjCelasD6T/8pd6qxt17wblybSGE+DcptiIDX19f3It7\nYOfgiFfZcty+fTvTuLcqV2HTjt24VKzJroNHKVe+IlqtNkNMSkoK3Xv0YNTXC/l02Vamr9rNl1/N\n5MqVK3q/j9q1a7Fn1a+kpaYS/TiS4zs2UKd2Lb1fVwghMiPFVjwTGhpKsxYtqfF2Rz6c/xtFylSm\navWaOk+tx44d41ZgIF8s30anoe/x+dLNPAgNZceOHRniwsPDQTGibNX0IlfQ2YWS5d8iICBA7/cy\nf95ciH3EkEYVGNu6Jh3fac27776r9+sKIURmZOmPeOaPP/7AqUhROg15D4DiUyoyuFEFzp07R82a\nNZ/FhYaGYmVti7mlFQBmFhZY29kTFhaWoT0nJycUBfxPHaVizfqE37/HzUsXKFOmjN7vpUCBAuzf\nu4fo6GhMTU0xNzfX+zXzo6tXr+Lv74+HhwfVq1c3dDpCvLak2IpnrK2tSYiNRZuWhpFGQ3JSImkp\nKVhbW2eIa968OQlxMWxd9hN1W7bn9IGdREU85J133skQZ2Jiwro1a+jctSt2BQsRHhrClzNm5Eqx\n/ZutrW2uXet1kZSUxOxvv+Gy30UqVq7C+x+Mz3RZ1C+/LOGj98dR3tmGGxFx9Bs8jK9mzjJAxkK8\n/mQ2sngmMTERN/diuJepSJX6TTi0ZQ2a1CQCrl3ViT1y5Aidu3YjOjoaGxtbVq5YTrNmzTJtNzo6\nmlu3buHq6oqzs7O+byNfCgoK4uHDh5QpU+aVvkBotVpaNm1MbNBlqjkZc+phCk5e3mzbuTvDeu/Y\n2FhcnJ2Y5eNKEVtTYpLSGLf/AQePnaBChQo5cUtC5EsyG1m8kLm5OQHXruJkocF38x9UK++V5daN\nDRo04GHoAxLj4wgPe5BloYX0p8sqVapIoc2mTyZNoHqligzt0g6vEh6cPn06221duXKFyxfP82GN\ngjQtacdHNRw5c+oEN27cyBD38OFDrM1MKWKb/sRrY6ahuIMVwcEyo1uI7JBuZJGBg4MDu3fvNnQa\n4qmjR4+yYvEiTveoTkELU7beDKNn507cvHsvW+0lJydjaqzB6On3bmMjMDM2Jjk5OUOcm5sbGJvg\nezeauu62XI9I4GZErDzVCpFNUmyFyMOuXbtGPTd7ClqkP2G2KVmIfjv8SElJwcTE5D+3V6FCBawd\nnPjN/zE1Xcw5fj+RgoVddcbRTU1N2b5zN+3atGLB+SCMNBpWrFydXoSFEP/ZC4utoihmQCeg+P/H\nq6o6TX9pCSEAypYty4zgSMLjk3CyNGPLjTA83Ipkq9BCehHdf/go748ZxepL/lSoVJu93/+AsbHu\nr4KqVatyN+QBjx49wt7ePtMYIcTLeeEEKUVRdgFPgHNA2t/vq6o6+wWfkwlSIsfdvXuXug0acj/4\nHsYmpvTt3YtFixYZOi29mvrpJ3w/Zw6udtZEJqWyZccuqlWrZui0hBCZyPbeyIqiXFJV9T8P1Eix\nFQCqqrJy5UpOnj6NR7FijBgx4pXWvBYvUYqiZSsxYNKXhN4LYsbQbsyeNZOhQ4fmYNZ5T3BwMA8f\nPsTT01NnKZYQIu94ldnIxxVFqaiHnEQeFRwczM6dO4mMjHxuTJt32uJRshQtW7UmKCgo07j3PxjP\n51/O5LHGltXbdtGsRcss91F+Gffvh9Bz3GTMLa0o7lUenw49WbduXbbbe124ubnh7e0thfZfduzY\nQdOmTWnVqhV//fWXodMRIktZFltFUfwVRfED6gF/KYpyXVEUv/97X+RDI0eNokTJkvTs0w+XIkWY\nN2+eTkxycjJNm7fAokgpRn6zBPuSFWnStBkJCQkZ4mJiYli4cCEf/bSS1r2HMvbbJTx4GMHRo0ez\nnZ+pqSn3bl0H0p+a71y/JEuK3lC//PILHdu2wfLeX6RcP06dmtXx9fU1dFpCZOp5Mx7a5FoWItu2\nb9/O6LHjSEhIpEqVSmzZtCnbh6T7+vryy69Lmb7iT4qWKoP/ySN89P5A+vbti52d3bO469evE5+Y\nRIch76EoCkU8RnFm/3b8/f2pUaPGs7iEhARMTU2xtE7fhMHIyIgCDo7ExcVl+34/Gv8BX30wiFrN\n2/LgTiChQTc59OfmbLcnXl+fThjPkGrONC2R/rNpa6Zh1LAhnPe/bODMhNCV5ZOtqqp3VFW9A0z/\n++///17upZi/JCYmsmvXLrZt20ZUVNQrtXXq1Ck6d+lKnbY96DfpK64F3qVe/YbZbu/QoUMULeVF\n0VLpy0Aq1mqAmYWlTveclZUVsTHRJCcmApCSnETskyisrKwyxDk5OVG+QgV+/+ZTggMD2Lv2N+7f\nvkGdOnWyneOnn37K6pUrKahJoVGNKgQF3sLBwSHb7YmsxcbG0tSnEcVdClGjahXu3Llj6JQySE5K\nxtnqn1nZLjamxMXGGDAjIbL2MhOk/lJV1fv/XmsAf1VVy73gczJB6l+io6Np0MiHZFXB3NKK8HtB\nHD1ymOLFi2ervc6dOxOWqDLii/Su3siHD3ivbT2Sk5Ky1d7u3bvp0Kkz32w4gEMhF4KuX+azvu0I\nvneXQoUKPYtTVZXeffpy/sp1Kjdohv/xg5R2L8L6tWsybPkHEBkZyfCRIzl79hxFixZlwY8/ULZs\n2WzlJ3KPVqulRFFX7LSxNC9ZgHP34zgXlsTdB2F5Zr/pFk2bcPOv43xY15XEFC3TjwTTd9hoZn/3\nnaFTE2+w/zwbWVGUicAkwAKI//ttIBlYpKrqxBdcUIrtv0yeMoXj/gEM/mw2iqKw5df5JN2/xaYN\n67PVXvfu3bnzOJ4xX/8EwMOQu3zYuQnJiQkv+GTW3mnXnn379uFavBQhgQGMGDGc72brrvJKS0tj\n6dKl+Ptfoly5sgwaNAiNRpPt64q8xc/PjxrelfmjU2lMNEaoqsrIHbf5cNosxo4da+j0gPS5Az4N\n6nH+r3MoikKbtu1Zsy57/5aEyClZFdssx2xVVf0K+EpRlK9eVFjFy7l9OwjPKjWfPf2VrVqbLScO\nZLu9KVOmUK16Ddb+9A1uJT3Z8PMc6tSp/Uo5btuymf3793Pu3Dl8fHyyPFZNo9EwaNCgV7qWyLuS\nk5NRFCVDT4WxorzSTPKcZmpqiu/J7O8TLURuet6TrXem/8dTqqo+d569PNnq+uGHH1iw9Hfen7sM\nUzNzFk/9gArFXfnpxx+y3aavry+Dhw4jLi6eenVr8/vy5RgZyfkS4tVotVrcnB0pbp5Ki1J2nLsf\ny8G7sdwJCZUxciGeIzvdyAef/tUcqAZcJL0b+S3grKqqz32EkmKrS6vVMnTYcFb8/jsaY2Pq1a/P\n+rVrZO2kyJMiIiJo1aIpdwMDcXB0Yt3mrZQvX97QaeVLqqqyf/9+7t27R7Vq1ahYUbY2eF29yg5S\nG4HPVFX1f/q6AvC5qqqdX/A5KbZZiI2NJSUlBXt7e0OnIoQwMFVVGdC3N4d2/0kpB3PO34/lu+9/\noE/fvoZOTWTDq+wg5fV3oQVQVfUSINNJX4G1tbUUWiHyoLS0NM6cOcPRo0eJj4/PMm7Z0qW4OhXE\n2sKC7p06EBsbm+1r+vr6sn/ndmY1LMToKnZMrefEiOHDSElJyXabIu95mWM8/BRFWQKsePr6XUB2\nkBI8evSIoKAgihUrhqOjo6HTESJL169fZ/lvy1BVlV69+1CunO7KxcTERFq1aEbg1UtYmhqTpDHn\n0LHjFC1aNEPc4cOHmfzBe6x9uxzuthaMP/oXI4cM4reVq7OV2/379ylmb4GZcfqzT9ECZhiRvlSw\nYMGC2WpT5D0v82TbH7gMjH3658rT90Q+lJaWxooVK5g2bRrbtm3LMm7Dhg2ULF2a7r37UrJ0aVau\nWpWLWQrx8vz8/KhTozrXty7h5rZfqFe7JmfPntWJmzd3Lon3rjGnsTNfN3Ckln0Ko4frHnCxb99e\nenk68lYhW+zMTfisZnH27d2X7fyqVavG5dAYrkckoKoqfwZEUbiws0xEy2de+GSrqmoiMOfpH5GP\nqapKtx49uXIjEC/vWvy6fDzHT5zkqy9nZIh7/PgxAwcP5qMf/sCjbEXu3bzGiKFdaezjQ+HChQ2U\nvXgd+Pv7c+jQIRwcHOjcuTNmZmZ6v+bMGV/QvqQF7cqkFy8Hi8d8Oe0zNm79M0NcwLUrVCqoQWOU\nPtxWtbAFv94I0GnP0dGJg9HJqKqKoihcj4zFwSH7w0IlSpRg2YqV9OvTi9i4eEqX8GD7rj91NogR\nr7fnHUSw9un/+j89gCDDn9xLUeSWs2fPcvL0GSYsXE3XUR8zcdE65s+fr3P6T1BQEAULueBRNn3G\nZNFSZXAp6kFgYKAh0hYGFh8fT+OGDXAtWACvEsU4ceJEpnFbt26lUb3a7Fowg28njaNxg3okZXO3\ns/8iJiYae/N/niscLIyJiY7WiatctTonQlNIStWiVVUO3YunchXdFZADBw7kNpZ023WV8UdvMPRA\nALPn//hKObZt25ZHj58Q9SSay9dv4Onp+UrtibzneU+2f28TIwcSvCGioqIo6OyCiWn604aNnT2W\n1tZER0dn6NIqVqwYEWH3uRNwhWKe5QgJvMGDe7fx8PAwVOrCgKq+VQHjmDAGVXQgICKOJg3rc/n6\nDZ2fh9HDhzK+ekHKF7JEVVW+OBHEqlWr6Nevn17z69qjF5PeH0UhaxM0Cqy6FsukGb114kaMGMFJ\n36MM+fNPzIw1FC3uwdIfF+jEWVtbc/zMOdasWUN0dDRjmzXLkSVRiqJgaWn5yu2IvOl5O0g9ePrX\npsARVVVv5E5KwlCqVq1K6J1AjmxbR6U6jTi4aSUFHRx0Jog4ODiw+OefGTKsO85F3AkLucv38+bh\n4uJioMyFocTGxhIQeJuVnTyxMDGimqs11yISmTt3rs7xjI8eR1HMLv1nSVEU3KyNiIiI0HuOvXr3\n5kn0E+Z/NxtV1TJ2wicMGjRYJ06j0fDH6rWEhISQlJREsWLFstyC1MrKigEDBug7dZGPvMxsZHfg\nZ0VRigPngCPAUVVVL+gxL2EADg4O7N61kwGDBrN63nQqVa7M7p07Mv2F07VrV3x8fAgMDKR48eJy\npmw23blzh6ioKLy8vDA3Nzd0Otmm8s+aei1qpruYNWpQnz8uX6BfRXvuRSfjGxzP5IbZP6Xqvxg5\nchQjR456qdgiRYroORvxJnrhphbPAhXFAhgMjAeKqKr63F3nZVMLIbKmqipjR45g1YoVONlYkGhk\nwq79B1/LsbqKZb1ICb9H+7L2XI9IYE9gDFdv3MLd3T1DXGRkJO9268L+Q0ews7Vh7vwf6Nmzp4Gy\nFkI/XmUHqSlAXcAaOA8cI/3J9sELPifFVogsbNy4kc/GDGNH24oUMDPh54t32Rxjge+Zc4ZO7T9L\nTk6mY/t2+P11Bjv7gqxYs4633nory/i/Z/EKkR/951N//k9HIBX4EzgMnFBVVf9TCIXIx65cuUKz\nIrYUMEs//LxT6cJ8ufL1PMHG1NSU7Tt2vnS8FFrxJnrhphZPD45vCpwGmgH+iqIc03diQuRnXl5e\nHLgfQ2xy+pF1W289pIxnaQNnJYTQlxcW26cHD7wL9AW6ASFA9g9hFULQuXNnajZvhffK0zTceJE5\nVyL4dcVKQ6f1xlu2bBmuhZxwLmjP6NGjDZ2OyEdeZsx2O+kzkI8BZ1RVfandsWXMVogXCwgIICoq\ninLlyslRi9kQGxvLsmXLsLa2pnfv3lku1XkZy5YtY+igAbTzcsDGTMPqSxG06dCZNWvW5GDGIr/L\n9gSpV7igFNvXUGBgICNGjuZWYCDe3pX5cf58OWRA5Ennzp2jQe2amGogVaui0ZgQGHw/23sKuxZy\nop6TSq9KTgCcCo5h/ukwohPl9B3x8l7liD3xhoiJiaFhIx8KelVh0BffE6NY8nbrNmi1WkOnJvI4\nVVV5/PgxqampuXbNt5v40KCYDcval2JZ+9IUt9VQp2aNbLeXlpaKjdk/T8Y2ZhqQBwaRQ6TYGoCq\nqqSlpT03Zs2aNXiUKk0R92L07dsvVwre6dOnKeBUmDZ9h+Feuiy9xn/O7dtBBAcH6/3a4vV148YN\nypYuSVFXF+wL2LDi999z5bpJifE0LG6LoiiYaBQaFrflUdhzVyQ+V9eevVh9KYJTwTFcCY/nh1Oh\nFC1RKgczFm8yKba5SFVVJk2egpWVNZZWVvTu2y/Tjdh37NhBn379adCxN13HTGHn/gN07NRJ7/lZ\nWloS+ySCFT3zAAAgAElEQVQK7dMvAonxcSQlJWBhYaH3a4vXV/s2rWhgl8DKDh581ciFsaOGc/ny\nZb1f19TMnBN3Y9K/vGpVjt+Lwc6xULbbmz9/Pm06dGb+6VC+PBKCvVsJ/C7p/z7EmyHLMVtFUbYB\nWfahqKra9rkN55Mx26NHj7Jy1WoszM0ZMWI4pUpl/5vur7/+ypffzuH9ucswt7RiweTRNKrpzTez\nZmaIq1e/PvYlKtBz3GQAAq9c5OsRvYh58viV7uVF0tLSaNaiJbGpUKZ6Pc7s3Ubd6t4sWbxIr9cV\nr6+4uDgc7O1Y26nks/Wz8/6Kou+kmfTp00ev1z5y5Agtm/pgY6ohJU1LKkYE3L4rxzwa2M6dO9my\ncQMF7O0ZM3bcG7f9ZXbGbL8FZj/nT773559/0r5TZ6LN7Ln9JJladeoQEKB7vuXL2n/wII279MXe\nyRkLK2ta9xvBgYMHdeIURcmw12z6lxb9f3HRaDTs/HM73d5piWVsGO+PHMqinxfq/bri9WVpaYmF\nuTk3IxMBSErVEvg4ETc3N71fu0GDBgSHhjPiw8lMnj6TiCexUmj1JCYmht49u1PE2YlK5cpwMJPf\nWwDLli6lf89upJ7bxvU/f6O6d2UePMh+135+IrORn6N23XrU7tSP6j4tAVj30zcUtVSYN29uttr7\nYPyHXA4Op+/H0wHYvXop4VdO8+e2bRnidu/eTdv2Heg6fDwOzi78MXcGjerWYv369a92Q0LowaZN\nmxjYtzcVXWwIepyAT4vWLF2+QnaKykc6tXuHx1dO0qOsLUFRSSy4+IQTp8/i5eWVIa5ksaIM9dJQ\nxjF96OmncxE06v8BEyZMMETaBpHt7RoVRSkNfAWUA54dS6KqaokczTAPSkxMxKaA/bPX1gXsSYgJ\nzXZ7Ez7+iFq16zDnvQFYWFlx5exxDmfyDbFFixb88ftyJkyaTFJyMu1bt2TRIunKFXlThw4dqFjx\nAmfPnsXV1ZX69etLoc2GtLQ0YmJiKFCgQJ7676eqKtt37mJ5uxJYmBjhbG3KX+Gp7N+/X6fYJiUl\nYWNq8+y1tQkkJMTndsp50svsjbwU+AyYA/gA/XlDJlb1ercnC2Z/Tq8PpxEX/YQdyxeybs2qbLfn\n5OTEX+fOsm3bNpKTk2m5+EdcXV0zje3cuTOdO3fO9rWEyE2lSpV6pfkM+ZVWq2X//v2Eh4dTq1Yt\nSpTI/Bll/fr1DOzfj9TUFFwLF2brjl2ULVs2l7PNnKIoWFlYEB6fgnsBM1RV5VGiFhsbG53YHu/2\nYuHa3+hdzobwuBQO3E3gk476n9z5OniZHaTOqapaVVEUf1VVK/7/ey/43GvfjayqKt/Ons2KP1Zi\nZmbGpAkf0759e0OnJYR4DWi1Wjq3b4vfmeO42ZrhFxrLqrXradGiRYa4mzdvUsO7Mp/WcaKEgzl7\nbkWxK8yEG7fv5Jkn3IULFzB10sf4uJlxLw6izR05ceYclpaWGeJSU1OZ+tknbN6wHltbW6bP/BYf\nHx8DZW0Yr3LE3nGgHrCe9D2RQ4CvVVX1esHnXvtiK4TIH1JTU7lw4QKqqlK5cmVMTEz0fs2NGzcy\nefRgptdzwkSj4B8Wx4IrSQQ/CMsQt3btWn6YMpbx1eyevddrcxC37wVTsGBBvef5svbt28eB/fso\n5FyYQYMGyfaiWXiVI/bGApbAGOALoDHphxIIIYRBbdiwgblz5mBiasqMGTOoXbu2Tkx0dDTNGzci\n7F4QigL2zkXYd+gI9vb2ug3moJCQEEramWCiSf+9W8bRgtCHITrn+bq5uXE7MoGEFFssTIwIikpE\nVaBAgQJ6ze+/atq0KU2bNjV0Gq+tlzli74yqqrFANDBGVdWOqqqe1H9qIic9efKERo0a4enpydix\nYw2djnhDJScns3379ucuoUtLS2Pqp59Suawndat5s3v37kzjFixYwLvdu1Io8irm987j06Aehw4d\n0omb+uknWEcHM6dxIeb4FMI5OYzJEz7OqVvKUq1atTgVEkdIdDKqqrLpehTVvSvrdA3Xrl2bVu07\nMv5QGN+di2Kqbzg/L1qCsfHLPAuJ18XLdCNXI32S1N+j4U+AAaqqnnvB56QbORf4+/szafIUIh49\n4u0WLZg0aaLOP9LY2FhcirjhXNQDz8rVOfbnekqXKsm5M2cMlLV4E61bt46+PbujKJCcpqVkCQ+u\n3QjUiftk0kT2rlzGV7WL8SA2iXFHb/Hn3v1Ur149Q5yroz1dS1vQ2CP9CXDFxXACFGcu/GvXp1bN\nmlAp8Tq1i6b/Cjt7P5ajqW4cPHZCT3f6j8WLFzFuzBhUVYtn6VJs/XMX7u7uOnGqqnLs2DGCg4Op\nUqUKZcqU0XtuQj9epRv5V2CEqqpHnzZUj/Ti+1bOpvhmiI+PZ+/evaSkpODj4/NKYzJ37tyhUePG\ntBs4lkoepVn3yzwiHj3i+3+tAx4yZAi2BZ2YumwzRhoNzbv1ZXyHRiQkyFaMIvcM6NWTzuUd6FzO\nkciEVN7fdZshQ4boLGtbs/IPfqtfgvKO6cXxUkQs69et1Sm2KSnJOFr+MyO2kJUJfhGxOtet5O2N\n7wY/ahRJH2P0vZ9EpRbPnd+ZYwYPHsKAAQOJj4/PdPbu3xRFoX79+rmSkzCMlym2aX8XWgBVVY8p\nipJ7R3vkI1FRUdSt3wCNuRXmVtaMGTuOo0cOU7JkyWy1t3XrVrwbNKdZt34AuHqUYmLXpjrFNiws\njMLuHhg9PevTybUoKiqhoaF4eHi80j2JvGXTpk1cvHiRFi1aZDp+aUgJKam0Kp0+TupgYUxdd1uO\nHDmiE2duZkZkwj/H2j1KSsPlX7NeAWrUqc+SE4f4oI4riala/vCPYMjo93TiPvlsKm1PnWLY7vMo\nioJnmXJM/+rrHLyz59NoNM8ttOLN8DLF9rCiKD8Dq0jfM7AbcEhRFG8AVVX/0mN++crMWbNwKV2e\nAVNmoSgK239byPvjP2TLpo3Zas/Y2JjkpMRnr5MTE9AY6x6ePWTIEPr068f5YwcoXbEKm3+Zj7mF\nlRTa18STJ0+YN3cO94ODadysOV26dMl0SYhPg3qcOXWCYgXMmTnjC0aOGce3s/POzqqmxkZcDI2n\ndlEbUtK0+IXFUbGe7trciZ9PY/DIYYyq6MKD+BS2B8dwatBgnbgt27bTrLEPE/efxMhIoX3nbsyc\nNUsnztLSkr0HD3Pz5k1UVaVUqVIYGb0RWwWIPORlxmwz3wQznaqqauMsPidjtv/ybu/eWJeoRKP2\n3QG4fuEM2376inNnTmervfDwcCpX8aZ6s7a4eJRi1x+LGNinF1MmT9aJHTFiBEt/W05qSjIWltZs\n3byRRo0avcrtiFc0Y8YMvvnyC1JSUylZoiSHfE/oHHweFxdHDe/KuKhReNgYse9eEgNGvccnn36W\nIW7t2rUM7N2TBW08sDUzJjAykY/23iHicRS2tra5eVtZmjhxIt99M5MS9uY8jEtBa2RCWGQUpqam\nOrH79u1j8/p1WNnYMnL06EzHOYXIi7K9zvYVLijF9l9+WrCA7xcuYfz85ZiamfPzp+Pw9irB/O/n\nZbvN4OBgvvp6JuEREbRq2YK+ffvmmYXw+U1YWBhr164lLS2N9u3bU7x48Wy3tXbtWvq+24MP67ri\namPKknNhJFgX5nLAzQxxa9as4ZsJo/mklgOKohARn8LIXfeIi0/I8HQ2adIktv/6PdMbF332Xo/1\nARzyPakz1mlIR44c4eeff8bd3Z3p06ej0ej2xIjnS0pK4uHDhxQuXDhX1guL/+ZVNrVwBr4EXFVV\nfVtRlHJAbVVVf3nB56TY/otWq2XM2HEsXrwIBYUWb7/Nqj9W6OzCInLXqJEj2bJ6JYqRQr/ho5g2\nbZpOzN27d6lTvRoNnK0w0yjsuPOYfYePULFixWxds2XLlpgFnWagtzMAj+JTGLY9kKRUbYa4pUuX\n8vvXExlXNX2sMylVy7ubAolPSMgw6/zAgQO0btGMb5oXw72AGSeDY5hz4gGPY+IwNzdH5A/btm2j\nz7s9MTYCRWPM+k1baNCggaHTEv/nVYrtTtJnH09WVbWSoijGwPm/t258zuek2GYhMTGRtLQ0rKys\nDJ3KaykkJARfX1/s7Oxo0qTJKz0djR41irVLFzOvSXmS0rSM2X+ZCZ9NY+LEiRniRgwZjJX/IT6p\nlT6ZbdHFexy3LMbG7Tuydd0uXboQfHIXE+unH0V3PSKBzw8HE5eUce5hcHAwlStWoIeXJSXtzdh0\nIxbH8rXYsGWbTpsjhw9n8aKfMTM2IkWrsmDREvr375+t/ETeExoaSlnPUkyq5YiXowV/PYjlx4sx\nBN0Lkd8leUh2zrP9m6OqqmsBLYCqqqlAWg7n90YxNzeXfxzZdPz4capUKMeKaR/z4cDetG7elJSU\nlBd/MAtbVv/BnMblaFWyEB08C/NFPU9++/knnbjIiHBKF/hnmVQpOwsiIyKyfd1vv/0W//AkZh+/\nz2r/cL44HEzvfgN14tzc3Nh74CD+mmIsDFAp69OW5StXZ9rmjwsWcD/sIbsOHCbySYwU2nzm2rVr\nuDtY4fX0+DpvF2ssTYy4c+eOgTMTL+Nlim2coigFeXp6uaIotUjf2EKIXDdsQD++q1eCFc28ONSx\nMvFB11mxYkW221OMjEj8v67bxFQtipHuk3KLNm2Z5/eAW4/jCI5JYNb5EFq80zbb1y1WrBj+V69j\nWaY2t61KMG3mtyz8+edMY6tUqcIh3xNcuxXEwsW/PPeLmqOjI3Xr1pWhiXzI3d2de5FxPIpP/3J5\nPyaZx7GJuLi4GDgz8TJeZunP+8BWoKSiKL6AEyBnv+VTqamp/Pbbb9y6dQtvb286deqUpyZchTwI\npXbD9A0JNEYK1R0tCQ4OznZ7g8e8x7gvphKVlEJympYvTtzk+wW6Ra9f//48uH+flt/NJk2bRr/+\nA/howsRMWnx5Hh4e7Ny165XayK9SU1O5desW1tbWFClSxNDp5AklSpRgwqTJfPT1l5RytCbgYQzf\nzZun9z2eRc54qdnIT8dpvQAFuK6q6gv77WTMNvc8fPiQ+/fvU6FChSz3U71y5QrvvT+ekPsh1Ktb\nl9nffqPzhKTVamnfsRNB98PwqlqHcwd30KFNa2Z/+01u3MZLad2sCaWj7zK1dglCYhNps9WfRSvX\nvtIG6bNnz2bJj/MxMjJi/ORPpPvVwEJCQni7aWNiHkUQnZBIpy5dWLjkV1kb+9Tly5e5desWZcuW\npXTp0oZOR/zLf54gpShKdeCeqqqhT1/3AToBd4DPVVWNfMEFpdjmgtZt2rB79x5MzMww1mjYteNP\n6tatmyEmLCyMtypVpnW/kZSsWIVdKxZjb6awdfOmDHGnT5+mU7cefLlmL8YmpsQ+ecx779TlTtDt\nPHPUV1hYGB1at8Lv8mW0qsq0adMY/5H+N5V/U6mqSlxcHFZWVln2cCQlJTH10yn4HjqEi5sbX34z\nO8tD0l9G25bNKRMdxOQaHsSlpNFu2yVGfTGTvn3lsDGR92VngtTPQPLTDzcAvgaWkz5eu+g5nxO5\n5LvvvsP3xCnmbffllyNXaNV7KO06dNSJO3DgACUrVqFZt36UKFeJIVO/Y/eunSQkJGSIi46Oxt7J\nGWOT9E0GrGztsLSyIjZWd79ZQ3F2dub42XOEhIbxJCZWCm02XbhwgTVr1nDx4sUsY/z8/ChZzB1H\nB3ucHOzZs2dPpnGD+vbhr82r+MANPB9epWGd2jx69Cjbufn7+9PD0xlFUbA2NaaNuy0Xz8tGdeL1\n9rxiq/m/p9duwCJVVTeoqvoJoLvHWj4VGxvL/v378fX1JTU1b20JfeDAAWo2a4NDIRcURaF5t35E\nPdbtcDA3Nycu+gl/9zTEx8YA6HQ5V6tWjfDgOxzYuJKIByFsXDibQoUK4ebmpv+b+Y8KFCggC/r/\nJSYmhimTJtK7WxfmzZ1LWlrmiwZmfjmDVo0bsmr6RFr6NGDO7G91YlJSUmjdsjnt3LSs6VSSD6oW\noHuXTty/fz9DXFJSEms3bGRZs7I0ci/I+GrFecvBgr1792b7PkqXLs3OoPSZ3slpWg48iMWrbLls\ntydy1+PHj9m7dy+nTp1Cq9W++ANviOdNkNIoimL8dKlPE2DIS37utbBv3z62bttGAVtbRo4cSeHC\nhXVi7ty5Q8NGPlg7OBEfG4OrsxN7d+/KMyfleHp6sm7rDlKSkzAxNePK2eOYW+rOVG3ZsiWffj6V\nRZ+/T4nylTm6dQ3jxr2nU6zs7OzYu2c3g4cMY9uSuVSuUoXdO3fILj8Gtn//fkYPHczDiAjq16vH\nkt9+1+nWT0pKokmDepRMe0L9wjas/N4Xv/N/8ctvyzPE3bt3j1lffcWJHtUpbGVGcEwCdT/7jB7v\n9srwbyA4OJjUxAR8irsCUL6QJSULJuHn54erq+uzOCMjI3h6ZJ6lSfrPSVJa2iv9zPy05FeaN2rI\npjt+RMQlUqVmbQYO1F0WJfKeS5cu0dSnIS5WxjyKS6JKjdps2LJNzubl+WO2k4FWQATgDnirqqoq\nilIK+E1V1bqZfvCfz+fZMdsVK1bwwUcf06RrfyJDg7nke4BzZ8/g7OycIa5dh46Yu5ak/aCxaLVa\nfpgwnDaN6jA5k72HDSE1NZXSXmWIjounUBF3Aq/48f3cOQwdOlQnNjo6mtnffUdwyH0a1KtLnz59\n8tQsY5G5mzdvUruaNwsblaZSIVtmnbvLHVs3dh88nCFu3759fDywNwc6vIWiKMQmp+L16zHu3n+Q\nYbbqyZMnGdm9Iwc7/HNCZp31F/h9606qVKny7L2YmBhcnAvxXVNXClubEpecxrj9D9h72Je33sp4\nuuaYEcM5vWMzg8s6cTY8jgOP0jjr5/9KJ93ExsZy4cIFrK2tqVSpkvysZkNcXBx3796lSJEiubY/\ndu3q3lQ1CqV5yQKkpKl8cSKC0Z/NfKO+LP3n82xVVZ2hKMp+wAXY83+V0wgYrZ80c8fUL6YzfMYP\neFVO3zN2yRcf8ttvv/HRRx9liLt16xY9Oqf/kBgZGVGuRj1u3Lqp056hGBsbc+tGAIsXLyY4OJgu\nSxfp/CL8m62tLVM//zx3E3xNqKpKZGQkGo0GOzu7LOPWrFnDvFlfkZaaRv9hIxg6bJjei8Dhw4dp\nVtyJZh5OAHxVtyQuCw6QkpKSoWciJSUFK1PjZ/mYGxthrDHS2fDDy8uLe0/iOXj3ET7uBdkXFE54\nXBKlSmUcGbKxsWHWN98wefJE3ipsTUBEAj1798v052vuDz/yw3wvdh/Yj2vtYhz79LNXPlLO2tqa\nevXqvVIbb7K9e/fSvUsnrE01RMUns2jJL3Tr3l3v1w28HcSguumHaZhoFMrZKdy6mXd+ZxrSc5/t\nVVU9mcl7AfpLJ3fEx8dj6+D47LWtvSNxcXE6cd7e3hzevJriZSqSnJTIqd1bGNavV26m+kJGRkaZ\nPsmKlxMfH0/3Th05fOQwWq1Khw7t+XX5Cp1urz///JMPRgxlXv2SmBmb8P7nUzAxMWHgoEF6zc/O\nzo7b0QloVRUjReFOdALmpqY6+dWtW5eQRJVZZ25T39WOZdfCqFGjBk5OThni7O3tWbtpM907dyQx\n4TIWlhas37I10+I4YuQo6tarz8WLFylRokSWxc/IyIgxY8cxZuy4nLtxkW1xcXF079KJ8dXsKV/I\nkqDHiQwbMoh69evrfc1y5UqV2Bt0lR7l7IhN1nL6YSodq1bV6zVfF2/kqT/vvf8BB46fpvu4KYTf\nv8eyLyeyb89uvL29M8Q9fvyY1u+05XpAAMlJiXTo0JGlvyyRMcx8ZPx7Y7m1ZxOLm5QhVavSc9cV\nWgwayccTJmWIe7dLJ2pHXadPhfTJYrsCH7LokTn7jh3Xa37Jyck0bVgf00chvGVvzrqbEUyZ8TVD\nhw3Tib179y4fjh3DnduBVK1Zi6+/nZ3lE6ZWq+Xx48fY29vL+tV85urVq7RsWIf5Tf8Zg//0eCRz\nlq7Gx8dHr9cOCQmhZdPGhIXeJz4xhaHDhvHtd3PeqGGA/9yNnJ/Nmvk1Uz75lGVT38fW1pY1q1bq\nFFpIfwo4duQwISEhmJmZUahQIQNkK/TpzPHjfFDGGVONEaYa6FXakZ3HdQuouYUFUWH/zEZ/kpSK\nWS6cpmNqasreQ0f4/fffCQ0NZUX9+jRs2DDTWHd3d9Zs2vxS7RoZGeWZtdMiZxUpUoSo+GSCHidS\n3N6csNhk7kbG4eHhkSvXvnDpCsHBwVhbW8vP2P95I59shfhbnx7dcA46zye1SqCqKmMPB1CwYVtm\nz814xvDFixdp2rA+Q8s5Y6YxYr7/A1Zv3Ezjxo0NlLl4Waqq8vWXX/LT93NRVRg8fDiffj41Xz9t\nrV61iuFDB1PMwYq7kXFMm/Elo0aPMXRabwQ5PF6ITISEhNCoTm0Km2hJTtOSaGHLId8Tme436+/v\nz6IFP5GWmkrv/gOoXbu2ATIW/9XiRYuYN3UyS5t6oVEUBu4PYMD4SYweO9bQqelVSEgIAQEBeHh4\nULx4cUOn88aQYitEFmJiYjh69CgajYaGDRvKYevZkJqamn4EnLt7lstMUlJS+GzKJPbs+BOHggWZ\n8c13VK9eXe+5tX+7Be2NwujomT6GuTPwIUufWLPr0BG9X1u8eV7lPFsh8jUbGxtatWpFixYtpND+\ny9mzZ1myZAkHDhwgqy/P27dvp4CVBdWrVKKgvR39+2W+h/G4USM5tWElX5ezpb1pFK2aNeXWrVv6\nTB8AO4eC3I7+Z2vSwCcJ2DnIWKLIXfJkK8QbJjY2li+mfs4V/4tUrFyFKZ9+nun5t9/Pm8v0zz6h\nios11yISeKdzN37I5PhBa3NT+lR0oGVpe4Kjk/hwzx2Wr1pLp06dMsTZ21hzukd1nK3MAHjvcAAV\n+4xh3Dj9LhkKCAigQe1atHa3Q2MEmwMjOXjMl/Lly+v1uuLNJE+22bR//34GDBrMiJGjuHr1qqHT\nESJLP/zwA7bmpmiMFAoVsObIEd1u0rS0NN5u1oQzm5dTLvYKvuuX0qZlc509bKOjo5k4YQJfNnRm\nZBU7ZjYqxPrVq7hw4UKGuIcPH5KYnEKLUumbgbjZmlGpsBW7Mjmn19TEhCdJ/2yy8SRFi6mpaU7c\n+nN5enpy+vwFvHoMo2TXoZz667wUWpHrpNg+x5YtW+jesxfagu6Ea82p16CBFFyRJ124cIGP3hvH\nkhYVCBnRlJFvudK2ZXOdwzMuXbpE0M3rjK7qQF13W8ZVK8jVS34EBGTcq+bRo0fYWphRyCp9lypL\nEw1u9paEhoZmiHN0dMTISCHgUSIACSlabjxKyLSYfTxpMj13X2XJxbt8fPQG5x4n061bt5z8z5Al\nd3d3PvroIz7++ONcWQIjxL9JsX2Or2bOou/EL2nRvT8dh4yjUYdeLFi40NBpCQPRarWcOXMGX1/f\n555msnHjRiqV86J0cXemTJqQ5ek7OWndunV4O9vSskQhLE00jKvqQVpaGn5+fhni1Kc7URk97eRS\nFNAYGencj5ubG6aWVuwLTD8t6mJoHIGP4qhUqVKGOCMjIyZMmsInB+/xyYG7DN8eSGH3EowZo7vM\n5P3x45n+/QKuFK2KrU8HTp47L+swxRvjjdzU4mUlJydjaW397LWFjQ1JceEGzEjoQ3BwMFu3bkWj\n0dCxY0edLQ4hfZzzLa/ShIeHozFSsLC04tylKxlOwAE4cuQIQ/v3ZZS3PXbmGpYsX4yiKHwx4yu9\n3oO7uzu3n8STmJqGubGGkNhEklLTdI5HrFChAoWKuPPzhVBqFDblxP0kinqUxMvLK0OciYkJO3bv\npVO7d1hw9gaO9vas37QFFxcXnWtPmzaN5s2bs23bNjw9Penfv3+Wu1J17dqVrl275tyNC/GakAlS\nzzHv+++Z++MC3n3/c+JinrB81idsWr+OBg0aGDo1kUOuXLlCo3p1qVzIlBQtBESrnDxzjqJFi2aI\na9G0CUkBF1jXrirGRgpDd/tzy8Sec/6XM8SNHT2KqKNr6VQu/YntVmQiP99QuXYrSK/3odVqKVOi\nOEpMJPWK2LP15kPqNmnG5m3bdWKjoqKY+NGHXLnkR8VKlfly5jfPPRUmKSkJMzMzfaYvRL4h2zVm\nw5jRozFSFH5fOhczMzN+X7ZUCm0+M/njD3nHw4x2XumbWKzwf8T0qZ/z85JfMsQFXr/KhApFMDNO\nf2LrXaEIQ/df12nP2saGoOR/vmQ+TkjFykr/x5sZGRlx5WYgkydPJiAggM+GNGHUqFGZxtrZ2bFg\n0eKXblsKrRCvTp5sxRutXs1qNLd8SGUXKwAOBT3hjqM3G7ZmfCKsV6smzo/v8svbb6EAE49c53C0\nEZdvBmaICw4OplqVytRwVLA1hZ1BCfz2x2ratGmTW7ckhDAgebIVIhMtWrVh/ZL5FLMzI0Wrsv12\nAh/01S2M6zdv4S2v0lRedhQzjRFhCakcO31WJ87NzY2z5y+weNEi4uJi2dapM3Xq1MmNWxFC5GHy\nZCveaKmpqbw3djTLli7DSGPE2LHjmPrF9Ew3qY+Pj+f3338nJSWFXr16PfegeSHEm0n2RhZC6M2p\nU6c4efIkLi4udOrUSc58Fm8sKbZCiP8kPj6eTyZP4tzpk5Ty9OKrWd9muixqyZLFTBz/PrWKWBH4\nJBWPCt5s+XOnFFzxRpJiK4R4aaqq8nbzpiTc9sPHzZwLEclcT7LmL79LWFhYPIvTarXYWlsxq7EL\nbrZmpGlVJhwJZ+4vK3j77bcNeAdCGIbsjSzEayoyMpLJEycwsE8vli9fnuXpOznpwYMHnDp5grHV\nCuLtak3/ivYYJUZz8uTJDHGJiYmkpKbiapO+x7HGSKGIrSkRERF6z1HkPbGxsXTt1AErC3OcHR1Y\n+uuvhk4pz5BiK0QOS01N5euvv2bcuHE6ewn/V7GxsdSrUZ2QnWuoEHKebyZ+wGdTpuRQpllL/3ZO\nhn5rZasAABXSSURBVMKeplV1Jo5ZWlpSqUJ5Vl9+TFKqFv+wOC7cj6V27dp6z1HkPSOGDCbc7xiL\nW7szsXoBJnwwlkOHDhk6rTxBupGFyEGxsbEULmgPqhYLEyNik9P4+Zdl9OnTJ1vtrVq1il8//4gN\nrdI39g+NS6Ly8uPExidkuSViTlBVlY5t2xB66TQNipjiF5HCA2NHTp07r7PJxf379+neuSMnz5zD\n2dGBxUt/o2XLlnrLTeRdLk4FmVbbDmfr9J6OVf4RlGw3hOnTZxg4s9wj3chCZGHpr79Sv3pVGtWq\nwcaNG1+prfLly1HExphlHUqxpG1J2ng6MHLwgGy3l5SURAGzf5bD25oak6bVPvcghJygKAqr12+k\nRe9hXLEuR5nm3Th07Himu0m5urpy5H/t3Wl4FFWixvH/SSchJIQsgCAIhD0kQRNAEMOmwogIERCR\nGRF1vCOCC26ICiMgzqg412XGDYHxARRGUUG4iCiIFxhkC4IEBTEsCQTCEkII2ejuuh/oizABhZCi\nkvT7+6JdqT79to+dN+dUddWq1RSXlJC5L1tF68eio6PIzCsBTv7BtrcQatcufVKdP9LMVvzajOnT\nmTDqUV5ObkqJx8tjK9OZ+v5sevfuXabxIkJDuC02gpTYaAAyjhYz+qvdHC8p251/srKyaNsmgScT\nLyfpsnBe3ZhFjYQOzJrzcZnGE7HTl19+yeCBA0huWIODhR6OV4tm1dr1hIeHOx3tktHZyCJn0bNL\nMvdGFdK72WUAzEjbw6qIlsz6uGwz3EYNGxJWdJjnr29IkCuAWd8fZOH2XI4Vu3/7yeeQlpbGk488\nzP59WXS7vgd/nfTyGWcEi1QkW7Zs4auvviI8PJzBgwcTFhbmdKRLSpdrFEfs3r2bO2+/jfUbN9G4\nQX2mznif5ORkp2OdEhQcTP6JY6ceHzvhIfgiLrz//ebNXFG3NvfM+5nqQS7yit288c75X/T/bBIS\nEvh8ydcXNYbIpRIfH098fLzTMSoczWzFNl6vl6viYulfJ4BhV17B8swcRq7YwaYffjzrfVGdsGTJ\nEu4YOIBHr6pPsdfijc37WLTka9q3b1/mMd1uN8899xwHDx5k9OjRxMTElF9gEanQtIwsl9z+/ftp\n07IFP//x2lNfGbnti6088NLfSUlJcTjdL1auXMn0qVNwBQYy7IEHSUpKcjqSiFRSWkaWS65mzZoU\nlJwgK7+YBuEhFLu97MrNp1atWk5HO0Pnzp3p3Lmz0zFEpApT2YptQkNDmfDcBG56+UVujqnF2gP5\nJHbqolvOiYjf0TKy2G7ZsmWkpqbSqFEjBg4caOvFGPxZbm4uzz83gZ0/b6fDtck8/sQoAgP197TI\npaRjtiJVWFFRER3bJVHPfZiEWoEs21tC3LU38P7sD52OJuJXdMxW5BLZt28fU6dOpeB4Pv0H3EqH\nDh1sf80VK1bgzjvEiM61MMZwzRVe7p43j5ycHKKjo21/fRH5dVrPEzlPM2fMoH2bOBJbt+Tvr712\n1rvvZGVl0SEpkd1z34P//ZS+v+vBokWLbM/m8XgICjCnzvp2BRhcJgCPp2xXrhKR8qVlZKmydu7c\nyeLFiwkNDWXAgAHUqFGjzGN99tlnPHTvXbzZvQUhrgAeXp7OyGcncv/w4WfsN3bMMxz64l9M6toS\ngEU7DvDKbjdrNn5/Ue/lt+Tn53NlXGs6RJ4gvnYwSzMLCW1yFQsXf1XqTj0iYh/diED8ytq1a+mQ\nlMjKdyYx669juaZtIrm5uWUe78OZ0xnd9gq6NaxFx/pRTOzYmDnvzyi1X35eHvVDg049rl8jhPz8\n/DK/7vmqUaMGK75djat1FxYdq01ir9v5eN58Fa1IBaFjtlIlPTnyIf5yTWMGt64PwIivt/L311/n\n2XHjyjReaFgNDmWfOPX4UGEJ1UMjS+3X79aB/OGDmbStW5PLQqvx9Lc76X/7XWV7ExeoQYMGfPDh\nnEvyWiJyYVS2UiUdOHCA+Ga/XBIyPjKErOyy38j9kVFPcl3nZI6f8FDNZXgnLZtPFrxZar/u3bvz\n6tvv8syzYyksLGTg4CGMn/h8mV9Xfl1RUREbN24kKCiIxMREXC6X05FEzkrHbKVKemDYfexd8QVv\nXteSQ4UlDFy4hUnvTKV///5lHnPr1q1MfXcyHo+HIUPvol27duWYWC5UdnY213VJxp2fS5HbTbNW\n8Xz+5RLdEUkcpe/Zil8pKCjgT3cP5ZN586kWHMSfxz3LE6NGOx1LytEdt99G0Q/LGZoQhdeCV9Yf\npseQEYwbP8HpaOLH9D1b8SuhoaF88NHHzPR6McboRKEymjx5MsuWLSMhIYFnnnmmQl39a9vWHxlY\nLwRjDC4DbWsHsnVLmtOxRM6q4nxyRGwQEBCgoi2jW/r05smRD3Bo3Re88dJzXBnXCq/X63SsU65M\nTGLFniK8lsUJj5dv958gUUv7UkFpGVlESsnIyKBZkxgm921K7dAgitxe7pufzpvTpnPHHXc4HQ+A\nnJwcbup5Axk7d3DC46Fzl6589Ok8goODnY4mfkzLyCJy3nbt2kX1wABq+74zHBIYQL3wYHbu3Olw\nsl9ER0ezau16fv75Z4KDg4mJidEqhlRYWkYWOU+WZZGens5PP/1UoZZT7dChQwc8GD7/6Qger8X6\nrHx25RbTr18/p6OdweVy0apVK5o0aaKilQpNM1vxe7t27WLu3Lm4XC4GDRpEvXr1Su1TXFzMgJQ+\nrF2zmsCAAJo0b8Gir5YSERHhQGL7hYSE8PG8+fx+YH/eTc2merCLv736OgkJCU5HE6mUdMxWqiTL\nsnj7rTf5aOYMqoeG8tS4CXTr1q3Ufps3b+aGrl3o0ySaEo/Fsn35/HvtOmJiYs7Yb8L4cSye8RZP\ndKhFgIG3v8uhSZe+vDNl2iV6R85xu926L67IedK1kcWvvPbKK7z5lwk8fLmH/sE5DEzpw/r160vt\nN+7p0YxKrM9r3Vry1vWtGNIsihefn1hqv+83pNKpXhCBAYYAY+hSP4TvN353Kd6K41S0IhdPZStV\n0vSpk3m9SzN6xNRhcOv63J9Qj1nvzyy135HDh2keFXrqcYvIEHIOHii1X2x8AqkHT+C1LCzLYn12\nMbFx8ba+BxGpOlS2Uibbtm3j2vZtiQwPo0PilWzZssXpSGcIDAyk8LR7uRa6LQIDg0rt16vvLby4\nYQ8ZeYX8fOQ4r27aR6+U0icBPTP2z3hqN2Xk0mwe/+YgO6woJv33q7a+BxGpOnTM1k8UFRUxbdo0\n9u7JJLlzF26++eaLGiuuRTNGtIhkYMu6zE8/yN/SDvDD9vSLumdseXp/5kyeefRhRiU14GDhCd7Z\nks2K1Wto1arVGft5PB6efnIU7/1zGq4AFw8/+hhPjxlz1jNbPR4PGzduxOPxkJiYqO9zikgpujay\nHyspKeG6Lsl4DuyiaTisyCrhwcefYvTTT5dpvE2bNvH73j1ZPajtqW3dP/2etz+aS8eOHcsr9kVb\nsGABH70/g+phYTzyxJPExcU5HUlEqjiVrR9bsGABTw+/h+e71CbAGA4VnGDEogzyjxeU6eSXjIwM\n2ibE8d2QjkRUCyK/xE27Wev4ZvXaUjNHERF/oitI+bG8vDxqhQUR4FsajQoJxLIsSkpKylS2jRo1\nYsjQu+g19yN6NohgWVYeKQMGqGhFRM5BM1s/kJmZSWKbeO6JD6dVrRDmbT9GYZ2WfL18ZZnHtCyL\n+fPnk5aWRmxsLAMGDNAVfETE72kZ2c+tWbOGEffdy759+0lOTmbytPeIjo52OpaISJWishUREbGZ\nriAlfsfr9ZKWlsaePXucjiIifk5lK1VSeno6l0dH0jEpkWaNG9H+qjZV/k49IlJxqWylSrq5x/Xc\n1DCCvQ/cwPb7rqNg7y6GDRvmdCwR8VMqW6mSDmRnMyyxEQHGEBkSxND4BqxfvcrpWCLip1S2UiWF\nhYWxPPMwAF7LYlnGYRo0inE2lIj4LZ2N7CdSU1MZOXwY+7Ky6JSczBuTpxAZGel0LNssWrSI225J\nIa52DXIKSzhGINt27q7S71lEnKev/vixvXv30rZNAhM7NKR9vQhe37SH7OgYvli6zOlottq5cycz\nZswgPDycESNGEBIS4nSkSsftdnPgwAFq166tGy+InAeVrR+bNWsWc14Yw/QeJy+n6PZ6qf/2MnJy\njxIaGvobzxZ/tWbNGm5N6cuJ4kJKPBb/nDGT/v37Ox1LpELTtZH9WFhYGPvzi7EsC2MMBwtKMMZo\npiLnVFJSwoC+fXi54xX0aV6X77KPcuvdQ2nf/gcaNmzodDyRSkcnSPmBXr16Qa163Ln4R15dt4OU\nBWmMGTOmTDchEP+wZ88eXF43fZrXBSCpbgRX1otmy5YtDicTqZz029YPVKtWjaXLVzJ58mT2ZOzm\nxce7aTlQftVll13G0aJituXk0yq6BocLS/jhwBHNakXKSMdsReSsZkyfzhMjH6J9g1ps2n+EP414\nkPETn3c6lkiFphOkxO9s376dzz//nNDQUAYNGkRERITTkSqd7du3k5aWRpMmTUhMTHQ6jkiFp7IV\nv7Jq1Spu6d2LlKZ1OFjkZluh4dvUDbqtoIjYSmUrfqVrx6u5O7qEgbGXA/Dg11tp2u8uxk+Y4HAy\nEanKdIs98SuHDx+mVa2wU49bR4Zw+OABBxOJiD9T2coZNm3aRMekq7gsKpLfde9KZmam05HKpMeN\nvfjL+kxyCkvYlpPPlB8P0LPXTU7HEhE/pWVkOeXIkSPEt2zOn9vW54bGtZmxJYt5h7xs+mErLpfL\n6XgXpKioiOF/upePP/mU6iHVGDtuPA+PfMTpWCJSxemYrfymJUuWMGH4PSzsEw+AZVnEzVzDv1M3\nEhMT42w4EZFKQMdsbZaTk8OsWbOYPXs2ubm5Tscpk4iICLLyCih2ewE4UnSCY4XF1KxZ0+FkUp5y\nc3N5YNh93JDciYeG38/Ro0edjiRS5WlmWw4yMjJI7tyFBs1b47W8ZO/czrer/k39+vWdjnZBLMti\nUP9+7Nm0li51w1iQcZR+Q+7mhUkvOx3Nr3m9XqZMmUJ6ejr9+/enU6dOZR7L7XZz7dXtiOcYfWOi\nmbfzMOnBtVj+7ZpKd6hApCLSMrKN7rzrbgpDorj1/scB+PAfL1In8ART3p3scLIL5/F4+OCDD0hP\nT6dt27akpKRgTKn/b+QS8Xq9xLdszsGsTK6IqMbWQ4WMHTeBsWPHlmm8jRs3MqhXD9YNbocxBq9l\nkTRrHQu/WUlcXFw5pxfxP7rrj4327dtH0s3dTz1uEncV276Z71ygi+ByuRg6dKjTMcTnhRde4Ej2\nHt7u05RqgQFsOVDAhAnjyly2xhg8loXXApcBr2Xh8VoEBOiIkoid9AkrB926duHL2dMoyD9GwbE8\nlnz0Ht27dXU6llQB27ZtI7Z2daoFnvyotq5TnRK3l4KCgjKNl5CQQMPmLblv6Vbm/rSf/1qyjVYJ\nbWjZsmV5xhaR/6CyLQdPP/UUV7dpzfAeiQzvmURyu0Qef+wxp2NJFXDjjTeyPiuffcdKAFj40xHC\nQ6sRGhpapvFcLhcLv1xC896DmGfVIy7lD8xftFgzWxGb6ZhtOXK73QC6T6yUqzuH3MHs2bMJCjC4\nXC7mLlhIz549nY4lImehE6REKrG8vDwyMjKIjY3VH3MiFZjKVkRExGa6qIWIiIhDVLYiIiI2U9mK\niIjYTGUrIiJiM5WtiIiIzVS2IiIiNlPZioiI2ExlKyIiYjOVrYiIiM1UtiIiIjZT2YqIiNhMZSsi\nImIzla2IiIjNVLYiIiI2U9mKiIjYTGUrIiJiM5WtiIiIzVS2IiIiNlPZioiI2ExlKyIiYjOVrYiI\niM1UtiIiIjZT2YqIiNhMZStnSE1NJTEulogaYXTr1JFdu3Y5HUlEpNJT2cophw8fps+NPXmocTCb\n7uzI9UF59O5xPW632+loIiKVmspWTtmwYQMtosK4LbY+0SHBPNquMXlHcsjMzHQ6mohIpaaylVOi\noqLIPHqcIrcHgIMFxeQVFhMZGelwMhGRyi3Q6QBScbRr145ru99A7/krSK5bg0UZR3j0sceIiopy\nOpqISKVmLMuyZ2BjLLvGFvt4vV7mzJnDjh07SEpKolevXk5HEhGpNIwxWJZlSm1X2YqIiJSPc5Wt\njtmKiIjYTGUrIiJiM5WtiIiIzVS2IiIiNlPZioiI2ExlKyIiYjOVrYiIiM1UtiIiIjZT2YqIiNhM\nZSsiImIzla2IiIjNVLYiIiI2U9mKiIjYTGUrIiJiM5WtiIiIzVS2IiIiNlPZioiI2ExlKyIiYjOV\nrYiIiM1UtiIiIjZT2YqIiNgs0M7BjTF2Di8iIlIpGMuynM4gIiJSpWkZWURExGYqWxEREZupbEVE\nRGymshWxiTFmjDEmzRizyRizwRhzdTmP380Ys+B8t5fD691ijIk97fEyY0zb8n4dkarI1rORRfyV\nMeYaoDeQaFmW2xgTDQTb8FLnOsPRjjMf+wH/A2y1YWyRKk0zWxF7XA4csizLDWBZVo5lWfsBjDFt\njTHfGGPWGWMWGWPq+rYvM8a8Zoz5zhjzvTGmvW/71caYVcaYVGPMSmNMi/MNYYwJNcZMM8as9j2/\nr2/7XcaYT3yvv80Y89Jpz7nXt221MeZdY8w/jDGdgBRgkm+W3tS3+yBjzBpjzFZjTHJ5/IcTqYpU\ntiL2+BJo5CuhN40xXQGMMYHAP4BbLcu6GngP+Otpz6tuWVYS8IDvZwA/Ap0ty2oHjANeuIAcY4Cl\nlmVdA1wP/M0YU933s6uA24ArgduNMQ2MMZcDY4EOQDIQC1iWZX0LzAdGWZbV1rKsHb4xXJZldQQe\nBcZfQC4Rv6JlZBEbWJZ13Hc8swsnS+5fxpingFQgAfjKnLzqSwCQddpTZ/uev8IYE26MqQnUBGb4\nZrQWF/a5/R3Q1xgzyvc4GGjk+/ellmXlAxhjtgCNgTrAN5ZlHfVtnwP82kz6U98/U33PF5GzUNmK\n2MQ6ecWY5cByY8xmYCiwAUizLOtcS67/eazVAiYCX1uWNcAY0xhYdgExDCdn0dvP2HjymHLxaZu8\n/PL74EIu/fb/Y3jQ7xORc9IysogNjDEtjTHNT9uUCOwGtgF1fGWHMSbQGBN32n63+7Z3Bo5alnUM\niAD2+n5+zwVGWQw8fFquxN/Yfx3Q1RgT4VvyvvW0nx3j5Cz7XHR9VpFzUNmK2KMGMN331Z+NQGtg\nvGVZJ4CBwEu+7d8BnU57XpExZgPwFvBH37ZJwIvGmFQu/DM7EQjynXCVBjx3jv0sAMuysjh5DHkt\nsALYCRz17fMvYJTvRKumnH0WLiJnoWsji1QQxphlwOOWZW1wOEeY75izC5gLTLMs6zMnM4lUdprZ\nilQcFeUv3/HGmO+AzcAOFa3IxdPMVkRExGaa2YqIiNhMZSsiImIzla2IiIjNVLYiIiI2U9mKiIjY\nTGUrIiJis/8Ddjbvid3BBjsAAAAASUVORK5CYII=\n", - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - } - ], - "source": [ - "plot(X_lsml, Y)" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "### Neighborhood Components Analysis\n", - "\n", - "NCA is an extrememly popular metric-learning algorithm, and one of the first few (published back in 2005).\n", - "\n", - "Neighbourhood components analysis aims at \"learning\" a distance metric by finding a linear transformation of input data such that the average leave-one-out (LOO) classification performance is maximized in the transformed space. The key insight to the algorithm is that a matrix $A$ corresponding to the transformation can be found by defining a differentiable objective function for $A$, followed by use of an iterative solver such as conjugate gradient descent. One of the benefits of this algorithm is that the number of classes $k$ can be determined as a function of $A$, up to a scalar constant. This use of the algorithm therefore addresses the issue of model selection.