OptWok-0.3.1/0000755ڛ|׿-0000000000012140420542017667 5ustar cheng.ongUNIMELB\Domain Users00000000000000OptWok-0.3.1/COPYING0000644ڛ|׿-0000010451311731540564020742 0ustar cheng.ongUNIMELB\Domain Users00000000000000 GNU GENERAL PUBLIC LICENSE Version 3, 29 June 2007 Copyright (C) 2007 Free Software Foundation, Inc. Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it is not allowed. Preamble The GNU General Public License is a free, copyleft license for software and other kinds of works. The licenses for most software and other practical works are designed to take away your freedom to share and change the works. By contrast, the GNU General Public License is intended to guarantee your freedom to share and change all versions of a program--to make sure it remains free software for all its users. We, the Free Software Foundation, use the GNU General Public License for most of our software; it applies also to any other work released this way by its authors. You can apply it to your programs, too. When we speak of free software, we are referring to freedom, not price. Our General Public Licenses are designed to make sure that you have the freedom to distribute copies of free software (and charge for them if you wish), that you receive source code or can get it if you want it, that you can change the software or use pieces of it in new free programs, and that you know you can do these things. To protect your rights, we need to prevent others from denying you these rights or asking you to surrender the rights. Therefore, you have certain responsibilities if you distribute copies of the software, or if you modify it: responsibilities to respect the freedom of others. For example, if you distribute copies of such a program, whether gratis or for a fee, you must pass on to the recipients the same freedoms that you received. You must make sure that they, too, receive or can get the source code. And you must show them these terms so they know their rights. Developers that use the GNU GPL protect your rights with two steps: (1) assert copyright on the software, and (2) offer you this License giving you legal permission to copy, distribute and/or modify it. For the developers' and authors' protection, the GPL clearly explains that there is no warranty for this free software. For both users' and authors' sake, the GPL requires that modified versions be marked as changed, so that their problems will not be attributed erroneously to authors of previous versions. Some devices are designed to deny users access to install or run modified versions of the software inside them, although the manufacturer can do so. This is fundamentally incompatible with the aim of protecting users' freedom to change the software. The systematic pattern of such abuse occurs in the area of products for individuals to use, which is precisely where it is most unacceptable. Therefore, we have designed this version of the GPL to prohibit the practice for those products. If such problems arise substantially in other domains, we stand ready to extend this provision to those domains in future versions of the GPL, as needed to protect the freedom of users. Finally, every program is threatened constantly by software patents. States should not allow patents to restrict development and use of software on general-purpose computers, but in those that do, we wish to avoid the special danger that patents applied to a free program could make it effectively proprietary. To prevent this, the GPL assures that patents cannot be used to render the program non-free. The precise terms and conditions for copying, distribution and modification follow. TERMS AND CONDITIONS 0. Definitions. "This License" refers to version 3 of the GNU General Public License. "Copyright" also means copyright-like laws that apply to other kinds of works, such as semiconductor masks. "The Program" refers to any copyrightable work licensed under this License. Each licensee is addressed as "you". "Licensees" and "recipients" may be individuals or organizations. To "modify" a work means to copy from or adapt all or part of the work in a fashion requiring copyright permission, other than the making of an exact copy. The resulting work is called a "modified version" of the earlier work or a work "based on" the earlier work. A "covered work" means either the unmodified Program or a work based on the Program. To "propagate" a work means to do anything with it that, without permission, would make you directly or secondarily liable for infringement under applicable copyright law, except executing it on a computer or modifying a private copy. Propagation includes copying, distribution (with or without modification), making available to the public, and in some countries other activities as well. To "convey" a work means any kind of propagation that enables other parties to make or receive copies. Mere interaction with a user through a computer network, with no transfer of a copy, is not conveying. An interactive user interface displays "Appropriate Legal Notices" to the extent that it includes a convenient and prominently visible feature that (1) displays an appropriate copyright notice, and (2) tells the user that there is no warranty for the work (except to the extent that warranties are provided), that licensees may convey the work under this License, and how to view a copy of this License. If the interface presents a list of user commands or options, such as a menu, a prominent item in the list meets this criterion. 1. Source Code. The "source code" for a work means the preferred form of the work for making modifications to it. "Object code" means any non-source form of a work. A "Standard Interface" means an interface that either is an official standard defined by a recognized standards body, or, in the case of interfaces specified for a particular programming language, one that is widely used among developers working in that language. The "System Libraries" of an executable work include anything, other than the work as a whole, that (a) is included in the normal form of packaging a Major Component, but which is not part of that Major Component, and (b) serves only to enable use of the work with that Major Component, or to implement a Standard Interface for which an implementation is available to the public in source code form. A "Major Component", in this context, means a major essential component (kernel, window system, and so on) of the specific operating system (if any) on which the executable work runs, or a compiler used to produce the work, or an object code interpreter used to run it. The "Corresponding Source" for a work in object code form means all the source code needed to generate, install, and (for an executable work) run the object code and to modify the work, including scripts to control those activities. However, it does not include the work's System Libraries, or general-purpose tools or generally available free programs which are used unmodified in performing those activities but which are not part of the work. For example, Corresponding Source includes interface definition files associated with source files for the work, and the source code for shared libraries and dynamically linked subprograms that the work is specifically designed to require, such as by intimate data communication or control flow between those subprograms and other parts of the work. The Corresponding Source need not include anything that users can regenerate automatically from other parts of the Corresponding Source. The Corresponding Source for a work in source code form is that same work. 2. Basic Permissions. All rights granted under this License are granted for the term of copyright on the Program, and are irrevocable provided the stated conditions are met. This License explicitly affirms your unlimited permission to run the unmodified Program. The output from running a covered work is covered by this License only if the output, given its content, constitutes a covered work. This License acknowledges your rights of fair use or other equivalent, as provided by copyright law. You may make, run and propagate covered works that you do not convey, without conditions so long as your license otherwise remains in force. You may convey covered works to others for the sole purpose of having them make modifications exclusively for you, or provide you with facilities for running those works, provided that you comply with the terms of this License in conveying all material for which you do not control copyright. Those thus making or running the covered works for you must do so exclusively on your behalf, under your direction and control, on terms that prohibit them from making any copies of your copyrighted material outside their relationship with you. Conveying under any other circumstances is permitted solely under the conditions stated below. Sublicensing is not allowed; section 10 makes it unnecessary. 3. Protecting Users' Legal Rights From Anti-Circumvention Law. No covered work shall be deemed part of an effective technological measure under any applicable law fulfilling obligations under article 11 of the WIPO copyright treaty adopted on 20 December 1996, or similar laws prohibiting or restricting circumvention of such measures. When you convey a covered work, you waive any legal power to forbid circumvention of technological measures to the extent such circumvention is effected by exercising rights under this License with respect to the covered work, and you disclaim any intention to limit operation or modification of the work as a means of enforcing, against the work's users, your or third parties' legal rights to forbid circumvention of technological measures. 4. Conveying Verbatim Copies. You may convey verbatim copies of the Program's source code as you receive it, in any medium, provided that you conspicuously and appropriately publish on each copy an appropriate copyright notice; keep intact all notices stating that this License and any non-permissive terms added in accord with section 7 apply to the code; keep intact all notices of the absence of any warranty; and give all recipients a copy of this License along with the Program. You may charge any price or no price for each copy that you convey, and you may offer support or warranty protection for a fee. 5. Conveying Modified Source Versions. You may convey a work based on the Program, or the modifications to produce it from the Program, in the form of source code under the terms of section 4, provided that you also meet all of these conditions: a) The work must carry prominent notices stating that you modified it, and giving a relevant date. b) The work must carry prominent notices stating that it is released under this License and any conditions added under section 7. This requirement modifies the requirement in section 4 to "keep intact all notices". c) You must license the entire work, as a whole, under this License to anyone who comes into possession of a copy. This License will therefore apply, along with any applicable section 7 additional terms, to the whole of the work, and all its parts, regardless of how they are packaged. This License gives no permission to license the work in any other way, but it does not invalidate such permission if you have separately received it. d) If the work has interactive user interfaces, each must display Appropriate Legal Notices; however, if the Program has interactive interfaces that do not display Appropriate Legal Notices, your work need not make them do so. A compilation of a covered work with other separate and independent works, which are not by their nature extensions of the covered work, and which are not combined with it such as to form a larger program, in or on a volume of a storage or distribution medium, is called an "aggregate" if the compilation and its resulting copyright are not used to limit the access or legal rights of the compilation's users beyond what the individual works permit. Inclusion of a covered work in an aggregate does not cause this License to apply to the other parts of the aggregate. 6. Conveying Non-Source Forms. You may convey a covered work in object code form under the terms of sections 4 and 5, provided that you also convey the machine-readable Corresponding Source under the terms of this License, in one of these ways: a) Convey the object code in, or embodied in, a physical product (including a physical distribution medium), accompanied by the Corresponding Source fixed on a durable physical medium customarily used for software interchange. b) Convey the object code in, or embodied in, a physical product (including a physical distribution medium), accompanied by a written offer, valid for at least three years and valid for as long as you offer spare parts or customer support for that product model, to give anyone who possesses the object code either (1) a copy of the Corresponding Source for all the software in the product that is covered by this License, on a durable physical medium customarily used for software interchange, for a price no more than your reasonable cost of physically performing this conveying of source, or (2) access to copy the Corresponding Source from a network server at no charge. c) Convey individual copies of the object code with a copy of the written offer to provide the Corresponding Source. This alternative is allowed only occasionally and noncommercially, and only if you received the object code with such an offer, in accord with subsection 6b. d) Convey the object code by offering access from a designated place (gratis or for a charge), and offer equivalent access to the Corresponding Source in the same way through the same place at no further charge. You need not require recipients to copy the Corresponding Source along with the object code. If the place to copy the object code is a network server, the Corresponding Source may be on a different server (operated by you or a third party) that supports equivalent copying facilities, provided you maintain clear directions next to the object code saying where to find the Corresponding Source. Regardless of what server hosts the Corresponding Source, you remain obligated to ensure that it is available for as long as needed to satisfy these requirements. e) Convey the object code using peer-to-peer transmission, provided you inform other peers where the object code and Corresponding Source of the work are being offered to the general public at no charge under subsection 6d. A separable portion of the object code, whose source code is excluded from the Corresponding Source as a System Library, need not be included in conveying the object code work. A "User Product" is either (1) a "consumer product", which means any tangible personal property which is normally used for personal, family, or household purposes, or (2) anything designed or sold for incorporation into a dwelling. In determining whether a product is a consumer product, doubtful cases shall be resolved in favor of coverage. For a particular product received by a particular user, "normally used" refers to a typical or common use of that class of product, regardless of the status of the particular user or of the way in which the particular user actually uses, or expects or is expected to use, the product. A product is a consumer product regardless of whether the product has substantial commercial, industrial or non-consumer uses, unless such uses represent the only significant mode of use of the product. "Installation Information" for a User Product means any methods, procedures, authorization keys, or other information required to install and execute modified versions of a covered work in that User Product from a modified version of its Corresponding Source. The information must suffice to ensure that the continued functioning of the modified object code is in no case prevented or interfered with solely because modification has been made. If you convey an object code work under this section in, or with, or specifically for use in, a User Product, and the conveying occurs as part of a transaction in which the right of possession and use of the User Product is transferred to the recipient in perpetuity or for a fixed term (regardless of how the transaction is characterized), the Corresponding Source conveyed under this section must be accompanied by the Installation Information. But this requirement does not apply if neither you nor any third party retains the ability to install modified object code on the User Product (for example, the work has been installed in ROM). The requirement to provide Installation Information does not include a requirement to continue to provide support service, warranty, or updates for a work that has been modified or installed by the recipient, or for the User Product in which it has been modified or installed. Access to a network may be denied when the modification itself materially and adversely affects the operation of the network or violates the rules and protocols for communication across the network. Corresponding Source conveyed, and Installation Information provided, in accord with this section must be in a format that is publicly documented (and with an implementation available to the public in source code form), and must require no special password or key for unpacking, reading or copying. 7. Additional Terms. "Additional permissions" are terms that supplement the terms of this License by making exceptions from one or more of its conditions. Additional permissions that are applicable to the entire Program shall be treated as though they were included in this License, to the extent that they are valid under applicable law. If additional permissions apply only to part of the Program, that part may be used separately under those permissions, but the entire Program remains governed by this License without regard to the additional permissions. When you convey a copy of a covered work, you may at your option remove any additional permissions from that copy, or from any part of it. (Additional permissions may be written to require their own removal in certain cases when you modify the work.) You may place additional permissions on material, added by you to a covered work, for which you have or can give appropriate copyright permission. Notwithstanding any other provision of this License, for material you add to a covered work, you may (if authorized by the copyright holders of that material) supplement the terms of this License with terms: a) Disclaiming warranty or limiting liability differently from the terms of sections 15 and 16 of this License; or b) Requiring preservation of specified reasonable legal notices or author attributions in that material or in the Appropriate Legal Notices displayed by works containing it; or c) Prohibiting misrepresentation of the origin of that material, or requiring that modified versions of such material be marked in reasonable ways as different from the original version; or d) Limiting the use for publicity purposes of names of licensors or authors of the material; or e) Declining to grant rights under trademark law for use of some trade names, trademarks, or service marks; or f) Requiring indemnification of licensors and authors of that material by anyone who conveys the material (or modified versions of it) with contractual assumptions of liability to the recipient, for any liability that these contractual assumptions directly impose on those licensors and authors. All other non-permissive additional terms are considered "further restrictions" within the meaning of section 10. If the Program as you received it, or any part of it, contains a notice stating that it is governed by this License along with a term that is a further restriction, you may remove that term. If a license document contains a further restriction but permits relicensing or conveying under this License, you may add to a covered work material governed by the terms of that license document, provided that the further restriction does not survive such relicensing or conveying. If you add terms to a covered work in accord with this section, you must place, in the relevant source files, a statement of the additional terms that apply to those files, or a notice indicating where to find the applicable terms. Additional terms, permissive or non-permissive, may be stated in the form of a separately written license, or stated as exceptions; the above requirements apply either way. 8. Termination. You may not propagate or modify a covered work except as expressly provided under this License. Any attempt otherwise to propagate or modify it is void, and will automatically terminate your rights under this License (including any patent licenses granted under the third paragraph of section 11). However, if you cease all violation of this License, then your license from a particular copyright holder is reinstated (a) provisionally, unless and until the copyright holder explicitly and finally terminates your license, and (b) permanently, if the copyright holder fails to notify you of the violation by some reasonable means prior to 60 days after the cessation. Moreover, your license from a particular copyright holder is reinstated permanently if the copyright holder notifies you of the violation by some reasonable means, this is the first time you have received notice of violation of this License (for any work) from that copyright holder, and you cure the violation prior to 30 days after your receipt of the notice. Termination of your rights under this section does not terminate the licenses of parties who have received copies or rights from you under this License. If your rights have been terminated and not permanently reinstated, you do not qualify to receive new licenses for the same material under section 10. 9. Acceptance Not Required for Having Copies. You are not required to accept this License in order to receive or run a copy of the Program. Ancillary propagation of a covered work occurring solely as a consequence of using peer-to-peer transmission to receive a copy likewise does not require acceptance. However, nothing other than this License grants you permission to propagate or modify any covered work. These actions infringe copyright if you do not accept this License. Therefore, by modifying or propagating a covered work, you indicate your acceptance of this License to do so. 10. Automatic Licensing of Downstream Recipients. Each time you convey a covered work, the recipient automatically receives a license from the original licensors, to run, modify and propagate that work, subject to this License. You are not responsible for enforcing compliance by third parties with this License. An "entity transaction" is a transaction transferring control of an organization, or substantially all assets of one, or subdividing an organization, or merging organizations. If propagation of a covered work results from an entity transaction, each party to that transaction who receives a copy of the work also receives whatever licenses to the work the party's predecessor in interest had or could give under the previous paragraph, plus a right to possession of the Corresponding Source of the work from the predecessor in interest, if the predecessor has it or can get it with reasonable efforts. You may not impose any further restrictions on the exercise of the rights granted or affirmed under this License. For example, you may not impose a license fee, royalty, or other charge for exercise of rights granted under this License, and you may not initiate litigation (including a cross-claim or counterclaim in a lawsuit) alleging that any patent claim is infringed by making, using, selling, offering for sale, or importing the Program or any portion of it. 11. Patents. A "contributor" is a copyright holder who authorizes use under this License of the Program or a work on which the Program is based. The work thus licensed is called the contributor's "contributor version". A contributor's "essential patent claims" are all patent claims owned or controlled by the contributor, whether already acquired or hereafter acquired, that would be infringed by some manner, permitted by this License, of making, using, or selling its contributor version, but do not include claims that would be infringed only as a consequence of further modification of the contributor version. For purposes of this definition, "control" includes the right to grant patent sublicenses in a manner consistent with the requirements of this License. Each contributor grants you a non-exclusive, worldwide, royalty-free patent license under the contributor's essential patent claims, to make, use, sell, offer for sale, import and otherwise run, modify and propagate the contents of its contributor version. In the following three paragraphs, a "patent license" is any express agreement or commitment, however denominated, not to enforce a patent (such as an express permission to practice a patent or covenant not to sue for patent infringement). To "grant" such a patent license to a party means to make such an agreement or commitment not to enforce a patent against the party. If you convey a covered work, knowingly relying on a patent license, and the Corresponding Source of the work is not available for anyone to copy, free of charge and under the terms of this License, through a publicly available network server or other readily accessible means, then you must either (1) cause the Corresponding Source to be so available, or (2) arrange to deprive yourself of the benefit of the patent license for this particular work, or (3) arrange, in a manner consistent with the requirements of this License, to extend the patent license to downstream recipients. "Knowingly relying" means you have actual knowledge that, but for the patent license, your conveying the covered work in a country, or your recipient's use of the covered work in a country, would infringe one or more identifiable patents in that country that you have reason to believe are valid. If, pursuant to or in connection with a single transaction or arrangement, you convey, or propagate by procuring conveyance of, a covered work, and grant a patent license to some of the parties receiving the covered work authorizing them to use, propagate, modify or convey a specific copy of the covered work, then the patent license you grant is automatically extended to all recipients of the covered work and works based on it. A patent license is "discriminatory" if it does not include within the scope of its coverage, prohibits the exercise of, or is conditioned on the non-exercise of one or more of the rights that are specifically granted under this License. You may not convey a covered work if you are a party to an arrangement with a third party that is in the business of distributing software, under which you make payment to the third party based on the extent of your activity of conveying the work, and under which the third party grants, to any of the parties who would receive the covered work from you, a discriminatory patent license (a) in connection with copies of the covered work conveyed by you (or copies made from those copies), or (b) primarily for and in connection with specific products or compilations that contain the covered work, unless you entered into that arrangement, or that patent license was granted, prior to 28 March 2007. Nothing in this License shall be construed as excluding or limiting any implied license or other defenses to infringement that may otherwise be available to you under applicable patent law. 12. No Surrender of Others' Freedom. If conditions are imposed on you (whether by court order, agreement or otherwise) that contradict the conditions of this License, they do not excuse you from the conditions of this License. If you cannot convey a covered work so as to satisfy simultaneously your obligations under this License and any other pertinent obligations, then as a consequence you may not convey it at all. For example, if you agree to terms that obligate you to collect a royalty for further conveying from those to whom you convey the Program, the only way you could satisfy both those terms and this License would be to refrain entirely from conveying the Program. 13. Use with the GNU Affero General Public License. Notwithstanding any other provision of this License, you have permission to link or combine any covered work with a work licensed under version 3 of the GNU Affero General Public License into a single combined work, and to convey the resulting work. The terms of this License will continue to apply to the part which is the covered work, but the special requirements of the GNU Affero General Public License, section 13, concerning interaction through a network will apply to the combination as such. 14. Revised Versions of this License. The Free Software Foundation may publish revised and/or new versions of the GNU General Public License from time to time. Such new versions will be similar in spirit to the present version, but may differ in detail to address new problems or concerns. Each version is given a distinguishing version number. If the Program specifies that a certain numbered version of the GNU General Public License "or any later version" applies to it, you have the option of following the terms and conditions either of that numbered version or of any later version published by the Free Software Foundation. If the Program does not specify a version number of the GNU General Public License, you may choose any version ever published by the Free Software Foundation. If the Program specifies that a proxy can decide which future versions of the GNU General Public License can be used, that proxy's public statement of acceptance of a version permanently authorizes you to choose that version for the Program. Later license versions may give you additional or different permissions. However, no additional obligations are imposed on any author or copyright holder as a result of your choosing to follow a later version. 15. Disclaimer of Warranty. THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 16. Limitation of Liability. IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. 