From 5bfd4e835b995e75bebaf2d7db6fa08c91a85c64 Mon Sep 17 00:00:00 2001 From: Gently Date: Thu, 24 Mar 2022 14:00:10 +0000 Subject: [PATCH 1/4] =?UTF-8?q?=E6=96=B0=E5=BB=BA=20DeepSort=5FID0505=5Ffo?= =?UTF-8?q?r=5FTensorFlow?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/.keep | 0 1 file changed, 0 insertions(+), 0 deletions(-) create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/.keep diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/.keep b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/.keep new file mode 100644 index 000000000..e69de29bb -- Gitee From 35f78a063b400f75a44dcfbd478aab93b7d71a34 Mon Sep 17 00:00:00 2001 From: Gently Date: Thu, 24 Mar 2022 14:01:19 +0000 Subject: [PATCH 2/4] add contrib/TensorFlow/Research/cv/ascend/ID0505 YOLACT. --- .../DeepSort_ID0505_for_TensorFlow/LICENSE | 674 +++++++++++++++++ .../DeepSort_ID0505_for_TensorFlow/README.md | 165 +++++ .../boot_modelarts.py | 72 ++ .../DeepSort_ID0505_for_TensorFlow/cfg.py | 46 ++ .../datasets/__init__.py | 1 + .../datasets/market1501.py | 214 ++++++ .../datasets/mars.py | 190 +++++ .../datasets/util.py | 173 +++++ .../help_modelarts.py | 80 ++ .../DeepSort_ID0505_for_TensorFlow/losses.py | 145 ++++ .../DeepSort_ID0505_for_TensorFlow/metrics.py | 223 ++++++ .../nets/__init__.py | 0 .../queued_trainer.py | 616 ++++++++++++++++ .../DeepSort_ID0505_for_TensorFlow/run_1p.sh | 25 + .../scripts/run_1p.sh | 25 + .../scripts/run_apulis.sh | 30 + .../train_app.py | 690 ++++++++++++++++++ .../train_market1501.py | 139 ++++ .../train_mars.py | 124 ++++ .../vis_tools.py | 82 +++ 20 files changed, 3714 insertions(+) create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/LICENSE create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/README.md create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/boot_modelarts.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/cfg.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/__init__.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/market1501.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/mars.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/util.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/help_modelarts.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/losses.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/metrics.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/__init__.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/queued_trainer.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/run_1p.sh create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/scripts/run_1p.sh create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/scripts/run_apulis.sh create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/train_app.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/train_market1501.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/train_mars.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/vis_tools.py diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/LICENSE b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/LICENSE new file mode 100644 index 000000000..94a9ed024 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/LICENSE @@ -0,0 +1,674 @@ + 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 +. diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/README.md b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/README.md new file mode 100644 index 000000000..a4923fa2c --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/README.md @@ -0,0 +1,165 @@ +# cosine_metric_learning + +## Introduction + +This repository contains code for training a metric feature representation to be +used with the [deep_sort tracker](https://github.com/nwojke/deep_sort). The +approach is described in + + @inproceedings{Wojke2018deep, + title={Deep Cosine Metric Learning for Person Re-identification}, + author={Wojke, Nicolai and Bewley, Alex}, + booktitle={2018 IEEE Winter Conference on Applications of Computer Vision (WACV)}, + year={2018}, + pages={748--756}, + organization={IEEE}, + doi={10.1109/WACV.2018.00087} + } + +Pre-trained models used in the paper can be found +[here](https://drive.google.com/open?id=13HtkxD6ggcrGJLWaUcqgXl2UO6-p4PK0). +A preprint of the paper is available [here](http://elib.dlr.de/116408/). +The repository comes with code to train a model on the +[Market1501](http://www.liangzheng.org/Project/project_reid.html) +and [MARS](http://www.liangzheng.com.cn/Project/project_mars.html) datasets. + +## Training on Market1501 + +The following description assumes you have downloaded the Market1501 dataset to +``./Market-1501-v15.09.15``. The following command starts training +using the cosine-softmax classifier described in the above paper: +``` +python train_market1501.py \ + --dataset_dir=./Market-1501-v15.09.15/ \ + --loss_mode=cosine-softmax \ + --log_dir=./output/market1501/ \ + --run_id=cosine-softmax +``` +This will create a directory `./output/market1501/cosine-softmax` where +TensorFlow