\n", - "\n", - "You can read more about it in the paper here: [NCA](https://papers.nips.cc/paper/2566-neighbourhood-components-analysis.pdf). " - ] - }, - { - "cell_type": "code", - "execution_count": 13, - "metadata": { - "collapsed": false - }, - "outputs": [], - "source": [ - "nca = metric_learn.NCA(max_iter=1000, learning_rate=0.01)\n", - "X_nca = nca.fit_transform(X, Y)" - ] - }, - { - "cell_type": "code", - "execution_count": 14, - "metadata": { - "collapsed": false - }, - "outputs": [ - { - "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAAdsAAAFsCAYAAACEtRP5AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nQmCC\n", - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - } - ], - "source": [ - "plot(X_nca, Y)" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "### Local Fischer Discriminant Analysis\n", - "\n", - "LFDA is a linear supervised dimensionality reduction method. It is particularly useful when dealing with multimodality, where one ore more classes consist of separate clusters in input space. The core optimization problem of LFDA is solved as a generalized eigenvalue problem.\n", - "\n", - "Link to paper: [LFDA](http://www.machinelearning.org/proceedings/icml2006/114_Local_Fisher_Discrim.pdf)" - ] - }, - { - "cell_type": "code", - "execution_count": 15, - "metadata": { - "collapsed": false - }, - "outputs": [], - "source": [ - "lfda = metric_learn.LFDA(k=2, dim=2)\n", - "X_lfda = lfda.fit_transform(X, Y)" - ] - }, - { - "cell_type": "code", - "execution_count": 16, - "metadata": { - "collapsed": false - }, - "outputs": [ - { - "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAAdsAAAFsCAYAAACEtRP5AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n", - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - } - ], - "source": [ - "plot(X_lfda, Y)" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "### Relative Components Analysis\n", - "\n", - "RCA is another one of the older algorithms.\n", - "It learns a full rank Mahalanobis distance metric based on a weighted sum of in-class covariance matrices. It applies a global linear transformation to assign large weights to relevant dimensions and low weights to irrelevant dimensions. Those relevant dimensions are estimated using “chunklets”, subsets of points that are known to belong to the same class.\n", - "\n", - "Link to paper: [RCA](https://www.aaai.org/Papers/ICML/2003/ICML03-005.pdf)" - ] - }, - { - "cell_type": "code", - "execution_count": 17, - "metadata": { - "collapsed": false - }, - "outputs": [], - "source": [ - "rca = metric_learn.RCA_Supervised(num_chunks=30, chunk_size=2)\n", - "X_rca = rca.fit_transform(X, Y)" - ] - }, - { - "cell_type": "code", - "execution_count": 18, - "metadata": { - "collapsed": false - }, - "outputs": [ - { - "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAAdsAAAFsCAYAAACEtRP5AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nKbZCCCGEmUmxFUIIIczs/wCB5ulx+0dRCQAAAABJRU5ErkJggg==\n", - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - } - ], - "source": [ - "plot(X_rca, Y)" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "### Manual Constraints\n", - "\n", - "Some of the algorithms we've mentioned have alternate ways to pass constraints.\n", - "So far we've been passing them as just class labels - and letting the internals of metric-learn deal with creating our constrints.\n", - "\n", - "We'll be looking at one other way to do this - which is to pass a Matrix X such that - (a,b,c,d) indices into X, such that $d(X[a],X[b]) < d(X[c],X[d])$. \n", - "\n", - "This kind of input is possible for ITML and LSML.\n", - "\n", - "We're going to create these constraints through the labels we have, i.e $Y$.\n", - "\n", - "This is done internally through metric learn anyway (do check out the `constraints` class!) - but we'll try our own version of this. I'm going to go ahead and assume that two points labelled the same will be closer than two points in different labels.\n", - "\n", - "Do keep in mind that we are doing this method because we know the labels - we can actually create the constraints any way we want to depending on the data!" - ] - }, - { - "cell_type": "code", - "execution_count": 19, - "metadata": { - "collapsed": false - }, - "outputs": [], - "source": [ - "def create_constraints(labels):\n", - " import itertools\n", - " import random\n", - " \n", - " # aggregate indices of same class\n", - " zeros = np.where(Y==0)[0]\n", - " ones = np.where(Y==1)[0]\n", - " twos = np.where(Y==2)[0]\n", - " # make permutations of all those points in the same class\n", - " zeros_ = list(itertools.combinations(zeros, 2))\n", - " ones_ = list(itertools.combinations(ones, 2))\n", - " twos_ = list(itertools.combinations(twos, 2))\n", - " # put them together!\n", - " sim = np.array(zeros_ + ones_ + twos_)\n", - " \n", - " # similarily, put together indices in different classes\n", - " dis = []\n", - " for zero in zeros:\n", - " for one in ones:\n", - " dis.append((zero, one))\n", - " for two in twos:\n", - " dis.append((zero, two))\n", - " for one in ones:\n", - " for two in twos:\n", - " dis.append((one, two))\n", - " \n", - " # pick up just enough dissimilar examples as we have similar examples\n", - " dis = np.array(random.sample(dis, len(sim)))\n", - " \n", - " # return a four-tuple of arrays with d(X[a],X[b])" - ] - }, - "metadata": {}, - "output_type": "display_data" - } - ], - "source": [ - "plot(X_itml, Y)" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "And that's the result of ITML after being trained on our manual constraints! A bit different from our old result but not too different. We can also notice that it might be better to rely on the randomised algorithms under the hood to make our constraints if we are not very sure how we want our transformed space to be.\n", - "\n", - "RCA and SDML also have their own specific ways of taking in inputs - it's worth one's while to poke around in the constraints.py file to see how exactly this is going on.\n", - "\n", - "This brings us to the end of this tutorial!\n", - "Have fun Metric Learning :)" - ] - } - ], - "metadata": { - "kernelspec": { - "display_name": "Python 2", - "language": "python", - "name": "python2" - }, - "language_info": { - "codemirror_mode": { - "name": "ipython", - "version": 2 - }, - "file_extension": ".py", - "mimetype": "text/x-python", - "name": "python", - "nbconvert_exporter": "python", - "pygments_lexer": "ipython2", - "version": "2.7.12" - } - }, - "nbformat": 4, - "nbformat_minor": 1 -} diff --git a/examples/plot_metric_examples.py b/examples/plot_metric_examples.py new file mode 100644 index 00000000..82f0ebc7 --- /dev/null +++ b/examples/plot_metric_examples.py @@ -0,0 +1,398 @@ +""" +Metric Learning and Plotting +~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + +This is a small walkthrough which illustrates all the Metric Learning +algorithms implemented in metric\_learn, and also does a quick +visualisation which can help understand which algorithm might be best +suited for you. + +Of course, depending on the data set and the constraints your results +will look very different; you can just follow this and change your data +and constraints accordingly. + +""" + +# License: BSD 3 clause +# Authors: Bhargav Srinivasa Desikan + +###################################################################### +# Imports +# ^^^^^^^ +# + +import metric_learn +import numpy as np +from sklearn.datasets import load_iris + +# visualisation imports +import matplotlib.pyplot as plt +from mpl_toolkits.mplot3d import Axes3D + + +###################################################################### +# Loading our data-set and setting up plotting +# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# +# We will be using the IRIS data-set to illustrate the plotting. You can +# read more about the IRIS data-set here: +# `link `__. +# +# We would like to point out that only two features - Sepal Width and +# Sepal Length are being plotted. This is because it is tough to visualise +# more features than this. The purpose of the plotting is to understand +# how each of the new learned metrics transform the input space. +# + +# loading our dataset + +iris_data = load_iris() +# this is our data +X = iris_data['data'] +# these are our constraints +Y = iris_data['target'] + +# function to plot the results +def plot(X, Y): + x_min, x_max = X[:, 0].min() - .5, X[:, 0].max() + .5 + y_min, y_max = X[:, 1].min() - .5, X[:, 1].max() + .5 + plt.figure(2, figsize=(8, 6)) + + # clean the figure + plt.clf() + + plt.scatter(X[:, 0], X[:, 1], c=Y, cmap=plt.cm.Paired) + plt.xlabel('Sepal length') + plt.ylabel('Sepal width') + + plt.xlim(x_min, x_max) + plt.ylim(y_min, y_max) + plt.xticks(()) + plt.yticks(()) + + plt.show() + +# plotting the dataset as is. +plot(X, Y) + + +###################################################################### +# Metric Learning +# ~~~~~~~~~~~~~~~ +# +# Why is Metric Learning useful? We can, with prior knowledge of which +# points are supposed to be closer, figure out a better way to understand +# distances between points. Especially in higher dimensions when Euclidean +# distances are a poor way to measure distance, this becomes very useful. +# +# Basically, we learn this distance: +# :math:`D(x,y)=\sqrt{(x-y)\,M^{-1}(x-y)}`. And we learn this distance by +# learning a Matrix :math:`M`, based on certain constraints. +# +# Some good reading material for the same can be found +# `here `__. It serves as a good +# literature review of Metric Learning. +# +# We will briefly explain the metric-learning algorithms implemented by +# metric-learn, before providing some examples for it's usage, and also +# discuss how to go about doing manual constraints. +# +# Metric-learn can be easily integrated with your other machine learning +# pipelines, and follows (for the most part) scikit-learn conventions. +# + + +###################################################################### +# Large Margin Nearest Neighbour +# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# +# LMNN is a metric learning algorithm primarily designed for k-nearest +# neighbor classification. The algorithm is based on semidefinite +# programming, a sub-class of convex programming (as most Metric Learning +# algorithms are). +# +# The main intuition behind LMNN is to learn a pseudometric under which +# all data instances in the training set are surrounded by at least k +# instances that share the same class label. If this is achieved, the +# leave-one-out error (a special case of cross validation) is minimized. +# +# You can find the paper +# `here `__. +# + + +###################################################################### +# Fit and then transform! +# ^^^^^^^^^^^^^^^^^^^^^^^ +# + +# setting up LMNN +lmnn = metric_learn.LMNN(k=5, learn_rate=1e-6) + +# fit the data! +lmnn.fit(X, Y) + +# transform our input space +X_lmnn = lmnn.transform(X) + + +###################################################################### +# So what have we learned? The matrix :math:`M` we talked about before. +# Let's see what it looks like. +# + +lmnn.metric() + + +###################################################################### +# Now let us plot the transformed space - this tells us what the original +# space looks like after being transformed with the new learned metric. +# + +plot(X_lmnn, Y) + + +###################################################################### +# Pretty neat, huh? +# +# The rest of this notebook will briefly explain the other Metric Learning +# algorithms before plottting them. Also, while we have first run ``fit`` +# and then ``transform`` to see our data transformed, we can also use +# ``fit_transform`` if you are using the bleeding edge version of the +# code. The rest of the examples and illustrations will use +# ``fit_transform``. +# + + +###################################################################### +# Information Theoretic Metric Learning +# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# +# ITML uses a regularizer that automatically enforces a Semi-Definite +# Positive Matrix condition - the LogDet divergence. It uses soft +# must-link or cannot like constraints, and a simple algorithm based on +# Bregman projections. +# +# Link to paper: +# `ITML `__. +# + +itml = metric_learn.ITML_Supervised(num_constraints=200) +X_itml = itml.fit_transform(X, Y) + +plot(X_itml, Y) + + +###################################################################### +# Sparse Determinant Metric Learning +# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# +# Implements an efficient sparse metric learning algorithm in high +# dimensional space via an :math:`l_1`-penalised log-determinant +# regularization. Compare to the most existing distance metric learning +# algorithms, the algorithm exploits the sparsity nature underlying the +# intrinsic high dimensional feature space. +# +# Link to paper here: +# `SDML `__. +# +# One feature which we'd like to show off here is the use of random seeds. +# Some of the algorithms feature randomised algorithms for selecting +# constraints - to fix these, and get constant results for each run, we +# pass a numpy random seed as shown in the example below. +# + +sdml = metric_learn.SDML_Supervised(num_constraints=200, sparsity_param=0.1) +X_sdml = sdml.fit_transform(X, Y, random_state = np.random.RandomState(1234)) + +plot(X_sdml, Y) + + +###################################################################### +# Least