17. Interpretation of Sections 15 and 16. If the disclaimer of warranty and limitation of liability provided above cannot be given local legal effect according to their terms, reviewing courts shall apply local law that most closely approximates an absolute waiver of all civil liability in connection with the Program, unless a warranty or assumption of liability accompanies a copy of the Program in return for a fee. END OF TERMS AND CONDITIONS How to Apply These Terms to Your New Programs If you develop a new program, and you want it to be of the greatest possible use to the public, the best way to achieve this is to make it free software which everyone can redistribute and change under these terms. To do so, attach the following notices to the program. It is safest to attach them to the start of each source file to most effectively state the exclusion of warranty; and each file should have at least the "copyright" line and a pointer to where the full notice is found. Copyright (C) This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with this program. If not, see . Also add information on how to contact you by electronic and paper mail. If the program does terminal interaction, make it output a short notice like this when it starts in an interactive mode: Copyright (C) This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. This is free software, and you are welcome to redistribute it under certain conditions; type `show c' for details. The hypothetical commands `show w' and `show c' should show the appropriate parts of the General Public License. Of course, your program's commands might be different; for a GUI interface, you would use an "about box". You should also get your employer (if you work as a programmer) or school, if any, to sign a "copyright disclaimer" for the program, if necessary. For more information on this, and how to apply and follow the GNU GPL, see . The GNU General Public License does not permit incorporating your program into proprietary programs. If your program is a subroutine library, you may consider it more useful to permit linking proprietary applications with the library. If this is what you want to do, use the GNU Lesser General Public License instead of this License. But first, please read . OptWok-0.3.1/examples/0000755ڛ|׿-0000000000012140420542021505 5ustar cheng.ongUNIMELB\Domain Users00000000000000OptWok-0.3.1/examples/__init__.py0000644ڛ|׿-0000000000012127456646023627 0ustar cheng.ongUNIMELB\Domain Users00000000000000OptWok-0.3.1/examples/classifier_demo.py0000644ڛ|׿-0000002457212055542136025232 0ustar cheng.ongUNIMELB\Domain Users00000000000000""" Some code for testing and also as a demo """ from optwok.classifier import SVMPrimal, OneNormSVM, CapOneNormSVM, ZeroNormSVM from optwok.classifier import RLSVM2, RLSVM1, RLSVMC1, RLSVM0 from optwok.classifier import OneNormSVM_dense from optwok.evaluation import accuracy from optwok.simdata import create_data from optwok.testtools import plot_2d_pred, plot_2d import pylab import numpy from pprint import pprint from numpy import dot, eye, array, sign, sum, arange from numpy.linalg import norm, cholesky, LinAlgError, eig, svd def example_cloud_2d(): """Use 2d gaussian clouds for training, plot the resulting classifier""" numpy.random.seed(1) (train_ex, train_lab, test_ex, test_labels) = create_data(num_train=200, num_test=10, num_feat=2, num_noise_feat=0, width=0.7, frac_flip=0.2) mu = 1.0 reg_param = 1.0 verbose = False kdict = {'name':'linear', 'param': 0} #kdict = {'name':'poly', 'param': 3, 'inhomog': True} # Train svm = SVMPrimal(reg_param, kdict, verbose=verbose) svm.train(train_ex, train_lab) rlsvm = RLSVM2(reg_param, kdict, mu, rand_init=False, verbose=verbose) rlsvm.train(train_ex, train_lab) # Predict on grid plot_limit = 4.0 axis_limit = [-plot_limit, plot_limit, -plot_limit, plot_limit] x1 = numpy.linspace(-plot_limit, plot_limit, 100) x2 = numpy.linspace(-plot_limit, plot_limit, 100) x,y = numpy.meshgrid(x1,x2) pred_ex = numpy.array((numpy.ravel(x), numpy.ravel(y))) svmout = svm.predict(pred_ex) svmout.shape = (100,100) rlsvmout = rlsvm.predict(pred_ex) rlsvmout.shape = (100,100) print('=====================================================================') print('SVM Primal objective: %f' % svm.primal_obj) print('SVM weights: ') print(svm.w) print('SVM bias: %f' % svm.b) print('=====================================================================') print('RLSVM Primal objective: %f' % rlsvm.primal_obj) print('RLSVM weights: ') print(rlsvm.w) print('RLSVM bias: %f' % rlsvm.b) print('Number of passes through data: %d' % rlsvm.num_iter) plot_2d_pred(x, y, svmout, train_ex, train_lab, axis_limit, 'svm') plot_2d_pred(x, y, rlsvmout, train_ex, train_lab, axis_limit, 'rlsvm') pylab.show() def demo_helper(classifier, train_ex, train_labels, test_ex, test_labels): """Helper script to train, predict and report performance""" classifier.train(train_ex, train_labels) svmout = classifier.predict(test_ex) acc = accuracy(svmout, test_labels) print('Accuracy:\t\t\t %1.2f' % acc) #print('Margin:\t\t\t\t %f' % classifier.margin) #print('Empirical Risk:\t\t\t %f' % sum(classifier.emp_risk)) #print('Primal Objective:\t\t %f' % classifier.primal_obj) #print(numpy.where(abs(classifier.w)>classifier.epsilon)[0]) #print('Fraction of features used:\t %1.4f' % classifier.frac_feat) #print('Fraction support vectors:\t %1.4f' % classifier.frac_sv) #print('Total fraction:\t\t\t %1.4f' % classifier.frac_total) pprint(classifier.perf) def demo_svm(reg_param, kdict, verbose, train_ex, train_labels, test_ex, test_labels): """Create a SVM, train and predict""" print('=== Running SVM ===') svm = SVMPrimal(reg_param, kdict, verbose=verbose) demo_helper(svm, train_ex, train_labels, test_ex, test_labels) def demo_onenormsvm(reg_param, kdict, verbose, train_ex, train_labels, test_ex, test_labels): """Create a one norm SVM, train and predict""" print('=== Running one norm SVM ===') svm = OneNormSVM(reg_param, kdict, verbose=verbose) demo_helper(svm, train_ex, train_labels, test_ex, test_labels) def demo_onenormsvm_dense(reg_param, kdict, verbose, train_ex, train_labels, test_ex, test_labels): """Create a one norm SVM, train and predict""" print('=== Running one norm SVM with dense matrices ===') svm = OneNormSVM_dense(reg_param, kdict, verbose=verbose) demo_helper(svm, train_ex, train_labels, test_ex, test_labels) def demo_caponenormsvm(reg_param, kdict, a, verbose, train_ex, train_labels, test_ex, test_labels): """Create a capped one norm SVM, train and predict""" print('=== Running cap one norm SVM ===') svm = CapOneNormSVM(reg_param, kdict, a, rand_init=False, verbose=verbose) demo_helper(svm, train_ex, train_labels, test_ex, test_labels) def demo_zeronormsvm(reg_param, kdict, a, verbose, train_ex, train_labels, test_ex, test_labels): """Create a one norm SVM, train and predict""" print('=== Running zero norm SVM ===') svm = ZeroNormSVM(reg_param, kdict, a, rand_init=False, verbose=verbose) demo_helper(svm, train_ex, train_labels, test_ex, test_labels) def demo_rlsvm2(reg_param, kdict, mu, verbose, train_ex, train_labels, test_ex, test_labels): """Create an RLSVM, train and predict""" print('=== Running RLSVM2 ===') rlsvm = RLSVM2(reg_param, kdict, mu, rand_init=False, verbose=verbose) demo_helper(rlsvm, train_ex, train_labels, test_ex, test_labels) def demo_rlsvm1(reg_param, kdict, mu, verbose, train_ex, train_labels, test_ex, test_labels): """Create an RLSVM, train and predict""" print('=== Running RLSVM1 ===') rlsvm = RLSVM1(reg_param, kdict, mu, rand_init=False, verbose=verbose) demo_helper(rlsvm, train_ex, train_labels, test_ex, test_labels) def demo_rlsvmc1(reg_param, kdict, mu, a, verbose, train_ex, train_labels, test_ex, test_labels): """Create an RLSVM, train and predict""" print('=== Running RLSVMC1 ===') rlsvm = RLSVMC1(reg_param, kdict, mu, a, rand_init=False, verbose=verbose) demo_helper(rlsvm, train_ex, train_labels, test_ex, test_labels) def demo_rlsvm0(reg_param, kdict, mu, a, verbose, train_ex, train_labels, test_ex, test_labels): """Create an RLSVM, train and predict""" print('=== Running RLSVM0 ===') rlsvm = RLSVM0(reg_param, kdict, mu, a, rand_init=False, verbose=verbose) demo_helper(rlsvm, train_ex, train_labels, test_ex, test_labels) def example_toy(): """Training and classification using two gaussian clouds""" #numpy.random.seed(1) (train_ex, train_labels, test_ex, test_labels) = create_data(num_train=200, num_test=1000, num_feat=5, num_noise_feat = 45, width=1.5, frac_flip=0.1, symm_flip=True) mu = 1.0 a = 4.0 # the exponent in zero norm cap = 0.5 # the cap in capped one norm reg_param = 1.0 verbose = False #kdict = {'name':'poly', 'param': 2, 'inhomog': True} kdict = {'name':'linear', 'param': 0} demo_svm(reg_param, kdict, verbose, train_ex, train_labels, test_ex, test_labels) demo_onenormsvm(reg_param, kdict, verbose, train_ex, train_labels, test_ex, test_labels) demo_caponenormsvm(reg_param, kdict, cap, verbose, train_ex, train_labels, test_ex, test_labels) demo_zeronormsvm(reg_param, kdict, a, verbose, train_ex, train_labels, test_ex, test_labels) demo_rlsvm2(reg_param, kdict, mu, verbose, train_ex, train_labels, test_ex, test_labels) demo_rlsvm1(reg_param, kdict, mu, verbose, train_ex, train_labels, test_ex, test_labels) demo_rlsvmc1(reg_param, kdict, mu, cap, verbose, train_ex, train_labels, test_ex, test_labels) demo_rlsvm0(reg_param, kdict, mu, a, verbose, train_ex, train_labels, test_ex, test_labels) def example_debug(): """Tiny example to try to check global optimality""" plot_limit = 4.0 axis_limit = [-plot_limit, plot_limit, -plot_limit, plot_limit] train_ex = numpy.array([[-1.0, -1.0, -1.0, -1.0, -1.0, 1.0, 1.0, 1.0, 1.0, 1.0], [-1.0, 0.5, 0.0, -0.5, 1.0, 0.5, -0.5, 0.0, 1.0, -1.0]]) train_lab = numpy.array([-1., -1., -1., -1., 1., 1., 1., 1.,1.,1.]) x1 = numpy.linspace(-plot_limit, plot_limit, 100) x2 = numpy.linspace(-plot_limit, plot_limit, 100) x,y = numpy.meshgrid(x1,x2) pred_ex = numpy.array((numpy.ravel(x), numpy.ravel(y))) verbose = False mu = 1.0 C = 1.0 kdict = {'name':'linear', 'param': 0} svm = OneNormSVM(C, kdict, verbose=verbose) svm.train(train_ex, train_lab) print('=====================================================================') print('OneNormSVM objective: %f' % svm.primal_obj) print('SVM weights: ') print(svm.w) print('SVM bias: %f' % svm.b) print('=====================================================================') svmout = svm.predict(pred_ex) svmout.shape = (100,100) del svm svm = RLSVM1(C, kdict, mu, verbose=verbose) svm.train(train_ex, train_lab) print('=====================================================================') print('RLSVM1 Primal objective: %f' % svm.primal_obj) print('RLSVM1 weights: ') print(svm.w) print('RLSVM1 bias: %f' % svm.b) print('Number of passes through data: %d' % svm.num_iter) print('=====================================================================') rlsvmout = svm.predict(pred_ex) rlsvmout.shape = (100,100) del svm plot_2d_pred(x, y, svmout, train_ex, train_lab, axis_limit, 'svm') plot_2d_pred(x, y, rlsvmout, train_ex, train_lab, axis_limit, 'dcsvm') pylab.show() def example_sparse(): """Training and classification using two gaussian clouds, use sparse classifiers """ #numpy.random.seed(1) (train_ex, train_labels, test_ex, test_labels) = create_data(num_train=200, num_test=1000, num_feat=5, num_noise_feat = 45, width=1.5, frac_flip=0.1, symm_flip=True) mu = 1.0 a = 4.0 # the exponent in zero norm cap = 0.5 # the cap in capped one norm reg_param = 1.0 verbose = False #kdict = {'name':'poly', 'param': 2, 'inhomog': True} kdict = {'name':'linear', 'param': 0} demo_onenormsvm_dense(reg_param, kdict, verbose, train_ex, train_labels, test_ex, test_labels) demo_onenormsvm(reg_param, kdict, verbose, train_ex, train_labels, test_ex, test_labels) if __name__ == '__main__': example_debug() example_toy() example_cloud_2d() #example_sparse() OptWok-0.3.1/examples/contextucb_demo.py0000644ڛ|׿-0000000603612127456765025273 0ustar cheng.ongUNIMELB\Domain Users00000000000000""" Demo to show how to use the contextual bandits optimization problem using the upper confidence bound Gaussian process """ from pylab import show from numpy import ones, eye from optwok.kernel import GaussKernel from optwok.testtools import list2matrix from optwok.io_pickle import load from optwok.gaussian_proc import GaussianProcess from optwok import datatools from optwok.context_ucb import expt_cucb, expt_cucb_batch from optwok.context_ucb import plot_min_regret, plot_average_regret from examples.gaussianproc_demo import plot_prediction def demo_expt_cucb(): raw_data = load('tests/gaussproc_06_20_0.1.pkl') data = datatools.Data(raw_data.examples, raw_data.context, raw_data.labels) print(data.examples(1)) num_context = 6 reg_param = 0.1 width = 0.5 num_epoch = 10 Ky = list2matrix([[0,1],[2],[3,4,5]], neg_corr=True) print(Ky) gp = GaussianProcess(GaussKernel, Ky, kparam=width, reg_param=reg_param) gp.train(raw_data.examples[:,:1], raw_data.context[:1], raw_data.labels[:1]) expt_cucb(data, gp, 'expt_cucb.pkl', num_context, num_epoch) Ky = ones((num_context, num_context)) gp = GaussianProcess(GaussKernel, Ky, kparam=width, reg_param=reg_param) gp.train(raw_data.examples[:,:1], raw_data.context[:1], raw_data.labels[:1]) expt_cucb(data, gp, 'expt_cucb_cat.pkl', num_context, num_epoch) Ky = eye(num_context) gp = GaussianProcess(GaussKernel, Ky, kparam=width, reg_param=reg_param) gp.train(raw_data.examples[:,:1], raw_data.context[:1], raw_data.labels[:1]) expt_cucb(data, gp, 'expt_cucb_ind.pkl', num_context, num_epoch) plot_min_regret('expt_cucb.pkl') plot_min_regret('expt_cucb_cat.pkl') plot_min_regret('expt_cucb_ind.pkl') plot_average_regret(['expt_cucb.pkl','expt_cucb_cat.pkl','expt_cucb_ind.pkl'], ['CGP-UCB','merge','ignore']) def demo_expt_cucb_batch(): raw_data = load('tests/gaussproc_06_20_0.1.pkl') data = datatools.Data(raw_data.examples, raw_data.context, raw_data.labels) num_context = 6 reg_param = 0.1 width = 0.5 num_epoch = 3 batch_size = 3 Ky = list2matrix([[0,1],[2],[3,4,5]], neg_corr=True) print(Ky) gp = GaussianProcess(GaussKernel, Ky, kparam=width, reg_param=reg_param) gp.train(raw_data.examples[:,:1], raw_data.context[:1], raw_data.labels[:1]) expt_cucb_batch(data, gp, 'expt_cucb_kbest.pkl', num_context, num_epoch, batch_size=batch_size, mode='kbest') del gp gp = GaussianProcess(GaussKernel, Ky, kparam=width, reg_param=reg_param) gp.train(raw_data.examples[:,:1], raw_data.context[:1], raw_data.labels[:1]) expt_cucb_batch(data, gp, 'expt_cucb_batch.pkl', num_context, num_epoch, batch_size=batch_size) del gp plot_min_regret('expt_cucb_batch.pkl') plot_min_regret('expt_cucb_kbest.pkl') plot_average_regret(['expt_cucb_batch.pkl','expt_cucb_kbest.pkl'], ['greedy','kbest']) if __name__ == '__main__': demo_expt_cucb() demo_expt_cucb_batch() show() OptWok-0.3.1/examples/emil_demo.py0000644ڛ|׿-0000000467512140420225024023 0ustar cheng.ongUNIMELB\Domain Users00000000000000""" Some code for testing and also as a usage demo See bioweb.me/emil for benchmark experiments """ from numpy import r_ from numpy.random import seed from optwok.mil.mi_data import EllipsoidData, MIMeanData from optwok.utility import create_perm, permfile_name from emil_synthetic_ellipsoids import random_ellipsoids from emil_plotting import plot_ellipsoids, plot_hyperplane from optwok.mil.mi_classifier import MeanSVM from optwok.mil.mi_classifier_emil import eMILDcaSocp, eMILDcaBfgs, eMILBfgs from matplotlib.pyplot import show, figure from optwok.performance_measures import get_conf_matrix, accuracy def example_ellipsoid_2d(): """Create some random ellipsoids, train and plot the resulting classifier""" reg_param = 1.5 num_pos = 13 num_neg = 13 num_ex = num_pos + num_neg num_feat = 2 seed(1) # Create some data perm_desc = 'example' perm_dataset = 'synthetic_ellipsoid_2d' create_perm('.', perm_dataset, perm_desc, 10, num_ex) means, covariances, labels = random_ellipsoids(num_neg=num_neg, num_pos=num_pos, n=num_feat, mean_radius=3) data = EllipsoidData(means, covariances, labels) meandata = MIMeanData(means, labels) train_data, test_data = data.get_train_pred_data(permfile=permfile_name('.', perm_dataset, perm_desc), perm_idx=1, model_selection=False, frac_ms=0.5, fold_cv=None, num_cv=None) print('Training mean SVM for initialization') meanclassifier = MeanSVM() mean_train_data, _ = meandata.get_train_pred_data(permfile=permfile_name('.', perm_dataset, perm_desc), perm_idx=1, model_selection=False, frac_ms=0.5, fold_cv=None, num_cv=None) meanclassifier.train(mean_train_data, reg_param) classifier = eMILDcaSocp() print('Training eMIL') classifier.train(train_data, reg_param, v0=r_[meanclassifier.w, meanclassifier.b], huber_delta=0.1) print('Predicting') train_pred = classifier.predict(train_data) conf = get_conf_matrix(train_pred, train_data.labels) acc = accuracy(conf) print('Accuracy = %1.3f' % acc) # Plot training fig = figure() plot_ellipsoids(train_data.means, train_data.covariances, train_data.labels, color=True) plot_hyperplane(classifier.w, classifier.b) fig.savefig('%s-reg_param_%d' % (classifier.__class__.__name__, reg_param)) show() if __name__ == '__main__': example_ellipsoid_2d() OptWok-0.3.1/examples/emil_plotting.py0000644ڛ|׿-0000000463212140417273024741 0ustar cheng.ongUNIMELB\Domain Users00000000000000''' Created on 05.09.2012 @author: kgabriel ''' from numpy import linalg from sys import float_info from math import atan2 from matplotlib.patches import Ellipse from numpy import array, logical_not, sqrt, pi from numpy.linalg import norm from matplotlib.pyplot import plot, xlim, gca, Polygon def plot_ellipsoids(q, S, y, color=False): ''' plot 2d-ellipsoids ''' for qi, Si, yi in zip(q, S, y): plot_ellipse(qi, Si, label=yi, color=color) def plot_ellipse(q, Q, edge = 'k', face = 'none', scale = 1, label=None, color=False, alpha=0.7): # Plots the ellipse defined by: U, s , _ = linalg.svd(Q) # print 's0: ', s[0] # print 's1: ', s[1] orient = atan2(U[1, 0], U[0, 0]) * 180 / pi ellipsePlot = Ellipse(xy = q, width = 2 * scale * sqrt(s[0]), height = 2 * scale * sqrt(s[1]), angle = orient, facecolor = face, edgecolor = edge) ax = gca() ax.set_aspect('equal') if label is None: l = '.' elif label > 0: l = '+' c = 'g' else: l = 'o' c = 'r' if color: ellipsePlot.set_facecolor(c) ellipsePlot.set_alpha(alpha) ax.add_patch(ellipsePlot); plot(q[0], q[1], 'k' + l) return ellipsePlot def plot_vector(origin, vector, color='k'): ''' plot vector from origin ''' dest = origin + vector plot([origin[0], dest[0]], [origin[1], dest[1]], '-', color = color) plot(origin[0], origin[1], 'o', color = color) plot(dest[0], dest[1], 'x', color = color) def plot_hyperplane(w,b,lcolor='black',pcolor='green',lstyle='solid',alpha=0.5): ''' plot hyperplane defined by w and b, plus margin 1/|w| ''' xl = array(xlim()) if w[1] == 0: w[1] = float_info.min y = (-w[0] * xl - b) / w[1] plot(xl, y, color=lcolor, linestyle=lstyle) margin = 1 / norm(w) patch = Polygon([[xl[0], y[0] - margin], [xl[0], y[0] + margin], [xl[1], y[1] + margin], [xl[1], y[1]-margin]], fill=True, closed=True, color=pcolor,alpha=alpha) gca().add_patch(patch) def plot_hyperplane_nomargin(w, b): # plot hyperplane w'x + b = 0 in 2d xl = array(xlim()) plot(xl, (-w[0] * xl - b) / w[1]) def plot_points(p, y): p, y = array(p), array(y) pos = y > 0 plot(p[pos,0], p[pos,1], 'go') plot(p[logical_not(pos),0], p[logical_not(pos),1], 'ro') OptWok-0.3.1/examples/emil_synthetic_ellipsoids.py0000644ڛ|׿-0000000331112140417273027333 0ustar cheng.ongUNIMELB\Domain Users00000000000000''' Created on 05.09.2012 @author: kgabriel ''' from __future__ import division from scipy import random from numpy import diag, ones, dot, pi, cos, sin, array from numpy.random import rand def random_ellipsoids(num_pos, num_neg, n, var_pos=1, var_neg=1, mean_radius=2, axisp=False): ''' generate num_pos positive and num_neg negative random ellipsoids of dimensionality n ''' offset_p = random.randn(n) + 1 offset_n = random.randn(n) - 1 qp = random.randn(num_pos, n) * var_pos + offset_p qn = random.randn(num_neg, n) * var_neg + offset_n S = [] for _ in range(num_pos + num_neg): #print 'next random covariance...' if not axisp: S.append(random_covariance_matrix(n, mean_radius)) else: S.append(diag(rand(n))*mean_radius) return array(list(qp) + list(qn)), array(S), array(list(ones(num_pos))+ list(-1*ones(num_neg))) def random_covariance_matrix(n, scale=1): ''' generate random covariance matrix of dimension n ''' # Use Householder reflections to build orthonormal matrix and then C = QSQ' # I = eye(n) # Q = I # K = 2 # for i in range(K*n*n): # v = random.rand(n) # v = v / linalg.norm(v) # P = I - outer(v,v) # Q = dot(P, Q) # # S = diag(random.rand(n)) # # C = dot(Q,dot(S, Q.transpose())) # Use C=Q'Q Q = random.rand(n,n) C = dot(Q, Q.transpose()) if n==2: if rand(1) > 0.5: theta = 0 else: theta = pi/2 C = dot(dot(rotate(theta).T,C),rotate(theta)) return scale * C def rotate(theta): return array([[cos(theta), -sin(theta)], [sin(theta), cos(theta)]]) OptWok-0.3.1/examples/gaussianproc_demo.py0000644ڛ|׿-0000000746112055534750025605 0ustar cheng.ongUNIMELB\Domain Users00000000000000""" Some code for testing Gaussian process regression, also as a demo """ import os from matplotlib.pyplot import plot, fill, show, axis, figure from numpy import concatenate, arange, ones, sqrt from numpy import flatnonzero from optwok.testtools import pair2vec, flatten_list, list2matrix from optwok.kernel import GaussKernel from optwok.io_pickle import load from optwok.gaussian_proc import GaussianProcess, GaussianProcessScalar def _plot_1D_gp(data_ex, data_lab, test_range, preds, pred_var, colour): plot(test_range.flatten(), preds, colour, hold=True) sd = 2.0*sqrt(pred_var) var_x = concatenate((test_range.flatten(), test_range.flatten()[::-1])) var_y = concatenate((preds+sd, (preds-sd)[::-1])) fill(var_x, var_y, edgecolor='w', facecolor=colour, alpha=0.3) plot(data_ex, data_lab, colour+'o') def plot_prediction(X, Z, Y, gp): """Plot the data points, the predictions and the predictive variance""" colour_list = ['b','g','r','c','m','k'] min_val = min(X) max_val = max(X) test_range = arange(min_val-1, max_val+1, 0.1) num_test = len(test_range) test_range.shape = (1,num_test) num_context = int(max(Z))+1 for cur_cont in range(num_context): (preds, pred_var) = gp.predict((test_range, cur_cont*ones(num_test))) data_ex = X[flatnonzero(Z==cur_cont)] data_lab = Y[flatnonzero(Z==cur_cont)] _plot_1D_gp(data_ex, data_lab, test_range, preds, pred_var, colour_list[cur_cont]) def generate_toy_data(num_ex=10, domain=5, noise_level=0.1, mix_list=[[0,1],[2]]): """Generate a multi-output Gaussian process dataset with dependencies between outputs.""" from optwok.simdata import gp_gen_pair, save_gp_context_data num_class = len(flatten_list(mix_list)) dataname = 'gaussproc_%02d_%d_%1.1f' % (num_class, num_ex, noise_level) print('Generating %s' % dataname) (X,Z,Y) = gp_gen_pair(num_ex, 1, domain, noise_level, mix_list) save_gp_context_data(X, Z, Y, dataname, data_dir='tests') def toy_data_gp(num_ex=10, noise_level=0.1, mix_list=[[0,1],[2]]): """Estimate and plot the 1D Gaussian process""" num_class = len(flatten_list(mix_list)) data = load('tests/gaussproc_%02d_%d_%1.1f.pkl' % (num_class, num_ex, noise_level)) #Ky = list2matrix(mix_list, noise=0.1) # so that can observe difference bewteen 0 and 1 Ky = list2matrix(mix_list) print('Estimating mean and variance of Gaussian Process') gp = GaussianProcess(GaussKernel, Ky, reg_param=0.1) gp.train(data.examples, data.context, data.labels) plot_prediction(data.examples.flatten(), data.context, data.labels, gp) def toy_data_gp_scalar(num_ex=10, num_class=3, noise_level=0.1): """Estimate and plot the 1D scalar output GP""" colour_list = ['b','g','r','c','m','k'] test_ex = arange(-1, 6, 0.1) num_test = len(test_ex) test_ex.shape = (1,num_test) data = load('tests/gaussproc_%02d_%d_%1.1f.pkl' % (num_class, num_ex, noise_level)) for context in range(num_class): gp = GaussianProcessScalar(GaussKernel, reg_param=0.1) idx = flatnonzero(data.context == context) gp.train(data.examples[:,idx], data.labels[idx]) (preds, pred_var) = gp.predict(test_ex) _plot_1D_gp(data.examples[:,idx].flatten(), data.labels[idx], test_ex, preds, pred_var, colour_list[context]) if __name__ == '__main__': num_ex = 20 #generate_toy_data(num_ex, mix_list=[[0,1],[2],[3,4,5]]) toy_data_gp(num_ex, mix_list=[[0,1],[2],[3,4,5]]) print('mix_list=[[0,1],[2],[3,4,5]]') figure() toy_data_gp(num_ex, mix_list=[[0,1,2,3,4,5]]) print('mix_list=[[0,1,2,3,4,5]]') figure() toy_data_gp(num_ex, mix_list=[[0],[1],[2],[3],[4],[5]]) print('mix_list=[[0],[1],[2],[3],[4],[5]]') figure() toy_data_gp_scalar(num_ex, 6) print('scalar GP') show() OptWok-0.3.1/examples/kernelopt_demo.py0000644ڛ|׿-0000001472012055551176025107 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Demo for learning the output kernel""" from numpy import vstack from matplotlib.pyplot import matshow, gray, colorbar, savefig, show from optwok.kernelopt import KernelOptLS, KernelOptLin def show_kernel(kmat): """Use matshow with gray colormap""" matshow(kmat) gray() colorbar() def show_kernel_dev(kmat): """Use matshow with a nice colormap""" minval = min(kmat.flatten()) maxval = max(kmat.flatten()) absval = max([abs(minval),abs(maxval)]) near_zero = 0.1*absval if minval < 0.0: zeroval = -minval/(2.0*absval) else: zeroval = 0.0 small_neg = zeroval-near_zero small_pos = min([1.0, zeroval+near_zero]) if small_neg >= 0.0: blueval = (small_neg-minval)/(small_neg + absval) assert(blueval>=0.0 and blueval<=1.0) print(blueval) cdict = {'red': ((0.0, blueval, blueval), (small_neg, 1.0, 1.0), (zeroval, 1.0, 1.0), (small_pos, 1.0, 1.0), (1.0, 1.0, 1.0)), 'green': ((0.0, 0.8-blueval, 0.8-blueval), (small_neg, 1.0, 1.0), (zeroval, 1.0, 1.0), (small_pos, 1.0, 1.0), (1.0, 0.2, 0.2)), 'blue': ((0.0, 1-blueval, 1-blueval), (small_neg, 1.0, 1.0), (zeroval, 1.0, 1.0), (small_pos, 1.0, 1.0), (1.0, 0.0, 0.0)) } else: small_neg = 0.0 cdict = {'red': ((0.0, 1.0, 1.0), (small_neg, 1.0, 1.0), (zeroval, 1.0, 1.0), (small_pos, 1.0, 1.0), (1.0, 1.0, 1.0)), 'green': ((0.0, 1.0, 1.0), (small_neg, 1.0, 1.0), (zeroval, 1.0, 1.0), (small_pos, 1.0, 1.0), (1.0, 0.2, 0.2)), 'blue': ((0.0, 1.0, 1.0), (small_neg, 1.0, 1.0), (zeroval, 1.0, 1.0), (small_pos, 1.0, 1.0), (1.0, 0.2, 0.2)) } blue_red = LinearSegmentedColormap('BlueRedZero', cdict) matshow(kmat, cmap=blue_red) colorbar() def test_kernelopt(): """Small test for class KernelOpt""" from numpy.random import randn X = randn(10,50) Kx = array(matrix(X).T*matrix(X)) alpha = randn(50,4) print(normalize_unit_diag(matrix(alpha).T*matrix(Kx)*matrix(alpha))) kern_mach = KernelOpt() kern_mach.classifier.Kx = Kx print(kern_mach.solve_opt(alpha)) def _create_test_data(mix_list=[[0,1],[2],[3],[4]], num_point=[100, 50, 50, 50], num_feat=20, edge_length=3): """Create an multiclass dataset with mixed classes""" from optwok.simdata import multi_cloudgen_mix num_blobs = len(mix_list) train_data = multi_cloudgen_mix(num_point, num_feat, num_blobs, edge_length, mix_list) (metadata, examples, labels) = train_data return (metadata, examples, labels) def test_kerneloptls(): """Small test for class KernelOptLS""" from optwok.kernel import LinearKernel (metadata, examples, labels) = _create_test_data() print(metadata) svm = KernelOptLS(reg_param=1000.0, kernel=LinearKernel(), verbose=True) svm.train(examples, labels) train_out = svm.predict(examples) print(vstack([train_out, labels])) show_kernel(svm.Ky) savefig('testKy.pdf') def test_gp(mix_list=[[0,1],[2]]): """Small test for class KernelOptLS, for Gaussian process regression""" from optwok.kernel import LinearKernel from optwok.simdata import gp_gen from optwok.evaluation import rmse (examples, labels) = gp_gen(100, mix_list) svm = KernelOptLSGP(reg_param=10.0, kernel=LinearKernel(), verbose=True) svm.train(examples, labels) train_out = svm.predict(examples) #print(vstack([train_out, labels])) print('Error:') print(rmse(train_out, labels)) show_kernel(svm.Ky) def test_reg_path_Ky(): """The regularization path""" from optwok.kernel import LinearKernel (metadata, examples, labels) = _create_test_data() print(metadata) svm = KernelOptLS(kernel=LinearKernel(), verbose=True) svm.train_reg_path(examples, labels) for ix in range(svm.all_Ky.shape[2]): show_kernel(normalize_unit_diag(svm.all_Ky[:,:,ix])) title(ix) def test_reg_path_acc(): """Check accuracy for whole regularization path""" from optwok.kernel import LinearKernel from optwok.evaluation import accuracy from pylab import semilogx, xlabel, ylabel, legend (metadata, examples, labels) = _create_test_data() print(metadata) svm = KernelOptLS(kernel=LinearKernel(), verbose=True) svm.train_reg_path(examples, labels) (metadata, examples, labels) = _create_test_data() all_preds = svm.predict_reg_path(examples) all_acc = zeros(len(svm.all_reg_param)) for ix in range(all_preds.shape[1]): all_acc[ix] = accuracy(all_preds[:,ix], labels) svm.reg_param = 1.0 svm.infer_sv(eye(svm.num_classes)) diag_acc = accuracy(svm.predict(examples), labels) semilogx(svm.all_reg_param, all_acc, hold=True, label='learn Ky') semilogx(svm.all_reg_param, all_acc, 'rx') semilogx([svm.all_reg_param[0],svm.all_reg_param[-1]], [diag_acc, diag_acc], 'g-', label='identity') xlabel('regularization parameter') ylabel('accuracy (higher is better)') legend() def test_kerneloptlin(): """Small test for class KernelOptLS""" from optwok.kernel import LinearKernel (metadata, examples, labels) = _create_test_data() print(metadata) svm = KernelOptLS(reg_param=1000.0, kernel=LinearKernel(), verbose=True) svm.train(examples, labels) train_out = svm.predict(examples) print(train_out.shape) print(vstack([train_out, labels])) show_kernel(svm.Ky) savefig('testKy.pdf') svm = KernelOptLin(reg_param=1000.0, verbose=True) svm.train(examples, labels) train_out = svm.predict(examples) print(train_out.shape) print(vstack([train_out, labels])) show_kernel(svm.Ky) savefig('testLin.pdf') if __name__ == '__main__': #test_kerneloptls() #test_reg_path_Ky() #test_reg_path_acc() #test_gp() test_kerneloptlin() show() OptWok-0.3.1/examples/multiclass_demo.py0000644ڛ|׿-0000001044312055534622025257 0ustar cheng.ongUNIMELB\Domain Users00000000000000""" Some code for testing and also as a demo """ import pylab from numpy import eye, vstack, ones, matrix from numpy import nonzero, abs, zeros, sum, sqrt, unique from optwok.simdata import multi_cloudgen, multi_cloudgen_mix from optwok.testtools import plot_confusion from optwok.multiclass import StructPredictor from optwok.evaluation import accuracy, accuracy_mix from optwok.kernelopt import KernelOpt from optwok.kernel import JointKernel, LinearKernel, GaussKernel from optwok.kernel import normalize_unit_diag def demo_identity_Ky(): print('demo_identity_Ky') num_point = 100 num_feat = 10 num_class = 3 edge_length = 3 train_data = multi_cloudgen(num_point, num_feat, num_class, edge_length) (metadata, examples, labels) = train_data print(metadata) kernel = JointKernel(LinearKernel(), num_classes=num_class) svm = StructPredictor(1.0, kernel) svm.train(examples, labels) train_out = svm.predict(examples) train_acc = accuracy(train_out, labels) print('Training accuracy:\t %1.2f' % train_acc) test_data = multi_cloudgen(num_point, num_feat, num_class, edge_length) (metadata, test_ex, test_lab) = test_data svmout = svm.predict(test_ex) acc = accuracy(svmout, test_lab) print('Test accuracy:\t\t %1.2f' % acc) def demo_merged_classes(): """The examples come from different Gaussians, but have the same label.""" print('demo_merged_classes') num_point = 100 num_feat = 10 num_class = 4 edge_length = 3 mix_list = [[0,1],[2],[3]] train_data = multi_cloudgen_mix(num_point, num_feat, num_class, edge_length, mix_list) (metadata, examples, labels) = train_data print(metadata) kernel = JointKernel(GaussKernel(), num_classes=3) svm = StructPredictor(1.0, kernel) svm.train(examples, labels) train_out = svm.predict(examples) train_acc = accuracy(train_out, labels) print('Training accuracy:\t %1.2f' % train_acc) test_data = multi_cloudgen_mix(num_point, num_feat, num_class, edge_length, mix_list) (metadata, test_ex, test_lab) = test_data svmout = svm.predict(test_ex) acc = accuracy(svmout, test_lab) print('Test accuracy:\t\t %1.2f' % acc) def demo_split_classes(): """The examples from the same Gaussians may have different labels.""" print('demo_split_classes') num_point = 100 num_feat = 6 num_class = 3 edge_length = 2 mix_list = [[0,1],[2],[3]] Ky = eye(4) Ky[0,1] = 0.99 Ky[1,0] = 0.99 print(Ky) train_data = multi_cloudgen_mix(num_point, num_feat, num_class, edge_length, mix_list) (metadata, examples, labels) = train_data print(metadata) test_data = multi_cloudgen_mix(num_point, num_feat, num_class, edge_length, mix_list) (metadata, test_ex, test_lab) = test_data kernel = JointKernel(LinearKernel(), num_classes=4) print('============== With identity Ky ==============') svm = StructPredictor(1.0, kernel) svm.train(examples, labels) train_out = svm.predict(examples) #print(vstack([train_out, labels])) print('Training accuracy:\t\t %1.2f' % accuracy(train_out, labels)) print('Training accuracy (mix):\t %1.2f' % accuracy_mix(train_out, labels, mix_list)) svmout = svm.predict(test_ex) #print(vstack([svmout, test_lab])) print('Test accuracy:\t\t\t %1.2f' % accuracy(svmout, test_lab)) print('Test accuracy (mix):\t\t %1.2f' % accuracy_mix(svmout, test_lab, mix_list)) Ktarget = normalize_unit_diag(matrix(svm.alpha).T*matrix(svm.kernel.Kx.kmat)*matrix(svm.alpha)) print(Ktarget) print('============== With mix Ky ==============') svm.kernel.Ky = Ky svm.train(examples, labels) train_out = svm.predict(examples) #print(vstack([train_out, labels])) print('Training accuracy:\t\t %1.2f' % accuracy(train_out, labels)) print('Training accuracy (mix):\t %1.2f' % accuracy_mix(train_out, labels, mix_list)) svmout = svm.predict(test_ex) #print(vstack([svmout, test_lab])) print('Test accuracy:\t\t\t %1.2f' % accuracy(svmout, test_lab)) print('Test accuracy (mix):\t\t %1.2f' % accuracy_mix(svmout, test_lab, mix_list)) print(normalize_unit_diag(matrix(svm.alpha).T*matrix(svm.kernel.Kx.kmat)*matrix(svm.alpha))) if __name__ == '__main__': demo_identity_Ky() demo_merged_classes() demo_split_classes() OptWok-0.3.1/examples/named_array_demo.py0000644ڛ|׿-0000000554012140417273025360 0ustar cheng.ongUNIMELB\Domain Users00000000000000from optwok.named_array import NArray import subprocess if __name__ == '__main__': res = NArray(['CV','Task','Algorithm']) print('Initial conditions') print(res.ndim) print(res.dim_names) print(res.shape) res.add_keys(range(5), 'CV') print('Added 5 CV elements') print(res.shape) for task in ['A','B','A']: res.add_key(task, 'Task') for alg in ['svm','mtl','ky']: res.add_key(alg, 'Algorithm') print('Added tasks and methods') print(res.shape) print('Still nothing inside') print(res.content) # Note that you can fill in values in any order values = range(5,100) val_ix = 0 for cv in range(3): for task in ['A','B']: for alg in ['ky','svm','mtl']: res.fill_value(values[val_ix], [task, cv, alg], ['Task','CV','Algorithm']) val_ix += 1 print('After CV=%d' % cv) print(res.content) res2 = NArray(['CV','Task','Algorithm']) res2.add_keys(range(5), 'CV') res2.add_keys(['A','B'], 'Task') res2.add_keys(['svm','mtl','ky'], 'Algorithm') for cv in [2,3,4]: for task in ['A','B']: for alg in ['ky','svm','mtl']: res.fill_value(values[val_ix], [task, cv, alg], ['Task','CV','Algorithm']) val_ix += 1 print(res2.content) print('Merging') res.merge(res2) del res2 print(res.content) print('Get some entries') print(res.get_value(['B','mtl',2], ['Task','Algorithm','CV'])) print(res.get_value(['mtl',2,'B'], ['Algorithm','CV','Task'])) print(res.get_value([0,'A','ky'], ['CV', 'Task', 'Algorithm'])) print('Get all keys') print(res.keys('Task')) print(res.keys('Algorithm')) print('Get a slice') print(res.slice('Task','B')) print('A thinner slice') print(res.slice(['Task','Algorithm'],['B','svm'])) print('Bad slices') try: print(res.slice('Task','Z')) except IndexError as msg: print(msg) try: print(res.slice('Tas','B')) except IndexError as msg: print(msg) try: print(res.slice(['Task','Alg'],['B','svm'])) except IndexError as msg: print(msg) try: print(res.slice(['Task','Algorithm'],['Z','svm'])) except IndexError as msg: print(msg) print('Saving and loading') res.save('na_demo.json') new_res = NArray() new_res.load('na_demo.json') print('Get all keys') print(new_res.keys('Task')) print(new_res.keys('Algorithm')) print('Get a slice') print(new_res.slice('Task','B')) print('A thinner slice') print(new_res.slice(['Task','Algorithm'],['B','svm'])) print('Save in matlab format') new_res.savemat('na_demo.mat') subprocess.check_call('rm na_demo.json na_demo.mat', shell=True) OptWok-0.3.1/examples/score_demo.py0000644ڛ|׿-0000000247412055534135024216 0ustar cheng.ongUNIMELB\Domain Users00000000000000from numpy import array from optwok.evaluation import accuracy, balanced_accuracy from optwok.evaluation import stats_empirical, stats_binormal def _demo_show(tpr, fpr, ppv, auc, apr): from pylab import plot, figure, axis, title, xlabel, ylabel print('AUC=%f, APR=%f' % (auc, apr)) figure() if len(tpr) < 100: plot(fpr, tpr, 'bo', hold=True) plot(fpr, tpr, 'b-') axis([0,1,0,1]) xlabel('False positive rate') ylabel('True positive rate') title('ROC') figure() if len(ppv) < 100: plot(tpr, ppv, 'bo', hold=True) plot(tpr, ppv, 'b-') axis([0,1,0,1]) ylabel('Positive predictive value') xlabel('True positive rate') title('PRC') def demo_stats_binary(): from numpy.random import rand from pylab import show labels = array([-1, 1., -1., 1., -1., -1., -1., 1., -1., -1.]) output = array([-0.4, 0.9, 0.4, 0.3, -0.1, -0.6, -0.7, -0.2, -1., -0.9,]) acc = accuracy(output, labels) balacc = balanced_accuracy(output, labels) print('ACC=%f, BAL=%f' % (acc, balacc)) tpr, fpr, ppv, auc, apr = stats_empirical(output, labels) _demo_show(tpr, fpr, ppv, auc, apr) tpr, fpr, ppv, auc, apr = stats_binormal(output, labels) _demo_show(tpr, fpr, ppv, auc, apr) show() if __name__ == '__main__': demo_stats_binary() OptWok-0.3.1/examples/socp_demo.py0000644ڛ|׿-0000000107512055534115024041 0ustar cheng.ongUNIMELB\Domain Users00000000000000''' Created on 16.12.2011 @author: kgabriel ''' from optwok.socp import SOCP from numpy import array if __name__ == '__main__': socp = SOCP(3, verbose = True) socp.set_objective(array([-2, 1, 5.])) A = array([[-13., 3, 5], [-12, 12, -6]]) b = array([-3, -2.]) c = array([-12., -6, 5]) d = -12 socp.add_ineq_constraints(A, b, c, d) A = array([[-3., 6, 2], [1, 9, 2], [-1, -19, 3]]) b = array([0, 3, -42.]) c = array([-3, 6, -10.]) d = 27 socp.add_ineq_constraints(A, b, c, d) socp.solve() print(socp.solution) OptWok-0.3.1/optwok/0000755ڛ|׿-0000000000012140420542021212 5ustar cheng.ongUNIMELB\Domain Users00000000000000OptWok-0.3.1/optwok/__init__.py0000644ڛ|׿-0000000000011731540564023325 0ustar cheng.ongUNIMELB\Domain Users00000000000000OptWok-0.3.1/optwok/classifier.py0000644ڛ|׿-0000010216612055565522023733 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Implementation of the optimization methods for the empirical loss + regularizer framework, for binary clssification. The algorithms are primal variable algorithms. We consider the following losses: - hinge loss - ramp loss We consider the following regularizers: - l2 - l1 - capped l1 - l0 (approximated by exponent) """ from time import time import numpy from numpy.random import randn from numpy import matrix, array, vstack, hstack from numpy import zeros, ones, transpose, nonzero from numpy import sign, eye, where, dot, diag from numpy.linalg import norm from numpy import newaxis, outer, inner, mean, dot from optwok.sip import SILP, SILPS, SIQP from optwok.poly_feature import get_poly_features from optwok.kernel import KernelFactory from optwok.classifier_base import Classifier, ConvexClassifier from optwok.classifier_base import Loss, HingeLoss, HingeLossKernel from optwok.classifier_base import HuberHingeLoss, RampLoss, SlantLoss from optwok.classifier_base import Regularizer from optwok.classifier_base import Reg_l2, Reg_l1, Reg_l1cap, Reg_l0 from optwok.classifier_base import Reg_elastic_net, Reg_graph_net from cvxopt import sparse from cvxopt import matrix as cm class SVMPrimal(ConvexClassifier): """ Standard l2 regularized SVM, in the primal """ def __init__(self, reg_param, kdict, verbose=False): ConvexClassifier.