checkpoints are stored and which can be monitored using +``tensorboard``: +``` +tensorboard --logdir ./output/market1501/cosine-softmax --port 6006 +``` +The code splits off 10% of the training data for validation. +Concurrently to training, run the following command to run CMC evaluation +metrics on the validation set: +``` +CUDA_VISIBLE_DEVICES="" python train_market1501.py \ + --mode=eval \ + --dataset_dir=./Market-1501-v15.09.15/ \ + --loss_mode=cosine-softmax \ + --log_dir=./output/market1501/ \ + --run_id=cosine-softmax \ + --eval_log_dir=./eval_output/market1501 +``` +The command will block indefinitely to monitor the training directory for saved +checkpoints and each stored checkpoint in the training directory is evaluated on +the validation set. The results of this evaluation are stored in +``./eval_output/market1501/cosine-softmax`` to be monitored using +``tensorboard``: +``` +tensorboard --logdir ./eval_output/market1501/cosine-softmax --port 6007 +``` + +## Training on MARS + +To train on MARS, download the +[evaluation software](https://github.com/liangzheng06/MARS-evaluation) and +extract ``bbox_train.zip`` and ``bbox_test.zip`` from the +[dataset website](http://www.liangzheng.com.cn/Project/project_mars.html) +into the evaluation software directory. The following description assumes they +are stored in ``./MARS-evaluation-master/bbox_train`` and +``./MARS-evaluation-master/bbox_test``. Training can be started with the following +command: +``` +python train_mars.py \ + --dataset_dir=./MARS-evaluation-master \ + --loss_mode=cosine-softmax \ + --log_dir=./output/mars/ \ + --run_id=cosine-softmax +``` +Again, this will create a directory `./output/mars/cosine-softmax` where +TensorFlow checkpoints are stored and which can be monitored using +``tensorboard``: +``` +tensorboard --logdir ./output/mars/cosine-softmax --port 7006 +``` +As for Market1501, 10% of the training data are split off for validation. +Concurrently to training, run the following command to run CMC evaluation +metrics on the validation set: +``` +CUDA_VISIBLE_DEVICES="" python train_mars.py \ + --mode=eval \ + --dataset_dir=./MARS-evaluation-master/ \ + --loss_mode=cosine-softmax \ + --log_dir=./output/mars/ \ + --run_id=cosine-softmax \ + --eval_log_dir=./eval_output/mars +``` +Evaluation metrics on the validation set can be monitored with ``tensorboard`` +``` +tensorboard --logdir ./eval_output/mars/cosine-softmax +``` + +## Testing + +Final model testing has been carried out using evaluation software provided by +the dataset authors. The training scripts can be used to write features of the +test split. The following command exports MARS test features to +``./MARS-evaluation-master/feat_test.mat`` +``` +python train_mars.py \ + --mode=export \ + --dataset_dir=./MARS-evaluation-master \ + --loss_mode=cosine-softmax .\ + --restore_path=PATH_TO_CHECKPOINT +``` +where ``PATH_TO_CHECKPOINT`` the checkpoint file to evaluate. Note that the +evaluation script needs minor adjustments to apply the cosine similarity metric. +More precisely, change the feature computation in +``utils/process_box_features.m`` to average pooling (line 8) and apply +a re-normalization at the end of the file. The modified file should look like +this: +``` +function video_feat = process_box_feat(box_feat, video_info) + +nVideo = size(video_info, 1); +video_feat = zeros(size(box_feat, 1), nVideo); +for n = 1:nVideo + feature_set = box_feat(:, video_info(n, 1):video_info(n, 2)); +% video_feat(:, n) = max(feature_set, [], 2); % max pooling + video_feat(:, n) = mean(feature_set, 2); % avg pooling +end + +%%% normalize train and test features +sum_val = sqrt(sum(video_feat.^2)); +for n = 1:size(video_feat, 1) + video_feat(n, :) = video_feat(n, :)./sum_val; +end +``` +The Market1501 script contains a similar export functionality which can be +applied in the same way as described for MARS: +``` +python train_market1501.py \ + --mode=export \ + --dataset_dir=./Market-1501-v15.09.15/ + --sdk_dir=./Market-1501_baseline-v16.01.14/ + --loss_mode=cosine-softmax \ + --restore_path=PATH_TO_CHECKPOINT +``` +This command creates ``./Market-1501_baseline-v16.01.14/feat_query.mat`` and +``./Market-1501_baseline-v16.01.14/feat_test.mat`` to be used with the +Market1501 evaluation code. + +## Model export + +To export your trained model for use with the +[deep_sort tracker](https://github.com/nwojke/deep_sort), run the following +command: +``` +python train_mars.py --mode=freeze --restore_path=PATH_TO_CHECKPOINT +``` +This will create a ``mars.pb`` file which can be supplied to Deep SORT. Again, +the Market1501 script contains a similar function. +##kernel_meta&ckpt&resources （three missed doc） \ No newline at end of file diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/boot_modelarts.