Squares Metric Learning +# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# +# LSML is a simple, yet effective, algorithm that learns a Mahalanobis +# metric from a given set of relative comparisons. This is done by +# formulating and minimizing a convex loss function that corresponds to +# the sum of squared hinge loss of violated constraints. +# +# Link to paper: +# `LSML `__ +# + +lsml = metric_learn.LSML_Supervised(prior=np.eye(X.shape[1]), num_constraints=200, tol=0.0001, max_iter=10000, verbose=True) +X_lsml = lsml.fit_transform(X, Y) + +plot(X_lsml, Y) + + +###################################################################### +# Neighborhood Components Analysis +# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# +# NCA is an extrememly popular metric-learning algorithm, and one of the +# first few (published back in 2005). +# +# Neighbourhood components analysis aims at "learning" a distance metric +# by finding a linear transformation of input data such that the average +# leave-one-out (LOO) classification performance is maximized in the +# transformed space. The key insight to the algorithm is that a matrix +# :math:`A` corresponding to the transformation can be found by defining a +# differentiable objective function for :math:`A`, followed by use of an +# iterative solver such as conjugate gradient descent. One of the benefits +# of this algorithm is that the number of classes :math:`k` can be +# determined as a function of :math:`A`, up to a scalar constant. This use +# of the algorithm therefore addresses the issue of model selection. +# +# You can read more about it in the paper here: +# `NCA `__. +# + +nca = metric_learn.NCA(max_iter=1000) +X_nca = nca.fit_transform(X, Y) + +plot(X_nca, Y) + + +###################################################################### +# Local Fischer Discriminant Analysis +# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# +# LFDA is a linear supervised dimensionality reduction method. It is +# particularly useful when dealing with multimodality, where one ore more +# classes consist of separate clusters in input space. The core +# optimization problem of LFDA is solved as a generalized eigenvalue +# problem. +# +# Link to paper: +# `LFDA `__ +# + +lfda = metric_learn.LFDA(k=2, num_dims=2) +X_lfda = lfda.fit_transform(X, Y) + +plot(X_lfda, Y) + + +###################################################################### +# Relative Components Analysis +# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# +# RCA is another one of the older algorithms. It learns a full rank +# Mahalanobis distance metric based on a weighted sum of in-class +# covariance matrices. It applies a global linear transformation to assign +# large weights to relevant dimensions and low weights to irrelevant +# dimensions. Those relevant dimensions are estimated using “chunklets”, +# subsets of points that are known to belong to the same class. +# +# Link to paper: +# `RCA `__ +# + +rca = metric_learn.RCA_Supervised(num_chunks=30, chunk_size=2) +X_rca = rca.fit_transform(X, Y) + +plot(X_rca, Y) + + +###################################################################### +# Manual Constraints +# ~~~~~~~~~~~~~~~~~~ +# +# Some of the algorithms we've mentioned have alternate ways to pass +# constraints. So far we've been passing them as just class labels - and +# letting the internals of metric-learn deal with creating our constrints. +# +# We'll be looking at one other way to do this - which is to pass a Matrix +# X such that - (a,b,c,d) indices into X, such that +# :math:`d(X[a],X[b]) < d(X[c],X[d])`. +# +# This kind of input is possible for ITML and LSML. +# +# We're going to create these constraints through the labels we have, i.e +# :math:`Y`. +# +# This is done internally through metric learn anyway (do check out the +# ``constraints`` class!) - but we'll try our own version of this. I'm +# going to go ahead and assume that two points labelled the same will be +# closer than two points in different labels. +# +# Do keep in mind that we are doing this method because we know the labels +# - we can actually create the constraints any way we want to depending on +# the data! +# + +def create_constraints(labels): + import itertools + import random + + # aggregate indices of same class + zeros = np.where(Y==0)[0] + ones = np.where(Y==1)[0] + twos = np.where(Y==2)[0] + # make permutations of all those points in the same class + zeros_ = list(itertools.combinations(zeros, 2)) + ones_ = list(itertools.combinations(ones, 2)) + twos_ = list(itertools.combinations(twos, 2)) + # put them together! + sim = np.array(zeros_ + ones_ + twos_) + + # similarily, put together indices in different classes + dis = [] + for zero in zeros: + for one in ones: + dis.append((zero, one)) + for two in twos: + dis.append((zero, two)) + for one in ones: + for two in twos: + dis.append((one, two)) + + # pick up just enough dissimilar examples as we have similar examples + dis = np.array(random.sample(dis, len(sim))) + + # return an array of pairs of indices of shape=(2*len(sim), 2), and the corresponding labels, array of shape=(2*len(sim)) + # Each pair of similar points have a label of +1 and each pair of dissimilar points have a label of -1 + return (np.vstack([np.column_stack([sim[:,0], sim[:,1]]), np.column_stack([dis[:,0], dis[:,1]])]), + np.concatenate([np.ones(len(sim)), -np.ones(len(sim))])) + +pairs, pairs_labels = create_constraints(Y) + + +###################################################################### +# Now that we've created our constraints, let's see what it looks like! +# + +print(pairs) +print(pairs_labels) + + +###################################################################### +# Using our constraints, let's now train ITML again. We should keep in +# mind that internally, ITML\_Supervised does pretty much the same thing +# we are doing; I was just giving an example to better explain how the +# constraints are structured. +# + +itml = metric_learn.ITML(preprocessor=X) +itml.fit(pairs, pairs_labels) + +X_itml = itml.transform(X) + +plot(X_itml, Y) + + +###################################################################### +# And that's the result of ITML after being trained on our manual +# constraints! A bit different from our old result but not too different. +# We can also notice that it might be better to rely on the randomised +# algorithms under the hood to make our constraints if we are not very +# sure how we want our transformed space to be. +# +# RCA and SDML also have their own specific ways of taking in inputs - +# it's worth one's while to poke around in the constraints.py file to see +# how exactly this is going on. +# +# This brings us to the end of this tutorial! Have fun Metric Learning :) +# From 8aadc42f09d4b856a234b41fb8f8b55f9f5592d9 Mon Sep 17 00:00:00 2001 From: William de Vazelhes Date: Thu, 14 Mar 2019 16:58:34 +0100 Subject: [PATCH 2/9] Fix the figure number in order to get the image printed in the logo of the example --- examples/plot_metric_examples.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/examples/plot_metric_examples.py b/examples/plot_metric_examples.py index 82f0ebc7..dc0ddbb1 100644 --- a/examples/plot_metric_examples.py +++ b/examples/plot_metric_examples.py @@ -56,7 +56,7 @@ def plot(X, Y): x_min, x_max = X[:, 0].min() - .5, X[:, 0].max() + .5 y_min, y_max = X[:, 1].min() - .5, X[:, 1].max() + .5 - plt.figure(2, figsize=(8, 6)) + plt.figure(figsize=(8, 6)) # clean the figure plt.clf() From 8af4ff578efb7af8fa8341d767f9b39dcd34c98b Mon Sep 17 00:00:00 2001 From: William de Vazelhes Date: Wed, 3 Apr 2019 09:54:41 +0200 Subject: [PATCH 3/9] wip replace dataset by faces --- examples/plot_metric_examples.py | 23 +++++++++++++---------- 1 file changed, 13 insertions(+), 10 deletions(-) diff --git a/examples/plot_metric_examples.py b/examples/plot_metric_examples.py index dc0ddbb1..01edde54 100644 --- a/examples/plot_metric_examples.py +++ b/examples/plot_metric_examples.py @@ -19,11 +19,12 @@ ###################################################################### # Imports # ^^^^^^^ -# +# +from sklearn.manifold import TSNE import metric_learn import numpy as np -from sklearn.datasets import load_iris +from sklearn.datasets import fetch_olivetti_faces # visualisation imports import matplotlib.pyplot as plt @@ -46,27 +47,29 @@ # loading our dataset -iris_data = load_iris() +faces_data = fetch_olivetti_faces() # this is our data -X = iris_data['data'] +X = faces_data['data'] # these are our constraints -Y = iris_data['target'] +Y = faces_data['target'] # function to plot the results def plot(X, Y): - x_min, x_max = X[:, 0].min() - .5, X[:, 0].max() + .5 - y_min, y_max = X[:, 1].min() - .5, X[:, 1].max() + .5 + # x_min, x_max = X[:, 0].min() - .5, X[:, 0].max() + .5 + # y_min, y_max = X[:, 1].min() - .5, X[:, 1].max() + .5 plt.figure(figsize=(8, 6)) # clean the figure plt.clf() - plt.scatter(X[:, 0], X[:, 1], c=Y, cmap=plt.cm.Paired) + tsne = TSNE() + X_embedded = tsne.fit_transform(X) + plt.scatter(X_embedded[:, 0], X_embedded[:, 1], c=Y, cmap=plt.cm.Paired) plt.xlabel('Sepal length') plt.ylabel('Sepal width') - plt.xlim(x_min, x_max) - plt.ylim(y_min, y_max) + # plt.xlim(x_min, x_max) + # plt.ylim(y_min, y_max) plt.xticks(()) plt.yticks(()) From 27fbd8b205152348f78f2ea3fba25669b63d3e4a Mon Sep 17 00:00:00 2001 From: William de Vazelhes Date: Wed, 10 Apr 2019 14:55:13 +0200 Subject: [PATCH 4/9] Finalize notebook --- examples/plot_metric_examples.py | 82 ++++++++++++++++---------------- 1 file changed, 40 insertions(+), 42 deletions(-) diff --git a/examples/plot_metric_examples.py b/examples/plot_metric_examples.py index 01edde54..1bebe8bd 100644 --- a/examples/plot_metric_examples.py +++ b/examples/plot_metric_examples.py @@ -1,9 +1,9 @@ """ -Metric Learning and Plotting -~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +Algorithms walkthrough +~~~~~~~~~~~~~~~~~~~~~~ This is a small walkthrough which illustrates all the Metric Learning -algorithms implemented in metric\_learn, and also does a quick +algorithms implemented in metric-learn, and also does a quick visualisation which can help understand which algorithm might be best suited for you. @@ -20,43 +20,43 @@ # Imports # ^^^^^^^ # + from sklearn.manifold import TSNE import metric_learn import numpy as np -from sklearn.datasets import fetch_olivetti_faces +from sklearn.datasets import load_iris # visualisation imports import matplotlib.pyplot as plt -from mpl_toolkits.mplot3d import Axes3D ###################################################################### # Loading our data-set and setting up plotting -# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # # We will be using the IRIS data-set to illustrate the plotting. You can # read more about the IRIS data-set here: # `link `__. -# -# We would like to point out that only two features - Sepal Width and -# Sepal Length are being plotted. This is because it is tough to visualise -# more features than this. The purpose of the plotting is to understand -# how each of the new learned metrics transform the input space. -# - -# loading our dataset -faces_data = fetch_olivetti_faces() +iris_data = load_iris() # this is our data -X = faces_data['data'] +X = iris_data['data'] # these are our constraints -Y = faces_data['target'] +Y = iris_data['target'] + +######################################################################## +# To make the task more difficult, we will add five noise columns to the +# data, of variance 5. +rng = np.random.RandomState(42) +X = np.hstack([X, rng.randn(X.shape[0], 5)*5]) + +########################################################################### +# Note that the dimensionality of the data is now 4 + 5 = 9, so to plot the +# transformed data in 2D, we will use the t-sne algorithm. (See +# `sklearn.manifold.TSNE`). -# function to plot the results def plot(X, Y): - # x_min, x_max = X[:, 0].min() - .5, X[:, 0].max() + .5 - # y_min, y_max = X[:, 1].min() - .5, X[:, 1].max() + .5 plt.figure(figsize=(8, 6)) # clean the figure @@ -65,23 +65,21 @@ def plot(X, Y): tsne = TSNE() X_embedded = tsne.fit_transform(X) plt.scatter(X_embedded[:, 0], X_embedded[:, 1], c=Y, cmap=plt.cm.Paired) - plt.xlabel('Sepal length') - plt.ylabel('Sepal width') - # plt.xlim(x_min, x_max) - # plt.ylim(y_min, y_max) plt.xticks(()) plt.yticks(()) plt.show() -# plotting the dataset as is. +################################### +# Let's now plot the dataset as is. + plot(X, Y) ###################################################################### # Metric Learning -# ~~~~~~~~~~~~~~~ +# ^^^^^^^^^^^^^^^ # # Why is Metric Learning useful? We can, with prior knowledge of which # points are supposed to be closer, figure out a better way to understand @@ -107,7 +105,7 @@ def plot(X, Y): ###################################################################### # Large Margin Nearest Neighbour -# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # # LMNN is a metric learning algorithm primarily designed for k-nearest # neighbor classification. The algorithm is based on semidefinite @@ -126,7 +124,7 @@ def plot(X, Y): ###################################################################### # Fit and then