__init__(self, reg_param, kdict, verbose) self.loss = HingeLoss() self.regularizer = Reg_l2() def solve_opt(self): """Solve the SVM QP Use cvxopt qp solver which expects the following format: min_x 1/2 x'Px + q'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving QP') n = len(self.y) d = self.train_ex.shape[0] m = n+d+1 svm = SIQP(m, self.verbose) P = zeros((m, m)) P[:d,:d] = 2.0*self.reg_param*eye(d) q = vstack([zeros((d+1,1)), ones((n,1))]) svm.set_objective(P,q) svm.add_ineq_constraints(hstack([zeros((n,d+1)), -eye(n)]), zeros((n,1))) G = hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n)]) svm.add_ineq_constraints(G, -ones((n,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] class SVMDual(ConvexClassifier): """ Standard l2 regularized SVM, in the dual, not finished yet. TODO: test time, arbitrary kernel """ def __init__(self, reg_param, kdict, verbose=False): ConvexClassifier.__init__(self, reg_param, kdict, verbose) self.loss = HingeLoss() self.regularizer = Reg_l2() self.kernel = KernelFactory(kdict) self.alpha = array([]) def train(self, train_ex, train_lab): """Wrapper to convert data to feature matrix, and solve the optimization""" self.train_ex = train_ex self.kernel.compute(train_ex) self.y = train_lab t_start = time() self.solve_opt() t_end = time() # Collect statistics about optimization self.train_time = t_end-t_start def predict(self, pred_ex): """Wrapper to do prediction""" Kpred = self.kernel.compute_test(pred_ex) Ymat = diag(self.y) scores = dot(dot(Kpred.T,Ymat),self.alpha) + self.b return scores def solve_opt(self): """Solve the SVM QP Use cvxopt qp solver which expects the following format: min_x 1/2 x'Px + q'x st Gx <= h Ax = b """ n = len(self.y) qp = SIQP(n, use_mosek = True) q = -ones((n, 1), dtype = 'double') ll = outer(self.y, self.y) #K = inner(self.train_ex, self.train_ex) K = self.kernel.kmat P = ll * K qp.set_objective(P, q) qp.add_eq_constraints(array([self.y]), array([[0]])) qp.add_ineq_constraints(eye(n), (self.reg_param * ones(n))[..., newaxis]) qp.add_ineq_constraints(-1. * eye(n), zeros((n, 1))) qp.solve(warn = True) self.alpha = qp.solution sel = (qp.solution > self.epsilon) & (qp.solution < self.reg_param-self.epsilon) pred = dot(dot(K[sel,:],diag(self.y)),self.alpha) b = self.y[sel] - pred self.b = mean(b) class OneNormSVM_dense(ConvexClassifier): """ One norm SVM """ def __init__(self, reg_param, kdict, verbose=False): ConvexClassifier.__init__(self, reg_param, kdict, verbose) self.loss = HingeLoss() self.regularizer = Reg_l1() def solve_opt(self): """Solve the linear program corresponding to the one norm SVM. Use cvxopt lp solver which expects the following format: min_x c'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi,eta). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving LP') n = len(self.y) d = self.train_ex.shape[0] m = n+2*d+1 svm = SILP(m, self.verbose) svm.set_objective(vstack([zeros((d+1,1)), ones((n,1)), self.reg_param*ones((d,1))])) G = hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n), zeros((n,d))]) svm.add_ineq_constraints(G, -ones((n,1))) svm.add_ineq_constraints(hstack([zeros((n,d+1)),-eye(n),zeros((n,d))]), zeros((n,1))) svm.add_ineq_constraints(hstack([zeros((d,d+1)),zeros((d,n)),-eye(d)]), zeros((d,1))) svm.add_ineq_constraints(hstack([eye(d), zeros((d,n+1)), -eye(d)]), zeros((d,1))) svm.add_ineq_constraints(hstack([-eye(d), zeros((d,n+1)), -eye(d)]), zeros((d,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] class OneNormSVM(ConvexClassifier): """ One norm SVM, with sparse matrices as constraints Should behave exactly the same as the non sparse version """ def __init__(self, reg_param, kdict, verbose=False): ConvexClassifier.__init__(self, reg_param, kdict, verbose) self.loss = HingeLoss() self.regularizer = Reg_l1() def solve_opt(self): """Solve the linear program corresponding to the one norm SVM. Use cvxopt lp solver which expects the following format: min_x c'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi,eta). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving LP') n = len(self.y) d = self.train_ex.shape[0] m = n+2*d+1 svm = SILPS(m, self.verbose) svm.set_objective(vstack([zeros((d+1,1)), ones((n,1)), self.reg_param*ones((d,1))])) G = sparse(cm(hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n), zeros((n,d))]))) svm.add_ineq_constraints(G, -ones((n,1))) svm.add_ineq_constraints(sparse(cm(hstack([zeros((n,d+1)),-eye(n),zeros((n,d))]))), zeros((n,1))) svm.add_ineq_constraints(sparse(cm(hstack([zeros((d,d+1)),zeros((d,n)),-eye(d)]))), zeros((d,1))) svm.add_ineq_constraints(sparse(cm(hstack([eye(d), zeros((d,n+1)), -eye(d)]))), zeros((d,1))) svm.add_ineq_constraints(sparse(cm(hstack([-eye(d), zeros((d,n+1)), -eye(d)]))), zeros((d,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] class DiffConvexClassifier(Classifier): """ Classifiers which have difference of convex training methods """ def __init__(self, reg_param, kdict, verbose=False, rand_init=False): Classifier.__init__(self, reg_param, kdict, verbose) self.rand_init = rand_init # Use SVM or random initialization self.num_iter = 0 self.max_iter = 100 # Maximum number of DC iterations self.min_diff = 1e-2 # Stopping criteria, change in parameters self.all_objective = [] # The objective value for all iterations @property def perf(self): """Summary of results""" perf = super(DiffConvexClassifier, self).perf perf['num_iter'] = self.num_iter perf['all_objective'] = self.all_objective return perf def _init_param(self, norm): """Initialize values of w and b. The parameter denotes whether to use l1 or l2 norm. """ assert(norm == 1 or norm==2) if self.rand_init: self.num_iter = 0 if self.verbose: print('Random initialization') w = self.reg_param*randn(self.train_ex.shape[0],1) b = -randn() else: self.num_iter = 1 if self.verbose: print('SVM initialization') if norm == 1: svm = OneNormSVM(self.reg_param, self.kdict, self.verbose) elif norm == 2: svm = SVMPrimal(self.reg_param, self.kdict, self.verbose) # train computes the feature mapping explicitly svm.train_ex = self.train_ex svm.y = self.y svm.solve_opt() (w, b) = (svm.w, svm.b) return (w, b) def _subgradient_concave_ramp(self, mu): """Compute the subgradient of the concave part of the ramp loss""" (clipped, sloped, correct) = self.loss.get_loss_case(self.y*self.train_out) #sw = zeros(self.train_ex.shape[0]) #sb = zeros((1,1)) #for idx in clipped: # sw += -self.y[idx]*self.train_ex[:,idx] # sb += -self.y[idx] sw = numpy.sum(-self.y[clipped]*self.train_ex[:,clipped], axis=1) sw.shape = (self.train_ex.shape[0], 1) sb = sum(-self.y[clipped]) sw /= mu sb /= mu return (sw, sb) def _subgradient_concave_cap_l1(self, a): """Compute the subgradient of the concave part of the capped l1 norm""" grad = sign(self.w) grad[nonzero(abs(self.w) < a)[0]] = 0.0 return grad def _subgradient_concave_l0(self, a): """Compute the subgradient of the concave part of the l0 approximation""" return a*sign(self.w)*(1.0 - numpy.exp(-a*abs(self.w))) class CapOneNormSVM(DiffConvexClassifier): """ SVM with capped one norm regularizer. """ def __init__(self, reg_param, kdict, a=1.0, verbose=False, rand_init=False): DiffConvexClassifier.__init__(self, reg_param, kdict, verbose, rand_init) self.a = a # The value of the cap self.loss = HingeLoss() self.regularizer = Reg_l1cap(a) #@property #def margin(self): # """The value of the regularizer, specialized because of parameter a""" # return self.regularizer(self.w, self.a) def _solve_linprog(self, u): """Solve the linear program corresponding to the convex part of the capped one norm SVM. Use cvxopt lp solver which expects the following format: min_x c'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi,eta). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving LP') n = len(self.y) d = self.train_ex.shape[0] m = n+2*d+1 svm = SILPS(m, self.verbose) svm.set_objective(vstack([-self.reg_param*u, zeros((1,1)), ones((n,1)), self.reg_param*ones((d,1))])) G = sparse(cm(hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n), zeros((n,d))]))) svm.add_ineq_constraints(G, -ones((n,1))) svm.add_ineq_constraints(sparse(cm(hstack([zeros((n,d+1)),-eye(n),zeros((n,d))]))), zeros((n,1))) svm.add_ineq_constraints(sparse(cm(hstack([zeros((d,d+1)),zeros((d,n)),-eye(d)]))), zeros((d,1))) svm.add_ineq_constraints(sparse(cm(hstack([eye(d), zeros((d,n+1)), -eye(d)]))), zeros((d,1))) svm.add_ineq_constraints(sparse(cm(hstack([-eye(d), zeros((d,n+1)), -eye(d)]))), zeros((d,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] def _solve_linprog_dense(self, u): """Solve the linear program corresponding to the convex part of the capped one norm SVM. Uses dense matrices. Use cvxopt lp solver which expects the following format: min_x c'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi,eta). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving LP') n = len(self.y) d = self.train_ex.shape[0] m = n+2*d+1 svm = SILP(m, self.verbose) svm.set_objective(vstack([-self.reg_param*u, zeros((1,1)), ones((n,1)), self.reg_param*ones((d,1))])) G = hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n), zeros((n,d))]) svm.add_ineq_constraints(G, -ones((n,1))) svm.add_ineq_constraints(hstack([zeros((n,d+1)),-eye(n),zeros((n,d))]), zeros((n,1))) svm.add_ineq_constraints(hstack([zeros((d,d+1)),zeros((d,n)),-eye(d)]), zeros((d,1))) svm.add_ineq_constraints(hstack([eye(d), zeros((d,n+1)), -eye(d)]), zeros((d,1))) svm.add_ineq_constraints(hstack([-eye(d), zeros((d,n+1)), -eye(d)]), zeros((d,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] def solve_opt(self): """Use a DCA to solve the ramp loss SVM""" if self.verbose: print('Solving DCA') (self.w, self.b) = self._init_param(1) self.all_objective.append(self.primal_obj) u = self._subgradient_concave_cap_l1(self.a) diff_param = numpy.inf while (self.num_iter < self.max_iter)\ and (diff_param > self.min_diff): self._solve_linprog(u) old_u = u u = self._subgradient_concave_cap_l1(self.a) diff_param = sum(abs(old_u-u)) self.num_iter += 1 self.all_objective.append(self.primal_obj) class ZeroNormSVM(DiffConvexClassifier): """ SVM with zero norm regularizer. Uses exponential approximation. """ def __init__(self, reg_param, kdict, a=4.0, verbose=False, rand_init=False): DiffConvexClassifier.__init__(self, reg_param, kdict, verbose, rand_init) self.a = a # The exponent in approximation self.loss = HingeLoss() self.regularizer = Reg_l0(a) #@property #def margin(self): # """The value of the regularizer, specialized because of parameter a""" # return self.regularizer(self.w, self.a) def _solve_linprog(self, u): """Solve the linear program corresponding to the convex part of the zero norm SVM. Use cvxopt lp solver which expects the following format: min_x c'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi,eta). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving LP') n = len(self.y) d = self.train_ex.shape[0] m = n+2*d+1 svm = SILPS(m, self.verbose) svm.set_objective(vstack([-self.reg_param*u, zeros((1,1)), ones((n,1)), self.reg_param*self.a*ones((d,1))])) G = sparse(cm(hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n), zeros((n,d))]))) svm.add_ineq_constraints(G, -ones((n,1))) svm.add_ineq_constraints(sparse(cm(hstack([zeros((n,d+1)),-eye(n),zeros((n,d))]))), zeros((n,1))) svm.add_ineq_constraints(sparse(cm(hstack([zeros((d,d+1)),zeros((d,n)),-eye(d)]))), zeros((d,1))) svm.add_ineq_constraints(sparse(cm(hstack([eye(d), zeros((d,n+1)), -eye(d)]))), zeros((d,1))) svm.add_ineq_constraints(sparse(cm(hstack([-eye(d), zeros((d,n+1)), -eye(d)]))), zeros((d,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] def _solve_linprog_dense(self, u): """Solve the linear program corresponding to the convex part of the zero norm SVM. Uses dense matrices. Use cvxopt lp solver which expects the following format: min_x c'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi,eta). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving LP') n = len(self.y) d = self.train_ex.shape[0] m = n+2*d+1 svm = SILP(m, self.verbose) svm.set_objective(vstack([-self.reg_param*u, zeros((1,1)), ones((n,1)), self.reg_param*self.a*ones((d,1))])) G = hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n), zeros((n,d))]) svm.add_ineq_constraints(G, -ones((n,1))) svm.add_ineq_constraints(hstack([zeros((n,d+1)),-eye(n),zeros((n,d))]), zeros((n,1))) svm.add_ineq_constraints(hstack([zeros((d,d+1)),zeros((d,n)),-eye(d)]), zeros((d,1))) svm.add_ineq_constraints(hstack([eye(d), zeros((d,n+1)), -eye(d)]), zeros((d,1))) svm.add_ineq_constraints(hstack([-eye(d), zeros((d,n+1)), -eye(d)]), zeros((d,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] def solve_opt(self): """Use a DCA to solve the ramp loss SVM""" if self.verbose: print('Solving DCA') (self.w, self.b) = self._init_param(1) self.all_objective.append(self.primal_obj) u = self._subgradient_concave_l0(self.a) diff_param = numpy.inf while (self.num_iter < self.max_iter)\ and (diff_param > self.min_diff): self._solve_linprog(u) old_u = u u = self._subgradient_concave_l0(self.a) diff_param = sum(abs(old_u-u)) self.num_iter += 1 self.all_objective.append(self.primal_obj) class RLSVM2(DiffConvexClassifier): """ Ramp Loss SVM with l2 regularizer. """ def __init__(self, reg_param, kdict, mu, verbose=False, rand_init=False): DiffConvexClassifier.__init__(self, reg_param, kdict, verbose, rand_init) self.mu = mu # the slope of the ramp self.loss = RampLoss(mu) self.regularizer = Reg_l2() def _solve_quadprog(self, sw, sb): """Solve the quadratic program given by linearizing the concave part Use cvxopt qp solver which expects the following format: min_x 1/2 x'Px + q'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving QP') n = len(self.y) d = self.train_ex.shape[0] m = n+d+1 svm = SIQP(m, self.verbose) P = zeros((m,m)) P[:d,:d] = 2.0*self.reg_param*eye(d) q = vstack([-sw, -sb, ones((n,1))/self.mu]) svm.set_objective(P, q) svm.add_ineq_constraints(hstack([zeros((n,d+1)),-eye(n)]), zeros((n,1))) G = hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n)]) svm.add_ineq_constraints(G, -self.mu*ones((n,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] def solve_opt(self): """Use a DCA to solve the ramp loss SVM""" if self.verbose: print('Solving DCA') (self.w, self.b) = self._init_param(2) self.all_objective.append(self.primal_obj) (sw, sb) = self._subgradient_concave_ramp(self.mu) #print(sw, sb) diff_param = numpy.inf while (self.num_iter < self.max_iter)\ and (diff_param > self.min_diff): self._solve_quadprog(sw, sb) #print('weights') #print(self.w, self.b) (old_sw, old_sb) = (sw, sb) (sw, sb) = self._subgradient_concave_ramp(self.mu) #print(sw, sb) diff_param = sum(abs(old_sw-sw)) + abs(old_sb-sb) #print(diff_param) self.num_iter += 1 self.all_objective.append(self.primal_obj) class RLSVM1(DiffConvexClassifier): """ Ramp Loss SVM with l1 regularizer. """ def __init__(self, reg_param, kdict, mu, verbose=False, rand_init=False): DiffConvexClassifier.__init__(self, reg_param, kdict, verbose, rand_init) self.mu = mu # the slope of the ramp self.loss = RampLoss(mu) self.regularizer = Reg_l1() def _solve_linprog(self, sw, sb): """Solve the linear program corresponding to the convex part of the ramp loss SVM with l1 penalty. Use cvxopt lp solver which expects the following format: min_x c'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi,eta). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving LP') n = len(self.y) d = self.train_ex.shape[0] m = n+2*d+1 svm = SILPS(m, self.verbose) svm.set_objective(vstack([-sw, -sb, ones((n,1))/self.mu, self.reg_param*ones((d,1))])) G = sparse(cm(hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n), zeros((n,d))]))) svm.add_ineq_constraints(G, -self.mu*ones((n,1))) svm.add_ineq_constraints(sparse(cm(hstack([zeros((n,d+1)),-eye(n),zeros((n,d))]))), zeros((n,1))) svm.add_ineq_constraints(sparse(cm(hstack([zeros((d,d+1)),zeros((d,n)),-eye(d)]))), zeros((d,1))) svm.add_ineq_constraints(sparse(cm(hstack([eye(d), zeros((d,n+1)), -eye(d)]))), zeros((d,1))) svm.add_ineq_constraints(sparse(cm(hstack([-eye(d), zeros((d,n+1)), -eye(d)]))), zeros((d,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] def _solve_linprog_dense(self, sw, sb): """Solve the linear program corresponding to the convex part of the ramp loss SVM with l1 penalty. Uses dense matrices Use cvxopt lp solver which expects the following format: min_x c'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi,eta). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving LP') n = len(self.y) d = self.train_ex.shape[0] m = n+2*d+1 svm = SILP(m, self.verbose) svm.set_objective(vstack([-sw, -sb, ones((n,1))/self.mu, self.reg_param*ones((d,1))])) G = hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n), zeros((n,d))]) svm.add_ineq_constraints(G, -self.mu*ones((n,1))) svm.add_ineq_constraints(hstack([zeros((n,d+1)),-eye(n),zeros((n,d))]), zeros((n,1))) svm.add_ineq_constraints(hstack([zeros((d,d+1)),zeros((d,n)),-eye(d)]), zeros((d,1))) svm.add_ineq_constraints(hstack([eye(d), zeros((d,n+1)), -eye(d)]), zeros((d,1))) svm.add_ineq_constraints(hstack([-eye(d), zeros((d,n+1)), -eye(d)]), zeros((d,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] def solve_opt(self): """Use a DCA to solve the ramp loss SVM with one norm penalty""" if self.verbose: print('Solving DCA') (self.w, self.b) = self._init_param(1) self.all_objective.append(self.primal_obj) (sw, sb) = self._subgradient_concave_ramp(self.mu) diff_param = numpy.inf while (self.num_iter < self.max_iter)\ and (diff_param > self.min_diff): self._solve_linprog(sw, sb) (old_sw, old_sb) = (sw, sb) (sw, sb) = self._subgradient_concave_ramp(self.mu) diff_param = sum(abs(old_sw-sw)) + abs(old_sb-sb) self.num_iter += 1 self.all_objective.append(self.primal_obj) class RLSVMC1(DiffConvexClassifier): """ Ramp Loss SVM with capped l1 regularizer. """ def __init__(self, reg_param, kdict, mu, a=1.0, verbose=False, rand_init=False): DiffConvexClassifier.__init__(self, reg_param, kdict, verbose, rand_init) self.mu = mu # the slope of the ramp self.a = a # The value of the cap self.loss = RampLoss(mu) self.regularizer = Reg_l1cap(a) #@property #def margin(self): # """The value of the regularizer, specialized because of parameter a""" # return self.regularizer(self.w, self.a) def _solve_linprog(self, sw, sb, u): """Solve the linear program corresponding to the convex part of the ramp loss SVM with capped l1 penalty. Use cvxopt lp solver which expects the following format: min_x c'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi,eta). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving LP') n = len(self.y) d = self.train_ex.shape[0] m = n+2*d+1 svm = SILPS(m, self.verbose) svm.set_objective(vstack([-sw-self.reg_param*u, -sb, ones((n,1))/self.mu, self.reg_param*ones((d,1))])) G = sparse(cm(hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n), zeros((n,d))]))) svm.add_ineq_constraints(G, -self.mu*ones((n,1))) svm.add_ineq_constraints(sparse(cm(hstack([zeros((n,d+1)),-eye(n),zeros((n,d))]))), zeros((n,1))) svm.add_ineq_constraints(sparse(cm(hstack([zeros((d,d+1)),zeros((d,n)),-eye(d)]))), zeros((d,1))) svm.add_ineq_constraints(sparse(cm(hstack([eye(d), zeros((d,n+1)), -eye(d)]))), zeros((d,1))) svm.add_ineq_constraints(sparse(cm(hstack([-eye(d), zeros((d,n+1)), -eye(d)]))), zeros((d,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] def _solve_linprog_dense(self, sw, sb, u): """Solve the linear program corresponding to the convex part of the ramp loss SVM with capped l1 penalty. Uses dense matrices Use cvxopt lp solver which expects the following format: min_x c'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi,eta). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving LP') n = len(self.y) d = self.train_ex.shape[0] m = n+2*d+1 svm = SILP(m, self.verbose) svm.set_objective(vstack([-sw-self.reg_param*u, -sb, ones((n,1))/self.mu, self.reg_param*ones((d,1))])) G = hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n), zeros((n,d))]) svm.add_ineq_constraints(G, -self.mu*ones((n,1))) svm.add_ineq_constraints(hstack([zeros((n,d+1)),-eye(n),zeros((n,d))]), zeros((n,1))) svm.add_ineq_constraints(hstack([zeros((d,d+1)),zeros((d,n)),-eye(d)]), zeros((d,1))) svm.add_ineq_constraints(hstack([eye(d), zeros((d,n+1)), -eye(d)]), zeros((d,1))) svm.add_ineq_constraints(hstack([-eye(d), zeros((d,n+1)), -eye(d)]), zeros((d,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] def solve_opt(self): """Use a DCA to solve the ramp loss SVM with capped l1 regularizer""" if self.verbose: print('Solving DCA') (self.w, self.b) = self._init_param(1) self.all_objective.append(self.primal_obj) (sw, sb) = self._subgradient_concave_ramp(self.mu) u = self._subgradient_concave_cap_l1(self.a) diff_param = numpy.inf while (self.num_iter < self.max_iter)\ and (diff_param > self.min_diff): self._solve_linprog(sw, sb, u) (old_sw, old_sb, old_u) = (sw, sb, u) (sw, sb) = self._subgradient_concave_ramp(self.mu) u = self._subgradient_concave_cap_l1(self.a) diff_param = sum(abs(old_sw-sw)) + abs(old_sb-sb) + sum(abs(old_u-u)) self.num_iter += 1 self.all_objective.append(self.primal_obj) class RLSVM0(DiffConvexClassifier): """ Ramp Loss SVM with l0 regularizer. """ def __init__(self, reg_param, kdict, mu, a=4.0, verbose=False, rand_init=False): DiffConvexClassifier.__init__(self, reg_param, kdict, verbose, rand_init) self.mu = mu # the slope of the ramp self.a = a # The exponent in approximation self.loss = RampLoss(mu) self.regularizer = Reg_l0(a) #@property #def margin(self): # """The value of the regularizer, specialized because of parameter a""" # return self.regularizer(self.w, self.a) def _solve_linprog(self, sw, sb, u): """Solve the linear program corresponding to the convex part of the ramp loss SVM with l0 penalty. Use cvxopt lp solver which expects the following format: min_x c'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi,eta). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving LP') n = len(self.y) d = self.train_ex.shape[0] m = n+2*d+1 svm = SILPS(m, self.verbose) svm.set_objective(vstack([-sw-self.reg_param*u, -sb, ones((n,1))/self.mu, self.reg_param*self.a*ones((d,1))])) G = sparse(cm(hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n), zeros((n,d))]))) svm.add_ineq_constraints(G, -self.mu*ones((n,1))) svm.add_ineq_constraints(sparse(cm(hstack([zeros((n,d+1)),-eye(n),zeros((n,d))]))), zeros((n,1))) svm.add_ineq_constraints(sparse(cm(hstack([zeros((d,d+1)),zeros((d,n)),-eye(d)]))), zeros((d,1))) svm.add_ineq_constraints(sparse(cm(hstack([eye(d), zeros((d,n+1)), -eye(d)]))), zeros((d,1))) svm.add_ineq_constraints(sparse(cm(hstack([-eye(d), zeros((d,n+1)), -eye(d)]))), zeros((d,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] def _solve_linprog_dense(self, sw, sb, u): """Solve the linear program corresponding to the convex part of the ramp loss SVM with l0 penalty. Uses dense matrices Use cvxopt lp solver which expects the following format: min_x c'x st Gx <= h Ax = b We organise the variables of the SVM as (w,b,xi,eta). """ yvec = self.y.copy() yvec.shape = (len(self.y), 1) if self.verbose: print('Solving LP') n = len(self.y) d = self.train_ex.shape[0] m = n+2*d+1 svm = SILP(m, self.verbose) svm.set_objective(vstack([-sw-self.reg_param*u, -sb, ones((n,1))/self.mu, self.reg_param*self.a*ones((d,1))])) G = hstack([-yvec*transpose(self.train_ex), -yvec, -eye(n), zeros((n,d))]) svm.add_ineq_constraints(G, -self.mu*ones((n,1))) svm.add_ineq_constraints(hstack([zeros((n,d+1)),-eye(n),zeros((n,d))]), zeros((n,1))) svm.add_ineq_constraints(hstack([zeros((d,d+1)),zeros((d,n)),-eye(d)]), zeros((d,1))) svm.add_ineq_constraints(hstack([eye(d), zeros((d,n+1)), -eye(d)]), zeros((d,1))) svm.add_ineq_constraints(hstack([-eye(d), zeros((d,n+1)), -eye(d)]), zeros((d,1))) svm.solve() if svm.optimal: self.w = svm.solution[:d] self.w.shape = (d,1) self.b = svm.solution[d] def solve_opt(self): """Use a DCA to solve the ramp loss SVM with zero norm (approx) regularizer""" if self.verbose: print('Solving DCA') (self.w, self.b) = self._init_param(1) self.all_objective.append(self.primal_obj) (sw, sb) = self._subgradient_concave_ramp(self.mu) u = self._subgradient_concave_l0(self.a) diff_param = numpy.inf while (self.num_iter < self.max_iter)\ and (diff_param > self.min_diff): self._solve_linprog(sw, sb, u) (old_sw, old_sb, old_u) = (sw, sb, u) (sw, sb) = self._subgradient_concave_ramp(self.mu) u = self._subgradient_concave_l0(self.a) diff_param = sum(abs(old_sw-sw)) + abs(old_sb-sb) + sum(abs(old_u-u)) self.num_iter += 1 self.all_objective.append(self.primal_obj) OptWok-0.3.1/optwok/classifier_base.py0000644ڛ|׿-0000002753312055566021024724 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Base classes for classifiers""" from time import time import numpy from numpy import array, matrix, zeros, nonzero class Classifier(object): """ Implementation of the optimization methods for the empirical loss + regularizer framework, for binary clssification. We only consider linear classifiers. """ def __init__(self, reg_param=1.0, kdict=None, verbose=False): assert(reg_param > 0.0) self.reg_param = reg_param # Tradeoff regularizer vs loss self.regularizer = None self.loss = None self.kdict = kdict # kernel self.train_ex = array([]) # training examples self.y = array([]) # labels self.train_time = 0.0 # time used for training self.w = array([]) # weight self.b = array([]) # bias self.epsilon = 1e-6 # how small is zero? self.verbose = verbose # how much output @property def num_examples(self): return len(self.y) @property def num_feat(self): return self.train_ex.shape[0] @property def train_out(self): """Output of the classifier on the training set""" return self.get_svm_out(self.train_ex) @property def margin(self): """The value of the regularizer""" return self.regularizer.margin(self.w) @property def emp_risk(self): """A vector of the empirical loss""" return self.loss.get_emp_risk(self.y*self.train_out) @property def primal_obj(self): """Directly compute the primal objective function.""" if self.verbose: print('Margin: %f, Risk: %f' % (self.margin, self.reg_param*sum(self.emp_risk))) obj = self.margin + self.reg_param*sum(self.emp_risk) return obj @property def max_weight(self): """The maximum absolute value of the weight.""" return max(abs(self.w))[0] @property def frac_sv(self): """The fraction of examples used as support vectors.""" return self.loss.get_frac_sv(self.y*self.train_out) @property def nonzero_feat(self): """The features corresponding to non-zero parameters""" return numpy.where(abs(self.w)>self.epsilon)[0] @property def frac_feat(self): """The fraction of features used in the final classifier, i.e. the number of non-zero parameters""" return float(len(self.nonzero_feat))/float(len(self.w)) @property def frac_total(self): """The fraction of training data used in the final classifier, taking both features and support vectors into account.""" num_sv = float(self.loss.get_num_sv(self.y*self.train_out)) num_feat = float(len(numpy.where(abs(self.w)>self.epsilon)[0])) return (num_sv*num_feat) / (float(self.train_ex.shape[1])*float(len(self.w))) @property def perf(self): """Summary of results""" perf = {'max_weight': self.max_weight, 'nonzero_feat': self.nonzero_feat, 'frac_feat': self.frac_feat, 'frac_sv': self.frac_sv, 'frac_total': self.frac_total, 'train_time': self.train_time, 'objective': self.primal_obj, } return perf def get_svm_out(self, X): """Compute K*alpha+b""" return array(matrix(self.w).T*matrix(X) + self.b).flatten() def solve_opt(self): """solve_opt should the implemented in child class""" pass def train(self, train_ex, train_lab): """Wrapper to convert data to feature matrix, and solve the optimization""" self.train_ex = compute_features(train_ex, self.kdict) self.y = train_lab t_start = time() self.solve_opt() t_end = time() # Collect statistics about optimization self.train_time = t_end-t_start if self.verbose: print('%f support vectors' % self.frac_sv) if self.verbose: print('Primal Objective: %f' % self.primal_obj) def predict(self, pred_ex): """Wrapper to do prediction""" return self.get_svm_out(compute_features(pred_ex, self.kdict)) class ConvexClassifier(Classifier): """ Classifiers which have convex training methods """ def __init__(self, reg_param, kdict, verbose): Classifier.__init__(self, reg_param, kdict, verbose) self.num_iter = 1 class Loss(object): """A loss function, for classification""" epsilon = 0.01 def gradient(self, w): """The vector of gradients""" raise NotImplementedError class HingeLoss(Loss): """ Standard SVM loss """ def get_emp_risk(self, t): """Compute the empirical risk, for each example.""" (sloped, correct) = self.get_loss_case(t) emp_risk = zeros(len(t)) emp_risk[sloped] = 1.0 - t[sloped] return emp_risk def get_num_sv(self, t): """Estimate the number of support vectors""" (sloped, correct) = self.get_loss_case(t) return len(sloped) def get_frac_sv(self, t): """Estimate the fraction of support vectors""" num_sv = self.get_num_sv(t) frac_sv = float(num_sv)/float(len(t)) return frac_sv def get_loss_case(self, t): """Detect whether we are on the slope, or classified correctly (larger than margin).""" sloped = numpy.where(t <= 1.0+self.epsilon)[0] correct = numpy.where(t > 1.0+self.epsilon)[0] return (sloped, correct) def gradient(self, x, y, t): """The gradient of a single example""" if t <= 1.: return -y*x else: return zeros(x.shape) class HingeLossKernel(Loss): """ Standard SVM loss, with matrix inner product """ def __init__(self, M): self.M = M class HuberHingeLoss(Loss): """ Huberized hinge loss, to make things differentiable """ def __init__(self, delta=0.5): """The loss is quadratic in window delta""" self.delta = delta def get_loss_case(self, t): """Detect the three cases, correct, curved, sloped """ sloped = numpy.where(t <= 1.0-self.delta)[0] curved = numpy.where(abs(1-t)<=self.delta)[0] correct = numpy.where(t > 1.0+self.delta)[0] return (sloped, curved, correct) def gradient(self, x, y, t): """The gradient of a single example""" if t <= 1.0-self.delta: return -y*x elif t > 1.0+self.delta: return zeros(x.shape) else: return (-0.5*y/self.delta)*(1+self.delta-t)*x class RampLoss(Loss): """ Truncated hinge loss """ def __init__(self, mu): """The parameter mu controls the slope between [0,2]""" assert((mu > 0.0) and (mu <= 1.0)) self.mu = mu def get_emp_risk(self, t): """Compute the empirical risk, for each example.""" (clipped, sloped, correct) = self.get_loss_case(t) emp_risk = zeros(len(t)) emp_risk[clipped] = self.mu + 1.0 emp_risk[sloped] = 1.0 - (t[sloped]/self.mu) return emp_risk def get_num_sv(self, t): """Estimate the number of support vectors""" (clipped, sloped, correct) = self.get_loss_case(t) return len(sloped) def get_frac_sv(self, t): """Estimate the fraction of support vectors""" num_sv = self.get_num_sv(t) frac_sv = float(num_sv)/float(len(t)) return frac_sv def get_loss_case(self, t): """Detect whether we are at saturation, on the slope, or classified correctly (larger than margin).""" clipped = numpy.where(t<-(self.mu*self.mu)-self.epsilon)[0] lb = numpy.where(t>=-(self.mu*self.mu)-self.epsilon, 1, 0) ub = numpy.where(t<=self.mu+self.epsilon, 1, 0) sloped = numpy.where((lb+ub)>1)[0] correct = numpy.where(t>self.mu+self.epsilon)[0] return (clipped, sloped, correct) class SlantLoss(RampLoss): """ Hinge loss with double kink """ def get_emp_risk(self, t): """Compute the empirical risk, for each example.""" (clipped, sloped, correct) = self.get_loss_case(t) emp_risk = zeros(len(t)) emp_risk[clipped] = 0.5*(3.0-t[clipped]) emp_risk[sloped] = 1.0 - t[sloped] return emp_risk class Regularizer(object): """ A regularizer class """ def margin(self, w): """The value of the regularizer, intuitively the width of the margin""" raise NotImplementedError def gradient(self, w): """The vector of gradients""" raise NotImplementedError class Reg_l2(Regularizer): """The l2 norm of a vector""" def margin(self, w): """The value of the regularizer, intuitively the width of the margin""" return 0.5*numpy.linalg.norm(w)*numpy.linalg.norm(w) def gradient(self, w): """The vector of gradients""" return w class Reg_l1(Regularizer): """The l1 norm of a vector""" def margin(self, w): """The value of the regularizer, intuitively the width of the margin""" return sum(abs(w)) def gradient(self, w): """The vector of gradients""" return sign(w) class Reg_l1cap(Regularizer): """The l1 norm of a vector""" def __init__(self, a=1.0): """'a' is the value of the cap""" assert(a > 0.0) self.a = a def margin(self, w, a=None): """The value of the regularizer, intuitively the width of the margin""" if a is not None: assert(a > 0.0) self.a = a cap_l1 = abs(w) cap_l1[nonzero(cap_l1 > self.a)[0]] = self.a return sum(cap_l1) def gradient(self, w, a=None): """The vector of gradients""" if a is not None: assert(a > 0.0) self.a = a cap_l1 = abs(w) grad = sign(w) grad[nonzero(cap_l1 > self.a)[0]] = 0. return grad class Reg_l0(Regularizer): """The l1 norm of a vector""" def __init__(self, a=1.0): """'a' is the coefficient of the exponent""" assert(a > 0.0) self.a = a def margin(self, w, a=None): """The value of the regularizer, intuitively the width of the margin""" if a is not None: assert(a > 0.0) self.a = a return sum(1.0 - numpy.exp(-self.a*abs(w))) def gradient(self, w, a=None): """The vector of gradients""" if a is not None: assert(a > 0.0) self.a = a return -self.a*sign(w)*numpy.exp(-self.a*abs(w)) class Reg_elastic_net(Regularizer): """The elastic net regularizer (l1+l2)""" def margin(self, w): """The value of the regularizer""" return 0.5*numpy.linalg.norm(w)*numpy.linalg.norm(w) + sum(abs(w)) def gradient(self, w): """The vector of gradients""" return w + sign(w) class Reg_graph_net(Regularizer): """The elastic net regularizer, with graph laplacian in l2""" def __init__(self, M): self.M = M def margin(self, w): """The value of the regularizer""" return 0.5*dot(dot(w,self.M),w) + sum(abs(w)) def gradient(self, w): """The vector of gradients""" return dot(self.M,w) + sign(w) def compute_features(train_ex, kdict): """Explicitly compute the feature space of the kernel""" if kdict['name'] == 'linear': return train_ex elif kdict['name'] == 'poly': if kdict['inhomog'] == True: train_ex = vstack([train_ex, ones((1,train_ex.shape[1]))]) phi = get_poly_features(train_ex, kdict['param']) return phi else: print('Explicit feature construction for kernel %s not implemented' % kdict['name']) OptWok-0.3.1/optwok/context_ucb.py0000644ڛ|׿-0000002130312055557045024116 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Contextual Gaussian Process Upper Confidence Bound""" import os import copy from numpy import delete, arange, kron, unique, array from numpy import argmax, argsort from numpy import log, log2, log10, zeros, ones, matrix, eye, sqrt from numpy import max, min, mean from numpy import flatnonzero from numpy.random import randint, permutation, randn from matplotlib.pyplot import figure, plot, legend, title, show from matplotlib.pyplot import xlabel, ylabel, axis, savefig from optwok