transform! -# ^^^^^^^^^^^^^^^^^^^^^^^ +# ----------------------- # # setting up LMNN @@ -159,17 +157,16 @@ def plot(X, Y): # Pretty neat, huh? # # The rest of this notebook will briefly explain the other Metric Learning -# algorithms before plottting them. Also, while we have first run ``fit`` +# algorithms before plotting them. Also, while we have first run ``fit`` # and then ``transform`` to see our data transformed, we can also use -# ``fit_transform`` if you are using the bleeding edge version of the -# code. The rest of the examples and illustrations will use +# ``fit_transform``. The rest of the examples and illustrations will use # ``fit_transform``. # ###################################################################### # Information Theoretic Metric Learning -# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # # ITML uses a regularizer that automatically enforces a Semi-Definite # Positive Matrix condition - the LogDet divergence. It uses soft @@ -188,7 +185,7 @@ def plot(X, Y): ###################################################################### # Sparse Determinant Metric Learning -# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # # Implements an efficient sparse metric learning algorithm in high # dimensional space via an :math:`l_1`-penalised log-determinant @@ -205,15 +202,16 @@ def plot(X, Y): # pass a numpy random seed as shown in the example below. # -sdml = metric_learn.SDML_Supervised(num_constraints=200, sparsity_param=0.1) -X_sdml = sdml.fit_transform(X, Y, random_state = np.random.RandomState(1234)) +sdml = metric_learn.SDML_Supervised(num_constraints=200, sparsity_param=0.1, + balance_param=0.0015) +X_sdml = sdml.fit_transform(X, Y, random_state=np.random.RandomState(1234)) plot(X_sdml, Y) ###################################################################### # Least Squares Metric Learning -# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # # LSML is a simple, yet effective, algorithm that learns a Mahalanobis # metric from a given set of relative comparisons. This is done by @@ -224,7 +222,7 @@ def plot(X, Y): # `LSML `__ # -lsml = metric_learn.LSML_Supervised(prior=np.eye(X.shape[1]), num_constraints=200, tol=0.0001, max_iter=10000, verbose=True) +lsml = metric_learn.LSML_Supervised(num_constraints=200, tol=0.0001, max_iter=10000, verbose=True) X_lsml = lsml.fit_transform(X, Y) plot(X_lsml, Y) @@ -232,7 +230,7 @@ def plot(X, Y): ###################################################################### # Neighborhood Components Analysis -# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # # NCA is an extrememly popular metric-learning algorithm, and one of the # first few (published back in 2005). @@ -260,7 +258,7 @@ def plot(X, Y): ###################################################################### # Local Fischer Discriminant Analysis -# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # # LFDA is a linear supervised dimensionality reduction method. It is # particularly useful when dealing with multimodality, where one ore more @@ -280,7 +278,7 @@ def plot(X, Y): ###################################################################### # Relative Components Analysis -# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # # RCA is another one of the older algorithms. It learns a full rank # Mahalanobis distance metric based on a weighted sum of in-class @@ -301,11 +299,11 @@ def plot(X, Y): ###################################################################### # Manual Constraints -# ~~~~~~~~~~~~~~~~~~ +# ^^^^^^^^^^^^^^^^^^ # # Some of the algorithms we've mentioned have alternate ways to pass # constraints. So far we've been passing them as just class labels - and -# letting the internals of metric-learn deal with creating our constrints. +# letting the internals of metric-learn deal with creating our constraints. # # We'll be looking at one other way to do this - which is to pass a Matrix # X such that - (a,b,c,d) indices into X, such that From b0ee23b30c04949a80b97f7239f639e36fcd6dff Mon Sep 17 00:00:00 2001 From: William de Vazelhes Date: Thu, 16 May 2019 12:57:01 +0200 Subject: [PATCH 5/9] change dataset for make_classification --- examples/plot_metric_examples.py | 28 +++++++++++++++------------- 1 file changed, 15 insertions(+), 13 deletions(-) diff --git a/examples/plot_metric_examples.py b/examples/plot_metric_examples.py index 1bebe8bd..c8a20d8c 100644 --- a/examples/plot_metric_examples.py +++ b/examples/plot_metric_examples.py @@ -22,13 +22,15 @@ # from sklearn.manifold import TSNE +from sklearn.utils import shuffle import metric_learn import numpy as np -from sklearn.datasets import load_iris +from sklearn.datasets import make_classification # visualisation imports import matplotlib.pyplot as plt +np.random.seed(42) ###################################################################### @@ -39,20 +41,20 @@ # read more about the IRIS data-set here: # `link `__. -iris_data = load_iris() -# this is our data -X = iris_data['data'] -# these are our constraints -Y = iris_data['target'] +X, Y = make_classification(n_classes=3, n_clusters_per_class=2, + n_informative=3, class_sep=4., n_features=5, + n_redundant=0, shuffle=True) + +########################################################################## +# Let's make the noise of the non-informative features higher, and shuffle +# the dataset: + +X[:, 3:] *= 20 +X, Y = shuffle(X, Y) -######################################################################## -# To make the task more difficult, we will add five noise columns to the -# data, of variance 5. -rng = np.random.RandomState(42) -X = np.hstack([X, rng.randn(X.shape[0], 5)*5]) ########################################################################### -# Note that the dimensionality of the data is now 4 + 5 = 9, so to plot the +# Note that the dimensionality of the data is 5, so to plot the # transformed data in 2D, we will use the t-sne algorithm. (See # `sklearn.manifold.TSNE`). @@ -284,7 +286,7 @@ def plot(X, Y): # Mahalanobis distance metric based on a weighted sum of in-class # covariance matrices. It applies a global linear transformation to assign # large weights to relevant dimensions and low weights to irrelevant -# dimensions. Those relevant dimensions are estimated using “chunklets”, +# dimensions. Those relevant dimensions are estimated using "chunklets", # subsets of points that are known to belong to the same class. # # Link to paper: From 34e29a2676b4be86760f8c61c4895a88fe562db0 Mon Sep 17 00:00:00 2001 From: William de Vazelhes Date: Thu, 16 May 2019 14:06:31 +0200 Subject: [PATCH 6/9] Add comments on the properties of the algorithms --- examples/plot_metric_examples.py | 17 +++++++++++++---- 1 file changed, 13 insertions(+), 4 deletions(-) diff --git a/examples/plot_metric_examples.py b/examples/plot_metric_examples.py index c8a20d8c..3bd8bed7 100644 --- a/examples/plot_metric_examples.py +++ b/examples/plot_metric_examples.py @@ -118,6 +118,10 @@ def plot(X, Y): # all data instances in the training set are surrounded by at least k # instances that share the same class label. If this is achieved, the # leave-one-out error (a special case of cross validation) is minimized. +# You'll notice that the points from the same labels are closer together, +# but they are not necessary in a same cluster. This is particular to LMNN +# and we'll see that some other algorithms implicitly enforce points from +# the same class to cluster together. # # You can find the paper # `here `__. @@ -173,7 +177,8 @@ def plot(X, Y): # ITML uses a regularizer that automatically enforces a Semi-Definite # Positive Matrix condition - the LogDet divergence. It uses soft # must-link or cannot like constraints, and a simple algorithm based on -# Bregman projections. +# Bregman projections. Unlike LMNN, ITML will implicitly enforce points from +# the same class to belong to the same cluster, as you can see below. # # Link to paper: # `ITML `__. @@ -246,8 +251,11 @@ def plot(X, Y): # iterative solver such as conjugate gradient descent. One of the benefits # of this algorithm is that the number of classes :math:`k` can be # determined as a function of :math:`A`, up to a scalar constant. This use -# of the algorithm therefore addresses the issue of model selection. -# +# of the algorithm therefore addresses the issue of model selection. Like +# LMNN, this algorithm does not try to cluster points from the same class in +# a unique cluster, because it enforces conditions at a local +# neighborhood scale. +# # You can read more about it in the paper here: # `NCA `__. # @@ -266,7 +274,8 @@ def plot(X, Y): # particularly useful when dealing with multimodality, where one ore more # classes consist of separate clusters in input space. The core # optimization problem of LFDA is solved as a generalized eigenvalue -# problem. +# problem. Like LMNN, and NCA, this algorithm does not try to cluster points +# from the same class in a unique cluster. # # Link to paper: # `LFDA `__ From b6dacb8c1854c43fa048916d20f211f01a7b20f8 Mon Sep 17 00:00:00 2001 From: William de Vazelhes Date: Wed, 29 May 2019 12:11:52 +0200 Subject: [PATCH 7/9] Address https://github.com/metric-learn/metric-learn/pull/180#pullrequestreview-242334192 --- doc/introduction.rst | 4 + doc/metric_learn.constraints.rst | 7 + doc/metric_learn.rst | 1 + doc/weakly_supervised.rst | 1 + examples/plot_metric_examples.py | 326 ++++++++++++++++++------------- 5 files changed, 208 insertions(+), 131 deletions(-) create mode 100644 doc/metric_learn.constraints.rst diff --git a/doc/introduction.rst b/doc/introduction.rst index dad530b3..ef221971 100644 --- a/doc/introduction.rst +++ b/doc/introduction.rst @@ -1,3 +1,5 @@ +.. _intro_metric_learning: + ======================== What is Metric Learning? ======================== @@ -77,6 +79,8 @@ necessarily the identity of indiscernibles. parameterizations are equivalent. In practice, an algorithm may thus solve the metric learning problem with respect to either :math:`M` or :math:`L`. +.. _use_cases: + Use-cases ========= diff --git a/doc/metric_learn.constraints.rst b/doc/metric_learn.constraints.rst new file mode 100644 index 00000000..97d79002 --- /dev/null +++ b/doc/metric_learn.constraints.rst @@ -0,0 +1,7 @@ +metric_learn.constraints module +=============================== + +.. automodule:: metric_learn.constraints + :members: + :undoc-members: + :show-inheritance: diff --git a/doc/metric_learn.rst b/doc/metric_learn.rst index c2472408..eb606542 100644 --- a/doc/metric_learn.rst +++ b/doc/metric_learn.rst @@ -6,6 +6,7 @@ Module Contents .. toctree:: + metric_learn.constraints metric_learn.base_metric metric_learn.itml metric_learn.lfda diff --git a/doc/weakly_supervised.rst b/doc/weakly_supervised.rst index 93720ffc..351c4e3b 100644 --- a/doc/weakly_supervised.rst +++ b/doc/weakly_supervised.rst @@ -118,6 +118,7 @@ through the argument `preprocessor`. paths in the filesystem, name of records in a database etc...) See section :ref:`preprocessor_section` for more details on how to use the preprocessor. +.. _sklearn_compat_ws: Scikit-learn compatibility ========================== diff --git a/examples/plot_metric_examples.py b/examples/plot_metric_examples.py index 3bd8bed7..8057d0f5 100644 --- a/examples/plot_metric_examples.py +++ b/examples/plot_metric_examples.py @@ -2,19 +2,15 @@ Algorithms walkthrough ~~~~~~~~~~~~~~~~~~~~~~ -This is a small walkthrough which illustrates all the Metric Learning -algorithms implemented in metric-learn, and also does a quick -visualisation which can help understand which algorithm might be best -suited for you. - -Of course, depending on the data set and the constraints your results -will look very different; you can just follow this and change your data -and constraints accordingly. - +This is a small walkthrough which illustrates most of the Metric Learning +algorithms implemented in metric-learn by using them on synthetic data, +with some visualizations to provide intuitions into what they are designed +to achieve. """ # License: BSD 3 clause # Authors: Bhargav Srinivasa Desikan +# William de Vazelhes ###################################################################### # Imports @@ -22,11 +18,10 @@ # from sklearn.manifold import TSNE -from sklearn.utils import shuffle import metric_learn import numpy as np -from sklearn.datasets import make_classification +from sklearn.datasets import make_classification, make_regression # visualisation imports import matplotlib.pyplot as plt @@ -34,31 +29,28 @@ ###################################################################### -# Loading our data-set and setting up plotting +# Loading our dataset and setting up plotting # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ -# -# We will be using the IRIS data-set to illustrate the plotting. You can -# read more about the IRIS data-set here: -# `link `__. - -X, Y = make_classification(n_classes=3, n_clusters_per_class=2, +# +# We will be using a synthetic dataset to illustrate the plotting, +# using the function `sklearn.datasets.make_classification` from +# scikit-learn. The dataset will contain: +# - 100 points in 3 classes with 2 clusters per class +# - 5 features, among which 3 are informative (correlated with the class +# labels) and