import datatools from io_pickle import load, save class BanditResults(object): """A container to collect the data for a contextual bandit experiment""" def __init__(self, verbose=False): """Does nothing except set verbosity""" self.verbose = verbose pass def init_data(self, data, num_context, num_epoch): """Make a copy of the data and initialize results""" self.data = copy.deepcopy(data) self.r_sum = zeros(num_context*(num_epoch-1)+1) self.r_sum[0] = data.max_labels[0] - data.min_label self.r_max = zeros((num_context, num_epoch)) for context in range(num_context): self.r_max[context,0] = self.data.max_labels[context] self.selection = zeros((num_context, num_epoch)) self.observed = zeros((num_context, num_epoch)) self.num_expt = 0 def do_experiment(self, context, epoch, pred_idx, pred_val=None): """Simulate an experiment""" y = self.data.observe(pred_idx, context) regret = self.data.max_labels[context] - y self.r_sum[self.num_expt+1] = self.r_sum[self.num_expt] + regret self.num_expt += 1 self.r_max[context, epoch] = min([self.r_max[context, epoch-1], regret]) self.selection[context, epoch] = pred_idx self.observed[context, epoch] = y if pred_val: if self.verbose: print('chose = %d, UCB = %f, observation = %f' % (pred_idx, pred_val, y)) if pred_val < y: print('=== Observation %f higher than UCB %f!' % (y, pred_val)) return y def save_result(self, filename): """Save the results to file with pickle""" save(filename, (self.r_sum, self.r_max, self.selection, self.observed)) def load_result(self, filename): """Initialize results from file""" (self.r_sum, self.r_max, self.selection, self.observed) = load(filename) def compute_upper_bound(preds, pred_var, epoch, num_data): """Compute the upper confidence bound""" beta = 2.0 upper_bound = preds + beta*log(num_data*(epoch+1.0))*sqrt(pred_var) diff = max([(max(upper_bound) - min(upper_bound))/abs(mean(upper_bound)),0.01]) # add some random noise upper_bound += 0.01*diff*randn(num_data) diff = (max(upper_bound) - min(upper_bound))/abs(mean(upper_bound)) if diff < 1e-5: print('==== UCB very flat!') print('Context=%d, epoch=%d' % (context, epoch)) print(max(upper_bound), min(upper_bound), mean(upper_bound), diff) return upper_bound def ucb(examples, context, gp, epoch, verbose): """Upper confidence bound algorithm with context""" num_data = examples.shape[1] (preds, pred_var) = gp.predict((examples, context*ones(num_data))) upper_bound = compute_upper_bound(preds, pred_var, epoch, num_data) max_possible = argmax(upper_bound) if verbose: print('Preds(max)=%f, pred_sd(max)=%f' % (preds[max_possible], sqrt(pred_var[max_possible]))) #plot(examples.flatten(), preds, 'b.') #plot(examples.flatten(), upper_bound, 'r.') #plot(examples[:,max_possible], preds[max_possible], 'go') #show() return max_possible, upper_bound[max_possible] def ucb_kbest(examples, context, gp, epoch, batch_size, verbose): """Predict the k best experiments from the GP""" num_data = examples.shape[1] (preds, pred_var) = gp.predict((examples, context*ones(num_data))) upper_bound = compute_upper_bound(preds, pred_var, epoch, num_data) idx = argsort(upper_bound) kbest = idx[-batch_size:] return kbest def ucb_batch(examples, context, gp, epoch, batch_size, verbose): """Predict small batches, using ucb, of examples to test""" local_gp = copy.deepcopy(gp) local_gp.reg_param /= float(batch_size) all_pred_idx = [] for batch_idx in range(batch_size): (pred_idx, pred_val) = ucb(examples, context, local_gp, epoch, verbose) all_pred_idx.append(pred_idx) local_gp.update(examples[:,pred_idx], context, pred_val) return all_pred_idx def expt_cucb(data, gp, savefile, num_context=3, num_epoch=15, verbose=False): """Estimate the learning curves for ucb""" print(savefile) results = BanditResults() results.init_data(data, num_context, num_epoch) for context in range(num_context): for epoch in range(1,num_epoch): if verbose: print('Context=%d, epoch=%d' % (context, epoch)) (pred_idx, pred_val) = ucb(data.examples(context), context, gp, epoch, verbose) y = results.do_experiment(context, epoch, pred_idx, pred_val) gp.update(data.examples(context)[:,pred_idx], context, y) results.save_result(savefile) def expt_cucb_batch(data, gp, savefile, num_context=3, num_epoch=15, batch_size=5, mode='greedy', verbose=False): """Estimate the learning curves for batch ucb""" print(savefile) results = BanditResults() results.init_data(data, num_context, (num_epoch-1)*batch_size+1) for context in range(num_context): for epoch in range(1,num_epoch): if verbose: print('Context=%d, epoch=%d' % (context, epoch)) if mode == 'greedy': all_pred_idx = ucb_batch(data.examples(context), context, gp, epoch*batch_size, batch_size, verbose) elif mode == 'kbest': all_pred_idx = ucb_kbest(data.examples(context), context, gp, epoch*batch_size, batch_size, verbose) elif mode == 'random': idx = permutation(len(data.examples(context))) all_pred_idx = idx[:batch_size] for pred_idx in all_pred_idx: y = results.do_experiment(context, epoch, pred_idx) gp.update(data.examples(context)[:,pred_idx], context, y) results.save_result(savefile) def plot_min_regret(filename, leg_list=None): print(filename) (r_sum, r_max, selection, observed) = load(filename) (num_context, num_epoch) = r_max.shape if leg_list is None: leg_list = map(str,range(num_context)) for ix, task in enumerate(leg_list): pos = flatnonzero(observed[ix,:]>0) print('%s:%d unique positives from %d' % (task, len(unique(selection[ix,pos])), num_epoch)) figure() plot(range(0,num_epoch),r_max.T, hold=True) #plot(range(0,num_epoch),r_max.T, 'k.') plot([0,num_epoch],[0,0], 'k--') legend(leg_list, loc=4) title('minimum regret') xlabel('number of rounds') #axis([0,num_epoch-1,-3,3]) savefig('%s_min.pdf' % filename[:-4], dpi=300, bbox_inches='tight') def plot_average_regret(file_list, leg_list): """Plot all average regrets, assuming that the file and legend order is the same. """ colours = ['b','r','g','c','m','k','y'] assert len(file_list) == len(leg_list) figure() for (fidx, file_name) in enumerate(file_list): (r_sum, r_max, selection, observed) = load(file_name) (num_context, num_epoch) = r_max.shape plot(arange(1,len(r_sum)+1), r_sum/(arange(1,len(r_sum)+1)), colours[fidx]) #for ix in range(1,num_context): # plot([(num_epoch-1)*ix,(num_epoch-1)*ix], [0,5], 'k:') legend(leg_list, loc=1) title('average regret') xlabel('number of rounds') #axis([1,num_context*(num_epoch-1), 1, 4]) savefig('%s_avg.pdf' % file_list[0][:-4], dpi=300, bbox_inches='tight') def plot_average_regret_old(f_learn, f_cat, f_ind): figure() (r_sum, r_max, selection, observed) = load(f_learn) (num_context, num_epoch) = r_max.shape plot(arange(1,len(r_sum)+1), r_sum/(arange(1,len(r_sum)+1)), 'b') (r_sum, r_max, selection, observed) = load(f_cat) plot(arange(1,len(r_sum)+1), r_sum/(arange(1,len(r_sum)+1)), 'r') (r_sum, r_max, selection, observed) = load(f_ind) plot(arange(1,len(r_sum)+1), r_sum/(arange(1,len(r_sum)+1)), 'g') #for ix in range(1,num_context): # plot([(num_epoch-1)*ix,(num_epoch-1)*ix], [0,5], 'k:') legend(['CGP-UCB','merge','ignore'], loc=1) title('average regret') xlabel('number of rounds') #axis([1,num_context*(num_epoch-1), 1, 4]) savefig('%s_avg.pdf' % f_learn[:-4], dpi=300, bbox_inches='tight') OptWok-0.3.1/optwok/datatools.py0000644ڛ|׿-0000001203412055531443023566 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Data management""" from numpy import array, transpose, mean from numpy import arange, hstack, kron, ones, eye from numpy import sqrt, concatenate, zeros, log10 from numpy.random import permutation, seed from numpy import flatnonzero from numpy.linalg import norm class Data(object): """A container for contextual bandit data""" def __init__(self, examples, contexts, labels): self.num_context = int(max(contexts))+1 self._examples = [] self._contexts = [] self._labels = [] self.min_label = 10000 self.max_labels = -10000*ones(self.num_context) for con in range(self.num_context): idx = flatnonzero(contexts==con) self._examples.append(examples[:,idx]) self._contexts.append(contexts[idx]) self._labels.append(labels[idx]) self.min_label = min(self.min_label, min(labels[:,idx])) self.max_labels[con] = max(labels[:,idx]) def __repr__(self): return str((self._examples, self._contexts, self._labels)) def remove(self, idx, context): """Delete data item indexed by idx""" self._examples[context] = delete(self._examples[context], idx, 1) self._contexts[context] = delete(self._contexts[context], idx) self._labels[context] = delete(self._labels[context], idx) def examples(self, context): """Return all active data for given context""" return self._examples[context] def observe(self, idx, context): """Return the label for idx""" return self._labels[context][idx] def set_label(self, idx, context, val): """Set the value of label""" self._labels[context][idx] = val def data_stats(all_data, task_list=None, num_cv=5): """Print some statistics about each dataset""" if task_list: all_tasks = task_list else: all_tasks = all_data.keys() for task in all_tasks: data = all_data[task] print('Number of examples in %s : %d' % (task, data.nof_peptides)) def _get_data(data, splits, subsample, num_feat): examples = zeros((num_feat,0)) labels = [] for cv in splits: # Normalize to unit length new_ex = transpose(array(data.encoded_peptides[cv])) for ix in range(new_ex.shape[1]): new_ex[:,ix] /= norm(new_ex[:,ix]) examples = hstack([examples, new_ex]) labels.extend(-log10(data.ic50s[cv])+log10(500)) examples = array(examples) labels = array(labels) if subsample: idx = flatnonzero(labels<-1.6) perm = permutation(len(idx)) subidx = idx[perm[:20]] idx = concatenate([flatnonzero(labels>=-1.6),subidx]) examples = examples[:,idx] labels = labels[idx] return examples, labels def get_data(all_data, splits, all_task=None, subsample=False, num_feat=45): """Concatenate data from the splits, randomly subsample 20 examples from labels < -1.6 """ if all_task is None: return _get_data(all_data, splits, subsample, num_feat) else: examples = zeros((num_feat,0)) labels = array([]) contexts = array([]) for (task_id, task) in enumerate(all_task): data = all_data[task] (new_ex, new_lab) = _get_data(data, splits, subsample, num_feat) examples = hstack([examples, new_ex]) labels = hstack([labels, new_lab]) contexts = concatenate([contexts, task_id*ones(len(new_lab))]) return (examples, contexts, labels) def get_split(data, split, num_cv=5, verbose=False): """Concatenate the data not in 'split' into the test set. 'split' is the training set. """ test_idx = range(num_cv) test_idx.remove(split) (train_ex, train_lab) = get_data(data, [split]) (test_ex, test_lab) = get_data(data, test_idx) if verbose: print('Split %d' % split) print('features=%d, num_train=%d, num_test=%d'\ % (train_ex.shape[0], train_ex.shape[1], test_ex.shape[1])) return train_ex, train_lab, test_ex, test_lab def get_multitask(all_tasks, all_data, split): """Treat each task as a different label""" train_ex = [] train_lab_list = [] test_ex = [] test_lab_list = [] num_train = 0 num_test = 0 for task in all_tasks: data = all_data[task] tre, trl, tee, tel = get_split(data, split) num_train += len(trl) num_test += len(tel) train_ex.append(tre) train_lab_list.append(trl) test_ex.append(tee) test_lab_list.append(tel) train_ex = hstack(train_ex) test_ex = hstack(test_ex) # Construct the matrix of labels train_lab = zeros((len(all_tasks), num_train)) idx = 0 for (ix, trl) in enumerate(train_lab_list): train_lab[ix, idx:idx+len(trl)] = trl idx += len(trl) test_lab = zeros((len(all_tasks), num_test)) idx = 0 for (ix, tel) in enumerate(test_lab_list): test_lab[ix, idx:idx+len(tel)] = tel idx += len(tel) return train_ex, train_lab, test_ex, test_lab OptWok-0.3.1/optwok/evaluation.py0000644ڛ|׿-0000002022111756655247023760 0ustar cheng.ongUNIMELB\Domain Users00000000000000from numpy import sign, nonzero, equal, zeros, mean, std from numpy import array, concatenate, sqrt, diag, matrix from numpy import log, pi, trace from numpy.linalg import det, pinv from numpy import sum, arange, nan, unique, argsort, isnan, cumsum from scipy import stats def confusionMatrix(labels_test, labels_predicted): """Compute the matrix of predictions versus labels""" if len(labels_test) != len(labels_predicted): return 0 TP = 0; FP = 0; TN = 0; FN = 0 for i in range(0, len(labels_test)): if labels_test[i] == 0 or labels_predicted[i] == 0: return 0 if labels_test[i] > 0: if labels_predicted[i] > 0: TP += 1 else: FN +=1 else: if labels_predicted[i] > 0: FP += 1 else: TN += 1 return (TP, TN, FP, FN) def accuracy(output, labels_test): """How many correct predictions?""" if max(labels_test) > 1: return accuracy_multiclass(output, labels_test) else: TP, TN, FP, FN = confusionMatrix(labels_test, sign(output)) return float(TP + TN) / (TP + TN + FP + FN) def balanced_accuracy(output, labels): """How many correct predictions?, normalized by number of positives and negatives""" assert all(unique(labels)==array([-1,1])), 'Binary classification only' TP, TN, FP, FN = confusionMatrix(labels, sign(output)) return 0.5*TP/(TP+FN) + 0.5*TN/(FP+TN) def balanced_accuracy_multitask(output, labels): """Balanced accuracy applied to each task separately""" assert output.shape == labels.shape, 'Predictions and labels have different shape' num_task = output.shape[0] balacc = zeros(num_task) for ix in range(num_task): lab_idx = nonzero(labels[ix,:]) balacc[ix] = balanced_accuracy(output[ix,lab_idx].flatten(), labels[ix,lab_idx].flatten()) return balacc def accuracy_multiclass(output, labels_test): """Multiclass accuracy""" int_out = map(int, output) int_lab = map(int, labels_test) return 1.0*len(nonzero(equal(int_out, int_lab))[0])/len(output) def accuracy_mix(output, labels_test, mix_list): """Multiclass accuracy, but allowing class confusions, as defined by mix_list""" assert(output.shape == labels_test.shape) int_out = map(int, output) new_out = zeros(len(int_out), dtype=int) for cid in range(len(mix_list)): idx = array([], dtype=int) for mix in mix_list[cid]: idx = concatenate([idx, nonzero(equal(int_out, mix))[0]]) new_out[idx] = cid int_lab = map(int, labels_test) new_lab = zeros(len(int_lab), dtype=int) for cid in range(len(mix_list)): idx = array([], dtype=int) for mix in mix_list[cid]: idx = concatenate([idx, nonzero(equal(int_lab, mix))[0]]) new_lab[idx] = cid return 1.0*len(nonzero(equal(new_out, new_lab))[0])/len(output) def rmse(output, labels): """Root mean squared error""" if output.ndim == 1: num_ex = len(output) else: num_ex = output.shape[1] return sqrt(diag(matrix(output - labels)*matrix(output-labels).T))/num_ex def rmse_multitask(output, labels): """rmse for each task separately""" assert output.shape == labels.shape, 'Predictions and labels have different shape' num_task = output.shape[0] error = zeros(num_task) for ix in range(num_task): lab_idx = nonzero(labels[ix,:]) error[ix] = rmse(output[ix,lab_idx].flatten(), labels[ix,lab_idx].flatten()) return error def differential_entropy(K): """The differential entropy for a multivariate normal distribution. Assume K is a positive semidefinite matrix.""" d = K.shape[0] return (d/2)*(1+log(2*pi)) + 0.5*log(det(K)) def relative_entropy(K): """The relative entropy for a multivariate normal distribution, compared to another Gaussian with the same mean and identity covariance. """ d = K.shape[0] return 0.5*(log(det(pinv(K))) + trace(K) - d) def trapz(x, y, ylim0, ylim1): """Trapezoidal rule for integrating the curve defined by x-y pairs. Assume x is in the range [0,1] and y is constant to the ends of the interval """ assert len(x) == len(y), 'x and y need to be of same length' x = concatenate([x, array([0.0, 1.0])]) y = concatenate([y, array([ylim0, ylim1])]) sort_idx = argsort(x) sx = x[sort_idx] sy = y[sort_idx] area = 0.0 for ix in range(len(x)-1): area += 0.5*(sx[ix+1]-sx[ix])*(sy[ix+1]+sy[ix]) return area def stats_empirical(output, labels): """Compute some statistics for binary predictions. tpr - true positive rate (recall) fpr - false positive rate ppv - positive predictive value (precision) auc - area under the ROC curve ap - area under the precision-recall curve (average precision) """ assert len(output)==len(labels), 'Predictions and labels have different lengths' assert all(unique(labels)==array([-1,1])), 'Labels are not binary {-1,+1}' # Sort true labels according to predictions in ascending order n = len(output) sort_idx = argsort(output) sorted_labels = labels[sort_idx] tpr = [] fpr = [] ppv = [] for thres in range(n+1): tp = 1.0*sum(sorted_labels[thres:]>0) fn = 1.0*sum(sorted_labels[:thres]>0) tn = 1.0*sum(sorted_labels[:thres]<0) fp = 1.0*sum(sorted_labels[thres:]<0) if tp+fn > 0.0: tpr.append(tp/(tp+fn)) else: tpr.append(nan) if fp+tn > 0.0: fpr.append(fp/(fp+tn)) else: fpr.append(nan) if tp+fp > 0.0: ppv.append(tp/(tp+fp)) else: ppv.append(nan) tpr = array(tpr) fpr = array(fpr) ppv = array(ppv) auc = trapz(fpr, tpr, 0.0, 1.0) idx = -isnan(ppv) apr = trapz(tpr[idx], ppv[idx], 1.0, 0.0) return tpr, fpr, ppv, auc, apr def stats_binormal(output, labels, step=0.001): """Compute some statistics for binary predictions. tpr - true positive rate (recall) fpr - false positive rate ppv - positive predictive value (precision) auc - area under the ROC curve ap - area under the precision-recall curve (average precision) Use the binormal assumption. step gives the smoothness of curve. """ assert len(output)==len(labels), 'Predictions and labels have different lengths' assert all(unique(labels)==array([-1,1])), 'Labels are not binary {-1,+1}' # Estimate the binormal parameters pos = output[labels>0] neg = output[labels<0] mu_pos = mean(pos) mu_neg = mean(neg) std_pos = std(pos) std_neg = std(neg) alpha = 1.0*len(pos)/len(output) # Sanity checks assert mu_pos > mu_neg, 'positive Gaussian is not to the right of negative' assert (std_pos>0) and (std_neg>0), 'Variance is zero' # Use Gaussian cdf to estimate scores thres = arange(mu_neg-5.0*std_neg, mu_pos+5*std_pos, step) tp = alpha*(1.0-stats.norm.cdf(thres, mu_pos, std_pos)) fp = (1.0-alpha)*(1.0-stats.norm.cdf(thres, mu_neg, std_neg)) fn = alpha*stats.norm.cdf(thres, mu_pos, std_pos) tn = (1.0-alpha)*stats.norm.cdf(thres, mu_neg, std_neg) tpr = tp/(tp+fn) fpr = fp/(fp+tn) ppv = tp/(tp+fp) A = (mu_pos-mu_neg)/std_pos B = std_neg/std_pos auc = stats.norm.cdf(A/sqrt(1+B*B)) apr = trapz(tpr, ppv, 1.0, 0.0) return tpr, fpr, ppv, auc, apr def stats_binormal_multitask(output, labels): """stats_binormal applied to each row""" assert output.shape == labels.shape, 'Predictions and labels have different shape' num_task = output.shape[0] tpr = [] fpr = [] ppv = [] auc = zeros(num_task) apr = zeros(num_task) for ix in range(num_task): lab_idx = nonzero(labels[ix,:]) ctpr, cfpr, cppv, auc[ix], apr[ix] = stats_binormal(output[ix,lab_idx].flatten(), labels[ix,lab_idx].flatten()) tpr.append(ctpr) fpr.append(cfpr) ppv.append(cppv) return tpr, fpr, ppv, auc, apr def auc(output, labels): """The area under the ROC curve, estimated using the binormal approximation """ tpr, fpr, ppv, auc, apr = stats_empirical(output, labels) return auc OptWok-0.3.1/optwok/expttools.py0000644ڛ|׿-0000003704112140420225023630 0ustar cheng.ongUNIMELB\Domain Users00000000000000 """Some scripts to help run machine learning experiments. Based on Hans Peter Langtangen's scripts for multipleloop """ import re, operator import sys import types import time multiprocessing_present = True try: import multiprocessing except ImportError as detail: print("Error importing multiprocessing. Local computing limited to one CPU.") print("Please install python2.6 or the backport of the multiprocessing package") print(detail) multiprocessing_present = False # Turn off multiprocessing #multiprocessing_present = False drmaa_present = True pythongrid_present = False try: import drmaa except ImportError as detail: print("Error importing drmaa. Limited to local computing.") print("See http://code.google.com/p/drmaa-python/""") print(detail) drmaa_present = False if drmaa_present: pythongrid_present = True try: import pythongrid except ImportError as detail: print("Error importing pythongrid. Limited to local computing.") print("See http://bitbucket.org/chengsoon.ong/pythongrid/overview/") print(detail) pythongrid_present = False # Turn off pythongrid #pythongrid_present = False from optwok.expttools_helper import _wrap_kwargs from optwok.expttools_helper import _pickle_method, _unpickle_method python_version = sys.version_info[:2] if python_version >= (3, 0): import copyreg as cr else: import copy_reg as cr cr.pickle(types.MethodType, _pickle_method, _unpickle_method) def apply_to_combination(function_handle, param, fixed=dict(), verbose=False, multiproc=True, cluster=False, check_result=None, lsf_param=None, blocking=True, addpath=[]): """Apply the function defined by function handle to the tensor product of the parameters defined by dictionary param. param is a dictionary of lists, each key defining the keyword argument to be passed to the function. The function is applied to all combinations of all the settings. The dictionary defined by fixed is concatenated to the end of the keywords. Set blocking to False if you don't want to wait for all jobs to finish on the cluster addpath is a list of directories you want to have added to pythonpath before the execution of the job, the current directory for instance. """ assert(param.keys() != fixed.keys()) # Convert single parameters into singletons # param = dict([(item[0], [item[1]]) if not isinstance(item[1],list) else item for item in param.iteritems()]) all_values = [(name, input2values(param[name])) for name in param] (params, names, varied) = combine(all_values) function_inputs = [] for cur_param in params: kwargs = fixed.copy() kwargs.update(dict((names[ix], cur_param[ix]) for ix in range(len(names)))) if verbose: print('apply_to_combination') print(function_handle) print(kwargs) print('====================') function_inputs.append(kwargs) if drmaa_present and cluster: jobs = [] for kwargs in function_inputs: create_job = True if check_result: create_job = not check_result(**kwargs) if create_job: job = pythongrid.LSFJob(function_handle, kwlist=kwargs, cleanup=True) #print('adding ', addpath, ' in optwok atc') job.addpath = addpath if lsf_param: job.wall_time_h = lsf_param['wall_time_h'] job.wall_time_m = lsf_param['wall_time_m'] job.memory = lsf_param['memory'] jobs.append(job) print('Processing %d jobs' % len(jobs)) time.sleep(5) (sid, jobids) = pythongrid.submit_jobs(jobs) function_outputs = [] if blocking: processed_jobs = pythongrid.collect_jobs(sid, jobids, jobs, wait=True) for job in processed_jobs: function_outputs.append(job.ret) elif multiprocessing_present and multiproc: #cpus = multiprocessing.Pool(processes=3) cpus = multiprocessing.Pool() process_inputs = [] for kwargs in function_inputs: if not check_result(**kwargs): kwargs['_function_handle'] = function_handle process_inputs += [kwargs] function_outputs = cpus.map(_wrap_kwargs, process_inputs) else: function_outputs = [] for kwargs in function_inputs: cur_out = function_handle(**kwargs) function_outputs.append(cur_out) return function_outputs def input2values(s): """ Translate a string s with multiple loop syntax into a list of single values (for the corresponding parameter). Multiple loop syntax: '-1 & -3.4 & 20 & 70 & [0:10,1.3] & [0:10] & 11.76' That is, & is delimiter between different values, [0:10,1.3] generates a loop from 0 up to and including 10 with steps 1.3, [0:10] generates the integers 1,2,...,10. Interactive session:: >>> input2values('-1 & -3.4 & 20 & 70 & [0:10,1.3] & [0:10] & 11.76') [-1, -3.3999999999999999, 20, 70, 0, 1.3, 2.6000000000000001, 3.9000000000000004, 5.2000000000000002, 6.5, 7.7999999999999998, 9.0999999999999996, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11.76] >>> p = {'w': '[0.7:1.3,0.1]', 'b': '1 & 0.3 & 0', 'func': 'y & siny'} >>> print(input2values(p['w'])) [0.69999999999999996, 0.79999999999999993, 0.89999999999999991, 0.99999999999999989, 1.0999999999999999, 1.2, 1.3] >>> print(input2values(p['b'])) [1, 0.29999999999999999, 0] >>> print(input2values(p['func'])) ['y', 'siny'] >>> prm_values = [(name, input2values(p[name])) for name in p] >>> prm_values [('b', [1, 0.29999999999999999, 0]), ('func', ['y', 'siny']), ('w', [0.69999999999999996, 0.79999999999999993, 0.89999999999999991, 0.99999999999999989, 1.0999999999999999, 1.2, 1.3])] """ if not isinstance(s, type('')): return s items = s.split('&') values = [] for i in items: i = i.strip() # white space has no meaning # is i a loop? m = re.search(r'\[(.+):([^,]+),?(.*)\]',i) if m: # the group are numbers, take eval to get right type start = eval(m.group(1)) stop = eval(m.group(2)) try: incr = m.group(3).strip() incr_op = operator.add if incr[0] == '*': incr_op = operator.mul incr = incr[1:] elif incr[0] == '+' or incr[0] == '-': incr = incr[1:] incr = eval(incr) except: incr = 1 r = start while (r <= stop and start <= stop) or \ (r >= stop and start >= stop): values.append(r) r = incr_op(r, incr) else: # just an ordinary item, convert i to right type: values.append(str2obj(i)) # return list only if there are more than one item: if len(values) == 1: return values[0] else: return values def _outer(a, b): """ Return the outer product/combination of two lists. a is a multi- or one-dimensional list, b is a one-dimensional list, tuple, NumPy array or scalar (new parameter) Return: outer combination 'all'. The function is to be called repeatedly:: all = _outer(all, p) """ all = [] if not isinstance(a, list): raise TypeError('a must be a list') if isinstance(b, (float,int,complex,type(''))): b = [b] # scalar? if len(a) == 0: # first call: for j in b: all.append([j]) else: for j in b: for i in a: if not isinstance(i, list): raise TypeError('a must be list of list') # note: i refers to a list; i.append(j) changes # the underlying list (in a), which is not what # we want, we need a copy, extend the copy, and # add to all k = i + [j] # extend previous prms with new one all.append(k) return all def combine(prm_values): """ @param prm_values: nested list of (parameter_name, list_of_parameter_values) or dictionary prm_values[parameter_name] = list_of_parameter_values @type prm_values: list or dict @return: (all,names,varied) where - all contains all combinations (experiments) all[i] is the list of individual parameter values in experiment no i - names contains a list of all parameter names - varied holds a list of parameter names that are varied (i.e. where there is more than one value of the parameter) Code example:: >>> dx = array([1.0/2**k for k in range(2,5)]) >>> dt = 3*dx; dt = dt[:-1] >>> p = {'dx': dx, 'dt': dt} >>> p {'dt': [ 0.75 , 0.375,], 'dx': [ 0.25 , 0.125 , 0.0625,]} >>> all, names, varied = combine(p) >>> all [[0.75, 0.25], [0.375, 0.25], [0.75, 0.125], [0.375, 0.125], [0.75, 0.0625], [0.375, 0.0625]] """ if isinstance(prm_values, dict): # turn dict into list [(name,values),(name,values),...]: prm_values = [(name, prm_values[name]) \ for name in prm_values] all = [] varied = [] for name, values in prm_values: all = _outer(all, values) if isinstance(values, list) and len(values) > 1: varied.append(name) names = [name for name, values in prm_values] return all, names, varied def dump(all, names, varied): e = 1 for experiment in all: print('Experiment %4d:' % e,) for name, value in zip(names, experiment): print('%s:' % name, value,) print('')# newline e += 1 # experiment counter for experiment in all: cmd = ' '.join(['-'+name+' '+repr(value) for \ name, value in zip(names, experiment)]) print(cmd) def options(all, names, prefix='--'): """ Return a list of command-line options. @param all: all[i] holds a list of parameter values in experiment no i @param names: names[i] holds name of parameter no. i @return: cmd[i] holds -name value pairs of all parameters in experiment no. i """ cmd = [] for experiment in all: cmd.append(' '.join([prefix + name + ' ' + repr(str2obj(value)) \ for name, value in zip(names, experiment)])) return cmd def varied_parameters(parameters, varied, names): """ @param names: names of parameters. @param parameters: values of parameters. @param varied: subset of names (the parameters that are varied elsewhere). @return: a list of the items in parameters whose names are listed in varied. An example may help to show the idea. Assume we have three parametes named 'a', 'b', and 'c'. Their values are 1, 5, and 3, i.e., 'a' is 1, 'b' is 5, and 'c' is 3. In a loop elsewhere we assume that 'a' and 'c' are varied while 'b' is fixed. This function returns a list of the parameter values that correspond to varied parameters, i.e., [1,3] in this case, corresponding to the names 'a' and 'c':: >>> parameters = [1,5,3] >>> names = ['a','b','c'] >>> varied = ['a','c'] >>> varied_parameteres(parameters, varied, names) [1,3] """ indices_varied = [names.index(i) for i in varied] varied_parameters = [parameters[i] for i in indices_varied] return varied_parameters def remove(condition, all, names): """ Remove experiments that fulfill a boolean condition. Example: all = remove('w < 1.0 and p = 1.2) or (q in (1,2,3) and f < 0.1', all, names) (names of the parametes must be used) """ import copy for ex in copy.deepcopy(all): # iterate over a copy of all! c = condition for n in names: # replace names by actual values #print('replace "%s" by "%s"' % (n, repr(ex[names.index(n)]))) c = c.replace(n, repr(ex[names.index(n)])) # note the use of repr: strings must be quoted #print('remove ',remove) if eval(c): # if condition all.remove(ex) return all # modified list def str2obj(s, globals=globals(), locals=locals()): """ Turn string s into the corresponding object. eval(s) normally does this, but if s is just a string ready from file, GUI or the command-line, eval will not work when s really represents a string: >>> eval('some string') Traceback (most recent call last): SyntaxError: unexpected EOF while parsing It tries to parse 'some string' as Python code. In this function we try to eval(s), and if it works, we return that object. If it does not work, s probably has meaning as a string, and we return just s. Examples:: >>> from misc import str2obj >>> s = str2obj('0.3') >>> print(s, type(s)) 0.3 >>> s = str2obj('3') >>> print(s, type(s)) 3 >>> s = str2obj('(1,8)') >>> print(s, type(s)) (1, 8) >>> s = str2obj('some string') >>> s 'some string' If the name of a user defined function, class or instance is sent to str2obj, one must also provide locals() and globals() dictionaries as extra arguments. Otherwise, str2obj will not know how to "eval" the string and produce the right object (user defined types are not known inside str2obj). Here is an example:: >>> def myf(x): ... return 1+x ... >>> class A: ... pass ... >>> a = A() >>> >>> s = str2obj('myf') >>> print(s, type(s) # now s is simply the string 'myf') myf >>> # provide locals and globals such that we get the function myf: >>> s = str2obj('myf', locals(), globals()) >>> print(s, type(s)) >>> s = str2obj('a', locals(), globals()) >>> print(s, type(s)) <__main__.A instance at 0xb70f6fcc> Caveat: if the string argument is the name of a valid Python class (type), that class will be returned. For example, >>> str2obj('list') # returns class list You can normally safely apply eval on the output of this function. """ try: s = eval(s, globals, locals) return s except: return s def _inner_func(a=1.0, b=2.0, x=3.0): """For use with _test1()""" return (a+b-x) def _test1(): """Test apply_to_combination""" from optwok.expttools import _inner_func params = {'a': [-0.2, 0.2, 1.0], 'b': [0.3, 3.2], } fixed_param = {'x': 1.3} print("Testing apply_to_combination") print("Single CPU") results = apply_to_combination(_inner_func, params, fixed=fixed_param, multiproc=False, cluster=False) print(results) del results if multiprocessing_present: print("multiprocessing") results = apply_to_combination(_inner_func, params, fixed=fixed_param, multiproc=True, cluster=False) print(results) del results if pythongrid_present: print("cluster") results = apply_to_combination(_inner_func, params, fixed=fixed_param, multiproc=False, cluster=True) print(results) del results if __name__ == '__main__': _test1() OptWok-0.3.1/optwok/expttools_helper.py0000644ڛ|׿-0000000256212140420225025167 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Helper functions for expttools""" import sys python_version = sys.version_info[:2] def _wrap_kwargs(kwargs): """Small helper function to enable Pool""" function_handle = kwargs['_function_handle'] del kwargs['_function_handle'] return function_handle(**kwargs) if python_version >= (3, 0): def _pickle_method(method): func_name = method.__func__.__name__ obj = method.__self__ cls = method.__self__.__class__ return _unpickle_method, (func_name, obj, cls) def _unpickle_method(func_name, obj, cls): for cls in cls.mro(): try: func = cls.__dict__[func_name] except KeyError: pass else: break return func.__get__(obj, cls) else: def _pickle_method(method): """A workaround for inability to pickle instance methods""" func_name = method.im_func.__name__ obj = method.im_self cls = method.im_class return _unpickle_method, (func_name, obj, cls) def _unpickle_method(func_name, obj, cls): """A workaround for inability to pickle instance methods""" for cls in cls.mro(): try: func = cls.__dict__[func_name] except KeyError: pass else: break return func.__get__(obj, cls) OptWok-0.3.1/optwok/gauss_kernel.py0000644ڛ|׿-0000000704511754625156024276 0ustar cheng.ongUNIMELB\Domain Users00000000000000 class Kernel(object): """A wrapper to define the interfaces for kernels""" def __init__(self, normalize=True): self.train_ex = None self.test_ex = None self.km = None self.normalize = normalize # not implemented yet def init_data(self, train_ex, test_ex): """Copy data to internal variables""" self.train_ex = train_ex.copy() self.test_ex = test_ex.copy() def get_kernel_matrix(self, split='train'): pass class LinearKernel(Kernel): """Standard scalar product""" def __init__(self, normalize=True): Kernel.__init__(self, normalize) def get_kernel_matrix(self, split='train'): if split == 'train': return array(matrix(self.train_ex).T*matrix(self.train_ex)) elif split == 'test': return array(matrix(self.train_ex).T*matrix(self.test_ex)) class GaussianKernel(Kernel): """Gaussian Kernel with scalar width k(x,y) = exp(-width * ||x-y||^2) """ def __init__(self, width=1.0, normalize=True): Kernel.