two are random noise with large magnitude + +X, Y = make_classification(n_samples=100, n_classes=3, n_clusters_per_class=2, n_informative=3, class_sep=4., n_features=5, - n_redundant=0, shuffle=True) - -########################################################################## -# Let's make the noise of the non-informative features higher, and shuffle -# the dataset: - -X[:, 3:] *= 20 -X, Y = shuffle(X, Y) - + n_redundant=0, shuffle=True, + scale=[1, 1, 20, 20, 20]) ########################################################################### # Note that the dimensionality of the data is 5, so to plot the # transformed data in 2D, we will use the t-sne algorithm. (See # `sklearn.manifold.TSNE`). -def plot(X, Y): + +def plot_tsne(X, Y, colormap=plt.cm.Paired): plt.figure(figsize=(8, 6)) # clean the figure @@ -66,7 +58,7 @@ def plot(X, Y): tsne = TSNE() X_embedded = tsne.fit_transform(X) - plt.scatter(X_embedded[:, 0], X_embedded[:, 1], c=Y, cmap=plt.cm.Paired) + plt.scatter(X_embedded[:, 0], X_embedded[:, 1], c=Y, cmap=colormap) plt.xticks(()) plt.yticks(()) @@ -76,13 +68,21 @@ def plot(X, Y): ################################### # Let's now plot the dataset as is. -plot(X, Y) +plot_tsne(X, Y) -###################################################################### +######################################################################### +# We can see that the classes appear mixed up: this is because t-sne +# is based on preserving the original neighborhood of points in the embedding +# space, but this original neighborhood is based on the euclidean +# distance in the input space, in which the contribution of the noisy +# features is high. So even if points from the same class are close to +# each other in some subspace of the input space, this is not the case in the +# total input space. +# # Metric Learning # ^^^^^^^^^^^^^^^ -# +# # Why is Metric Learning useful? We can, with prior knowledge of which # points are supposed to be closer, figure out a better way to understand # distances between points. Especially in higher dimensions when Euclidean @@ -91,18 +91,19 @@ def plot(X, Y): # Basically, we learn this distance: # :math:`D(x,y)=\sqrt{(x-y)\,M^{-1}(x-y)}`. And we learn this distance by # learning a Matrix :math:`M`, based on certain constraints. -# -# Some good reading material for the same can be found -# `here `__. It serves as a good -# literature review of Metric Learning. -# +# +# For more information, check the :ref:`intro_metric_learning` section +# from the documentation. Some good reading material can also be found +# `here `__. It serves as a +# good literature review of Metric Learning. +# # We will briefly explain the metric-learning algorithms implemented by # metric-learn, before providing some examples for it's usage, and also # discuss how to go about doing manual constraints. -# +# # Metric-learn can be easily integrated with your other machine learning -# pipelines, and follows (for the most part) scikit-learn conventions. -# +# pipelines, and follows scikit-learn conventions. +# ###################################################################### @@ -122,10 +123,10 @@ def plot(X, Y): # but they are not necessary in a same cluster. This is particular to LMNN # and we'll see that some other algorithms implicitly enforce points from # the same class to cluster together. -# -# You can find the paper -# `here `__. -# +# +# - See more in the :ref:`User Guide ` +# - See more in the documentation of the class :py:class:`LMNN +# ` ###################################################################### @@ -145,18 +146,14 @@ def plot(X, Y): ###################################################################### # So what have we learned? The matrix :math:`M` we talked about before. -# Let's see what it looks like. -# - -lmnn.metric() ###################################################################### # Now let us plot the transformed space - this tells us what the original # space looks like after being transformed with the new learned metric. -# +# -plot(X_lmnn, Y) +plot_tsne(X_lmnn, Y) ###################################################################### @@ -167,8 +164,6 @@ def plot(X, Y): # and then ``transform`` to see our data transformed, we can also use # ``fit_transform``. The rest of the examples and illustrations will use # ``fit_transform``. -# - ###################################################################### # Information Theoretic Metric Learning @@ -179,41 +174,53 @@ def plot(X, Y): # must-link or cannot like constraints, and a simple algorithm based on # Bregman projections. Unlike LMNN, ITML will implicitly enforce points from # the same class to belong to the same cluster, as you can see below. -# -# Link to paper: -# `ITML `__. -# +# +# - See more in the :ref:`User Guide ` +# - See more in the documentation of the class :py:class:`ITML +# ` -itml = metric_learn.ITML_Supervised(num_constraints=200) +itml = metric_learn.ITML_Supervised() X_itml = itml.fit_transform(X, Y) -plot(X_itml, Y) +plot_tsne(X_itml, Y) +###################################################################### +# Mahalanobis Metric for Clustering +# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +# +# MMC is an algorithm that will try to minimize the distance between similar +# points, while ensuring that the sum of distances between dissimilar points is +# higher than a threshold. This is done by optimizing a cost function +# subject to an inequality constraint. +# +# - See more in the :ref:`User Guide ` +# - See more in the documentation of the class :py:class:`MMC +# ` + +itml = metric_learn.ITML_Supervised() +X_itml = itml.fit_transform(X, Y) + +plot_tsne(X_itml, Y) + ###################################################################### # Sparse Determinant Metric Learning # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ -# +# # Implements an efficient sparse metric learning algorithm in high # dimensional space via an :math:`l_1`-penalised log-determinant -# regularization. Compare to the most existing distance metric learning +# regularization. Compared to the most existing distance metric learning # algorithms, the algorithm exploits the sparsity nature underlying the # intrinsic high dimensional feature space. -# -# Link to paper here: -# `SDML `__. -# -# One feature which we'd like to show off here is the use of random seeds. -# Some of the algorithms feature randomised algorithms for selecting -# constraints - to fix these, and get constant results for each run, we -# pass a numpy random seed as shown in the example below. -# +# +# - See more in the :ref:`User Guide ` +# - See more in the documentation of the class :py:class:`SDML +# ` -sdml = metric_learn.SDML_Supervised(num_constraints=200, sparsity_param=0.1, - balance_param=0.0015) -X_sdml = sdml.fit_transform(X, Y, random_state=np.random.RandomState(1234)) +sdml = metric_learn.SDML_Supervised(sparsity_param=0.1, balance_param=0.0015) +X_sdml = sdml.fit_transform(X, Y) -plot(X_sdml, Y) +plot_tsne(X_sdml, Y) ###################################################################### @@ -224,47 +231,42 @@ def plot(X, Y): # metric from a given set of relative comparisons. This is done by # formulating and minimizing a convex loss function that corresponds to # the sum of squared hinge loss of violated constraints. -# -# Link to paper: -# `LSML `__ -# +# +# - See more in the :ref:`User Guide ` +# - See more in the documentation of the class :py:class:`LSML +# ` -lsml = metric_learn.LSML_Supervised(num_constraints=200, tol=0.0001, max_iter=10000, verbose=True) +lsml = metric_learn.LSML_Supervised(tol=0.0001, max_iter=10000) X_lsml = lsml.fit_transform(X, Y) -plot(X_lsml, Y) +plot_tsne(X_lsml, Y) ###################################################################### # Neighborhood Components Analysis # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ -# -# NCA is an extrememly popular metric-learning algorithm, and one of the -# first few (published back in 2005). -# +# +# NCA is an extremly popular metric-learning algorithm. +# # Neighbourhood components analysis aims at "learning" a distance metric # by finding a linear transformation of input data such that the average -# leave-one-out (LOO) classification performance is maximized in the -# transformed space. The key insight to the algorithm is that a matrix -# :math:`A` corresponding to the transformation can be found by defining a -# differentiable objective function for :math:`A`, followed by use of an -# iterative solver such as conjugate gradient descent. One of the benefits -# of this algorithm is that the number of classes :math:`k` can be -# determined as a function of :math:`A`, up to a scalar constant. This use -# of the algorithm therefore addresses the issue of model selection. Like -# LMNN, this algorithm does not try to cluster points from the same class in -# a unique cluster, because it enforces conditions at a local -# neighborhood scale. -# -# You can read more about it in the paper here: -# `NCA `__. -# +# leave-one-out (LOO) classification performance of a soft-nearest +# neighbors rule is maximized in the transformed space. The key insight to +# the algorithm is that a matrix :math:`A` corresponding to the +# transformation can be found by defining a differentiable objective function +# for :math:`A`, followed by use of an iterative solver such as +# `scipy.optimize.fmin_l_bfgs_b`. Like LMNN, this algorithm does not try to +# cluster points from the same class in a unique cluster, because it +# enforces conditions at a local neighborhood scale. +# +# - See more in the :ref:`User Guide ` +# - See more in the documentation of the class :py:class:`NCA +# ` nca = metric_learn.NCA(max_iter=1000) X_nca = nca.fit_transform(X, Y) -plot(X_nca, Y) - +plot_tsne(X_nca, Y) ###################################################################### # Local Fischer Discriminant Analysis @@ -276,15 +278,15 @@ def plot(X, Y): # optimization problem of LFDA is solved as a generalized eigenvalue # problem. Like LMNN, and NCA, this algorithm does not try to cluster points # from the same class in a unique cluster. -# -# Link to paper: -# `LFDA `__ -# +# +# - See more in the :ref:`User Guide ` +# - See more in the documentation of the class :py:class:`LFDA +# ` lfda = metric_learn.LFDA(k=2, num_dims=2) X_lfda = lfda.fit_transform(X, Y) -plot(X_lfda, Y) +plot_tsne(X_lfda, Y) ###################################################################### @@ -297,39 +299,89 @@ def plot(X, Y): # large weights to relevant dimensions and low weights to irrelevant # dimensions. Those relevant dimensions are estimated using "chunklets", # subsets of points that are known to belong to the same class. -# -# Link to paper: -# `RCA `__ -# +# +# - See more in the :ref:`User Guide ` +# - See more in the documentation of the class :py:class:`RCA +# ` rca = metric_learn.RCA_Supervised(num_chunks=30, chunk_size=2) X_rca = rca.fit_transform(X, Y) -plot(X_rca, Y) +plot_tsne(X_rca, Y) +###################################################################### +# Metric Learning for Kernel Regression +# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +# +# An algorithm very similar to NCA but for regression is Metric +# Learning for Kernel Regression (MLKR). It will optimize for the average +# leave-one-out *regression* performance from a soft-nearest neighbors +# regression. +# +# - See more in the :ref:`User Guide ` +# - See more in the documentation of the class :py:class:`MLKR +# ` +# +# To illustrate MLKR, let's use the dataset +# `sklearn.datasets.make_regression` the same way as we did with the +# classification before. The dataset will contain: 100 points of 5 features +# each, among which 3 are random noise but informative (used to generate the +# regression target from a linear model, and two are random noise with the +# same magnitude + +X_reg, Y_reg = make_regression(n_samples=100, n_informative=3, n_features=5, + shuffle=True) ###################################################################### -# Manual Constraints -# ^^^^^^^^^^^^^^^^^^ -# -# Some of the algorithms we've mentioned have alternate ways to pass -# constraints. So far we've been passing them as just class labels - and -# letting the internals of metric-learn deal with creating our constraints. -# -# We'll be looking at one other way to do this - which is to pass a Matrix -# X such that - (a,b,c,d) indices into X, such that -# :math:`d(X[a],X[b]) < d(X[c],X[d])`. -# -# This kind of input is possible for ITML and LSML. -# +# Let's plot the dataset as is + +plot_tsne(X_reg, Y_reg, plt.cm.Oranges) + +###################################################################### +# And let's plot the dataset after transformation by MLKR: +mlkr = metric_learn.MLKR() +X_mlkr = mlkr.fit_transform(X_reg, Y_reg) +plot_tsne(X_mlkr, Y_reg, plt.cm.Oranges) + +###################################################################### +# Points that have the same value to regress are now closer to each +# other ! This could improve the performance of +# `sklearn.neighbors.KNeighborsRegressor` for instance. + + +###################################################################### +# Constraints +# ^^^^^^^^^^^ +# +# To learn the right metric, so