__init__(self, normalize) self.width = width def get_kernel_matrix(self, split='train'): if split == 'train': m = self.train_ex.shape[1] a = self.train_ex unitvec = ones(m,1) dot_a = sum(self.width*a*a,1) K = matrix(sqrt(self.width)*a).T * matrix(sqrt(self.width)*a) for ix in range(m): K[:,ix] = exp(2*K[:,ix] - transpose(dot_a) - dot_a[ix]*unitvec) elif split == 'test': a = self.train_ex m = a.shape[1] b = self.test_ex n = b.shape[1] unitvec = ones(m,1) dot_a = sum(self.width*a*a,1) dot_a = sum(self.width*a*a,1) K = matrix(sqrt(self.width)*a).T * matrix(sqrt(self.width)*b) for ix in range(n): K[:,ix] = exp(2*K[:,ix] - transpose(dot_a) - dot_b[ix]*unitvec) return K def get_kernel_matrix_slow(self, split='train'): """Implement the double for loop""" if split == 'train': m = self.train_ex.shape[1] K = zeros(m,m) for ix in range(m): for iy in range(ix,m): K[ix,iy] = exp(-self.width * norm(self.train_ex[:,ix] - self.train_ex[:,iy])**2) K[iy,ix] = K[ix,iy] elif split == 'test': m = self.train_ex.shape[1] n = self.test_ex.shape[1] K = zeros(m,n) for ix in range(m): for iy in range(n): K[ix,iy] = exp(-self.width * norm(self.train_ex[:,ix] - self.test_ex[:,iy])**2) return K class StudentKernel(Kernel): """Kernel based on the unnormalized Student T distribution. k(x,y) = (1+(||x-y||^2)/nu)^(-(nu+1)/2) """ pass class IntegratedGaussianKernel(Kernel): """Integral of the Gaussian kernel with respect to the width. k(x,y) = sqrt(2)*sigma*exp(-1/2*sigma^2 ||x-y||^2) - sqrt(pi)*||x-y||*erfc(-1/sqrt(2)*sigma ||x-y||) """ pass class GaussianCovKernel(Kernel): """Gaussian kernel with the full covariance function. k(x,y) = exp(- (x-y).T * Cov * (x-y)) Consider also different shrinkage approaches. """ class MixGaussianKernel(Kernel): """A finite sum of Gaussian Kernels""" pass class TaylorGaussianKernel(Kernel): """A infinite sum of Gaussian Kernels""" pass OptWok-0.3.1/optwok/gaussian_proc.py0000644ڛ|׿-0000001317712055531247024444 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Gaussian process regression""" from numpy import array, matrix, kron, diag, ones, real, dot, ndarray from numpy import zeros, eye, hstack, concatenate, isnan, sum, log, pi, inner, outer, ones_like, trace from numpy.linalg import solve, inv, cholesky, lstsq from numpy.linalg import eig from scipy.linalg.matfuncs import sqrtm from optwok.kernel import ContextKernel from numpy.core.numeric import newaxis #import ipdb # TODO: general kernel function # gp.update to rank one updates class GaussianProcess(object): """Gaussian Process regression, for objects which are pairs of (data, context)""" def __init__(self, Ks, Kz, kparam = None, reg_param = 1.0): if kparam is not None: self.Ks = Ks(kparam) self.Ks_pred = Ks(kparam) else: self.Ks = Ks() self.Ks_pred = Ks() self.Kz = Kz self.Ktrain = ContextKernel(self.Ks, self.Kz) self.Kpred = ContextKernel(self.Ks_pred, self.Kz) self.train_ex = array([]) self.train_cont = array([], dtype = int) self.train_lab = array([]) self.reg_param = reg_param self.num_train = 0 self.num_context = Kz.shape[0] self.alpha = array([]) self.L = array([]) def train(self, train_ex, train_cont, train_lab): """Learn the parameters. train_ex is a tuple containing (example, context) """ self.train_ex = train_ex self.train_cont = array(train_cont, dtype = int) self.train_lab = array(train_lab) self.num_train = self.train_ex.shape[1] self._update() def _update(self): """Computations to estimate GP parameters""" self.Ktrain.compute(self.train_ex, self.train_cont) #self.L = real(sqrtm(self.Ktrain.kmat + self.reg_param*eye(self.num_train))) self.L = cholesky(self.Ktrain.kmat + self.reg_param * eye(self.num_train)) (x, residues, rank, sing) = lstsq(self.L, self.train_lab) (self.alpha, residues, rank, sing) = lstsq(self.L.T, x) def update(self, orig_ex, new_cont, new_lab): """Include the new example into training data""" new_ex = orig_ex.copy() new_ex.shape = (len(new_ex), 1) self.train_ex = hstack([self.train_ex, new_ex]) self.train_cont = concatenate([self.train_cont, array([new_cont], dtype = int)]) self.train_lab = concatenate([self.train_lab, array([new_lab])]) self.num_train += 1 assert len(self.train_cont) == self.num_train self._update() def predict(self, pred_data): """Estimate the outputs""" pred_ex = pred_data[0] pred_cont = array(pred_data[1], dtype = int) num_pred = pred_ex.shape[1] self.Kpred.compute(pred_ex, pred_cont) if self.num_train == 0: scores = zeros(num_pred) pred_var = diag(self.Kpred.kmat) else: Kmat_cross = self.Ktrain.compute_test(pred_ex, pred_cont) scores = dot(Kmat_cross.T, self.alpha).flatten() (v, residues, rank, sing) = lstsq(self.L, Kmat_cross) pred_var = diag(self.Kpred.kmat - dot(v.T, v)) if not (pred_var > 0).all(): print('=== Predictive variance negative!') #ipdb.set_trace() return scores, pred_var class GaussianProcessScalar(object): """Gaussian Process regression""" def __init__(self, Ks, kparam = None, reg_param = 1.0, m=0): ''' Ks: Kernel Object, kparam: Kernel Parameters, reg_param: noise in the data, m: constant mean ''' if kparam is not None: self.Ktrain = Ks(kparam) self.Kpred = Ks(kparam) else: self.Ktrain = Ks() self.Kpred = Ks() self.train_ex = array([]) self.train_lab = array([]) self.reg_param = reg_param self.num_train = 0 self.inv_cov = matrix([]) self.alpha = array([]) self.L = matrix([]) self.m = m self.kparam = kparam def train(self, train_ex, train_lab): """Learn the parameters.""" self.train_ex = train_ex self.train_lab = matrix(train_lab) try: self.num_train = self.train_ex.shape[1] except IndexError: # 1-dimensional self.num_train = self.train_ex.shape[0] self.train_ex = self.train_ex[newaxis, :] self.train_lab.shape = (self.num_train, 1) self.Ktrain.compute(self.train_ex) self.L = cholesky(self.Ktrain.kmat + self.reg_param * eye(self.num_train)) self.alpha = solve(self.L.T, solve(self.L, self.train_lab - self.m)) def predict(self, pred_ex): """Estimate the outputs""" try: num_pred = pred_ex.shape[1] except IndexError: # 1-d pred_ex = pred_ex[newaxis, :] self.Kpred.compute(pred_ex) Kmat_cross = self.Ktrain.compute_test(pred_ex) scores = array(Kmat_cross.T * self.alpha) v = matrix(solve(self.L, Kmat_cross)) pred_var = diag(self.Kpred.kmat - v.T * v) return self.m + scores.flatten(), pred_var def nl_likelihood(self): return array((self.train_lab - self.m).T.dot(self.alpha)/2 + sum(log(diag(self.L))) + self.num_train * log(2*pi)/2)[0][0] def grad_nl_likelihood(self): Q = solve(self.L, eye(self.num_train)) / self.reg_param - outer(self.alpha, self.alpha) dnoise = self.reg_param * trace(Q) dm = dot(ones_like(self.train_ex),array(self.alpha))[0][0] grad =[dnoise, dm] dkparam = self.Ktrain.compute_grad() for dkp in dkparam: grad.append(sum(sum(Q*dkp))/2) return array(grad) OptWok-0.3.1/optwok/io_pickle.py0000644ڛ|׿-0000000301112055561344023530 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""A Matlab like interface for loading and saving objects. Has an additional benefit of collecting all instances of pickle code here.""" import sys import bz2 import subprocess import os python_version = sys.version_info[:2] if python_version >= (3, 0): from pickle import load as pload from pickle import dump as pdump else: from cPickle import load as pload from cPickle import dump as pdump def save(filename, myobj): """save the myobj to filename using pickle""" try: f = bz2.BZ2File(filename, 'wb') except IOError: sys.stderr.write('Trying local path...') try: f = bz2.BZ2File(os.path.join(os.getcwd(), filename)) except IOError: sys.stderr.write('File ' + filename + ' cannot be written\n') return pdump(myobj, f, protocol = 2) f.close() def load(filename): """load the myobj from filename using pickle""" try: f = bz2.BZ2File(filename, 'rb') except IOError: sys.stderr.write('File ' + filename + ' cannot be read\n') return myobj = pload(f) f.close() return myobj def convert_v6(input_filename, output_filename): """ starts matlab and converts a structure to a matlab v6 structure """ cmd = "matlab -nojvm -nodisplay -r \"load %s; save -v6 '%s'; exit\"" % (input_filename, output_filename) obj = subprocess.Popen(cmd, shell = True, stdout = subprocess.PIPE, stderr = subprocess.PIPE) out, err = obj.communicate() assert err == '', 'An error occured!\n%s' % err OptWok-0.3.1/optwok/kernel.py0000644ڛ|׿-0000000452011731540564023061 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Implement some positive semidefinite kernels.""" from numpy import array, matrix from numpy import zeros, ones, eye, diag from numpy import average, exp, kron, sum from numpy import sqrt, divide, mean, multiply #try: # from optwok.kernelbase import Kernel, JointKernel, ContextKernel #except ImportError: # from optwok.kernel_pp import Kernel, JointKernel, ContextKernel from optwok.kernel_pp import * ##################################################################### # Some miscellaneaous functions for kernel matrix operations ##################################################################### def sqr_dist(a,b): """Compute the square distance between vectors""" assert(a.shape[0] == b.shape[0]) if len(a.shape)>1: # a is a matrix dot_a = sum(a*a, axis=0).T dot_b = sum(b*b ,axis=0).T unitvec = ones(dot_a.shape) D = array(2.0*matrix(a).T*matrix(b)) for ix,bval in enumerate(dot_b): D[:,ix] = dot_a - D[:,ix] + kron(bval,unitvec) else: D = zeros((len(a),len(b))) for (ixa, cur_a) in enumerate(a): for (ixb, cur_b) in enumerate(b): D[ixa,ixb] = (cur_a-cur_b)*(cur_a-cur_b) return D def center(K): """Center the kernel matrix, such that the mean (in feature space) is zero.""" one_mat = matrix(ones(K.shape)) one_vec = matrix(ones((K.shape[0],1))) row_sum = matrix(mean(K,axis=0)).T R = K - row_sum * one_vec.T - one_vec * row_sum.T +\ mean(row_sum.A)*one_mat return R def normalize_variance(K): """Normalize the kernel matrix, such that the variance (in feature space) is 1""" one_vec = matrix(ones((K.shape[0],1))) inv_sqrt_diag = divide(one_vec, matrix(sqrt(diag(K))).T) KN = multiply(kron(one_vec.T,inv_sqrt_diag),K) KN = multiply(kron(one_vec,inv_sqrt_diag.T),K) return KN def normalize_unit_diag(K): """Normalize the kernel matrix, such that all diagonal entries are 1.""" Kii = diag(K) Kii.shape = (len(Kii),1) return divide(K, sqrt(matrix(Kii)*matrix(Kii).T)) def kernel2distance(K): """Convert the kernel matrix into the corresponding Euclidean distance.""" # TODO: Vectorize D = zeros(K.shape) for ix in range(K.shape[0]): for iy in range(K.shape[1]): D[ix,iy] = sqrt(K[ix,ix] + K[iy,iy] - 2*K[ix,iy]) return D OptWok-0.3.1/optwok/kernel_pp.py0000644ڛ|׿-0000002443612055533446023571 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Implement some positive semidefinite kernels. These are in pure python. For the cython implementation, see kernelbase.pyx """ from numpy import array, matrix, max, ix_ from numpy import zeros, ones, eye, diag, empty from numpy import average, exp, kron, sum from numpy import sqrt, divide, mean, multiply from numpy import pi, sin, cos from numpy import flatnonzero class Kernel(object): """A Mercer kernel. Note that if Cython is available, then this is not used.""" def __init__(self): self.kmat = array([]) self.examples = None def __getitem__(self, idx): assert(len(idx)==2) return self.kmat[idx[0], idx[1]] def __repr__(self): rep = '<' + self.__class__.__name__ + ' instance>' return rep def dump(self) : """ returns a string that can be used to construct an equivalent object """ kstr = self.__module__ + '.' + self.__class__.__name__ + '(' + \ self.construction_params() + ')' return kstr def construction_params(self): raise NotImplementedError def compute(self, examples): """Evaluate the kernel function and store the value""" raise NotImplementedError def compute_test(self, test_ex): """Evaluate the kernel function on test examples""" raise NotImplementedError class JointKernel(object): """A kernel on input and output pairs.""" def __init__(self, input_kernel, output_kmat=None, num_classes=-1): self.Kx = input_kernel if output_kmat is not None: self.Ky = output_kmat else: assert(num_classes > 1) self.Ky = eye(num_classes) self.examples = None def __repr__(self) : rep = '<' + self.__class__.__name__ + ' instance>' return rep def __getitem__(self, idx): """Simulate matrix element. Order is assumed to be examples first then data. """ assert(len(idx)==4) return self.Kx[idx[0], idx[1]]*self.Ky[idx[2], idx[3]] @property def kmat(self): """Simulate the whole kernel matrix""" return kron(self.Kx.kmat, self.Ky) def compute(self, examples): self.examples = examples self.Kx.compute(examples) def compute_test(self, test_ex): """Returns a non-square matrix, of num_train by num_test size.""" Kx = self.Kx.compute_test(test_ex) return kron(Kx, self.Ky) class ContextKernel(Kernel): """A kernel of (data, context) pairs, mathematically the same as the JointKernel, but does not evaluate all combinations""" def __init__(self, input_kernel, output_kmat): self.Kx = input_kernel self.num_context = output_kmat.shape[0] self.Ky = output_kmat self.examples = None self.num_examples = 0 self.context = None def __repr__(self) : rep = '<' + self.__class__.__name__ + ' instance>' return rep def __getitem__(self, idx): """Simulate matrix element. Order is assumed to be (example,context,example,context) """ assert(len(idx)==4) return self.Kx[idx[0], idx[2]]*self.Ky[idx[1], idx[3]] @property def kmat(self): """Simulate the whole kernel matrix""" K = zeros((self.num_examples, self.num_examples)) for con_x in range(self.num_context): for con_y in range(self.num_context): ix = flatnonzero(self.context == con_x) iy = flatnonzero(self.context == con_y) if len(ix)>0 and len(iy)>0: Ky_val = self.Ky[self.context[ix[0]], self.context[iy[0]]] if abs(Ky_val) > 0.0: K[ix_(ix,iy)] = self.Kx[ix_(ix,iy)]*Ky_val K[ix_(iy,ix)] = self.Kx[ix_(iy,ix)]*Ky_val return K @property def kmat_forloop(self): """Simulate the whole kernel matrix""" K = zeros((self.num_examples, self.num_examples)) for ix in range(self.num_examples): for iy in range(ix, self.num_examples): val = self.Kx[ix,iy]*self.Ky[self.context[ix], self.context[iy]] K[ix,iy] = val K[iy,ix] = val return K def compute(self, examples, context): self.examples = examples self.context = context self.num_examples = examples.shape[1] self.Kx.compute(examples) def compute_test(self, examples, context): """Returns a non-square matrix, of num_train by num_test size.""" assert self.examples is not None assert self.context is not None num_test = examples.shape[1] Kx_cross = self.Kx.compute_test(examples) Kmat_cross = zeros((self.num_examples, num_test)) for con_x in range(self.num_context): for con_y in range(self.num_context): ix = flatnonzero(self.context == con_x) iy = flatnonzero(context == con_y) if len(ix)>0 and len(iy)>0: Ky_val = self.Ky[self.context[ix[0]], context[iy[0]]] if abs(Ky_val) > 0.0: Kmat_cross[ix_(ix,iy)] = Kx_cross[ix_(ix,iy)]*Ky_val return Kmat_cross def compute_test_forloop(self, examples, context): """Returns a non-square matrix, of num_train by num_test size.""" assert self.examples is not None assert self.context is not None num_test = examples.shape[1] Kx_cross = self.Kx.compute_test(examples) Kmat_cross = zeros((self.num_examples, num_test)) for ix in range(self.num_examples): for iy in range(num_test): Kmat_cross[ix,iy] = Kx_cross[ix,iy]*self.Ky[self.context[ix], context[iy]] return Kmat_cross class LinearKernel(Kernel): """Standard dot product""" def __init__(self): Kernel.__init__(self) def construction_params(self): return '' def compute(self, examples): """Evaluate the kernel function and store the value""" self.examples = examples self.kmat = array(matrix(examples).T*matrix(examples)) def compute_test(self, test_ex): """Evaluate the kernel function on test examples""" if test_ex.ndim == 1: test_ex.shape = (len(test_ex),1) return array(matrix(self.examples).T*matrix(test_ex)) class SEKernel(Kernel): ''' Squared Exponential Kernel ''' def sqr_dist(self, a, b): """Compute the square distance between vectors""" assert(a.shape[0] == b.shape[0]) if len(a.shape) > 1: # a is a matrix dot_a = sum(a * a, axis = 0).T dot_b = sum(b * b , axis = 0).T unitvec = ones(dot_a.shape) D = array(2.0 * matrix(a).T * matrix(b)) for ix, bval in enumerate(dot_b): D[:, ix] = dot_a - D[:, ix] + kron(bval, unitvec) else: D = zeros((len(a), len(b))) for (ixa, cur_a) in enumerate(a): for (ixb, cur_b) in enumerate(b): D[ixa, ixb] = (cur_a - cur_b) * (cur_a - cur_b) return D class GaussKernel(SEKernel): """Gaussian rbf kernel""" def __init__(self, gamma=1.0): Kernel.__init__(self) self.gamma = gamma def __repr__(self) : rep = '<' + self.__class__.__name__ + ' instance>\n' rep += 'gamma : ' + str(self.gamma) return rep def constructionParams(self) : return 'gamma = ' + str(self.gamma) def compute(self, examples): """Evaluate the kernel function and store the value""" self.examples = examples self.kmat = exp(-self.gamma * self.sqr_dist(examples, examples)) def compute_test(self, test_ex): """Evaluate the kernel function on test examples""" if test_ex.ndim == 1: test_ex.shape = (len(test_ex),1) return exp(-self.gamma * self.sqr_dist(self.examples, test_ex)) class PeriodicKernel(SEKernel): """ Periodic kernel k(x,y) = theta * exp( -2*sin^2( pi*||x-y||/P )/lamb^2 ) """ def __init__(self, theta=1.0, lamb=1.1, P=pi): Kernel.__init__(self) self.theta = theta self.lamb = lamb self.P = P def __repr__(self) : rep = '<' + self.__class__.__name__ + ' instance>\n' rep += 'theta : ' + str(self.theta) rep += ', lambda : ' + str(self.lamb) rep += ', P : ' + str(self.P) return rep def constructionParams(self) : return 'theta = ' + str(self.theta) + ', lambda : ' + str(self.lamb) + ', P : ' + str(self.P) def compute_pk(self, ex1, ex2): ''' Evaluate kernel for ex1 and ex2 ''' K = sin(pi * sqrt(self.sqr_dist(ex1, ex2)) / self.P) / self.lamb K = K * K return self.theta * exp(-2 * K) def compute(self, examples): """Evaluate the kernel function and store the value""" self.examples = examples self.kmat = self.compute_pk(examples, examples) def compute_test(self, test_ex): """Evaluate the kernel function on test examples""" if test_ex.ndim == 1: test_ex.shape = (len(test_ex), 1) return self.compute_pk(self.examples, test_ex) def compute_grad(self): K = pi * sqrt(self.sqr_dist(self.examples, self.examples)) / self.P R = sin(K) / self.lamb sKs = R * R dtheta = 2 * self.theta * exp(-2*sKs) dP = 4 * self.theta / self.lamb * exp(-2 * R*R) * R * cos(K) * K dlamb = 4 * self.theta * exp(-2*sKs) * sKs return dtheta, dlamb, dP class CustomKernel(Kernel): """Just a positive semidefinite matrix""" def __init__(self): Kernel.__init__(self) def compute(self, examples): """Assume input is a kernel matrix. Copy the training kernel matrix""" self.kmat = examples def compute_test(self, test_ex): """Assume input is a kernel matrix. Just return the test kernel""" return test_ex def KernelFactory(kdict): """Some if then statements to map between the strings describing the kernels to the actual implementation """ if kdict['name'] == 'gauss': return GaussKernel(kdict['param']) if kdict['name'] == 'custom': return CustomKernel() else: print('%s not parsed properly in KernelFactory' % kdict['name']) OptWok-0.3.1/optwok/kernelbase.pyx0000644ڛ|׿-0000002146311731540564024111 0ustar cheng.ongUNIMELB\Domain Users00000000000000# cython: profile=True """Implement a Cython interface.""" from numpy import array import numpy as np cimport numpy as np np.import_array() cimport cython DTYPE = np.float ctypedef np.float_t DTYPE_t @cython.boundscheck(False) # turn off bounds-checking for entire function cdef getitem2(np.ndarray[DTYPE_t, ndim=2, negative_indices=False] kmat, Py_ssize_t xi, Py_ssize_t xj): """Fast version of __getitem__ with 2 arguments""" return kmat[xi, xj] @cython.boundscheck(False) cdef getitem4(np.ndarray[DTYPE_t, ndim=2, negative_indices=False] Kx, np.ndarray[DTYPE_t, ndim=2, negative_indices=False] Ky, Py_ssize_t xi, Py_ssize_t xj, Py_ssize_t yi, Py_ssize_t yj): """Fast version of __getitem__ with 4 arguments""" return Kx[xi, xj]*Ky[yi, yj] @cython.boundscheck(False) cdef DTYPE_t getprod(np.ndarray[DTYPE_t, ndim=2, negative_indices=False] Kx, np.ndarray[DTYPE_t, ndim=2, negative_indices=False] Ky, np.ndarray[Py_ssize_t, ndim=1, negative_indices=False] context, Py_ssize_t ix, Py_ssize_t iy): """product of input and output kernels""" cdef DTYPE_t val = Kx[ix,iy]*Ky[context[ix], context[iy]] return val @cython.boundscheck(False) cdef DTYPE_t getprod_cross(np.ndarray[DTYPE_t, ndim=2, negative_indices=False] Kx, np.ndarray[DTYPE_t, ndim=2, negative_indices=False] Ky, np.ndarray[Py_ssize_t, ndim=1, negative_indices=False] context, np.ndarray[Py_ssize_t, ndim=1, negative_indices=False] context_cross, Py_ssize_t ix, Py_ssize_t iy): """product of input and output kernels""" cdef DTYPE_t val = Kx[ix,iy]*Ky[context[ix], context_cross[iy]] return val class Kernel(object): """A Mercer kernel, with Cython interface for getitem""" def __init__(self): self.kmat = array([], dtype=DTYPE) self.examples = None def __getitem__(self, idx): assert(len(idx)==2) (xi,xj) = idx return getitem2(self.kmat, xi, xj) def __repr__(self): rep = '<' + self.__class__.__name__ + ' instance>' return rep def dump(self) : """ returns a string that can be used to construct an equivalent object """ kstr = self.__module__ + '.' + self.__class__.__name__ + '(' + \ self.construction_params() + ')' return kstr def construction_params(self): raise NotImplementedError def compute(self, examples): """Evaluate the kernel function and store the value""" raise NotImplementedError def compute_test(self, test_ex): """Evaluate the kernel function on test examples""" raise NotImplementedError class JointKernel(object): """A kernel on input and output pairs.""" def __init__(self, input_kernel, output_kmat=None, num_classes=-1): self.Kx = input_kernel if output_kmat is not None: self.Ky = output_kmat else: assert(num_classes > 1) self.Ky = np.eye(num_classes) self.examples = None def __repr__(self) : rep = '<' + self.__class__.__name__ + ' instance>' return rep def __getitem__(self, idx): """Simulate matrix element. Order is assumed to be examples first then data. """ assert(len(idx)==4) (xi, xj, yi, yj) = idx return getitem4(self.Kx.kmat, self.Ky, xi, xj, yi, yj) @property def kmat(self): """Simulate the whole kernel matrix""" return np.kron(self.Kx.kmat, self.Ky) def compute(self, examples): self.examples = examples self.Kx.compute(examples) def compute_test(self, test_ex): """Returns a non-square matrix, of num_train by num_test size.""" Kx = self.Kx.compute_test(test_ex) return np.kron(Kx, self.Ky) class ContextKernel(Kernel): """A kernel of (data, context) pairs, mathematically the same as the JointKernel, but does not evaluate all combinations""" def __init__(self, input_kernel, output_kmat): self.Kx = input_kernel self.num_classes = output_kmat.shape[0] self.Ky = output_kmat self.examples = None self.num_examples = 0 self.context = None def __repr__(self) : rep = '<' + self.__class__.__name__ + ' instance>' return rep def __getitem__(self, idx): """Simulate matrix element. Order is assumed to be (example,context,example,context) """ assert(len(idx)==4) (xi, yi, xj, yj) = idx return getitem4(self.Kx.kmat, self.Ky, xi, xj, yi, yj) @property def kmat(self): """Simulate the whole kernel matrix""" cdef Py_ssize_t ix cdef Py_ssize_t iy cdef DTYPE_t val cdef np.ndarray K = np.zeros((self.num_examples, self.num_examples), dtype=DTYPE) for ix in range(self.num_examples): for iy in range(ix, self.num_examples): #val = self.Kx[ix,iy]*self.Ky[self.context[ix], self.context[iy]] val = getprod(self.Kx.kmat, self.Ky, self.context, ix, iy) K[ix,iy] = val K[iy,ix] = val return K def compute(self, examples, context): self.examples = examples self.context = context self.num_examples = examples.shape[1] self.Kx.compute(examples) def compute_test(self, examples, context): """Returns a non-square matrix, of num_train by num_test size.""" assert self.examples is not None assert self.context is not None cdef Py_ssize_t num_test = examples.shape[1] cdef np.ndarray[DTYPE_t, ndim=2, negative_indices=False] Kx_cross = self.Kx.compute_test(examples) cdef np.ndarray[DTYPE_t, ndim=2, negative_indices=False] Kmat_cross = np.empty((self.num_examples, num_test),dtype=DTYPE) cdef Py_ssize_t ix cdef Py_ssize_t iy for ix in range(self.num_examples): for iy in range(num_test): Kmat_cross[ix,iy] = Kx_cross[ix,iy]*self.Ky[self.context[ix], context[iy]] #Kmat_cross[ix,iy] = getprod_cross(Kx_cross, self.Ky, self.context, context, ix, iy) return Kmat_cross class LinearKernel(Kernel): """Standard dot product""" def __init__(self): Kernel.__init__(self) def construction_params(self): return '' def compute(self, np.ndarray[DTYPE_t, ndim=2] examples): """Evaluate the kernel function and store the value""" self.examples = examples self.kmat = np.dot(examples.T, examples) def compute_test(self, np.ndarray[DTYPE_t, ndim=2] test_ex): """Evaluate the kernel function on test examples""" cdef np.ndarray[DTYPE_t, ndim=2, negative_indices=False] k_cross = np.dot(self.examples.T, test_ex) return k_cross class GaussKernel(Kernel): """Gaussian rbf kernel""" def __init__(self, gamma=1.0): Kernel.__init__(self) self.gamma = gamma def __repr__(self) : rep = '<' + self.__class__.__name__ + ' instance>\n' rep += 'gamma : ' + str(self.gamma) return rep def constructionParams(self) : return 'gamma = ' + str(self.gamma) def compute(self, examples): """Evaluate the kernel function and store the value""" self.examples = examples self.kmat = np.exp(-self.gamma * self.sqr_dist(examples, examples)) def compute_test(self, test_ex): """Evaluate the kernel function on test examples""" if test_ex.ndim == 1: test_ex.shape = (len(test_ex),1) return np.exp(-self.gamma * self.sqr_dist(self.examples, test_ex)) def sqr_dist(a,b): """Compute the square distance between vectors""" assert(a.shape[0] == b.shape[0]) if len(a.shape)>1: # a is a matrix dot_a = np.sum(a*a, axis=0).T dot_b = np.sum(b*b ,axis=0).T unitvec = np.ones(dot_a.shape) D = np.array(2.0*np.matrix(a).T*np.matrix(b)) for ix,bval in enumerate(dot_b): D[:,ix] = dot_a - D[:,ix] + np.kron(bval,unitvec) else: D = np.zeros((len(a),len(b))) for (ixa, cur_a) in enumerate(a): for (ixb, cur_b) in enumerate(b): D[ixa,ixb] = (cur_a-cur_b)*(cur_a-cur_b) return D class CustomKernel(Kernel): """Just a positive semidefinite matrix""" def __init__(self): Kernel.__init__(self) def compute(self, examples): """Assume input is a kernel matrix. Copy the training kernel matrix""" self.kmat = examples def compute_test(self, test_ex): """Assume input is a kernel matrix. Just return the test kernel""" return test_ex OptWok-0.3.1/optwok/kernelopt.py0000644ڛ|׿-0000004156712055556203023615 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Optimize over the choice of kernel. For learning the kernel over labels. """ from time import time from numpy import zeros, ones, eye, sign from numpy import matrix, array, diag, trace from numpy import vstack, unique, maximum, argmax from numpy.linalg import eig, norm, solve, lstsq, inv from numpy import diag, sqrt, divide, logspace, kron from numpy import min, max, transpose, real from scipy.linalg.matfuncs import sqrtm from matplotlib.pyplot import matshow, title, colorbar, savefig, show, gray from matplotlib.colors import LinearSegmentedColormap from cvxopt import matrix as cm from cvxopt.solvers import sdp from optwok.multiclass import StructPredictor from optwok.multiclass import idx2binvec from optwok.kernel import normalize_unit_diag, JointKernel from optwok.io_pickle import load, save from optwok.mat_blas import solve_sylvester class KernelOptBase(object): """Base class for learning the kernel on labels""" def __init__(self, reg_param=1.0, verbose=False): assert(reg_param > 0.0) self.reg_param = reg_param # Tradeoff regularizer vs loss self.all_reg_param = logspace(-4,4,30) self.verbose = verbose # how much output self.Ky = array([]) # kernel on labels self.all_Ky = None # label kernel for whole reg path def train(self, train_ex, train_lab): raise NotImplementedError def predict(self, pred_ex): raise NotImplementedError def save_Ky(self, dataname, work_dir='.'): """Save the learned label kernel, as pickle and pdf image""" if self.all_Ky is None: save('%s/%s_Ky.pkl' % (work_dir, dataname), self.Ky) else: save('%s/%s_Ky.pkl' % (work_dir, dataname), self.all_Ky) show_kernel(self.Ky) title(dataname) savefig('%s/%s_Ky.pdf' % (work_dir, dataname)) def normalize_unit_diag(self): """Normalize Ky to unit diagonal, and appropriately scale alpha""" diag_Ky = diag(self.Ky) diag_Ky.shape = (self.num_classes, 1) self.Ky = divide(self.Ky, sqrt(matrix(diag_Ky)*matrix(diag_Ky).T)) diag_Ky.shape = (1, self.num_classes) diag_Ky = sqrt(diag_Ky) self.alpha = self.alpha*diag_Ky def _set_dimensions(self, train_lab): """Infer the number of examples and classes""" if len(train_lab.shape) == 1: self.y = idx2binvec(train_lab) self.num_train = len(train_lab) self.num_classes = len(unique(train_lab)) assert(self.y.shape == (self.num_train, self.num_classes)) else: self.y = train_lab.T (self.num_classes, self.num_train) = train_lab.shape class KernelOptLS(KernelOptBase): """Optimize the kernel on the labels, with the least squares loss, and also Frobenious norm regularizer. Uses slycot for solving the sylvester equation""" def __init__(self, reg_param=1.0, kernel=None, verbose=False): KernelOptBase.__init__(self, reg_param, verbose) self.max_iter = 1000 # maximum number of iterations self.rel_tol = 0.01 # relative stopping tolerance self.near_zero = 1e-5 # initialize a positive definite matrix self.train_ex = array([]) # training examples self.train_lab = array([]) # labels in original format self.y = array([]) # labels self.num_classes = 0 # number of classes self.num_train = 0 # number of training examples #self.objective = [] # objective value at each iteration self.Kx = kernel # kernel on examples self.alpha = array([]) # support vector coefficients def train(self, train_ex, train_lab, init_Ky=None): """Wrapper to convert data to feature matrix, and solve the optimization""" self.train_ex = train_ex self.train_lab = train_lab self._set_dimensions(train_lab) self.Kx.compute(train_ex) if init_Ky is not None: self.Ky = init_Ky else: #self.Ky = zeros((self.num_classes,self.num_classes)) self.Ky = self.near_zero*eye(self.num_classes) # shrink Ky towards this matrix #bias = matrix(ones((self.num_classes,self.num_classes))) #bias = matrix(eye(self.num_classes)) residual = norm(self.y) min_tol = self.rel_tol * residual num_iter = 0 print('Learning the kernel by coordinate descent') while residual > min_tol and num_iter < self.max_iter: num_iter += 1 self.alpha = solve_sylvester(self.Kx.kmat, self.Ky/self.reg_param, self.y/self.reg_param) #self.objective.append((trace(matrix(self.alpha).T*matrix(self.y))+ # norm(self.Ky)**2)/2.0) X = matrix(self.Kx.kmat)*matrix(self.alpha) #P = 0.5*X.T*matrix(self.alpha)-matrix(self.Ky)+bias P = 0.5*X.T*matrix(self.alpha)-matrix(self.Ky) P = (P+P.T)/2.0 Q = matrix(solve(X.T*X+self.reg_param*eye(self.num_classes), P)) #(Q, res, xrank, s) = lstsq(X.T*X+self.reg_param*eye(self.num_classes), P) Q = (Q+Q.T)/2.0 self.Ky += self.reg_param*Q yhat = X*matrix(self.Ky) residual = norm(yhat+self.reg_param*self.alpha-self.y) if self.verbose: print('Iteration: %d: residual = %f' % (num_iter, residual)) #self.normalize_unit_diag() print('Number of iterations: %d' % num_iter) if self.verbose: print('Kernel on labels') print(self.Ky) def predict(self, pred_ex, get_max=True): """Wrapper to do prediction. returns argmax of scores""" Ktest = self.Kx.compute_test(pred_ex) scores = array((matrix(self.Ky)*matrix(self.alpha).T*matrix(Ktest)).T) if get_max: return argmax(scores, axis=1) else: return transpose(scores) def train_reg_path(self, train_ex, train_lab, all_reg_param=None): """Solve for the whole regularization path""" if all_reg_param is not None: self.all_reg_param = all_reg_param self._set_dimensions(train_lab) self.all_Ky = zeros((self.num_classes, self.num_classes, len(self.all_reg_param))) self.Ky = zeros((self.num_classes, self.num_classes)) for (idx_reg, reg) in reversed(list(enumerate(self.all_reg_param))): self.reg_param = reg self.train(train_ex, train_lab, self.Ky) self.all_Ky[:,:,idx_reg] = self.Ky.copy() # Use the middle of the regularization range as default self.Ky = self.all_Ky[:,:,round(self.all_Ky.shape[2]/2)] def infer_sv(self, Ky=None): """Compute the support vector coefficients for a given Ky""" if Ky is not None: self.Ky = Ky self.alpha = solve_sylvester(self.Kx.kmat, self.Ky/self.reg_param, self.y/self.reg_param) def predict_reg_path(self, pred_ex, ident_Ky=False): """Compute prediction for each value in regularization path. Use learned Ky and retrain to get the corresponding support vector coefficients. """ all_preds = zeros((pred_ex.shape[1],len(self.all_reg_param))) for (idx_reg, reg) in reversed(list(enumerate(self.all_reg_param))): print('Predicting with reg param = %e' % reg) self.reg_param = reg if ident_Ky: self.infer_sv(eye(self.num_classes)) else: self.infer_sv(self.all_Ky[:,:,idx_reg]) all_preds[:,idx_reg] = self.predict(pred_ex) return all_preds class KernelOptLin(KernelOptBase): """Optimize the kernel on the labels, with the least squares loss, and also Frobenious norm regularizer. Uses slycot for solving the sylvester equation Use linear kernel on the inputs """ def __init__(self, reg_param=1.0, verbose=False): KernelOptBase.__init__(self, reg_param, verbose) self.max_iter = 1000 # maximum number of iterations self.rel_tol = 0.01 # relative stopping tolerance self.near_zero = 1e-5 # initialize a positive definite matrix self.train_ex = array([]) # training examples self.train_lab = array([]) # labels in original format self.y = array([]) # labels self.num_feat = 0 # number of features self.num_classes = 0 # number of classes self.num_train = 0 # number of training examples #self.objective = [] # objective value at each iteration self.alpha = array([]) # support vector coefficients def train(self, train_ex, train_lab, init_Ky=None): """Wrapper to convert data to feature matrix, and solve the optimization""" self.train_ex = train_ex self.train_lab = train_lab self.num_feat = train_ex.shape[0] self._set_dimensions(train_lab) if init_Ky is not None: self.Ky = init_Ky else: self.Ky = self.near_zero*eye(self.num_classes) residual = norm(self.y) min_tol = self.rel_tol * residual num_iter = 0 print('Learning the kernel by coordinate descent') while residual > min_tol and num_iter < self.max_iter: num_iter += 1 K = array(matrix(self.train_ex).T*matrix(self.train_ex)) self.alpha = solve_sylvester(K, self.Ky/self.reg_param, self.y/self.reg_param) X = K*matrix(self.alpha) P = 0.5*X.T*matrix(self.alpha)-matrix(self.Ky) P = (P+P.T)/2.0 Q = matrix(solve(X.T*X+self.reg_param*eye(self.num_classes), P)) Q = (Q+Q.T)/2.0 self.Ky += self.reg_param*Q yhat = X*matrix(self.Ky) residual = norm(yhat+self.reg_param*self.alpha-self.y) if self.verbose: print('Iteration: %d: residual = %f' % (num_iter, residual)) print('Number of iterations: %d' % num_iter) if self.verbose: print('Kernel on labels') print(self.Ky) def predict(self, pred_ex, get_max=True): """Wrapper to do prediction. returns argmax of scores""" Ktest = matrix(self.train_ex.T)*matrix(pred_ex) scores = array((matrix(self.Ky)*matrix(self.alpha).T*Ktest).T) if get_max: return argmax(scores, axis=1) else: return transpose(scores) class KernelOptLSGP(KernelOptLS): """Optimize the kernel on the labels, with the least squares loss, and also Frobenious norm regularizer. Uses slycot for solving the sylvester equation. Multiple output regression, also estimate predictive variance. """ def __init__(self, reg_param=1.0, kernel=None, verbose=False): KernelOptLS.