far we have always given the labels of the +# data to supervise the algorithms. However, in many applications, +# it is easier to obtain information about whether two samples are +# similar or dissimilar. For instance, when annotating a dataset of face +# images, it is easier for an annotator to tell if two faces belong to the same +# person or not, rather than finding the ID of the face among a huge database +# of every person's faces. +# +# Fortunately, one of the strength of metric learning is the ability to +# deal with such data. Indeed, some of the algorithms we've mentioned have +# alternate ways to pass some supervision about the metric we want to learn. +# The way to go is to pass a 3D array `pairs` of pairs, as well as an array +# of labels `y` such that for each `i` between `0` and `n_pairs` +# we want `X[i, 0, :]` and `X[i, 1, :]` to be similar if `y[i] == 1`, and we +# want them to be dissimilar if `y[i] == -1`. In other words, we want to +# enforce a metric that projects similar points closer together and +# dissimilar points further away from each other. +# (See also the section: :ref:`weakly_supervised_section`.) +# +# This kind of input is possible for ITML, SDML, and MMC. +# # We're going to create these constraints through the labels we have, i.e # :math:`Y`. -# +# # This is done internally through metric learn anyway (do check out the -# ``constraints`` class!) - but we'll try our own version of this. I'm +# `constraints` module!) - but we'll try our own version of this. I'm # going to go ahead and assume that two points labelled the same will be # closer than two points in different labels. -# +# # Do keep in mind that we are doing this method because we know the labels # - we can actually create the constraints any way we want to depending on # the data! @@ -392,7 +444,7 @@ def create_constraints(labels): X_itml = itml.transform(X) -plot(X_itml, Y) +plot_tsne(X_itml, Y) ###################################################################### @@ -401,10 +453,22 @@ def create_constraints(labels): # We can also notice that it might be better to rely on the randomised # algorithms under the hood to make our constraints if we are not very # sure how we want our transformed space to be. -# -# RCA and SDML also have their own specific ways of taking in inputs - +# +# RCA and LSML also have their own specific ways of taking in inputs - # it's worth one's while to poke around in the constraints.py file to see # how exactly this is going on. -# +# +# Finally, one of the main advantages of metric-learn is its out-of-the box +# compatibility with scikit-learn, for doing `model selection +# `__, +# cross-validation, and scoring for instance. Indeed, supervised algorithms are +# regular `sklearn.base.TransformerMixin` that can be plugged into any +# pipeline or cross-validation procedure. And weakly-supervised estimators are +# also compatible with scikit-learn, since their input dataset format described +# above allows to be sliced along the first dimension when doing +# cross-validations (see also this :ref:`section `). See +# also some :ref:`use cases ` where you could use scikit-learn +# estimators. + +######################################################################## # This brings us to the end of this tutorial! Have fun Metric Learning :) -# From edd3a54b74dc71038d40cf8b589ed4dc5af41249 Mon Sep 17 00:00:00 2001 From: William de Vazelhes Date: Mon, 3 Jun 2019 17:57:08 +0200 Subject: [PATCH 8/9] Address https://github.com/metric-learn/metric-learn/pull/180#pullrequestreview-243366233 --- examples/plot_metric_examples.py | 92 ++++++++++++++++---------------- 1 file changed, 47 insertions(+), 45 deletions(-) diff --git a/examples/plot_metric_examples.py b/examples/plot_metric_examples.py index 8057d0f5..bc66f183 100644 --- a/examples/plot_metric_examples.py +++ b/examples/plot_metric_examples.py @@ -76,9 +76,9 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # is based on preserving the original neighborhood of points in the embedding # space, but this original neighborhood is based on the euclidean # distance in the input space, in which the contribution of the noisy -# features is high. So even if points from the same class are close to -# each other in some subspace of the input space, this is not the case in the -# total input space. +# features is high. So even if points from the same class are close to each +# other in some subspace of the input space, this is not the case when +# considering all dimensions of the input space. # # Metric Learning # ^^^^^^^^^^^^^^^ @@ -89,8 +89,10 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # distances are a poor way to measure distance, this becomes very useful. # # Basically, we learn this distance: -# :math:`D(x,y)=\sqrt{(x-y)\,M^{-1}(x-y)}`. And we learn this distance by -# learning a Matrix :math:`M`, based on certain constraints. +# :math:`D(x,y)=\sqrt{(x-y)\,M^{-1}(x-y)}`. And we learn the parameters +# :math:`M` of this distance to satisfy certain constraints on the distance +# between points, for example requiring that points of the same class are +# close together and points of different class are far away. # # For more information, check the :ref:`intro_metric_learning` section # from the documentation. Some good reading material can also be found @@ -98,8 +100,9 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # good literature review of Metric Learning. # # We will briefly explain the metric-learning algorithms implemented by -# metric-learn, before providing some examples for it's usage, and also -# discuss how to go about doing manual constraints. +# metric-learn, before providing some examples for its usage, and also +# discuss how to perform metric learning with weaker supervision than class +# labels. # # Metric-learn can be easily integrated with your other machine learning # pipelines, and follows scikit-learn conventions. @@ -310,9 +313,11 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): plot_tsne(X_rca, Y) ###################################################################### -# Metric Learning for Kernel Regression -# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +# Regression example: Metric Learning for Kernel Regression +# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # +# The previous algorithms took as input a dataset with class labels. Metric +# learning can also be useful for regression, when the labels are real numbers. # An algorithm very similar to NCA but for regression is Metric # Learning for Kernel Regression (MLKR). It will optimize for the average # leave-one-out *regression* performance from a soft-nearest neighbors @@ -326,7 +331,7 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # `sklearn.datasets.make_regression` the same way as we did with the # classification before. The dataset will contain: 100 points of 5 features # each, among which 3 are random noise but informative (used to generate the -# regression target from a linear model, and two are random noise with the +# regression target from a linear model), and two are random noise with the # same magnitude X_reg, Y_reg = make_regression(n_samples=100, n_informative=3, n_features=5, @@ -350,42 +355,46 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): ###################################################################### -# Constraints -# ^^^^^^^^^^^ +# Metric Learning from Weaker Supervision +# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # -# To learn the right metric, so far we have always given the labels of the +# To learn the metric, so far we have always given the labels of the # data to supervise the algorithms. However, in many applications, # it is easier to obtain information about whether two samples are # similar or dissimilar. For instance, when annotating a dataset of face # images, it is easier for an annotator to tell if two faces belong to the same # person or not, rather than finding the ID of the face among a huge database # of every person's faces. -# +# Note that for some problems (e.g., in information +# retrieval where the goal is to rank documents by similarity to a query +# document), there is no notion of individual label but one can gather +# information on which pairs of points are similar or dissimilar. # Fortunately, one of the strength of metric learning is the ability to -# deal with such data. Indeed, some of the algorithms we've mentioned have -# alternate ways to pass some supervision about the metric we want to learn. -# The way to go is to pass a 3D array `pairs` of pairs, as well as an array -# of labels `y` such that for each `i` between `0` and `n_pairs` -# we want `X[i, 0, :]` and `X[i, 1, :]` to be similar if `y[i] == 1`, and we -# want them to be dissimilar if `y[i] == -1`. In other words, we want to -# enforce a metric that projects similar points closer together and -# dissimilar points further away from each other. -# (See also the section: :ref:`weakly_supervised_section`.) -# -# This kind of input is possible for ITML, SDML, and MMC. -# -# We're going to create these constraints through the labels we have, i.e -# :math:`Y`. -# -# This is done internally through metric learn anyway (do check out the -# `constraints` module!) - but we'll try our own version of this. I'm -# going to go ahead and assume that two points labelled the same will be -# closer than two points in different labels. +# learn from such weaker supervision. Indeed, some of the algorithms we've +# used above have alternate ways to pass some supervision about the metric +# we want to learn. The way to go is to pass a 3D array `pairs` of pairs, +# as well as an array of labels `y` such that for each `i` between `0` and +# `n_pairs` we want `X[i, 0, :]` and `X[i, 1, :]` to be similar if `y[i] == +# 1`, and we want them to be dissimilar if `y[i] == -1`. In other words, we +# want to enforce a metric that projects similar points closer together and +# dissimilar points further away from each other. This kind of input is +# possible for ITML, SDML, and MMC. See :ref:`weakly_supervised_section` for +# details on other kinds of weak supervision that some algorithms can work +# with. +# +# For the purpose of this example, we're going to explicitly create these +# pairwise constraints through the labels we have, i.e :math:`Y`. # # Do keep in mind that we are doing this method because we know the labels # - we can actually create the constraints any way we want to depending on # the data! -# +# +# Note that this is what metric-learn did under the hood in the previous +# examples (do check out the +# `constraints` module!) - but we'll try our own version of this. We're +# going to go ahead and assume that two points labelled the same will be +# closer than two points in different labels. + def create_constraints(labels): import itertools @@ -433,11 +442,7 @@ def create_constraints(labels): ###################################################################### -# Using our constraints, let's now train ITML again. We should keep in -# mind that internally, ITML\_Supervised does pretty much the same thing -# we are doing; I was just giving an example to better explain how the -# constraints are structured. -# +# Using our constraints, let's now train ITML again. itml = metric_learn.ITML(preprocessor=X) itml.fit(pairs, pairs_labels) @@ -450,9 +455,6 @@ def create_constraints(labels): ###################################################################### # And that's the result of ITML after being trained on our manual # constraints! A bit different from our old result but not too different. -# We can also notice that it might be better to rely on the randomised -# algorithms under the hood to make our constraints if we are not very -# sure how we want our transformed space to be. # # RCA and LSML also have their own specific ways of taking in inputs - # it's worth one's while to poke around in the constraints.py file to see @@ -466,9 +468,9 @@ def create_constraints(labels): # pipeline or cross-validation procedure. And weakly-supervised estimators are # also compatible with scikit-learn, since their input dataset format described # above allows to be sliced along the first dimension when doing -# cross-validations (see also this :ref:`section `). See -# also some :ref:`use cases ` where you could use scikit-learn -# estimators. +# cross-validations (see also this :ref:`section `). You +# can also look at some :ref:`use cases ` where you could combine +# metric-learning with scikit-learn estimators. ######################################################################## # This brings us to the end of this tutorial! Have fun Metric Learning :) From 0d623b529c5b787d25609aa0822b0239cd27488b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Aur=C3=A9lien?