__init__(self, reg_param, kernel(), verbose) self.test_kernel = JointKernel(kernel(), self.Ky) def predict(self, pred_ex): """Wrapper to do prediction and also estimate predictive variance.""" num_test = pred_ex.shape[1] Ktest = self.Kx.compute_test(pred_ex) scores = array((matrix(self.Ky)*matrix(self.alpha).T*matrix(Ktest)).T) self.test_kernel.Ky = self.Ky self.test_kernel.compute(pred_ex) A = matrix(inv(kron(self.Kx.kmat,self.Ky)\ +self.reg_param*eye(self.num_train*self.num_classes))) k_star = matrix(kron(Ktest, self.Ky)) pred_var = diag(self.test_kernel.kmat - k_star.T*A*k_star) pred_var.shape = (num_test, self.num_classes) return scores.T, pred_var.T def predict_reg_path(self, pred_ex, ident_Ky=False): """Compute prediction for each value in regularization path. Use learned Ky and retrain to get the corresponding support vector coefficients. """ all_preds = zeros((self.num_classes, pred_ex.shape[1], len(self.all_reg_param))) for (idx_reg, reg) in reversed(list(enumerate(self.all_reg_param))): print('Predicting with reg param = %e' % reg) self.reg_param = reg if ident_Ky: self.infer_sv(eye(self.num_classes)) else: self.infer_sv(self.all_Ky[:,:,idx_reg]) all_preds[:,:,idx_reg] = self.predict(pred_ex) return all_preds class KernelOpt(object): """Learn the kernel on the labels Uses semidefinite programming with cvxopt """ def __init__(self, reg_param=1.0, kernel=None, verbose=False): self.train_time = 0.0 # time used for training self.max_iter = 3 # maximum number of iterations self.verbose = verbose # how much output # Pointer to the classifier object self.classifier = StructPredictor(reg_param, kernel, verbose) def solve_opt(self, alpha): """Solve the optimization problem using an SDP. max tr(alpha' Kx alpha Ky) st diag(Ky) = I Ky > 0 cvxopt takes the problem as: sdp(c[, Gl, hl[, Gs, hs[, A, b]]]) minimize c'x subject to Gl x + sl = hl sl >= 0 Ax = b Gk x + vec(sk) = vec(hk) sk > 0 Gl and hl define the inequality constraints Gs and hs are lists, containing Gk and hk respectively, representing the LMIs. The inequality > mean positive semidefinite. We consider the dual form: max -hl'zl - sum_k tr(hk zk) - b'y st Gl'zl + sum_k Gk'vec(zk) + A'y + c = 0 zl >= 0 zk > 0 """ n = alpha.shape[1] # the number of classes m = n**2 # the size of Ky # TODO: How do we regularize? h = [cm(-array(normalize_unit_diag(matrix(alpha).T*self.classifier.Kx*matrix(alpha))))] #h = [cm(-array(matrix(alpha).T*self.classifier.Kx*matrix(alpha)))] #print(-h[0]) A = zeros((m,n)) for icol in range(n): A[(n+1)*icol,icol] = 1.0 G = [cm(A)] c = cm(-ones((n,1))) print('Solving SDP') sol = sdp(c, Gs=G, hs=h, solver='dsdp') Ky = array(sol['zs'][0]) Ky = Ky + Ky.T - diag(Ky.diagonal()) #Ky = maximum(Ky, 0.0) #print(trace(cm(Ky)*h[0])) return Ky def solve_opt_abspos(self, alpha): """Solve the optimization problem using an SDP. Also require that entries in Ky are positive. cvxopt takes the problem as: sdp(c[, Gl, hl[, Gs, hs[, A, b]]]) minimize c'x subject to Gl x + sl = hl sl >= 0 Ax = b Gk x + vec(sk) = vec(hk) sk > 0 Gl and hl define the inequality constraints Gs and hs are lists, containing Gk and hk respectively, representing the LMIs. The inequality > mean positive semidefinite. We consider the dual form: max -hl'zl - sum_k tr(hk zk) - b'y st Gl'zl + sum_k Gk'vec(zk) + A'y + c = 0 zl >= 0 zk > 0 """ n = alpha.shape[1] # the number of classes m = n**2 # the size of Ky nm = n+m # The number of variables in the SDP # TODO: How do we regularize? #h = [cm(-array(normalize_unit_diag(matrix(alpha).T*self.classifier.Kx*matrix(alpha))))] #h = [cm(-array(matrix(alpha).T*self.classifier.Kx*matrix(alpha)))] F = zeros((nm, nm)) F[:n,:n] = -array(normalize_unit_diag(matrix(alpha).T*self.classifier.Kx*matrix(alpha))) h = [cm(F)] #print(-h[0]) Fk = zeros((nm**2,nm)) for icol in range(n): Fk[(nm+1)*icol,icol] = 1.0 for irow in range(n): for icol in range(n): Fk[nm*irow+icol,n+n*irow+icol] = 1.0 Fk[n*nm+n+(nm+1)*(n*irow+icol),n+n*irow+icol] = -1.0 G = [cm(Fk)] c = cm(vstack([-ones((n,1)), zeros((m,1))])) print('Solving SDP') sol = sdp(c, Gs=G, hs=h, solver='dsdp') Ky = array(sol['zs'][0][:n,:n]) Ky = Ky + Ky.T - diag(Ky.diagonal()) #Ky = maximum(Ky, 0.0) #print(trace(cm(Ky)*h[0])) return Ky def train(self, train_ex, train_lab): """Wrapper to convert data to feature matrix, and solve the optimization""" self.classifier.train(train_ex, train_lab) for ix in range(self.max_iter): print('Iteration: %d' % ix) self.classifier.Ky = self.solve_opt(self.classifier.alpha) self.classifier.solve_opt() #print(self.classifier.Ky) #print(eig(self.classifier.Ky)[0]) def predict(self, pred_ex): """Wrapper to do prediction.""" return self.classifier.predict(pred_ex) OptWok-0.3.1/optwok/mat_blas.py0000644ڛ|׿-0000000102312055550713023353 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Collection of routines for solving matrix linear equations""" from numpy.linalg import eigh from numpy import dot, outer, ones def solve_sylvester(A,B,C,symm=False): """Solve the Sylvester equation: AXB + X = C where A and B are symmetric real valued, by using the eigenvalue decomposition """ LA, U = eigh(A) LB, V = eigh(B) frac = dot(dot(U.T,C),V)/(outer(LA,LB)+ones((len(LA),len(LB)))) sol = dot(dot(U,frac),V.T) if symm: return 0.5*(sol+sol.T) else: return sol OptWok-0.3.1/optwok/mil/0000755ڛ|׿-0000000000012140420542021773 5ustar cheng.ongUNIMELB\Domain Users00000000000000OptWok-0.3.1/optwok/mil/__init__.py0000644ڛ|׿-0000000000012115570207024100 0ustar cheng.ongUNIMELB\Domain Users00000000000000OptWok-0.3.1/optwok/mil/mi_classifier.py0000644ڛ|׿-0000000465612140420225025167 0ustar cheng.ongUNIMELB\Domain Users00000000000000''' Created on 21.02.2013 @author: Gabriel ''' from __future__ import division from numpy import dot, sign, array, squeeze, repeat from sklearn.svm import SVC class MIClassifier(object): ''' Multiple Instance Learning Classifier ''' def train(self, traindata, reg_param, v0 = None, verbose = False, *args, **kwargs): self.traindata = traindata self.reg_param = reg_param self.solve_opt(v0 = v0, verbose = verbose, *args, **kwargs) def predict(self, preddata): ''' implement classifier ''' y = [] for pd in preddata.iter(): if preddata.sparse: yi = self.one_predict_sparse(pd) else: yi = self.one_predict(pd) y.append(yi) return sign(array(y)) def read_data(self, data): ''' Return midata object in corresponding format, by calling .to_$ on data ''' raise NotImplementedError class MeanSVM(MIClassifier): ''' SVM on the means for MIL ''' def one_predict(self, data): ''' predict one datapoint ''' return dot(self.w,data['mean']) + self.b def read_data(self, data): ''' Read data and convert to EllipsoidData ''' return data.to_means() def solve_opt(self, v0, verbose, *args, **kwargs): svm = SVC(C=1/self.reg_param, kernel='linear') q, y = array(self.traindata.means), array(self.traindata.labels) svm.fit(q, y) self.w, self.b, self.primal_obj = squeeze(svm.coef_), svm.intercept_[0], None class InstanceSVM(MIClassifier): ''' Naive instance level SVM for MIL ''' def one_predict(self, data): ''' predict one datapoint ''' return reduce(max,[dot(self.w,inst)+self.b for inst in data['bag']]) def read_data(self, data): ''' Read data and convert to EllipsoidData ''' return data.to_instances() def solve_opt(self, v0, verbose, *args, **kwargs): svm = SVC(C=1/self.reg_param, kernel='linear') q = array([inst for qi in self.traindata.bags for inst in qi]) # flatten out all instances y = array([repeat(yi,len(qi)) for yi, qi in zip(self.traindata.labels,self.traindata.bags)]).flatten() # repeat labels for all instances svm.fit(q, y) self.w, self.b, self.primal_obj = squeeze(svm.coef_), svm.intercept_[0], None OptWok-0.3.1/optwok/mil/mi_classifier_emil.py0000644ڛ|׿-0000002764212140417274026207 0ustar cheng.ongUNIMELB\Domain Users00000000000000''' Created on 04.03.2013 @author: Gabriel ''' from __future__ import division from scipy.optimize.lbfgsb import fmin_l_bfgs_b from scipy.misc import logsumexp from numpy.linalg import norm, cholesky from sys import float_info from optwok.socp import SOCP from numpy.random import rand from scipy.sparse import issparse from numpy import sqrt, dot, zeros, array, squeeze, ones, r_, amax, eye, exp, diag from optwok.mil.mi_classifier import MIClassifier from optwok.mil.mi_loss import HingeLoss, HuberizedHingeLoss from scipy.sparse import csc_matrix class eMIL(MIClassifier): ''' Ellipsoidal Multiple Instance Learning Classifier ''' def one_predict_sparse(self, sparsedata): ''' Ellipsoidal MI classifier for one sparse ellipsoid ''' w = csc_matrix(self.w) return (w*sparsedata['mean'].transpose()).A[0][0] + self.b \ + sqrt((w * sparsedata['covariance'] * w.transpose()).A[0][0]) def one_predict(self, data): ''' Ellipsoidal MI classifier for one ellipsoid''' quad_form = dot(self.w,dot(data['covariance'],self.w)) return dot(self.w,data['mean']) + self.b + sqrt(quad_form) def read_data(self, data): ''' Read data and convert to EllipsoidData ''' return data.to_ellipsoids() def cost_ellipsoid(self, w, b, qi, Si): ''' Loss per ellipsoid''' if issparse(Si): return (w*qi.transpose()).A[0][0] + b + sqrt((w * Si * w.transpose()).A[0][0]) else: quad_form = dot(w,dot(Si,w)) return dot(w,qi) + b + sqrt(quad_form) class eMILDca(eMIL): ''' Ellipsoidal Multiple Instance Learning Classifier, Solve minimization with DCA''' def __init__(self): self.set_opt_param(1e10, 100) def solve_opt(self, v0, verbose, *args, **kwargs): self.w, self.b, self.primal_obj = self.min_dca(v0=v0, verbose=verbose, *args, **kwargs) def set_opt_param(self, factr, maxiter): self.maxiter = maxiter self.factr = factr def check_conv(self, v_old, v_new, f_old, f_new, iteration): return abs((f_old-f_new)/max([1,abs(f_old),abs(f_new)])) < self.factr * float_info.epsilon \ or iteration > self.maxiter def loss(self, w, b): ''' compute the loss for hyperplane (w,b) and bags with labels y and regularization parameter lamb ''' lossf = HingeLoss() loss = 0 if self.traindata.sparse: w = csc_matrix(w) for qi, si, yi in zip(self.traindata.means,self.traindata.covariances,self.traindata.labels): # sum over hinge loss for all distances fi = yi*(self.cost_ellipsoid(w,b,qi,si)) loss += lossf.loss(fi) if self.traindata.sparse: w = squeeze(w.A) # regularize with 2-norm squared: ww = dot(w,w) loss = loss + self.reg_param * ww return loss def min_dca(self, v0=None, verbose=False, check_convergence=None, *args, **kwargs): ''' Minimize with the CCCP, needs solve_dca_subproblem(w_k ,v0,verbose) defined check_convergence(v_old, v_new, f_old, f_new, iter) returns false to continue''' if check_convergence is None: check_convergence = self.check_conv if v0 is None: v = r_[rand(self.traindata.means[0].shape[0]), rand(1)] else: v = v0 converged = False i = 0 while not converged: v_old = v v, f = self.solve_dca_subproblem(v[:-1],v0=v, verbose=verbose) i += 1 f_old = self.loss(v_old[:-1],v_old[-1]) f_new = self.loss(v[:-1],v[-1]) if verbose: print 'current sub-problem loss: ', f print 'full loss: %f, old full loss: %f' %(f_new, f_old) print 'v diff: ', norm(v_old-v) converged = check_convergence(v_old, v, f_old, f_new, i) return v[:-1], v[-1], f class eMILDcaBfgs(eMILDca): ''' Ellipsoidal Multiple Instance Learning Classifier, Solve minimization with DCA and subproblem with bfgs and soft_max''' def check_convergence(self, v_old, v_new, f_old, f_new, iteration ): return self.check_conv(v_old, v_new, f_old, f_new, iteration) \ or norm(v_old-v_new)< self.factr * float_info.epsilon def loss_and_grad_dca(self, v, wk): ''' gradient and loss at w, b for the DCA loss''' w = v[:-1] b = v[-1] beta = 18 loss = 0 gradw = 0 gradb = 0 for qi, Pi, yi in zip(self.traindata.means, self.traindata.covariances, self.traindata.labels): quad_form = sqrt(dot(w,dot(Pi,w))) pred = quad_form + dot(w,qi) + b pred_grad = qi + dot(Pi,w)/quad_form if yi > 0: #positive smgp = self._soft_max_grad(-1+pred,beta) loss += self._soft_max(-1+pred,beta) gradw += smgp * pred_grad gradb += smgp # Linearization of concave part conc_linw = -qi - dot(Pi, wk) / sqrt(dot(wk,dot(Pi,wk))) loss += dot(w,conc_linw) - b gradw += conc_linw gradb += -1 else: #negative smgn = self._soft_max_grad(1+pred,beta) loss += self._soft_max(1+pred,beta) gradw += smgn * pred_grad gradb += smgn # Regularizer loss += self.reg_param * dot(w,w) gradw += self.reg_param * 2 * w return loss, r_[gradw, gradb] def solve_dca_subproblem(self, wk,v0,verbose): return self.solve_dca_bfgs(wk,v0,verbose) def solve_dca_bfgs(self, wk, v0, verbose=False): ''' Solve the same problem as solve_dca_socp, but instead of converting to a SOCP, use the soft max instead of the hinge loss ''' factr = 1e7 pgtol = 1e-05 v, f, info = fmin_l_bfgs_b(func = self.loss_and_grad_dca, x0 = v0, approx_grad = False, m=100, fprime = None, args = (wk,), factr = factr, pgtol=pgtol) if verbose: print 'loss: ', f print 'norm of gradient: ', norm(info['grad']) print 'warnflag: ', info['warnflag'] print 'stop reason: ', info['task'] print 'funcalls: ', info['funcalls'] return v, f def _soft_max(self,x,beta): ''' relax max(0,x) by 1/b (e^(b*x)+1) ''' return 1/beta * logsumexp([beta*x, 0]) def _soft_max_grad(self,x, beta): ''' return the gradient of the soft max (need to multiply this by the gradient of x...''' y = beta*x return exp(y-logsumexp([y,0])) class eMILDcaSocp(eMILDca): ''' Ellipsoidal Multiple Instance Learning Classifier Solve minimization with DCA and subproblem as SOCP''' #@memoized def chol_S(self,i): ''' return A, where A'A = Si ''' return cholesky(self.traindata.covariances[i]).T def solve_dca_subproblem(self, wk,v0,verbose): return self.solve_dca_socp(wk,verbose) def solve_dca_socp(self, wk, verbose=False): ''' x = [w, b, xi, theta], ''' num_vars = self.traindata.num_feat + 1 + self.traindata.num_ex + 1 num_feat = self.traindata.num_feat num_ex = self.traindata.num_ex socp = SOCP(num_vars= num_vars, verbose=verbose, use_mosek=True) pb_ind = array(self.traindata.labels) > 0 q = array(self.traindata.means) P = array(self.traindata.covariances) # Linearization of concave part conc_linw = 0 for qi, Pi in zip(q[pb_ind], P[pb_ind]): conc_linw += -qi - dot(Pi, wk) / sqrt(dot(wk,dot(Pi,wk))) conc_linb = -sum(pb_ind) # f = r_[conc_linw, conc_linb, ones(mibag.m), mibag.lamb] f = r_[zeros(num_vars-1),1] # only z socp.set_objective(f) # Constraints for original objective + regularization q = r_[-conc_linw, -conc_linb, -ones(num_ex), 1]/2 A = zeros((num_feat + 1,num_vars)) A[0,:] = -q A[1:,:num_feat] = diag(ones(num_feat) * sqrt(self.reg_param/2)) d = 1/2 b = zeros(num_feat+1) b[0] = 1/2 socp.add_ineq_constraints(A, b, q, d) for i, e in enumerate(zip(self.traindata.means, self.traindata.labels)): qi, yi = e A_i = zeros((num_feat, num_vars)) A_i[:,:num_feat] = self.chol_S(i) # Ai b_i = zeros(num_feat) c_i = r_[-qi,-1,zeros(num_ex+1)] c_i[num_feat+1+i] = 1 # xi d_i = yi socp.add_ineq_constraints(A_i, b_i, c_i, d_i) # Constraint for positive slack A_0 = zeros((num_vars, num_vars)) A_0[num_feat+1:-1,num_feat+1:-1] = -1*eye(num_ex) socp.set_component_ineq_constraints(A_0, zeros(num_vars)) socp.solve(warn=verbose) return socp.solution[:num_feat+1], socp.cur_obj class eMILBfgs(eMIL): ''' Ellipsoidal Multiple Instance Learning Classifier, Minimize with bfgs''' def solve_opt(self, v0, verbose, huber_delta, *args, **kwargs): self.lossf = HuberizedHingeLoss(huber_delta) self.w, self.b, self.primal_obj = self.min_bfgs(v0=v0, verbose=verbose, *args, **kwargs) def loss_and_grad_onepm(self, v): return self.loss_and_grad(v[:-1],v[-1]) def loss_and_grad(self, w,b): ''' compute the loss for hyperplane (w,b) and bag with labels y and regularization parameter lamb also compute the gradient with respect to (w,b)''' loss = 0 gradw, gradb = 0, 0 nw = norm(w) if self.traindata.sparse: w = csc_matrix(w) for qi, si, yi in zip(self.traindata.means,self.traindata.covariances,self.traindata.labels): # sum over hinge loss for all distances fi = yi*(self.cost_ellipsoid(w,b,qi,si)) loss += self.lossf.loss(fi) gradwi, gradbi = self.lossf.grad(fi, qi, si, yi, w, nw) gradb += gradbi gradw += gradwi if self.traindata.sparse: w = squeeze(w.A) # regularize with 2-norm squared: ww = dot(w,w) gradw = gradw + self.reg_param * 2* w loss = loss + self.reg_param * ww return loss, r_[gradw, gradb] def min_bfgs(self,v0=None, verbose=False): ''' Find optimal hyperplane for the balls (q,s) with labels y and the regularizer parameter lamb ''' if v0 is None: w0 = ones(amax(self.traindata.num_feat)) w0 = w0 / norm(w0) v0 = r_[w0, 0] if verbose: disp = 1 else: disp = 0 if self.traindata.sparse: factr = 1e7 pgtol = 1e-04 else: factr = 1e7 pgtol = 1e-05 v, f, info = fmin_l_bfgs_b(func = self.loss_and_grad_onepm, x0 = v0, approx_grad = False, m=100, fprime = None, args = (), factr = factr, pgtol=pgtol) if verbose: print 'loss: ', f print 'norm of gradient: ', norm(info['grad']) print 'warnflag: ', info['warnflag'] print 'stop reason: ', info['task'] print 'funcalls: ', info['funcalls'] return v[:-1], v[-1], f OptWok-0.3.1/optwok/mil/mi_data.py0000644ڛ|׿-0000000514212140417274023755 0ustar cheng.ongUNIMELB\Domain Users00000000000000''' Created on 28.02.2013 @author: Gabriel ''' from __future__ import division from optwok.mldata import ClassificationData from scipy.sparse import issparse from numpy import squeeze, array class MIData(ClassificationData): def iter(self): for i in range(self.num_ex): yield self.ith_data(i) class EllipsoidData(MIData): ''' Class for ellipsoids for multiple instance learning ''' def __init__(self, means, covariances, labels): self.means = means self.covariances = covariances self.labels = labels if issparse(self.means[0]): self.sparse = True else: self.sparse = False def split_off(self, idx, id_str): rbd = EllipsoidData(self.means[idx], self.covariances[idx], self.labels[idx]) rbd.id_str = id_str rbd.idx = idx return rbd def ith_data(self, i): return {'mean': self.means[i], 'covariance': self.covariances[i], 'label':self.labels[i]} @property def num_feat(self): return max(self.means[0].shape) class MultipleInstanceData(MIData): ''' Class for multiple instance bags: bags is a list of lists of instances label is on a bag level ''' def __init__(self, bags, labels): self.labels = labels if issparse(bags[0][0]): self.bags = [[squeeze(inst.A) for inst in qi] for qi in bags] else: self.bags = bags self.sparse = False def split_off(self, idx, id_str): rbd = MultipleInstanceData(self.bags[idx], self.labels[idx]) rbd.id_str = id_str rbd.idx = idx return rbd def ith_data(self, i): return {'bag': self.bags[i], 'label':self.labels[i]} @property def num_feat(self): return max(self.bags[0][0].shape) class MIMeanData(MIData): ''' Class for means of multiple instance bags ''' def __init__(self, means, labels): self.labels = labels if issparse(means[0]): self.means = [squeeze(qi.A) for qi in means] else: self.means = means self.sparse = False def split_off(self, idx, id_str): rbd = MIMeanData(self.means[idx], self.labels[idx]) rbd.id_str = id_str rbd.idx = idx return rbd def ith_data(self, i): return {'mean': self.means[i], 'label':self.labels[i]} @property def num_feat(self): return max(self.means[0].shape)OptWok-0.3.1/optwok/mil/mi_loss.py0000644ڛ|׿-0000000355212115570207024025 0ustar cheng.ongUNIMELB\Domain Users00000000000000''' Created on 04.03.2013 @author: Gabriel ''' from __future__ import division from numpy import squeeze, sqrt, dot from scipy.sparse import issparse class eMILLoss(object): def gradf(self, qi, Si, yi, w, nw): if issparse(Si): return -yi*(squeeze((qi+ (Si*w.transpose()).transpose()/sqrt((w*Si*w.transpose()).A[0,0])).A)) else: quad_form = dot(w,dot(Si,w)) return -yi*(qi + dot(Si,w)/sqrt(quad_form)) class HingeLoss(eMILLoss): def grad(self, fi, qi, Si, yi, w, nw): if 1 <= fi: gradbi, gradwi = 0, 0 else: gradbi, gradwi = -yi, self.gradf(qi,Si,yi,w,nw) return gradwi, gradbi def loss(self, x): return max(0,1-x) class HuberizedHingeLoss(eMILLoss): ''' hinge without the kink, delta >= 0 parametrizes how "squared" the loss is if delta is zero the original hinge loss is recovered ''' def __init__(self, delta): self.delta = delta def grad(self, fi, qi, Si, yi, w, nw): if fi <= 1: if fi > 1 - self.delta: fid = (1-fi) / self.delta gradbi = fid * -yi gradwi = fid * self.gradf(qi, Si, yi, w, nw) else: gradbi = -yi gradwi = self.gradf(qi, Si, yi, w, nw) else: gradbi, gradwi = 0, 0 return gradwi, gradbi def loss(self, x): if x <= 1: if x > 1 - self.delta: # 1 - delta < x <= 1 return ((1-x) **2) / (2*self.delta) else: # x <= 1 - delta return 1 - x - self.delta/2 else: # x > 1 return 0OptWok-0.3.1/optwok/mldata.py0000644ڛ|׿-0000004527212140420225023036 0ustar cheng.ongUNIMELB\Domain Users00000000000000#!/usr/bin/env python """Classes to encapsulate the idea of a dataset in machine learning, including file access. Currently this focuses on reading and writing transparently to different file formats. A dataset is modeled as an (example,label) tuple, each of which is an array. The base class doesn't know how to split, so just returns one array. The classes currently implemented use three different ways of iterating through files: - CSV uses the python module csv's iterator - Libsvm uses a hand crafted while loop - ARFF always reads the whole file, and does a slice - FASTA uses a hand crafted while loop that behaves like a generator The class DatasetFileARFF is in mldata-arff.py. """ ############################################################################################# # # # This program is free software; you can redistribute it and/or modify # # it under the terms of the GNU General Public License as published by # # the Free Software Foundation; either version 3 of the License, or # # (at your option) any later version. # # # # This program is distributed in the hope that it will be useful, # # but WITHOUT ANY WARRANTY; without even the implied warranty of # # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # # GNU General Public License for more details. # # # # You should have received a copy of the GNU General Public License # # along with this program; if not, see http://www.gnu.org/licenses # # or write to the Free Software Foundation, Inc., 51 Franklin Street, # # Fifth Floor, Boston, MA 02110-1301 USA # # # ############################################################################################# import sys import numpy from numpy import array, concatenate, unique, setdiff1d import csv from optwok.utility import jload try: import arff have_arff = True except ImportError: have_arff = False python_version = sys.version_info[:2] if python_version >= (3, 0): from io import FileIO as BUILTIN_FILE_TYPE else: BUILTIN_FILE_TYPE = file class DatasetFileBase(BUILTIN_FILE_TYPE): """A Base class defining barebones and common behaviour """ def __init__(self,filename,extype): """Just the normal file __init__, followed by the specific class corresponding to the file extension. """ self.extype = extype self.filename = filename def readlines(self,idx=None): """Read the lines defined by idx (a numpy array). Default is read all lines. """ if idx is None: data = self.readlines() else: data = self.readlines()[idx] #itertools.islice(open('tempx.txt'), 11, 12).next() #file("filename").readlines()[11] #linecache.getline( filename, lineno[, module_globals]) return data def writelines(self,data,idx=None): """Write the lines defined by idx (a numpy array). Default is write all lines. data is assumed to be a numpy array. """ if idx is None: self.writelines(data) else: self.writelines(data[idx]) class DatasetFileCSV(DatasetFileBase): """Comma Seperated Values file. Labels are in the first column. """ def __init__(self, filename, extype, verbose=False): self.verbose=verbose DatasetFileBase.__init__(self,filename,extype) def readlines(self,idx=None): """Read from file and split data into examples and labels""" reader = csv.reader(open(self.filename,'r'), delimiter=',', quoting=csv.QUOTE_NONE) labels = [] examples = [] for ix,line in enumerate(reader): if idx is None or ix in idx: labels.append(float(line[0])) if self.extype == 'vec': examples.append(array(map(float,line[1:]))) elif self.extype == 'seq': examples.append(line[1:][0]) elif self.extype == 'mseq': examples.append(array(line[1:])) if self.extype == 'vec': examples = array(examples).T if self.verbose: print('%d features, %d examples' % examples.shape) elif self.extype == 'seq': if self.verbose: print('sequence length = %d, %d examples' % (len(examples[0]),len(examples))) elif self.extype == 'mseq': printstr = 'sequence lengths = ' for seq in examples[0]: printstr += '%d, ' % len(seq) printstr += '%d examples' % len(examples) if self.verbose: print(printstr) return (examples,array(labels)) def writelines(self,examples,labels,idx=None): """Merge the examples and labels and write to file""" if idx==None: idx = range(len(labels)) if self.extype == 'seq': data = zip(labels[idx],list(array(examples)[idx])) if self.extype == 'mseq': data = [] for ix,curlab in enumerate(labels): data.append([curlab]+list(examples[ix])) elif self.extype == 'vec': data = [] for ix,curlab in enumerate(labels): data.append(concatenate((array([curlab]),examples[:,ix].T))) fp = open(self.filename,'w') writer = csv.writer(fp,delimiter=',',quoting=csv.QUOTE_NONE) for ix in idx: writer.writerow(data[ix]) fp.close() class DatasetFileLibsvm(DatasetFileBase): """LibSVM format, where the first column is assumed to be the label, and the entries are in sparse matrix format.""" def __init__(self, filename, extype='vec', verbose=False): assert(extype=='vec') self.verbose=verbose DatasetFileBase.__init__(self,filename,extype) def readlines(self,idx=None): """Read from file and split data into examples and labels""" if idx is not None: raise NotImplementedError infile = open(self.filename, 'r') num_dim = 0 parsed = [] for line in infile: (pline, max_idx) = self._parse_line(line) parsed.append(pline) num_dim = max(num_dim, max_idx) infile.close() labels = [] if self.verbose: print('%d features, %d examples' % (num_dim, len(parsed))) examples = numpy.zeros((int(num_dim), len(parsed))) for ix in range(len(parsed)): labels.append(parsed[ix]['label']) for v in parsed[ix]['variables']: examples[int(v[0])-1,ix] = float(v[1]) return (examples,array(labels)) def _parse_line(self, line): """Parse a LibSVM input line and return attributes.""" items = line.split() label = int(items[0])-1 variables = [] max_idx = 0 for entry in items[1:]: val = entry.split(':') max_idx = max(max_idx, int(val[0])) variables.append((int(val[0]),float(val[1]))) return ({'label':label, 'variables':variables}, max_idx) def writelines(self, examples, labels, idx=None): """Merge the examples and labels and write to file""" if idx==None: idx = range(len(labels)) num_feat = examples.shape[0] fp = open(self.filename,'w') for ix in idx: fp.write('%d ' % (labels[ix]+1)) for ix_feat in range(num_feat): if examples[ix_feat, ix] != 0.0: fp.write('%d:%f '% (ix_feat+1, examples[ix_feat, ix])) fp.write('\n') fp.close() class DatasetFileFASTA(DatasetFileBase): """Fasta format file, labels are in the comment after keyword 'label'. label=1 label=-1 """ def __init__(self, filename, extype): if extype != 'seq': print('Can only write fasta file for sequences!') raise IOError DatasetFileBase.__init__(self,filename,extype) self.fp = None def readlines(self,idx=None): """Read from file and split data into examples and labels""" self.fp = open(self.filename,'r') line = self.fp.readline() examples = [] labels = [] ix = 0 while True: if not line : break (ex,lab,line) = self.readline(line) if idx is None or ix in idx: examples.append(ex) labels.append(lab) ix += 1 self.fp.close() print('sequence length = %d, %d examples' % (len(examples[0]),len(examples))) return (examples,array(labels)) def writelines(self,examples,labels,idx=None,linelen=60): """Write the examples and labels and write to file""" if idx==None: idx = range(len(labels)) fp = open(self.filename,'w') for ix in idx: fp.write('> %d label=%d\n'%(ix,round(labels[ix]))) for lineidx in xrange(0, len(examples[ix]), linelen): fp.write(examples[ix][lineidx:lineidx+linelen] + '\n') fp.close() def readline(self,line): """Reads a fasta entry and returns the label and the sequence""" if line[0] == '' : return assert(line[0] == '>') # Use list comprehension to get the integer that comes after label= a = line.split() label = float([b.split('=')[1] for b in a if b.split('=')[0]=='label'][0]) lines = [] line = self.fp.readline() while True: if not line : break if line[0] == ">": break #Remove trailing whitespace, and any internal spaces lines.append(line.rstrip().replace(" ","")) line = self.fp.readline() return (''.join(lines),label,line) def init_datasetfile(filename, extype): """A factory that returns the appropriate class based on the file extension. recognised file extensions - .csv : Comma Separated Values - .tab : General tabular data - .arff : Attribute-Relation File Format (weka) - .fa : Fasta file format (seq only) - .fasta: same as above. Since the file type does not determine what type of data is actually being used, the user has to supply the example type. extype can be ('vec','seq','mseq') vec - array of floats seq - single sequence mseq - multiple sequences """ allowedtypes = ('vec','seq','mseq') assert(extype in allowedtypes) # map the file extensions to the relevant classes _format2dataset = {'csv' : DatasetFileCSV, 'fa' : DatasetFileFASTA, 'fasta' : DatasetFileFASTA, 'libsvm' : DatasetFileLibsvm, } if have_arff: from esvm.mldata_arff import DatasetFileARFF _format2dataset['arff'] = DatasetFileARFF extension = detect_extension(filename) return _format2dataset[extension](filename, extype) def detect_extension(filename): """Get the file extension""" #if filename.count('.') > 1: # print('WARNING: %s has more than one . using last one' % filename) detect_ext = filename.split('.')[-1] known_ext = ['csv','fasta','fa','libsvm'] if have_arff: known_ext.append('arff') if detect_ext not in known_ext: print('WARNING: %s is an unknown file extension, defaulting to csv' % detect_ext) detect_ext = 'csv' return detect_ext def convert(infile,outfile,extype): """Copy data from infile to outfile, possibly converting the file format.""" fp1 = init_datasetfile(infile,extype) (examples,labels) = fp1.readlines() fp2 = init_datasetfile(outfile,extype) fp2.writelines(examples,labels) class DatasetBase(object): """Encapsulate the data as well as permutations. This is to ensure consistent splitting of training and test sets. """ def __init__(self): self.name = '' self.filename = '' self.examples = None self.labels = None self.num_class = -1 self.perm_filename = '' self.perms = numpy.array([]) # all the permutations self.perm_idx = -1 self.data_dir = '' self.frac_train = 0.7 def get_id_str(self, split_idx, split_type, fold=None, num_cv=None): self.perm_idx = split_idx return self._param2string(split_idx, split_type, fold, num_cv) def get_perm(self, split_idx, split_type, fold=None, num_cv=None): """Returns the indices of the training and test examples.""" assert(split_type=='val' or split_type=='test') self.perm_idx = split_idx perm = self.perms[split_idx] split_train = int(self.frac_train*self.num_examples) if split_type == 'val': perm_train = perm[:split_train] idx_pred = perm_train[fold::num_cv] idx_train = [] for idx in perm_train: if not (idx in idx_pred): idx_train.append(idx) idx_train = numpy.array(idx_train) elif split_type == 'test': idx_train = perm[:split_train] idx_pred = perm[split_train:] id_str = self._param2string(split_idx, split_type, fold, num_cv) return (idx_train, idx_pred, id_str) def generate_perm(self, num_perm=50): """Generate permutations of the index of the examples""" self.perms = [] for iperm in range(num_perm): self.perms.append(permutation(self.num_examples)) self.perms = array(self.perms) def _param2string(self, split_idx, split_type, fold, num_cv): """Generate part of file name the split""" return param2string(split_idx, split_type, fold, num_cv, self.frac_train) def param2string(split_idx, split_type, fold=None, num_cv=None, frac_train=None): """Generate part of file name the split""" if split_type == 'test': id_str = '%d_r%1.2f' % (split_idx, float(frac_train)) elif split_type == 'val': id_str = '%d_cv%d:%d' % (split_idx, fold, num_cv) return id_str class ClassificationData(DatasetBase): ''' Dataset for classification ''' def __init__(self): self.labels = array([]) def split_off(self, idx, id_str): raise NotImplementedError def load_perm(self, permfile, i): ''' load permutation i for task ''' p = jload(permfile) return array(p[i]) def get_perm(self, permfile, perm_idx, model_selection, frac_ms=None, fold_cv=None, num_cv=None): """ Returns the indices of the training and test examples. model_selection: True: This is model selection False: This is test frac_ms: Fraction of model selection / test split (set to 1 for simple cv schemes) fold: None (when model_selection is False) or number of current cross validation fold num_cv: None (when model_selection is False) or total number of cross validation folds """ perm = self.load_perm(permfile, perm_idx) self.perm_idx = perm_idx split_train = int(frac_ms*self.num_ex) if model_selection: # use cross validation for model selection perm_train = perm[:split_train] idx_pred = perm_train[fold_cv::num_cv] idx_train = setdiff1d(perm_train, idx_pred) else: idx_train = perm[:split_train] idx_pred = perm[split_train:] id_str = self._str_perm(permfile, perm_idx, model_selection, frac_ms, fold_cv, num_cv) return (idx_train, idx_pred, id_str) def get_train_pred_data(self, permfile, perm_idx, model_selection, frac_ms, fold_cv, num_cv): (idx_train, idx_pred, id_str) = self.get_perm(permfile, perm_idx, model_selection, frac_ms, fold_cv, num_cv) traindata = self.split_off(idx_train, id_str) preddata = self.split_off(idx_pred, id_str) return traindata, preddata def _str_perm(self, permfile, perm_idx, model_selection, frac_ms, fold_cv, num_cv): test = '_validation' if not model_selection else '' cvstr = '_%d_of_%d-fold_cv' % (fold_cv, num_cv) if model_selection else '' return "perm_ms_%s_perm_idx_%d%s%s" % (permfile, perm_idx, test, cvstr) @property def num_ex(self): return len(self.labels) class Dataset(DatasetBase): """Encapsulate the data as well as permutations. This is to ensure consistent splitting of training and test sets. Initialize dataset from a file. """ def __init__(self, name, data_file=None, perm_file=None, data_dir='', frac_train=0.7, read_data=True): """Load the data into memory""" self.name = name self.data_dir = data_dir self.frac_train = frac_train if data_file: self.filename = data_file else: self.filename = '%s/%s.csv' % (self.data_dir, name) data = init_datasetfile(self.filename,'vec') if read_data: (self.examples, self.labels) = data.readlines() self.num_class = len(unique(self.labels)) if perm_file: self.perm_filename = perm_file else: self.perm_filename = '%s/%s_perm.txt' % (self.data_dir, name) if read_data: self.perms = numpy.loadtxt(self.perm_filename, dtype=int, delimiter=' ') @property def num_examples(self): return self.examples.shape[1] @property def num_features(self): return self.examples.shape[0] def generate_perm(datafile, permfile, num_perm=50): """Read in the data in datafile, and generate a file with permutations of the index of the examples. """ fp1 = init_datasetfile(datafile, 'vec') (examples,labels) = fp1.readlines() num_examples = len(labels) fp2 = open(permfile, 'w') for iperm in range(num_perm): cur_mix = permutation(num_examples) fp2.write(' '.join(map(str, cur_mix.tolist()))) fp2.write('\n') fp2.close() OptWok-0.3.1/optwok/mldata_pytables.py0000644ڛ|׿-0000001012512051075316024736 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Classes to encapsulate the idea of a dataset in machine learning, including file access. Similar to mldata.py. The file format is a pytables zlib compressed hdf5 file. A dataset is modeled as an (example,label) tuple, each of which is an array. The idea is to avoid loading the whole dataset into memory. """ from os.path import isfile import numpy from numpy import array, unique from numpy.random import permutation from tables import openFile class Dataset(object): """Encapsulate the data as well as permutations. This is to ensure consistent splitting of training and test sets. """ def __init__(self, name, data_file=None, perm_file=None, data_dir='', frac_train=0.7): """Initialize file names and parameters""" self.name = name self.data_dir = data_dir self.frac_train = frac_train if data_file: self.filename = data_file else: self.filename = '%s/%s_mat.h5' % (self.data_dir, name) self.data = openFile(self.filename, 'r') self.examples = self.data.root.examples self.labels = self.data.root.labels self.num_class = len(unique(array(self.labels))) self._get_set_permfile(perm_file) @property def num_examples(self): return self.examples.shape[1] @property def num_features(self): return self.examples.shape[0] def _generate_perm(self, num_perm): """Generate permutations of the index of the examples""" self.perms = [] for iperm in range(num_perm): self.perms.append(permutation(self.num_examples)) self.perms = array(self.perms) def _get_set_permfile(self, perm_file, num_perm=50): """If a permutation file exists, load it. Otherwise, generate it and save one. Sets self.perm_filename and self.perms """ if perm_file: self.perm_filename = perm_file else: self.perm_filename = '%s/%s_perm.txt' % (self.data_dir, self.name) if isfile(self.perm_filename): self.perms = numpy.loadtxt(self.perm_filename, dtype=int, delimiter=' ') else: self._generate_perm(num_perm) fp = open(self.perm_filename, 'w') for iperm in range(num_perm): cur_mix = permutation(self.num_examples) fp.write(' '.join(map(str, cur_mix.tolist()))) fp.write('\n') fp.close() def get_id_str(self, split_idx, split_type, fold=None, num_cv=None): self.perm_idx = split_idx return self._param2string(split_idx, split_type, fold, num_cv) def get_perm(self, split_idx, split_type, fold=None, num_cv=None): """Returns the indices of the training and test examples.""" assert(split_type=='val' or split_type=='test') self.perm_idx = split_idx perm = self.perms[split_idx] split_train = int(self.frac_train*self.num_examples) if split_type == 'val': perm_train = perm[:split_train] idx_pred = perm_train[fold::num_cv] idx_train = [] for idx in perm_train: if not (idx in idx_pred): idx_train.append(idx) idx_train = numpy.array(idx_train) elif split_type == 'test': idx_train = perm[:split_train] idx_pred = perm[split_train:] id_str = self._param2string(split_idx, split_type, fold, num_cv) return (idx_train, idx_pred, id_str) def close(self): """Close the HDF5 file""" self.data.close() def _param2string(self, split_idx, split_type, fold, num_cv): """Generate part of file name the split""" return param2string(split_idx, split_type, fold, num_cv, self.frac_train) def param2string(split_idx, split_type, fold=None, num_cv=None, frac_train=None): """Generate part of file name the split""" if split_type == 'test': id_str = '%d_r%1.2f' % (split_idx, float(frac_train)) elif split_type == 'val': id_str = '%d_cv%d:%d' % (split_idx, fold, num_cv) return id_str OptWok-0.3.1/optwok/multiclass.py0000644ڛ|׿-0000001161312055530734023760 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Implementation of a multiclass SVM in the dual. Uses the vector formulation of multiclass SVM, and not one-vs-one or one-vs-all. """ from time import time from numpy.random import randn from numpy import matrix, array, vstack, hstack from numpy import zeros, ones, transpose, nonzero from numpy import sign, eye, unique, kron, argmax from numpy import trace from optwok.kernel import kernel2distance from optwok.sip import SIQP class MulticlassSVM(object): """Base class for multiclass support vector machine, allowing a kernel on labels""" def __init__(self, reg_param, kernel, verbose=False): assert(reg_param > 0.0) self.reg_param = reg_param # Tradeoff regularizer vs loss self.kernel = kernel # kernel self.train_ex = array([]) # training examples self.y = array([]) # labels self.scale = 1.0 # the trace of Ktrain self.train_time = 0.0 # time used for training self.alpha = array([]) # weight vector self.num_train = 0 # number of examples self.num_classes = 0 # number of classes self.epsilon = 1e-6 # how small is zero? self.verbose = verbose # how much output @property def train_out(self): """Output of the classifier on the training set. return a column vector of scores for each example.""" return array((matrix(self.kernel.Ky)*matrix(self.alpha).T*matrix(self.kernel.Kx.kmat)).T) def solve_opt(self): """To be implemented in children""" raise NotImplementedError def train(self, train_ex, train_lab): """Wrapper to convert data to feature matrix, and solve the optimization""" self.train_ex = train_ex self.y = train_lab self.num_train = len(self.y) self.num_classes = len(unique(self.y)) self.kernel.compute(train_ex) t_start = time() self.solve_opt() t_end = time() # Collect statistics about optimization self.train_time = t_end-t_start if self.verbose: print('Trained for %f seconds' % self.train_time) def predict(self, pred_ex): """Wrapper to do prediction. Compute K*alpha, return a column vector of scores for each example """ Ktest = self.kernel.Kx.compute_test(pred_ex) scores = array((matrix(self.kernel.Ky)*matrix(self.alpha).T*matrix(Ktest)).T) return argmax(scores, axis=1) class StructPredictor(MulticlassSVM): """Structured output SVM in the dual""" def __init__(self, reg_param, kernel, verbose=False): MulticlassSVM.__init__(self, reg_param, kernel, verbose) def solve_opt(self): """Solve the multiclass SVM QP Use cvxopt qp solver which expects the following format: min_x 1/2 x'Px + q'x st Gx <= h Ax = b alpha has num_train rows and num_classes columns alphavec = self.alpha.T.reshape((m,1), order='F') """ m = self.num_train*self.num_classes # class indicator oneY = zeros((self.num_train, self.num_classes)) for ix in range(self.num_train): oneY[ix, self.y[ix]] = 1.0 oneY.shape = (m,1) # Margin term for margin rescaling distance = kernel2distance(self.kernel.Ky) distY = zeros((self.num_train, self.num_classes)) for ix in range(self.num_train): distY[ix,:] = distance[self.y[ix],:] distY.shape = (m,1) # Set up the optimization problem svm = SIQP(m, self.verbose) #svm.set_objective(kron(self.Kx, self.Ky), -oneY) #svm.set_objective(self.kernel.kmat, -oneY) svm.set_objective(self.kernel.kmat, distY) D = eye(m) idx = nonzero(oneY)[0] D[idx,idx] = -1.0 svm.add_ineq_constraints(D, zeros((m,1))) UB = zeros((m,m)) UB[idx,idx] = 1.0 svm.add_ineq_constraints(UB, self.reg_param*ones((m,1))) # sum_u alpha_iu = 0 for all i for ix in range(self.num_train): cur_const = zeros((1,m)) cur_const[0,self.num_classes*ix:self.num_classes*(ix+1)] = 1.0 svm.add_eq_constraints(cur_const, zeros((1,1))) # sum_i alpha_iu = 0 for all u #for ix in range(self.num_classes): # class_idx = zeros((1,self.num_classes)) # class_idx[0,ix] = 1.0 # svm.add_eq_constraints(kron(ones((1,self.num_train)), class_idx), zeros((1,1))) svm.solve() self.alpha = svm.solution.reshape((self.num_train, self.num_classes)) def idx2binvec(y): """Convert a vector y with class labels 0,1,... to a binary matrix with corresponding 1 in column label""" num_class = len(unique(y)) Y = zeros((len(y),num_class)) for ix in range(len(y)): Y[ix,y[ix]] = 1.0 return Y OptWok-0.3.1/optwok/named_array.py0000644ڛ|׿-0000002100012140417274024047 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""A thin wrapper around numpy arrays such that dimensions and coordinates can be named. This makes it easier to access slices and elements of the array without resorting to a particular convention. The idea is to have dictionary like interfaces. and have the array expand as needed. The code is not designed for efficiency, but ease of use. """ import os from numpy import array, ones, zeros, empty from numpy import flatnonzero import json class NArray(object): """A thin wrapper around numpy arrays such that dimensions and coordinates can be named. This makes it easier to access slices and elements of the array without resorting to a particular convention. The idea is to have dictionary like interfaces. and have the array expand as needed. The code is not designed for efficiency, but ease of use. """ def __init__(self, dim_names=None, max_elem=10, is_object=True): if dim_names is None: self.dim_names = array(['dummy']) else: self.dim_names = array(dim_names) # the name of each axis self.elem_names = [] # list of list of names of each coordinate self.max_elem = max_elem # size of the array for ix in range(len(self.dim_names)): self.elem_names.append([]) if is_object: self.data = empty([max_elem for ix in self.dim_names], dtype=object) # the contents else: self.data = zeros([max_elem for ix in self.dim_names]) # the contents @property def ndim(self): """Number of dimensions""" return self.data.ndim @property def shape(self): """The number of elements in each dimension""" cur_shape = [] for names in self.elem_names: cur_shape.append(len(names)) return tuple(cur_shape) @property def content(self): """Return the array corresponding to meaningful data""" last = self.shape slicer = '' for ix in last: slicer += ':%d,' % ix slicer = slicer[:-1] return eval('self.data['+slicer+']') def keys(self, dim_name): """Return the list of keys""" if dim_name not in self.dim_names: raise KeyError('%s not found' % dim_name) cur_dim = flatnonzero(self.dim_names==dim_name)[0] return self.elem_names[cur_dim] def num_keys(self, dim_name): return len(self.keys(dim_name)) def add_key(self, elem_name, dim_name, warn=True): """Add element elem_name to dimension dim_name""" cur_dim = flatnonzero(self.dim_names==dim_name)[0] if len(self.elem_names[cur_dim]) >= self.max_elem: raise IndexError('Exceeding size of array, allocate more using max_elem') if elem_name in self.elem_names[cur_dim]: if warn: print('Element name %s already exists, not added' % elem_name) else: self.elem_names[cur_dim].append(elem_name) def add_keys(self, elem_list, dim_name): """Add all elements in elem_list to dim_name""" for elem in elem_list: self.add_key(elem, dim_name) def fill_value(self, value, elem_names, dim_names): """Insert value into the array indexed by elem_names""" assert len(elem_names)==len(dim_names) assert len(dim_names)==self.ndim pointer = '' for (ix,dim) in enumerate(self.dim_names): cur_dim = flatnonzero(array(dim_names)==dim)[0] cur_elem = flatnonzero(array(self.elem_names[ix])==elem_names[cur_dim])[0] if cur_elem >= self.max_elem: raise IndexError('Exceeding size of array') pointer += '%d,' % cur_elem pointer = pointer[:-1] exec('self.data['+pointer+'] = value') def get_value(self, elem_names, dim_names): """Return the value from array""" assert len(elem_names)==len(dim_names) assert len(dim_names)==self.ndim pointer = '' for (ix,dim) in enumerate(self.dim_names): cur_dim = flatnonzero(array(dim_names)==dim)[0] cur_elem = flatnonzero(array(self.elem_names[ix])==elem_names[cur_dim])[0] pointer += '%d,' % cur_elem pointer = pointer[:-1] exec('value = self.data['+pointer+']') return locals()['value'] def _slice(self, dim_name, dim_value): """Return a slice of the array, i.e. conditioning on dim_name = dim_value Base case """ try: cur_dim = flatnonzero(self.dim_names==dim_name)[0] except IndexError: msg = 'Dimension name unknown: %s not found in ' % dim_name msg += str(self.dim_names) raise IndexError(msg) try: cur_val = flatnonzero(array(self.elem_names[cur_dim])==dim_value)[0] except IndexError: msg = 'Index name %s not found in dimension %s\n' % (dim_value, dim_name) msg += 'Known options:' msg += str(self.elem_names[cur_dim]) raise IndexError(msg) slicer = '' for ix in range(self.ndim): if ix == cur_dim: slicer += str(cur_val)+',' else: slicer += ':,' slicer = slicer[:-1] return eval('self.content['+slicer+']') def slice(self, dim_name, dim_value): """Return a slice of the array, i.e. conditioning on dim_name = dim_value """ if type(dim_name) == type(''): return self._slice(dim_name, dim_value) elif type(dim_name) == type(list()) and len(dim_name)==1: return self._slice(dim_name[0], dim_value[0]) else: assert len(dim_name)==len(dim_value) cur_dim = [] cur_val = [] for ix in range(len(dim_name)): try: idx = flatnonzero(self.dim_names==dim_name[ix])[0] except IndexError: msg = 'Dimension name unknown: %s not found in ' % dim_name[ix] msg += str(self.dim_names) raise IndexError(msg) cur_dim.append(idx) try: val = flatnonzero(array(self.elem_names[idx])==dim_value[ix])[0] except IndexError: msg = 'Index name %s not found in dimension %s\n' % (dim_value[ix], dim_name[ix]) msg += 'Known options:' msg += str(self.elem_names[idx]) raise IndexError(msg) cur_val.append(val) slicer = '' for ix in range(self.ndim): if ix in cur_dim: idx = flatnonzero(array(cur_dim)==ix)[0] slicer += str(cur_val[idx])+',' else: slicer += ':,' slicer = slicer[:-1] return eval('self.content['+slicer+']') def save(self, file_name, merge=False): """Save the named array in json format""" if merge and os.path.isfile(file_name): cur = NArray() cur.load(file_name) self.merge(cur) out = {} out['dim_names'] = self.dim_names.tolist() out['elem_names'] = self.elem_names out['max_elem'] = max(self.shape) out['content'] = self.content.tolist() fp = open(file_name, 'w') json.dump(out, fp, indent=4) fp.close() def load(self, file_name): """Load the named array from a file with json format""" fp = open(file_name, 'r') raw_data = json.load(fp) fp.close() self.dim_names = array(raw_data['dim_names']) self.elem_names = raw_data['elem_names'] self.max_elem = raw_data['max_elem'] self.data = array(raw_data['content']) def merge(self, other): """Overwrite non-zero elements in self""" assert((self.dim_names == other.dim_names).all()) assert(self.elem_names == other.elem_names) assert(self.max_elem == other.max_elem) to_copy = self.content == 0.0 self.content[to_copy] = other.content[to_copy].copy() def savemat(self, file_name): """Save the named array in matlab format""" from scipy.io import savemat out = {} out['dim_names'] = self.dim_names.tolist() out['elem_names'] = self.elem_names out['max_elem'] = max(self.shape) out['content'] = self.content.tolist() savemat(file_name, out, oned_as='column') OptWok-0.3.1/optwok/performance_measures.py0000644ڛ|׿-0000000233312140420225025770 0ustar cheng.ongUNIMELB\Domain Users00000000000000''' Created on 04.04.2013 @author: kgabriel ''' from __future__ import division from numpy import sum as asum, array def accuracy(conf): ''' compute the accuracy, given a confusion matrix conf format: [[TP, FP], [FN, TN]]''' return (conf[0,0] + conf[1,1]) / asum(conf) def recall(conf): ''' compute the true positive rate (sensitivity, recall), given a confusion matrix conf format: [[TP, FP], [FN, TN]]''' try: rec = conf[0,0] / (conf[0,0] + conf[1,0]) except RuntimeWarning: # division by zero rec = 0 return rec def precision(conf): ''' compute the precision (or PPV), given a confusion matrix conf format: [[TP, FP], [FN, TN]]''' try: prec = conf[0,0] / (conf[0,0] + conf[0,1]) except RuntimeWarning: # division by zero prec = 0 return prec def get_conf_matrix(ypred, ytrue): ''' get confusion matrix from [-1,+1] labels ''' ypred = array(ypred) ytrue = array(ytrue) yt = (ypred*ytrue) tp = asum((ytrue>0) & (yt>0)) tn = asum((ytrue<0) & (yt>0)) fp = asum((ytrue<0) & (yt<0)) fn = asum((ytrue>0) & (yt<0)) return array([[tp,fp], [fn,tn]])OptWok-0.3.1/optwok/poly_feature.py0000644ڛ|׿-0000000744112055525154024302 0ustar cheng.ongUNIMELB\Domain Users00000000000000try: import scipy except ImportError: print('poly_feature.py: scipy not available') import numpy def partitions(p, d): """Generate integer partitions of p with maximum d terms""" # base case of recursion: zero is the sum of the empty list if p == 0: yield () return # modify partitions of n-1 to form partitions of n for m in partitions(p-1, d): if len(m) < d: yield (1,) + m if m and (len(m) < 2 or m[1] > m[0]): yield (m[0] + 1,) + m[1:] def get_subsets(d, k): """Get all subsets of size k from set of size d. Assume items in d are integers from 0 to d-1. """ # check base cases if k == 0 or d < k: yield [] elif d == k: yield range(d) else: # Use recursive formula based on binomial coeffecients: # choose(d, k) = choose(d - 1, k - 1) + choose(d - 1, k) for s in get_subsets(d-1, k-1): s.append(d-1) yield s for s in get_subsets(d-1, k): yield s def compute_coeff(p,m): """Compute the coefficient for a polynomial kernel (of degree p) feature given the partition vector m""" denom = 1.0 for mi in m: denom *= scipy.factorial(mi) return numpy.sqrt(scipy.factorial(p)/denom) def get_feature_map(p,d): """Explicit feature map for polynomial kernel of degree p and original dimensions d""" for m in partitions(p, d): c = compute_coeff(p,m) for ss in get_subsets(d, len(m)): yield (c, m, ss) def get_poly_features(X, p): """Return the explicit feature vector for the polynomial kernel""" d = X.shape[0] n = X.shape[1] Phi = [] for ix in range(n): example = X[:,ix] cur_phi = [] for (c, m, ss) in get_feature_map(p, d): value = 1.0 for ixm in range(len(m)): value *= example[ss[ixm]]**m[ixm] cur_phi.append(c*value) Phi.append(cur_phi) Phi = (numpy.array(Phi)).T # How to normalize? for ix in range(Phi.shape[1]): Phi[:,ix] /= numpy.linalg.norm(Phi[:,ix]) return Phi ##################################################################### # Test scripts ##################################################################### def diff_kernel_matrix(train_ex, kdict): """Compute the difference between the kernel matrices between shogun and this implementation""" # shogun from shogun.Features import RealFeatures from shogun.Kernel import PolyKernel sfeat = RealFeatures(train_ex) kernel = PolyKernel(sfeat, sfeat, kdict['param'], kdict['inhomog']) K_shogun = kernel.get_kernel_matrix() # explicit feature map if kdict['inhomog']: X = numpy.vstack([train_ex, numpy.ones((1,train_ex.shape[1]))]) else: X = train_ex efeat = get_poly_features(X, kdict['param']) K_feat = numpy.matrix(efeat).T*numpy.matrix(efeat) # compare if train_ex.shape[1] < 10: print(K_shogun) print(K_feat) return numpy.sum(sum(numpy.abs(K_feat-K_shogun))) def compare_with_shogun(): """Compare the kernel matrix with shogun's""" from esvm.datafuncs import cloudgen #numpy.random.seed(1) num_train = 100 num_feat = 3 frac_pos = 0.5 width = 1.5 (metadata, train_ex, train_labels) = cloudgen(num_train, num_feat, frac_pos, width) for degree in range(1,7): kdict = {'name':'poly', 'param': degree, 'inhomog': False} print(kdict) print('Absolute difference: %f' % diff_kernel_matrix(train_ex, kdict)) kdict = {'name':'poly', 'param': degree, 'inhomog': True} print(kdict) print('Absolute difference: %f' % diff_kernel_matrix(train_ex, kdict)) if __name__ == '__main__': compare_with_shogun() OptWok-0.3.1/optwok/robust_svm.py0000644ڛ|׿-0000001040212140420225023762 0ustar cheng.ongUNIMELB\Domain Users00000000000000''' Created on 21.02.2013 @author: Gabriel ''' from optwok.socp import SOCPS from numpy import r_, zeros, sqrt, ones, eye, diag, sign, array, dot, squeeze from optwok.mldata import ClassificationData try: from sklearn.svm import SVC except ImportError: print 'sklearn missing' class RobustData(ClassificationData): ''' Dataset that has a covariance matrix associated with every data point ''' def __init__(self, examples, covariances, labels): self.examples = examples self.covariances = covariances self.labels = labels def split_off(self, idx, id_str): rbd = RobustData(self.examples[idx], self.covariances[idx], self.labels[idx]) rbd.id_str = id_str rbd.idx = idx return rbd def ith_data(self, i): return {'example': self.examples[i], 'covariance': self.covariances[i], 'label':self.labels[i]} def iter(self): # TODO: this should be in common MIData ancestor for i in range(self.num_ex): yield self.ith_data(i) @property def num_feat(self): return self.examples[0].shape[0] class VanillaSVM(): def __init__(self, C, verbose=False): self.verbose = verbose self.C = C def one_predict(self, data): ''' predict one datapoint ''' return dot(self.w,data['example']) + self.b def predict(self, testset): # Could also be in common MIClassifier ancestor ''' implement classifier ''' y = [] for pd in testset.iter(): yi = self.one_predict(pd) y.append(yi) return sign(array(y)) def train(self, trainset): self.train_ex = trainset.examples self.train_lab = trainset.labels self.num_feat = trainset.num_feat self.num_ex = trainset.num_ex self.solve_opt() def solve_opt(self): svm = SVC(C = self.C, kernel='linear') svm.fit(self.train_ex, self.train_lab) self.w, self.b, self.primal_obj = squeeze(svm.coef_), svm.intercept_[0], None class RobustSVM(VanillaSVM): def train(self, trainset): self.train_ex = trainset.examples self.train_cov = trainset.covariances self.train_lab = trainset.labels self.num_feat = trainset.num_feat self.num_ex = trainset.num_ex self.solve_opt() def solve_opt(self): ''' x = [w, b, xi, theta], ''' ndim = self.num_feat nobj = self.num_ex num_vars = ndim + 1 + nobj + 1 socp = SOCPS(num_vars= num_vars, verbose=self.verbose, use_mosek=True) f = r_[zeros(num_vars-1),1] # only theta socp.set_objective(f) # Constraints for original objective + regularization # (see http://www-personal.umich.edu/~mepelman/teaching/IOE511/Handouts/511notes07-16.pdf for QCQP -> LPQP, # http://retis.sssup.it/~bini/teaching/optim2010/cone_programming.pdf for LPQP -> SOCP) c = r_[zeros(ndim+1),-ones(nobj)*self.C, 1]/2 A = zeros((ndim+1,num_vars)) #A[:ndim,:ndim] = eye(ndim,ndim) * sqrt(1/2) A[0,:] = -c A[1:,:ndim] = diag(ones(ndim) * sqrt(1/2)) d = 1/2 b = zeros(ndim+1) b[0] = 1/2 socp.add_ineq_constraints(A, b, c, d) for i, e in enumerate(zip(self.train_ex, self.train_lab, self.train_cov)): qi, yi, Pi = e A_i = zeros((ndim, num_vars)) A_i[:,:ndim] = Pi b_i = zeros(ndim) c_i = r_[yi*qi,yi,zeros(nobj+1)] c_i[ndim+1+i] = 1 # xi d_i = -1 socp.add_ineq_constraints(A_i, b_i, c_i, d_i) # Constraint for positive slack A_0 = zeros((num_vars, num_vars)) A_0[ndim+1:-1,ndim+1:-1] = -1*eye(nobj) socp.set_component_ineq_constraints(A_0, zeros(num_vars)) socp.solve(warn=self.verbose) self.w = socp.solution[:ndim] self.b = socp.solution[ndim+1] self.primal_obj = socp.cur_objOptWok-0.3.1/optwok/simdata.py0000644ڛ|׿-0000004125412055562256023232 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Some functions to generate simulated data""" try: from os import link except ImportError: # Hack for windows from shutil import copy2 def link(src, dst): copy2(src, dst) from numpy import ones, concatenate, real from numpy.random import randn, rand, permutation, randint, seed from numpy import where, nonzero, sqrt, real from numpy import zeros, eye, mean, kron, sign from numpy import array, equal, linspace from numpy import flatnonzero from numpy.linalg import norm from optwok.mldata import DatasetFileCSV from optwok.testtools import flatten_list, list2matrix from numpy import matrix, vstack, hstack, concatenate from optwok.kernel import GaussKernel, JointKernel from optwok.io_pickle import load, save from optwok.mldata import DatasetBase from scipy.linalg.matfuncs import sqrtm ##################################################################### # Binary classification ##################################################################### def cloudgen(numpoint, numfeat, numnoise, fracpos, width): """Generate two Gaussian point clouds, centered around one and minus one. Gaussian clouds are in numfeat dimensions. Generate uniform noise in numnoise dimensions. """ numpos = int(round(fracpos*numpoint)) numneg = numpoint - numpos metadata = 'cloudgen(%d,%d,%d,%d,%3.2f)' % (numpos, numneg, numfeat, numnoise, width) print(metadata) datapos = ones((numfeat, numpos)) + width*randn(numfeat, numpos) dataneg = -ones((numfeat, numneg)) + width*randn(numfeat, numneg) noise = (2.0+width)*(rand(numnoise, numpos+numneg) -0.5*ones((numnoise, numpos+numneg))) #noise = sign(rand(numnoise, numpos+numneg)-0.5*ones((numnoise, numpos+numneg)))\ # + width*randn(numnoise, numpos+numneg) pointcloud = 0.2*concatenate((concatenate((datapos,dataneg),axis=1), noise), axis=0) labels = concatenate((ones(numpos),-ones(numneg))) return metadata, pointcloud, labels def cloudgen_weak(numpoint, numfeat, numweak, fracpos, width): """Generate two Gaussian point clouds, centered around one and minus one. Gaussian clouds are in numfeat dimensions. Overlapping Gaussian clouds are in numweak dimensions. """ numpos = int(round(fracpos*numpoint)) numneg = numpoint - numpos metadata = 'cloudgen_weak(%d,%d,%d,%d,%3.2f)' %\ (numpos, numneg, numfeat, numweak, width) print(metadata) strongpos = ones((numfeat, numpos)) + width*randn(numfeat, numpos) strongneg = -ones((numfeat, numneg)) + width*randn(numfeat, numneg) weakpos = 0.1*ones((numweak, numpos)) + 2.0*width*randn(numweak, numpos) weakneg = -0.1*ones((numweak, numneg)) + 2.0*width*randn(numweak, numneg) datapos = concatenate((strongpos, weakpos), axis=0) dataneg = concatenate((strongneg, weakneg), axis=0) pointcloud = 0.2*concatenate((datapos,dataneg),axis=1) labels = concatenate((ones(numpos),-ones(numneg))) return metadata, pointcloud, labels def create_data(num_train=500, num_test=200, num_feat=10, num_noise_feat=10, width=0.8, frac_pos=0.5, frac_flip=0.0, symm_flip=True): """Create mixture of Gaussians distribution""" (metadata, train_ex, train_labels) = cloudgen(num_train, num_feat, num_noise_feat, frac_pos, width) (metadata, test_ex, test_labels) = cloudgen(num_test, num_feat, num_noise_feat, frac_pos, width) random_classification_noise(train_labels, frac_flip, symm_flip) return (train_ex, train_labels, test_ex, test_labels) def _get_hyperplane(num_feat, in_simplex, homogeneous, min_val=0.01): """Randomly create a hyperplane""" if in_simplex: w = rand(num_feat,1) w += min_val+1.0/num_feat w = w/sum(abs(w)) if homogeneous: b = 0.0 else: b = rand() else: w = rand(num_feat,1)-0.5 w = sign(w)*(abs(w)+min_val+1.0/num_feat) w = w/norm(w) if homogeneous: b = 0.0 else: b = 2.0*(rand()-0.5) return w,b def linear_separation(num_point, num_feat, num_noise, margin=0.5, homogeneous=True, in_simplex=False): """Generate uniformly distributed data in [-1,1]^d, pick a random hyperplane and create labels according to it. """ assert(margin < 1.0 and margin >= 0.0) margin /= float(num_feat) w, b = _get_hyperplane(num_feat, in_simplex, homogeneous) param = 'linear_separation(%d, %d, %d, %1.2f, %d)' %\ (num_point, num_feat, num_noise, margin, int(homogeneous)) metadata = (param, w, b) data = 2.0*(rand(num_feat, int((2.0/(1.0-margin))*num_point))-0.5) pred = array(matrix(w).T*matrix(data) + b) to_keep = nonzero(abs(pred)>margin)[1] data = data[:,to_keep] data = data[:,:num_point] # label based on w and b pred = array(matrix(w).T*matrix(data) + b) pos_idx = nonzero(pred>0.0)[1] neg_idx = nonzero(pred<0.0)[1] label = zeros(num_point) label[pos_idx] = 1.0 label[neg_idx] = -1.0 # generate noise dimensions if num_noise > 0: noise = 2.0*(rand(num_noise, num_point)-0.5) data = vstack([data, noise]) return metadata, data, label def linear_separation_example(): from optwok.testtools import plot_2d from matplotlib.pyplot import figure metadata, data, label = linear_separation(200, 2, 0) figure() plot_2d(data, label) metadata, data, label = linear_separation(200, 2, 0, margin=0.4, homogeneous=False) figure() plot_2d(data, label) metadata, data, label = linear_separation(200, 1, 1, margin=0.5) figure() plot_2d(data, label) metadata, data, label = linear_separation(200, 2, 0, in_simplex=True) figure() plot_2d(data, label) metadata, data, label = linear_separation(200, 1, 1, in_simplex=True) figure() plot_2d(data, label) ##################################################################### # Multiclass ##################################################################### def multi_cloudgen(num_point, num_feat, num_class, edge_length): """Generate numclass number of Gaussian point clouds, centered around the simplex of numclass.""" assert(num_feat >= num_class) centers = zeros((num_feat, num_class)) centers[:num_class, :num_class] = (edge_length/sqrt(2))*eye(num_class) midpoint = mean(centers,axis=1) midpoint.shape = (num_feat,1) centers = centers - kron(ones((1,num_class)),midpoint) if type(num_point) == type(0): num_point_list = [] for ix in range(num_class): num_point_list.append(num_point) num_point = num_point_list metadata = 'multi_cloudgen(%s,%d,%d,%3.2f)' % (num_point, num_feat, num_class, edge_length) pointcloud = zeros((num_feat,0)) labels = array([],dtype=int) for ix in range(num_class): cur_cloud = kron(ones((num_point[ix],1)),centers[:,ix]).T + randn(num_feat, num_point[ix]) pointcloud = concatenate((pointcloud, cur_cloud), axis=1) labels = concatenate((labels, ix*ones(num_point[ix]))) pointcloud /= edge_length/sqrt(2) return metadata, pointcloud, labels def multi_cloudgen_mix(num_point, num_feat, num_class, edge_length, mix_list): """Generate num_class Gaussian point clouds, but labels are mixed depending on mix_list. """ (metadata, pointcloud, labels) = multi_cloudgen(num_point, num_feat, num_class, edge_length) metadata += str(mix_list) new_labels = zeros(labels.shape, dtype=int) if len(mix_list) == num_class: # split classes for cid in range(num_class): idx = nonzero(equal(labels,cid))[0] if len(mix_list[cid]) > 1: new_labels[idx] = randint(min(mix_list[cid]), max(mix_list[cid])+1, len(idx)) else: new_labels[idx] = mix_list[cid][0] else: # merge classes assert(len(flatten_list(mix_list)) == num_class) for cid in range(len(mix_list)): idx = array([], dtype=int) for mix in mix_list[cid]: idx = concatenate([idx, nonzero(equal(labels, mix))[0]]) new_labels[idx] = cid return metadata, pointcloud, new_labels ##################################################################### # Simulated data for sparsity experiments ##################################################################### def random_classification_noise(labels, frac_flip, symm_flip): """Generate random classification noise by flipping a proportion of labels randomly. """ if frac_flip>0.0: num_ex = len(labels) if symm_flip: flip_idx = permutation(num_ex)[:round(frac_flip*num_ex)] labels[flip_idx] = -1.0*labels[flip_idx] else: flip_idx = permutation(num_ex)[:round(2.0*frac_flip*num_ex)] labels[flip_idx] = -1.0*labels[where(labels[flip_idx]>0.0)] def twoballs_filename(data_dir, num_examples, num_feat, num_noise_feat, frac_flip): """Generate filename to save data and permutations""" data_filename = data_dir + '/twoballs_n=%d_%d:%d_rcn=%1.1f.csv'\ % (num_examples, num_feat, num_noise_feat, frac_flip) perm_filename = data_dir + '/twoballs_n=%d_%d:%d_rcn=%1.1f_perm.txt'\ % (num_examples, num_feat, num_noise_feat, frac_flip) return (data_filename, perm_filename) def save_cloudgen(data_dir, num_noise_feat, frac_flip, symm_flip=True, num_ex=100, num_feat=10, frac_pos=0.5, width=1.5, num_perms=50): """Generate two gaussians, with noise dimensions and random classification noise, and save the dataset. """ assert(frac_flip>=0.0 and frac_flip<=0.5) (metadata, examples, labels) = cloudgen(num_ex, num_feat, num_noise_feat, frac_pos, width) random_classification_noise(labels, frac_flip, symm_flip) (data_filename, perm_filename) = twoballs_filename(data_dir, num_ex, num_feat, num_noise_feat, frac_flip) outfile = DatasetFileCSV(data_filename, 'vec', verbose=True) outfile.writelines(examples, labels) handle = open(perm_filename, 'w') for iperm in range(num_perms): cur_mix = permutation(num_ex) handle.write(' '.join(map(str, cur_mix.tolist()))) handle.write('\n') handle.close() def sparse_data(data_dir): """Generate data to investigate sparsity, with respect to both features and support vectors.""" seed(1) num_feat = 10 num_examples = [50, 100] noise_feats = [0, 10, 100, 200] rcns = [0.0, 0.1, 0.2, 0.3] num_datasets = len(noise_feats) assert(len(rcns)==num_datasets) for num_ex in num_examples: for num_noise_feat in noise_feats: save_cloudgen(data_dir, num_noise_feat, 0.0, num_ex=num_ex, num_feat=num_feat) (data_filename, perm_filename) = twoballs_filename(data_dir, num_ex, num_feat, 0, 0.0) # Add random classification noise in the training for rcn in rcns[1:]: (df, pf) = twoballs_filename(data_dir, num_ex, num_feat, 0, rcn) link(data_filename, df) link(perm_filename, pf) for ix in range(1,num_datasets): (dfs, pfs) = twoballs_filename(data_dir, num_ex, num_feat, noise_feats[ix], 0.0) (df, pf) = twoballs_filename(data_dir, num_ex, num_feat, noise_feats[ix], rcns[ix]) link(dfs, df) link(pfs, pf) ##################################################################### # Gaussian Process regression ##################################################################### def mean_func1(x, shift=0): from numpy import sin x += shift return 2*sin(x)+sin(3*x+1) def mean_func2(x, shift=0): from numpy import cos x += shift return 2*cos(x)+cos(3*x+1) def gp_gen(num_point, num_dim, domain, noise_level, mix_list=[[0,1],[2]]): """Generate matrix variate normally distributed data""" reg_param = 1e-8 num_class = len(flatten_list(mix_list)) X = domain*rand(num_dim, num_point) #X = linspace(-domain, domain, num_point) #X.shape = (num_dim, num_point) Kx = GaussKernel(1.0) Kx.compute(X) Ky = list2matrix(mix_list, neg_corr=True) K = JointKernel(Kx, Ky) #imshow(K.kmat) #colorbar() L = real(sqrtm(0.5*(K.kmat)+reg_param*eye(num_point*num_class)).T) #mu = vstack([mean_func1(X), mean_func1(X), mean_func2(X)]) mu = zeros((num_class, num_point)) Y = L*matrix(randn(num_point*num_class,1)) Y.shape = (num_point,num_class) Y = real(Y.T) Y += mu + noise_level*randn(num_class, num_point) Y = array(Y) return (X,Y) def gp_gen_pair(num_pairs, num_dim, domain, noise_level, mix_list=[[0,1],[2]], pos_label=False): """Generate data from a multioutput Gaussian process, The outputs are irregular, and are dependent on context Z""" num_class = len(flatten_list(mix_list)) num_ex = num_class*num_pairs num_ex = max(num_ex, num_pairs) (X_full,Y_full) = gp_gen(num_ex, num_dim, domain, noise_level, mix_list) X = zeros((X_full.shape[0],0)) Y = array([]) Z = array([]) for ix in range(num_class): perm = permutation(num_ex) X = hstack([X, X_full[:,perm[:num_pairs]]]) Y = hstack([Y, Y_full[ix,perm[:num_pairs]]]) Z = concatenate((Z, ix*ones(num_pairs))) if pos_label: min_y = min(Y) - 0.1 Y -= min_y return (X,Z,Y) def plot_missing_data(): """A demo of gp_gen_pair""" from matplotlib.pyplot import plot, show colour_list = ['b','g','r','c','m','k'] (X,Z,Y) = gp_gen_pair(20, 1, 5, 0.1) for zix in range(3): cur_idx = flatnonzero(Z == zix) plot(X[:,cur_idx].flatten(), Y[:,cur_idx].flatten(), colour_list[zix]+'o', hold=True) show() def save_gp_data(x, y, dataname, data_dir='.'): """Save Gaussian process data""" dataset = DatasetBase() dataset.name = dataname dataset.filename = dataset.name + '.pkl' dataset.examples = x dataset.labels = y (dataset.num_features, dataset.num_examples) = x.shape dataset.num_class = y.shape[0] dataset.perm_filename = '' dataset.perms = zeros((50, dataset.num_examples), dtype=int) for ix in range(50): dataset.perms[ix] = permutation(dataset.num_examples) dataset.perm_idx = -1 dataset.data_dir = data_dir dataset.frac_train = 0.7 save('%s/%s' % (data_dir, dataset.filename), dataset) def save_gp_context_data(x, z, y, dataname, data_dir='.'): """Save Gaussian process with context (z) data""" dataset = DatasetBase() dataset.name = dataname dataset.filename = dataset.name + '.pkl' dataset.examples = x dataset.context = z dataset.labels = y (dataset.num_features, dataset.num_examples) = x.shape dataset.num_class = y.shape[0] dataset.perm_filename = '' dataset.perms = zeros((50, dataset.num_examples), dtype=int) for ix in range(50): dataset.perms[ix] = permutation(dataset.num_examples) dataset.perm_idx = -1 dataset.data_dir = data_dir dataset.frac_train = 0.7 save('%s/%s' % (data_dir, dataset.filename), dataset) def gp_example(): """Plot some points""" from matplotlib.pyplot import plot, figure num_ex = 200 domain = 5 noise_level = 0.3 (X,Y) = gp_gen(num_ex, 1, domain, noise_level) dataname = 'gaussproc_03_%d_%1.1f' % (num_ex, noise_level) save_gp_data(X, Y, dataname, data_dir='/local/cong/Data/muloutreg') data = load('/local/cong/Data/muloutreg/gaussproc_03_%d_%1.1f.pkl' % (num_ex, noise_level)) figure() plot(data.examples.T, data.labels.T, '+', hold=True) ##################################################################### # learning with teacher ##################################################################### def teacher_data(num_train=100, num_test=300, num_feat=5, num_feat_teach=10, width=1.0): """Generate two Gaussians and then split features into normal features and teacher features """ (metadata, examples, labels) = cloudgen(num_train+num_test, num_feat+num_feat_teach, 0, 0.5, width) cur_mix = permutation(num_train+num_test) examples = examples[:,cur_mix] labels = labels[cur_mix] train_ex = examples[:num_feat,:num_train] teach_ex = examples[-num_feat_teach:,:num_train] pred_ex = examples[:num_feat,-num_test:] train_lab = labels[:num_train] pred_lab = labels[-num_test:] return (train_ex, teach_ex, train_lab, pred_ex, pred_lab) if __name__ == '__main__': data = teacher_data() (train_ex, teach_ex, train_lab, test_ex, test_lab) = data print(train_ex.shape) print(teach_ex.shape) print(train_lab.shape) print(test_ex.shape) print(test_lab.shape) OptWok-0.3.1/optwok/sip.py0000644ڛ|׿-0000001455512055533155022403 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""A collection of classes for semi-infinite programming""" import cvxopt.solvers from cvxopt import matrix as cm from cvxopt import sparse as cs from numpy import matrix, array, dot, vstack, hstack, kron from numpy import zeros, ones, mean, std, sign, eye, nonzero import sys #sys.path.append('/cluster/home/infk/cong/Install/mosek/6/tools/platform/linux64x86/python/2') have_mosek = True try: import mosek except ImportError as error: print(error) have_mosek = False class SIP(object): """A wrapper to cvxopt solver. Allow an iterative addition of constraints to the original problem. The cvxopt solver expects the following format: min_x c'x st Gx <= h Ax = b min_x 1/2 x'Px + q'x st Gx <= h Ax = b """ def __init__(self, num_vars, verbose=False, use_mosek=False): assert num_vars > 0.0, 'Need at least one variable' self.num_vars = num_vars self.stop_eps = 1e-7 self.old_w = None self.P = cm(0.0, (num_vars, num_vars)) self.q = cm(0.0, (num_vars, 1)) self.G = cm(0.0, (0, num_vars)) self.h = cm(0.0, (0, 1)) self.A = cm(0.0, (0, num_vars)) self.b = cm(0.0, (0, 1)) self.solution = None self.lag_eq = None self.cur_obj = None self.optimal = False self.verbose = verbose self.use_mosek = use_mosek def add_ineq_constraints(self, G_new, h_new): """Concatenate the new inequality constraints""" assert G_new.shape[1] == self.num_vars, 'Size mismatch in inequality constraint (lhs)' assert h_new.shape[1] == 1, 'Size mismatch in inequality constraint (rhs)' assert G_new.shape[0] == h_new.shape[0], 'Size mismatch in inequality constraint' self.G = cm(vstack([array(self.G), G_new])) self.h = cm(vstack([array(self.h, dtype = 'd'), array(h_new, dtype = 'd')])) def add_eq_constraints(self, A_new, b_new): """Concatenate the new equality constraints""" assert A_new.shape[1] == self.num_vars, 'Size mismatch in equality constraint (lhs)' assert b_new.shape[1] == 1, 'Size mismatch in equality constraint (rhs)' assert A_new.shape[0] == b_new.shape[0], 'Size mismatch in equality constraint' self.A = cm(vstack([array(self.A), A_new])) self.b = cm(vstack([array(self.b), b_new])) def solve(self, warn=True): """Check for empty constraints then invoke solver""" res = self.solve_cvxopt() if self.verbose: print(res) self.solution = array(res['x']).copy().flatten() self.lag_eq = array(res['y']).copy().flatten() self.cur_obj = res['primal objective'] if res['status'] == 'optimal': self.optimal = True else: self.optimal = False if warn: print('sip.py: cvxopt did not converge to optimality') print(res) del res class SIQP(SIP): """A wrapper to cvxopt QP solver.""" def __init__(self, num_vars, verbose=False, use_mosek=True): SIP.