= Date: Wed, 5 Jun 2019 11:17:07 +0200 Subject: [PATCH 9/9] a few updates and minor corrections --- ...es.py => plot_metric_learning_examples.py} | 115 ++++++++++-------- 1 file changed, 62 insertions(+), 53 deletions(-) rename examples/{plot_metric_examples.py => plot_metric_learning_examples.py} (84%) diff --git a/examples/plot_metric_examples.py b/examples/plot_metric_learning_examples.py similarity index 84% rename from examples/plot_metric_examples.py rename to examples/plot_metric_learning_examples.py index bc66f183..fd6cff20 100644 --- a/examples/plot_metric_examples.py +++ b/examples/plot_metric_learning_examples.py @@ -39,7 +39,7 @@ # - 5 features, among which 3 are informative (correlated with the class # labels) and two are random noise with large magnitude -X, Y = make_classification(n_samples=100, n_classes=3, n_clusters_per_class=2, +X, y = make_classification(n_samples=100, n_classes=3, n_clusters_per_class=2, n_informative=3, class_sep=4., n_features=5, n_redundant=0, shuffle=True, scale=[1, 1, 20, 20, 20]) @@ -50,7 +50,7 @@ # `sklearn.manifold.TSNE`). -def plot_tsne(X, Y, colormap=plt.cm.Paired): +def plot_tsne(X, y, colormap=plt.cm.Paired): plt.figure(figsize=(8, 6)) # clean the figure @@ -58,7 +58,7 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): tsne = TSNE() X_embedded = tsne.fit_transform(X) - plt.scatter(X_embedded[:, 0], X_embedded[:, 1], c=Y, cmap=colormap) + plt.scatter(X_embedded[:, 0], X_embedded[:, 1], c=y, cmap=colormap) plt.xticks(()) plt.yticks(()) @@ -69,7 +69,7 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # Let's now plot the dataset as is. -plot_tsne(X, Y) +plot_tsne(X, y) ######################################################################### # We can see that the classes appear mixed up: this is because t-sne @@ -84,12 +84,13 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # ^^^^^^^^^^^^^^^ # # Why is Metric Learning useful? We can, with prior knowledge of which -# points are supposed to be closer, figure out a better way to understand -# distances between points. Especially in higher dimensions when Euclidean -# distances are a poor way to measure distance, this becomes very useful. +# points are supposed to be closer, figure out a better way to compute +# distances between points for the task at hand. Especially in higher +# dimensions when Euclidean distances are a poor way to measure distance, this +# becomes very useful. # # Basically, we learn this distance: -# :math:`D(x,y)=\sqrt{(x-y)\,M^{-1}(x-y)}`. And we learn the parameters +# :math:`D(x, x') = \sqrt{(x-x')^\top M(x-x')}`. And we learn the parameters # :math:`M` of this distance to satisfy certain constraints on the distance # between points, for example requiring that points of the same class are # close together and points of different class are far away. @@ -99,7 +100,7 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # `here `__. It serves as a # good literature review of Metric Learning. # -# We will briefly explain the metric-learning algorithms implemented by +# We will briefly explain the metric learning algorithms implemented by # metric-learn, before providing some examples for its usage, and also # discuss how to perform metric learning with weaker supervision than class # labels. @@ -141,7 +142,7 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): lmnn = metric_learn.LMNN(k=5, learn_rate=1e-6) # fit the data! -lmnn.fit(X, Y) +lmnn.fit(X, y) # transform our input space X_lmnn = lmnn.transform(X) @@ -156,7 +157,7 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # space looks like after being transformed with the new learned metric. # -plot_tsne(X_lmnn, Y) +plot_tsne(X_lmnn, y) ###################################################################### @@ -183,9 +184,9 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # ` itml = metric_learn.ITML_Supervised() -X_itml = itml.fit_transform(X, Y) +X_itml = itml.fit_transform(X, y) -plot_tsne(X_itml, Y) +plot_tsne(X_itml, y) ###################################################################### @@ -202,16 +203,16 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # ` itml = metric_learn.ITML_Supervised() -X_itml = itml.fit_transform(X, Y) +X_itml = itml.fit_transform(X, y) -plot_tsne(X_itml, Y) +plot_tsne(X_itml, y) ###################################################################### # Sparse Determinant Metric Learning # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # # Implements an efficient sparse metric learning algorithm in high -# dimensional space via an :math:`l_1`-penalised log-determinant +# dimensional space via an :math:`l_1`-penalized log-determinant # regularization. Compared to the most existing distance metric learning # algorithms, the algorithm exploits the sparsity nature underlying the # intrinsic high dimensional feature space. @@ -221,9 +222,9 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # ` sdml = metric_learn.SDML_Supervised(sparsity_param=0.1, balance_param=0.0015) -X_sdml = sdml.fit_transform(X, Y) +X_sdml = sdml.fit_transform(X, y) -plot_tsne(X_sdml, Y) +plot_tsne(X_sdml, y) ###################################################################### @@ -240,18 +241,18 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # ` lsml = metric_learn.LSML_Supervised(tol=0.0001, max_iter=10000) -X_lsml = lsml.fit_transform(X, Y) +X_lsml = lsml.fit_transform(X, y) -plot_tsne(X_lsml, Y) +plot_tsne(X_lsml, y) ###################################################################### # Neighborhood Components Analysis # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # -# NCA is an extremly popular metric-learning algorithm. +# NCA is an extremly popular metric learning algorithm. # -# Neighbourhood components analysis aims at "learning" a distance metric +# Neighborhood components analysis aims at "learning" a distance metric # by finding a linear transformation of input data such that the average # leave-one-out (LOO) classification performance of a soft-nearest # neighbors rule is maximized in the transformed space. The key insight to @@ -267,12 +268,12 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # ` nca = metric_learn.NCA(max_iter=1000) -X_nca = nca.fit_transform(X, Y) +X_nca = nca.fit_transform(X, y) -plot_tsne(X_nca, Y) +plot_tsne(X_nca, y) ###################################################################### -# Local Fischer Discriminant Analysis +# Local Fisher Discriminant Analysis # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ # # LFDA is a linear supervised dimensionality reduction method. It is @@ -287,9 +288,9 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # ` lfda = metric_learn.LFDA(k=2, num_dims=2) -X_lfda = lfda.fit_transform(X, Y) +X_lfda = lfda.fit_transform(X, y) -plot_tsne(X_lfda, Y) +plot_tsne(X_lfda, y) ###################################################################### @@ -308,9 +309,9 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # ` rca = metric_learn.RCA_Supervised(num_chunks=30, chunk_size=2) -X_rca = rca.fit_transform(X, Y) +X_rca = rca.fit_transform(X, y) -plot_tsne(X_rca, Y) +plot_tsne(X_rca, y) ###################################################################### # Regression example: Metric Learning for Kernel Regression @@ -330,27 +331,27 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # To illustrate MLKR, let's use the dataset # `sklearn.datasets.make_regression` the same way as we did with the # classification before. The dataset will contain: 100 points of 5 features -# each, among which 3 are random noise but informative (used to generate the +# each, among which 3 are informative (i.e., used to generate the # regression target from a linear model), and two are random noise with the -# same magnitude +# same magnitude. -X_reg, Y_reg = make_regression(n_samples=100, n_informative=3, n_features=5, +X_reg, y_reg = make_regression(n_samples=100, n_informative=3, n_features=5, shuffle=True) ###################################################################### # Let's plot the dataset as is -plot_tsne(X_reg, Y_reg, plt.cm.Oranges) +plot_tsne(X_reg, y_reg, plt.cm.Oranges) ###################################################################### # And let's plot the dataset after transformation by MLKR: mlkr = metric_learn.MLKR() -X_mlkr = mlkr.fit_transform(X_reg, Y_reg) -plot_tsne(X_mlkr, Y_reg, plt.cm.Oranges) +X_mlkr = mlkr.fit_transform(X_reg, y_reg) +plot_tsne(X_mlkr, y_reg, plt.cm.Oranges) ###################################################################### # Points that have the same value to regress are now closer to each -# other ! This could improve the performance of +# other ! This would improve the performance of # `sklearn.neighbors.KNeighborsRegressor` for instance. @@ -372,10 +373,11 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # Fortunately, one of the strength of metric learning is the ability to # learn from such weaker supervision. Indeed, some of the algorithms we've # used above have alternate ways to pass some supervision about the metric -# we want to learn. The way to go is to pass a 3D array `pairs` of pairs, -# as well as an array of labels `y` such that for each `i` between `0` and -# `n_pairs` we want `X[i, 0, :]` and `X[i, 1, :]` to be similar if `y[i] == -# 1`, and we want them to be dissimilar if `y[i] == -1`. In other words, we +# we want to learn. The way to go is to pass a 2D array `pairs` of pairs, +# as well as an array of labels `pairs_labels` such that for each `i` between +# `0` and `n_pairs` we want `X[pairs[i, 0], :]` and `X[pairs[i, 1], :]` to be +# similar if `pairs_labels[i] == 1`, and we want them to be dissimilar if +# `pairs_labels[i] == -1`. In other words, we # want to enforce a metric that projects similar points closer together and # dissimilar points further away from each other. This kind of input is # possible for ITML, SDML, and MMC. See :ref:`weakly_supervised_section` for @@ -383,8 +385,7 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # with. # # For the purpose of this example, we're going to explicitly create these -# pairwise constraints through the labels we have, i.e :math:`Y`. -# +# pairwise constraints through the labels we have, i.e. `y`. # Do keep in mind that we are doing this method because we know the labels # - we can actually create the constraints any way we want to depending on # the data! @@ -392,7 +393,7 @@ def plot_tsne(X, Y, colormap=plt.cm.Paired): # Note that this is what metric-learn did under the hood in the previous # examples (do check out the # `constraints` module!) - but we'll try our own version of this. We're -# going to go ahead and assume that two points labelled the same will be +# going to go ahead and assume that two points labeled the same will be # closer than two points in different labels. @@ -401,9 +402,9 @@ def create_constraints(labels): import random # aggregate indices of same class - zeros = np.where(Y==0)[0] - ones = np.where(Y==1)[0] - twos = np.where(Y==2)[0] + zeros = np.where(y == 0)[0] + ones = np.where(y == 1)[0] + twos = np.where(y == 2)[0] # make permutations of all those points in the same class zeros_ = list(itertools.combinations(zeros, 2)) ones_ = list(itertools.combinations(ones, 2)) @@ -427,10 +428,10 @@ def create_constraints(labels): # return an array of pairs of indices of shape=(2*len(sim), 2), and the corresponding labels, array of shape=(2*len(sim)) # Each pair of similar points have a label of +1 and each pair of dissimilar points have a label of -1 - return (np.vstack([np.column_stack([sim[:,0], sim[:,1]]), np.column_stack([dis[:,0], dis[:,1]])]), + return (np.vstack([np.column_stack([sim[:, 0], sim[:, 1]]), np.column_stack([dis[:, 0], dis[:, 1]])]), np.concatenate([np.ones(len(sim)), -np.ones(len(sim))])) -pairs, pairs_labels = create_constraints(Y) +pairs, pairs_labels = create_constraints(y) ###################################################################### @@ -442,19 +443,27 @@ def create_constraints(labels): ###################################################################### -# Using our constraints, let's now train ITML again. +# Using our constraints, let's now train ITML again. Note that we are no +# longer calling the supervised class :py:class:`ITML_Supervised +# ` but the more generic +# (weakly-supervised) :py:class:`ITML `, which +# takes the dataset `X` through the `preprocessor` argument (see +# :ref:`this section ` of the documentation to learn +# about more advanced uses of `preprocessor`) and the pair information `pairs` +# and `pairs_labels` in the fit method. itml = metric_learn.ITML(preprocessor=X) itml.fit(pairs, pairs_labels) X_itml = itml.transform(X) -plot_tsne(X_itml, Y) +plot_tsne(X_itml, y) ###################################################################### -# And that's the result of ITML after being trained on our manual -# constraints! A bit different from our old result but not too different. +# And that's the result of ITML after being trained on our manually +# constructed constraints! A bit different from our old result, but not too +# different. # # RCA and LSML also have their own specific ways of taking in inputs - # it's worth one's while to poke around in the constraints.py file to see @@ -470,7 +479,7 @@ def create_constraints(labels): # above allows to be sliced along the first dimension when doing # cross-validations (see also this :ref:`section `). You # can also look at some :ref:`use cases ` where you could combine -# metric-learning with scikit-learn estimators. +# metric-learn with scikit-learn estimators. ######################################################################## # This brings us to the end of this tutorial! Have fun Metric Learning :)