__init__(self, num_vars, verbose, use_mosek) def set_objective(self, P, q): """Set P and q""" assert P.shape == (self.num_vars, self.num_vars), 'Size mismatch in quadratic objective term' assert q.shape == (self.num_vars, 1), 'Size mismatch in linear objective term' self.P = cm(P) self.q = cm(q) def solve_cvxopt(self): """Check for empty constraints then invoke solver""" if self.verbose: print('have mosek %s' % str(have_mosek)) print('use mosek %s' % str(self.use_mosek)) if have_mosek and self.use_mosek: if not self.verbose: cvxopt.solvers.options['MOSEK'] = {mosek.iparam.log:0} res = cvxopt.solvers.qp(self.P, self.q, self.G, self.h, self.A, self.b, 'mosek') else: if not self.verbose: cvxopt.solvers.options['show_progress'] = False res = cvxopt.solvers.qp(self.P, self.q, self.G, self.h, self.A, self.b) return res class SILP(SIP): """A wrapper to cvxopt LP solver.""" def __init__(self, num_vars, verbose=False, use_mosek=True): SIP.__init__(self, num_vars, verbose, use_mosek) def set_objective(self, q): """Set c""" assert q.shape == (self.num_vars, 1), 'Size mismatch in linear objective term' self.q = cm(q) def solve_cvxopt(self): """Check for empty constraints then invoke solver""" if self.verbose: print('have mosek %s' % str(have_mosek)) print('use mosek %s' % str(self.use_mosek)) if have_mosek and self.use_mosek: if not self.verbose: cvxopt.solvers.options['MOSEK'] = {mosek.iparam.log:0} res = cvxopt.solvers.lp(self.q, self.G, self.h, self.A, self.b, 'mosek') else: if not self.verbose: cvxopt.solvers.options['show_progress'] = False res = cvxopt.solvers.lp(self.q, self.G, self.h, self.A, self.b) return res class SILPS(SILP): """A wrapper to cvxopt LP solver with sparse matrices for the constraints.""" def __init__(self, num_vars, verbose=False, use_mosek=True): SILP.__init__(self, num_vars, verbose, use_mosek) def add_ineq_constraints(self, G_new, h_new): """Concatenate the new inequality constraints""" assert h_new.shape[1] == 1, 'Size mismatch in inequality constraint (rhs)' assert G_new.size[1] == self.num_vars, 'Size mismatch in inequality constraint (lhs)' assert G_new.size[0] == h_new.shape[0], 'Size mismatch in inequality constraint' self.G = cs([self.G,G_new]) self.h = cm(vstack([array(self.h,dtype='d'), array(h_new,dtype='d')])) def add_eq_constraints(self, A_new, b_new): """Concatenate the new equality constraints""" assert A_new.size[1] == self.num_vars, 'Size mismatch in equality constraint (lhs)' assert b_new.shape[1] == 1, 'Size mismatch in equality constraint (rhs)' assert A_new.size[0] == b_new.shape[0], 'Size mismatch in equality constraint' self.A = cs([self.A, A_new]) self.b = cm(vstack([array(self.b,dtype='d'), array(b_new,dtype='d')])) OptWok-0.3.1/optwok/socp.py0000644ڛ|׿-0000000717412115570207022547 0ustar cheng.ongUNIMELB\Domain Users00000000000000''' Created on 16.12.2011 @author: kgabriel ''' from numpy.core.numeric import newaxis from optwok.sip import SIP """Class for second order cone programming""" import cvxopt import cvxopt.solvers from cvxopt import matrix as cm, sparse as cs from numpy import matrix, array, dot, vstack, hstack, kron from numpy import zeros, ones, mean, std, sign, eye, nonzero have_mosek = True try: print("using mosek") import mosek except: print("not using mosek") have_mosek = False class SOCP(SIP): ''' A wrapper to cvxopt SOCP solver, converting the boyd notation into cvxopt matrices boyd + component wise constraints: min_x f'x st ||A_i x + b_i|| <= c_i'x + d_i, i = 1 ... m 0 <= b_0 - A_0 x Fx = g cvxopt: min_x c'x st G_i x + s_i = h_i, i = 0 ... M Ax = b s_0 >= 0 s_i0 >= ||s_i1|| i = 1 ... M ''' def __init__(self, num_vars, verbose = False, use_mosek=False): SIP.__init__(self, num_vars, verbose, use_mosek=use_mosek) self.G0 = None self.h0 = None self.G1 = [] self.h1 = [] def set_objective(self, f): """Set f""" assert f.shape[0] == self.num_vars, 'Size mismatch in linear objective term' self.c = cm(f) def set_component_ineq_constraints(self, A_0, b_0): ''' Set the non-cone constraints 0 <= b_0 - A_0 x ''' assert A_0.shape == (b_0.shape[0], self.num_vars), 'Size mismatch in inequality constraint (lhs)' assert b_0.shape[0] == self.num_vars self.G0 = cm(A_0) self.h0 = cm(b_0) def add_ineq_constraints(self, A_i, b_i, c_i, d_i): """Add the new inequality constraints to the list""" assert c_i.shape[0] == self.num_vars, 'Size mismatch in inequality constraint (rhs)' assert A_i.shape == (b_i.shape[0], self.num_vars), 'Size mismatch in inequality constraint (lhs)' G = cm(vstack((-c_i.T, -A_i))) h = cm(vstack((d_i, b_i[:, newaxis]))) self.G1.append(G) self.h1.append(h) def solve_cvxopt(self): """Check for empty constraints then invoke solver""" if have_mosek and self.use_mosek: if not self.verbose: cvxopt.solvers.options['MOSEK'] = {mosek.iparam.log:0} res = cvxopt.solvers.socp(self.c, self.G0, self.h0, self.G1, self.h1, solver = 'mosek') else: if not self.verbose: cvxopt.solvers.options['show_progress'] = False res = cvxopt.solvers.socp(self.c, self.G0, self.h0, self.G1, self.h1) return res class SOCPS(SOCP): ''' SOCP for sparse constraints converts matrices/numpy arrays into sparse matrices''' def set_component_ineq_constraints(self, A_0, b_0): ''' Set the non-cone constraints 0 <= b_0 - A_0 x ''' assert A_0.shape == (b_0.shape[0], self.num_vars), 'Size mismatch in inequality constraint (lhs)' assert b_0.shape[0] == self.num_vars self.G0 = cs(cm(A_0)) self.h0 = cm(b_0) def add_ineq_constraints(self, A_i, b_i, c_i, d_i): """Add the new inequality constraints to the list""" assert c_i.shape[0] == self.num_vars, 'Size mismatch in inequality constraint (rhs)' assert A_i.shape == (b_i.shape[0], self.num_vars), 'Size mismatch in inequality constraint (lhs)' G = cs(cm(vstack((-c_i.T, -A_i)))) h = cm(vstack((d_i, b_i[:, newaxis]))) self.G1.append(G) self.h1.append(h) OptWok-0.3.1/optwok/testtools.py0000644ڛ|׿-0000000555012055564636023653 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Some functions to help test the modules""" from numpy import log, exp, nonzero from numpy import linspace, square, median from numpy import zeros, eye, sort, ix_ from optwok.kernel import sqr_dist, normalize_unit_diag def plot_confusion(mix_list): """Plot the kernel matrix corresponding to a particular mix_list denoting the mixing of labels.""" from matplotlib.pyplot import matshow, colorbar Ky = list2matrix(mix_list) matshow(Ky) colorbar() def list2matrix(mix_list, noise=0.0, neg_corr=False): """Create the corresponding {0,1} matrix to the mixture list""" k = len(flatten_list(mix_list)) Ky = zeros((k,k)) off_diag = 1.0-noise for block in mix_list: Ky[ix_(block,block)] = off_diag for ix in range(k): Ky[ix,ix] = 1.0 if neg_corr: num_ones = sum(sum(Ky)) neg_val = -num_ones/(k*k) Ky[nonzero(Ky==0)] = neg_val Ky = normalize_unit_diag(Ky+1e-5*eye(k)) return Ky def flatten_list(general_list): """Convert a list of lists into one list. Works recursively.""" if isinstance(general_list, list): return sum((flatten_list(x) for x in general_list), []) else: return [general_list] def pair2vec(context, value): """Convert a pair of (context, value) into a vector with value in element context. Assume that context is in range() """ vec = zeros((max(context)+1, len(context))) for (ix,cur) in enumerate(zip(context, value)): vec[cur[0],ix] = cur[1] return vec def plot_2d(train_ex, train_lab): """Plot 2D data""" from matplotlib.pyplot import plot pos_lab = train_lab > 0.0 neg_lab = train_lab < 0.0 plot(train_ex[0,pos_lab], train_ex[1,pos_lab], 'r+') plot(train_ex[0,neg_lab], train_ex[1,neg_lab], 'bx') def plot_2d_pred(x, y, svmout, train_ex, train_lab, axis_limit, title_text): """Plot 2D data, along with prediction contours.""" from matplotlib.pyplot import figure, pcolor, contour, axis, title figure() pcolor(x, y, svmout) contour(x, y, svmout, linewidths=1, colors='black', hold=True) plot_2d(train_ex, train_lab) axis(axis_limit) title(title_text) def get_gauss_widths(examples, num_widths, start_quantile=0.1, stop_quantile=0.9): """Return a candidate set of possible widths for use in the Gaussian kernel. """ distance = sqr_dist(examples, examples) # get the sorted distances (ignoring distances to self) dm = sort(distance.flatten()) dm = square(dm[examples.shape[1]:]) if num_widths == 1: widths = median(dm) else: # find points uniformly spaces in log scale start_idx = round(start_quantile*len(dm)) stop_idx = round(stop_quantile*len(dm)) dm = log(dm) widths = exp(linspace(dm[stop_idx],dm[start_idx],num_widths)) widths = 1/widths print(widths) return widths OptWok-0.3.1/optwok/utility.py0000644ڛ|׿-0000001057012140420225023270 0ustar cheng.ongUNIMELB\Domain Users00000000000000''' Created on 06.03.2013 @author: Gabriel ''' from __future__ import division import os from numpy import ndarray, array from numpy.random import rand, permutation try: import jsonpickle.handlers from jsonpickle.handlers import BaseHandler from jsonpickle import encode, decode, set_encoder_options except ImportError as imperr: print imperr, ": json dump and load won't work" from types import FunctionType import string import datetime as dt import inspect from os.path import join as pjoin, isfile import numpy dtformat = '%Y-%m-%d_%H-%M-%S' dformat = '%Y-%m-%d' class NumpyFloatHandler(jsonpickle.handlers.BaseHandler): def flatten(self, obj, data): """ Converts and round to float and encode """ return round(obj,12) jsonpickle.handlers.registry.register(numpy.float, NumpyFloatHandler) jsonpickle.handlers.registry.register(numpy.float32, NumpyFloatHandler) jsonpickle.handlers.registry.register(numpy.float64, NumpyFloatHandler) class NumpyHandler(BaseHandler): def flatten(self,obj,data): data['array'] = obj.__repr__() return data def restore(self,obj): return eval(obj['array']) jsonpickle.handlers.registry.register(ndarray,NumpyHandler) set_encoder_options('json',indent=4) def ensure_dir(f): d = os.path.dirname(f) if d <> '': if not os.path.exists(d): os.makedirs(d) #def name_experiment(*args,**kwargs): # ''' Create a string for an experiment ''' # fname = slugify('-'.join(['-'.join(map(strify,args)), '-'.join(map(strify_item,sorted(kwargs.items())))])) # return fname def check_result(path, result_desc): ''' Check if result file exists: path: path to result folder result_desc: string that identifies the experiment result file ''' return isfile(pjoin(path, result_desc)) def jappend(fname): if fname[-5:] <> '.json': fname = fname + '.json' return fname def jexists(fname): fnamej = jappend(fname) return isfile(fnamej) def jdump(fname, obj): ensure_dir(fname) fname = jappend(fname) with open(fname, 'w') as f: f.write(encode(obj)) def jload(fname): fname = jappend(fname) with open(fname, 'r') as f: ftxt = f.read() return(decode(ftxt)) def create_perm(path, dataset, description, n_perm, num_ex): ''' create n_perm permutations for dataset with num_ex examples and save to file ''' p = [] for i in range(n_perm): p += [permutation(num_ex).tolist()] fname = permfile_name(path, dataset, description) jdump(fname, p) def permfile_name(path, dataset, description): return pjoin(path, description + '_' + dataset) def slugify(filename): ''' convert a string into a valid filename ''' valid_chars = "-_.() %s%s" % (string.ascii_letters, string.digits) return ''.join(c for c in filename if c in valid_chars) def strify_item(item): ''' strify dict items ''' return "%s(%s)" %(strify(item[0]),strify(item[1])) def strify(anything): ''' Recursive application of str to anything s.t. it qualifies as a filename ''' if isinstance(anything, list) or isinstance(anything, tuple): return '_'.join(map(strify, anything)) elif isinstance(anything, dict): return '_'.join(map(strify_item, anything.items())) elif isinstance(anything, ndarray): raise TypeError('Can not convert numpy array to string') elif isinstance(anything, FunctionType): if anything.__name__ == '': return str(rand(1)[0]) else: return anything.__name__ elif isinstance(anything, dt.datetime): return anything.strftime(dtformat) elif isinstance(anything, dt.date): return anything.strftime(dformat) elif inspect.isclass(anything): return anything.__name__ else: return str(anything) def _test_jdump(): a = array([[1,2],[3,4]]) obj = {'a': a, 'b': 4.3} objs = encode(obj) print objs objl = decode(objs) print objl jdump('test.txt', obj) print jload('test.txt') def _test_strify(): class FooBar: pass a=[FooBar, [1,2,3]] print FooBar.__name__ print strify(a) if __name__ == '__main__': _test_strify()OptWok-0.3.1/PKG-INFO0000644ڛ|׿-0000000772412140420542020776 0ustar cheng.ongUNIMELB\Domain Users00000000000000Metadata-Version: 1.0 Name: OptWok Version: 0.3.1 Summary: A collection of tools for prototyping optimization methods in machine learning. Home-page: http://www.ong-home.my/optwok.html Author: Cheng Soon Ong Author-email: chengsoon.ong@unimelb.edu.au License: GNU GPL version 3 or later Description: ====== OptWok ====== A collection of python code to perform research in optimization. The aim is to provide reusable components that can be quickly applied to machine learning problems. Dependencies ============ * python 2.5+ * cvxopt 1.0+ (for solving linear and quadratic programs) * pythongrid (for using a cluster) * cython 0.14.1 (for speeding up kernel computations) * numpy, scipy, matplotlib Description =========== The projects currently prototyped: - Ellipsoidal multiple instance learning - difference of convex functions algorithms for sparse classfication - learning the output kernel using block coordinate descent - Gaussian process regression - Contextual bandits upper confidence bound algorithm (using GP) Please look at the examples directory for usage. Ellipsoidal Multiple Instance Learning -------------------------------------- - The code for eMIL is contained in mil.mi_classifier_emil - emil_demo.py gives usage examples - mil.mi_data provides containers for multiple instance learning data Difference of convex functions algorithms for sparse classfication ------------------------------------------------------------------ - The workhorse is classifier.py, which implements classifiers for binary classification in the primal. The optimization methods use the framework empirical loss + regularizer. - classifier_demo.py gives two usage examples. - poly_feature.py computes the polynomial kernel features explicitly. - experiment_poly.py is an experimental framework using polynomial kernels. Learning the output kernel using block coordinate descent --------------------------------------------------------- - The workhorse is kernelopt.py, which implements a framework for minimizing the invex function for learning the kernel on outputs. - multiclass_demo.py gives usage examples. Utilities --------- - sip.py is a wrapper around cvxopt to allow addition of constraints. - evaluation.py calculates the accuracy - mldata.py defines a machine learning dataset and their files - expttools.py defines utilities to flatten nested for loops. - io_pickle.py for loading and saving objects. - testtools.py is some code to ease the demos. - simdata.py contains code to generate toy datasets. - kernel.py implements code for joint kernels - kernelbase.pyx speeds up kernel.py with cython Contributors ============ Cheng Soon Ong Gabriel Krummenacher License ======= OptWok is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. OptWok is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with OptWok. If not, see . Platform: UNKNOWN OptWok-0.3.1/README0000644ڛ|׿-0000000570512140417273020565 0ustar cheng.ongUNIMELB\Domain Users00000000000000====== OptWok ====== A collection of python code to perform research in optimization. The aim is to provide reusable components that can be quickly applied to machine learning problems. Dependencies ============ * python 2.5+ * cvxopt 1.0+ (for solving linear and quadratic programs) * pythongrid (for using a cluster) * cython 0.14.1 (for speeding up kernel computations) * numpy, scipy, matplotlib Description =========== The projects currently prototyped: - Ellipsoidal multiple instance learning - difference of convex functions algorithms for sparse classfication - learning the output kernel using block coordinate descent - Gaussian process regression - Contextual bandits upper confidence bound algorithm (using GP) Please look at the examples directory for usage. Ellipsoidal Multiple Instance Learning -------------------------------------- - The code for eMIL is contained in mil.mi_classifier_emil - emil_demo.py gives usage examples - mil.mi_data provides containers for multiple instance learning data Difference of convex functions algorithms for sparse classfication ------------------------------------------------------------------ - The workhorse is classifier.py, which implements classifiers for binary classification in the primal. The optimization methods use the framework empirical loss + regularizer. - classifier_demo.py gives two usage examples. - poly_feature.py computes the polynomial kernel features explicitly. - experiment_poly.py is an experimental framework using polynomial kernels. Learning the output kernel using block coordinate descent --------------------------------------------------------- - The workhorse is kernelopt.py, which implements a framework for minimizing the invex function for learning the kernel on outputs. - multiclass_demo.py gives usage examples. Utilities --------- - sip.py is a wrapper around cvxopt to allow addition of constraints. - evaluation.py calculates the accuracy - mldata.py defines a machine learning dataset and their files - expttools.py defines utilities to flatten nested for loops. - io_pickle.py for loading and saving objects. - testtools.py is some code to ease the demos. - simdata.py contains code to generate toy datasets. - kernel.py implements code for joint kernels - kernelbase.pyx speeds up kernel.py with cython Contributors ============ Cheng Soon Ong Gabriel Krummenacher License ======= OptWok is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. OptWok is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with OptWok. If not, see . OptWok-0.3.1/run_demo.sh0000755ڛ|׿-0000000173012140417274022047 0ustar cheng.ongUNIMELB\Domain Users00000000000000#!/bin/bash echo '===== Classifier Demo =====' python examples/classifier_demo.py echo '===== Context UCB Demo =====' python examples/contextucb_demo.py rm expt_cucb.pkl expt_cucb_cat.pkl expt_cucb_kbest.pkl rm expt_cucb_batch.pkl expt_cucb_ind.pkl rm expt_cucb_avg.pdf expt_cucb_cat_min.pdf expt_cucb_min.pdf rm expt_cucb_batch_avg.pdf expt_cucb_ind_min.pdf rm expt_cucb_batch_min.pdf expt_cucb_kbest_min.pdf echo '===== Ellipsoidal MIL Demo =====' python examples/emil_demo.py rm example_synthetic_ellipsoid_2d.json echo '===== Gaussian Process Regression Demo =====' python examples/gaussianproc_demo.py echo '===== Learning Output Kernel Demo =====' python examples/kernelopt_demo.py rm testKy.pdf testLin.pdf echo '===== Multiclass Demo =====' python examples/multiclass_demo.py echo '===== Named Array Demo =====' python examples/named_array_demo.py echo '===== Score Demo =====' python examples/score_demo.py echo '===== SOCP Demo =====' python examples/socp_demo.py OptWok-0.3.1/setup.py0000644ڛ|׿-0000000404012140420515021377 0ustar cheng.ongUNIMELB\Domain Users00000000000000#!/usr/bin/env python import os import numpy # Use numpy version of distutils to get f2py #from numpy.distutils.core import setup, Extension from distutils.core import setup, Extension from distutils.command.sdist import sdist from glob import glob try: #from Cython.Build import cythonize from Cython.Distutils import build_ext have_cython = True except ImportError: have_cython = False opt_modules = [] if have_cython: # cython code is used to speed up kronecker product kernel kron = Extension('optwok.kernelbase', sources = ['optwok/kernelbase.pyx'], include_dirs = [numpy.get_include(),'.'], ) opt_modules.append(kron) class sdist_hg(sdist): """Append Mercurial repository version if 'dev' option is used with sdist""" user_options = sdist.user_options + [ ('dev', None, "Add a dev marker") ] def initialize_options(self): sdist.initialize_options(self) self.dev = 0 def run(self): if self.dev: suffix = '.dev%d' % self.get_tip_revision() self.distribution.metadata.version += suffix sdist.run(self) def get_tip_revision(self, path=os.path.expanduser('~')+'/Mercurial/OptWok'): from mercurial.hg import repository from mercurial.ui import ui from mercurial import node repo = repository(ui(), path) tip = repo.changelog.tip() return repo.changelog.rev(tip) setup(name = 'OptWok', author = 'Cheng Soon Ong', author_email = 'chengsoon.ong@unimelb.edu.au', version = '0.3.1', description = 'A collection of tools for prototyping optimization methods in machine learning.', long_description = open('README').read(), url = 'http://www.ong-home.my/optwok.html', packages = ['optwok', 'optwok.mil'], cmdclass = {'sdist': sdist_hg, 'build_ext': build_ext}, #ext_modules = cythonize(opt_modules), ext_modules = opt_modules, license = 'GNU GPL version 3 or later', ) OptWok-0.3.1/tests/0000755ڛ|׿-0000000000012140420542021031 5ustar cheng.ongUNIMELB\Domain Users00000000000000OptWok-0.3.1/tests/__init__.py0000644ڛ|׿-0000000000011731540564023144 0ustar cheng.ongUNIMELB\Domain Users00000000000000OptWok-0.3.1/tests/all_tests.py0000644ڛ|׿-0000000057211731540564023415 0ustar cheng.ongUNIMELB\Domain Users00000000000000""" Run all tests """ import glob import unittest def create_test_suite(): test_file_strings = glob.glob('optwok/test/*_test.py') module_strings = ['optwok.test.'+str[12:len(str)-3] for str in test_file_strings] suites = [unittest.defaultTestLoader.loadTestsFromName(name) for name in module_strings] testSuite = unittest.TestSuite(suites) return testSuite OptWok-0.3.1/tests/classifier_test.py0000644ڛ|׿-0000000037111731540564024603 0ustar cheng.ongUNIMELB\Domain Users00000000000000import unittest from examples.classifier_demo import example_toy, example_cloud_2d class Demo(unittest.TestCase): """Run the demo for classifier""" def test_demo(self): example_toy() if __name__ == '__main__': unittest.main() OptWok-0.3.1/tests/gaussproc_06_20_0.1.pkl0000644ڛ|׿-0000002010011731540564024740 0ustar cheng.ongUNIMELB\Domain Users00000000000000BZh91AY&SY$ItT5D OAOiOMhނ Ty zS5=0A2$اzGQz0$Sez=SyM M) d:4T!0&2d`#Ʉ 0L(23Iz#y@1=43QLzhhL0&120F`#L&&hɀLJidɈIz55=14MO?Dh(5=4MMhɦ=Gئz55<zj1# d$qB|G孑V^;E_A"k^hR2k.rr~/V쭨1+ڈ,(ҫ@$D<?t>^e$9o$LʑG%f9L 9ky|/!T.[TCB<ě@(Jƿ0$9xY{bf%CfGmjxK=|b]zgwca"dB{7UNXoeFm y<G6E㋶1㥑uǐRKq.i6l1îDymM"Wg,"ٞ>c?#jy&NtS09ȪèNɼ-4 <# M}J@P;Z~+>VU&bT*\fsBШv7'Rvq~08Qő|;ŜXΩqdf#jY赪UT#ky)1!b saoFȇ*]Ֆ rgugK6\E d9 $Eจi3?ڍՉL.$M}f|Z{pع]֮)8e.#$9;}n{ (7!(^!;xzM3r\N'RܮH!Ӿs4EC.^#;?5R؁+1bLX}mWxXn7# i& ^X/>7.y卽ʘQћ""Q^A+qvN#vG)^/R9XHB,`*r򋑙(7^FFyuتJgőZp5ge8Bx9a0>tlg #s% 5EﲚQ Zh ē8Hwk>3m6we[$T-*9zhA^Hx̍Lv3K5C^)moAE֩ #+@|}9(+8 M`3iKLݱr-n`,NZ&o2;QKaڕMN䩨d–7W\d4rw 3aəÃCNSiRmFZV;C1Zy%sNjnanW8u:kY,n+0ீs@Q4NA#qے3?, JJl#ό>gpOʮΤ-ul%IO8JN,%1l0#[ۋ}Fx{U9G3>^e!"=VJƣ$| tc27rs_mXc̲ýHNkl<7cڧd"!vG3ofe) 5EkuRcBi"%,HuI."kߠd8ܺ]H`.Ǝ_/Tg\Df9EOR ëg>a4 FHh|a{-E9w\6ܜ5WBeVvɕPTF"t胀*PrDg߮bЮ8RqdSD)Yj&˃;ru1CfsRګ(︧` Z:)1ØR6DKֱIؒǿB=T"!%V]<^p%8525JPM_ -. gKO!cΡ;c,LJoYU' F|]HM|O;jb%CLKfs"A#ǺMmHZ F;Hu.Nv%ҨyJm[-:+U;#6eEf:hay0t E]ះ([%$U7nra!"c$Ť89c2؏uE{x^Z{(nbe\c TY{o~!Wry&y!HXme-Yk%L殮$l!5 y|)sufzgGon0p>Ñ'̚ϛEbhn ZNp8jr -p1WjGvYqzg&&Ԡ xQ4pݜkAr1QIky'CRq\R&91+(7yrFP:#L<&R=>}&)*1\'M!q!.hC`l\W8ze:ԙgqw#QUf{MB!T*]t(@I5!&B ϐB/<{m_:Y vUk 쁁-jfC>PbfA(1>*X@5>I J*CB[Ln[f%iqeBC$SF mV]eaxZ`=[xG2m0g!}]]~K&r'"8nɿBDӸdS ZqY ,²#ӆXY^ AkC>ׅg=uszv#>G0GJFР~ֺOR V&t |Mv O76)95@bxd!`XeGË麅$-M5DU93lVd*,q4q٦O 45HlK q%ȗhq@z1,M5*LMDe*]}ԥ %?[њU-]T;*f?h`7w[.y %s4~m2!9pq&B> F@CM-qJhF0p܂Q(|i2]5Tp[E<WFucщ=٬̹WD{RMpߠ#LbZi vZ+-Ov<ܷרsq+H/MS~ͳ/Ӑ6NF;"}Az;ҐG&*u?i\Zwig;?BE~n 뤓ǖ3|;C (B3>|m x`>ir_[1$bu>GugTvZ3ygN쟰<.v(o%(1٫+'5CebY7io%V8,]º—@6J{=jv1,$79Q!w{?ܭx ['dqϵ8~ԸDjtkg0e(\+Z/GY"OwQF|ۿ%.&WȸP_¼Dc\gKrl6Y#|,?gswI!$Gf@I# WriW}Mu9 _̬ܳ (Nw`p1]ĐZ: %$$I I$- I,$Un"'? "j/@JF-WM @cϼ S~!j`/0gpwj= J;$1XbZKvg9@/`b*:?F_q$7 IYE6:rI. Azǯ&~ @g$ I sbz܈眻"諉BSGzAfۛskz&褈{wfW+%wVI^U wq*D';ux7 z1^gz(P&dU72m82Hm. W[bh|x@I+8$d< OQ W &0l?O )lh& Z&a7SJ;-A|8 ,Wg2Ij,h u$A"(>snۙ'|~Di=Љ쐿*EǪ.A=3 sؼ Skˋ"87Ĺ i# UfM'eStx@}?SfybvN}+Jl> ^ οZ XG/ HZkROZ cΔR1'~x))fQRZ|)PiRT^Ku@cq<:C/o?K\o5P`h4zM${)@9jvoZC#pzOW$HpiŎ4yq4N"eb+;UH w1rx8%#(.W: [!;ndgN0@v{{3պ$Hdo W:Z Ramf_JcFE,1%2-Ш{d0YkEIM h3uHgc"^@e=Yuψ>"[!.sMg +<'aӛ .hͷ8$ GBK-?ǂk c\cax}IA.L5 + 4qݤ.'$ O.oK}rC*B8^%G2H)Hfq a:@-4wMk64Be$S$LbO绘 O_WookH'Y#)QUOR`,+:e!zW> VNfDŽKb7G*s߳!VgFwԝݞwD ?ACw:O' yTEvbQ:c"y] qRȇCl#OPJ" Qj-Im`W(O\]1u#::r~'n `DD}1GGWAăy+_(q/f6D"I$d7dTK-8 ixOR,čUivaOb} QFvlop]}lD7W Zǣ`Iۈ2צg;1~~x5Z؂K k[MJUQ0Ȕ6v%1>F^{%|6i'islk6[FRyO</JA须jÝy:?T%KZ ~޸*f4kϽpR-s2lbnۋHi0įԈyHhR`1XceAE!ٸxf JL^覆Tsr?_/1Y QӠ|N?jV˔'[  ȫBmz*Ii{kP]"9HD.VxóingH6þk +;yRcz0nq'2́oݻ0 LCl~_5\r]9BD.@1ć(KfX #W n5Я8/2vn{{\!Mjk&21-t{R7pd+e>ZeZK ԰Qæo/` +Ԙ_?QذKgu8ۗ";4k13yd;*#:? `v'~_[AAUa\5lM{SXKšGv4d}>Эُl>8-- +dνt{%N ]F-}!ynsyMyDQnOΑ&T>itB#GrE8P$OptWok-0.3.1/tests/import_test.py0000644ڛ|׿-0000000073111731540564023771 0ustar cheng.ongUNIMELB\Domain Users00000000000000"""Checking whether all the external libraries are installed""" try: import cvxopt print cvxopt.__file__ except ImportError: print 'cvxopt not found' try: import pythongrid print pythongrid.__file__ except ImportError: print 'pythongrid not found' try: import cython print cython.__file__ except ImportError: print 'cython not found' try: import slycot print slycot.__file__ except ImportError: print 'slycot not found' OptWok-0.3.1/tests/kernel_test.py0000644ڛ|׿-0000001126611731540564023744 0ustar cheng.ongUNIMELB\Domain Users00000000000000# cython: profile=True """Test the kernel module, including speedups due to cython""" import pstats, cProfile from numpy import sum from numpy.random import rand, randint, seed from optwok.testtools import list2matrix def test_kernel(kfunc, X): print '=== Test Kernel ===' kg = kfunc() print kg kg.compute(X) print kg.kmat for ix in range(X.shape[1]): for iy in range(X.shape[1]): print '(%d,%d) %f' % (ix, iy, kg[ix,iy]) def test_joint_kernel(kfunc, X): print '=== Test Joint Kernel ===' joint = JointKernel(kfunc(), num_classes=2) print joint print joint.Kx joint.compute(X) print joint.kmat for xi in range(X.shape[1]): for yi in range(2): for xj in range(X.shape[1]): for yj in range(2): print 'Kx(%d,%d),Ky(%d,%d): %f' % (xi,xj,yi,yj, joint[xi,xj,yi,yj]) def test_prediction(kfunc, Xtrain, Xtest): print '=== Test prediction ===' kg = kfunc() print kg kg.compute(Xtrain) print kg.compute_test(Xtest) joint = JointKernel(kfunc(), num_classes=2) joint.compute(Xtrain) print joint.compute_test(Xtest) def demo_context_kernel(kfunc, X, Z): print '=== Test Context Kernel ===' from optwok.testtools import list2matrix ck = ContextKernel(kfunc(), list2matrix([[0],[1,2]], noise=0.2)) print ck print ck.Kx ck.compute(X,Z) print ck.kmat def _test_linear_kernel(kernel): X1 = rand(45,5000) X2 = rand(45,12000) kernel.compute(X1) kernel.kmat kernel.compute_test(X1) kernel.compute_test(X2) def test_linear_kernel(): from optwok.kernelbase import LinearKernel kernel = LinearKernel() _test_linear_kernel(kernel) def test_linear_kernel_pp(): from optwok.kernel_pp import LinearKernel kernel = LinearKernel() _test_linear_kernel(kernel) def profile_linear_kernel(): cProfile.runctx('test_linear_kernel_pp()', globals(), locals(), 'kernel_pp.prof') cProfile.runctx('test_linear_kernel()', globals(), locals(), 'kernelbase.prof') s = pstats.Stats('kernel_pp.prof') s.strip_dirs().sort_stats('cumulative').print_stats(10) f = pstats.Stats('kernelbase.prof') f.strip_dirs().sort_stats('cumulative').print_stats(10) def _test_context_kernel(kernel): num_feat = 45 num1 = 1000 num2 = 2000 seed(0) X1 = rand(num_feat, num1) c1 = randint(3, size=num1) X2 = rand(num_feat, num2) c2 = randint(3, size=num2) kernel.compute(X1, c1) kernel.kmat kernel.compute_test(X1, c1) kernel.compute_test(X2, c2) def test_context_kernel(): from optwok.kernelbase import LinearKernel, ContextKernel ck = ContextKernel(LinearKernel(), list2matrix([[0],[1,2]], noise=0.1)) _test_context_kernel(ck) def test_context_kernel_pp(): from optwok.kernel_pp import LinearKernel, ContextKernel ck = ContextKernel(LinearKernel(), list2matrix([[0],[1,2]], noise=0.1)) _test_context_kernel(ck) def profile_context_kernel(): cProfile.runctx('test_context_kernel_pp()', globals(), locals(), 'kernel_pp.prof') cProfile.runctx('test_context_kernel()', globals(), locals(), 'kernelbase.prof') s = pstats.Stats('kernel_pp.prof') s.strip_dirs().sort_stats('cumulative').print_stats(10) f = pstats.Stats('kernelbase.prof') f.strip_dirs().sort_stats('cumulative').print_stats(10) def test_all(): """Execute all the functions""" X = rand(2,3) print X demo_context_kernel(LinearKernel, X, [0,1,2]) test_kernel(LinearKernel, X) test_kernel(GaussKernel, X) test_joint_kernel(LinearKernel, X) test_joint_kernel(GaussKernel, X) Xtest = rand(2,1) test_prediction(LinearKernel, X, Xtest) test_prediction(GaussKernel, X, Xtest) if __name__ == '__main__': from optwok.kernel_pp import LinearKernel, ContextKernel kernel = ContextKernel(LinearKernel(), list2matrix([[0],[1,2]], noise=0.9)) num_feat = 45 num1 = 1000 num2 = 2000 seed(0) X1 = rand(num_feat, num1) c1 = randint(2, size=num1) X2 = rand(num_feat, num2) c2 = randint(3, size=num2) kernel.compute(X1, c1) print sum(abs(kernel.kmat - kernel.kmat_forloop)) print sum(abs(kernel.compute_test(X2,c2) - kernel.compute_test_forloop(X2,c2))) cProfile.runctx('kernel.kmat;kernel.compute_test(X2,c2)', globals(), locals(), 'kmat.prof') cProfile.runctx('kernel.kmat_forloop;kernel.compute_test_forloop(X2,c2)', globals(), locals(), 'kmat_forloop.prof') s = pstats.Stats('kmat.prof') s.strip_dirs().sort_stats('cumulative').print_stats(10) f = pstats.Stats('kmat_forloop.prof') f.strip_dirs().sort_stats('cumulative').print_stats(10) OptWok-0.3.1/tests/multiclass_test.py0000644ڛ|׿-0000000062311731540564024637 0ustar cheng.ongUNIMELB\Domain Users00000000000000import unittest from examples.multiclass_demo import demo_identity_Ky, demo_merged_classes, demo_split_classes class Demo(unittest.TestCase): """Run the demo for multiclass SVM""" def test_identity(self): demo_identity_Ky() def test_merged(self): demo_merged_classes() def test_split(self): demo_split_classes() if __name__ == '__main__': unittest.main() OptWok-0.3.1/tests/run_tests.py0000644ڛ|׿-0000000031711731540564023446 0ustar cheng.ongUNIMELB\Domain Users00000000000000""" Use unittest to run all tests in optwok/test """ import unittest import optwok.test.all_tests testSuite = optwok.test.all_tests.create_test_suite() text_runner = unittest.TextTestRunner().run(testSuite)