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/boot_modelarts.py new file mode 100644 index 000000000..f8e0663c4 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/boot_modelarts.py @@ -0,0 +1,72 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +This is the boot file for ModelArts platform. +Firstly, the train datasets are copyed from obs to ModelArts. +Then, the string of train shell command is concated and using 'os.system()' to execute +""" +import os +import numpy as np +import argparse +from help_modelarts import obs_data2modelarts +from cfg import make_config +print(os.system('env')) + + +import argparse +import moxing as mox +# 解析输入参数data_url + + + +if __name__ == '__main__': + ## Note: the code dir is not the same as work dir on ModelArts Platform!!! + code_dir = os.path.dirname(__file__) + work_dir = os.getcwd() + print("===>>>code_dir:{}, work_dir:{}".format(code_dir, work_dir)) + + parser = argparse.ArgumentParser() + parser.add_argument("--data_url", type=str, default="./dataset") + parser.add_argument("--train_url", type=str, default="./output") + config = parser.parse_args() + # 在ModelArts容器创建数据存放目录 + data_dir = "/cache/dataset" + model_dir = "/cache/result" + os.makedirs(data_dir) + os.makedirs(model_dir) + # OBS数据拷贝到ModelArts容器内 + mox.file.copy_parallel(config.data_url, data_dir) + mox.file.copy_parallel(model_dir, config.train_url) + #parser.add_argument("--train_url", type=str, default="s3://boom/Market-1501-v15.09.15/") + #parser.add_argument("--data_url", type=str, default="s3://boom/Market-1501-v15.09.15/") + #parser.add_argument("--modelarts_data_dir", type=str, default="/cache/Market-1501-v15.09.15") + #parser.add_argument("--modelarts_result_dir", type=str, default="./output") + + + print("--------config----------") + for k in list(vars(config).keys()): + print("key:{}: value:{}".format(k, vars(config)[k])) + print("--------config----------") + + ## copy dataset from obs to modelarts + obs_data2modelarts(config) + + ## start to train on Modelarts platform + if not os.path.exists(config.modelarts_result_dir): + os.makedirs(config.modelarts_result_dir) + bash_header = os.path.join(code_dir, 'scripts/run_1p.sh') + arg_url = '%s %s %s %s' % (code_dir, config.modelarts_data_dir, config.modelarts_result_dir) + bash_command = 'bash %s %s' % (bash_header, arg_url) + print("bash command:", bash_command) + os.system(bash_command) \ No newline at end of file diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/cfg.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/cfg.py new file mode 100644 index 000000000..a07258f7a --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/cfg.py @@ -0,0 +1,46 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Here, define the configuration of tensorflow session +For different chips, the config is not the same. +""" +import tensorflow as tf +import os + + +def make_config(FLAGS): + chip = FLAGS.chip.lower() + tf.logging.info("chip is [%s]", chip) + + if chip == 'cpu': + config = tf.ConfigProto() + elif chip == 'gpu': + config = tf.ConfigProto(allow_soft_placement=True) + config.gpu_options.allow_growth = True + elif chip == 'npu': + from tensorflow.core.protobuf.rewriter_config_pb2 import RewriterConfig + + config = tf.ConfigProto() + custom_op = config.graph_options.rewrite_options.custom_optimizers.add() + custom_op.name = "NpuOptimizer" + config.graph_options.rewrite_options.remapping = RewriterConfig.OFF + config.graph_options.rewrite_options.memory_optimization = RewriterConfig.OFF # 必须显式关闭 + ## Performance Profiling + ## refer to link:https://support.huaweicloud.com/Development-tg-cann202training1/atlasprofilingtrain_16_0003.html + + + else: + raise RuntimeError('chip [%s] has not supported' % chip) + + return config \ No newline at end of file diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/__init__.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/__init__.py new file mode 100644 index 000000000..43e08fb8a --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/__init__.py @@ -0,0 +1 @@ +# vim: expandtab:ts=4:sw=4 diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/market1501.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/market1501.py new file mode 100644 index 000000000..65b31a438 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/market1501.py @@ -0,0 +1,214 @@ +# vim: expandtab:ts=4:sw=4 +from npu_bridge.npu_init import * +import os +import numpy as np +import cv2 +import scipy.io as sio + + +# The maximum person ID in the dataset. +MAX_LABEL = 1501 + +IMAGE_SHAPE = 128, 64, 3 + + +def _parse_filename(filename): + """Parse meta-information from given filename. + + Parameters + ---------- + filename : str + A Market 1501 image filename. + + Returns + ------- + (int, int, str, str) | NoneType + Returns a tuple with the following entries: + + * Unique ID of the individual in the image + * Index of the camera which has observed the individual + * Filename without extension + * File extension + + Returns None if the given filename is not a valid filename. + + """ + filename_base, ext = os.path.splitext(filename) + if '.' in filename_base: + # Some images have double filename extensions. + filename_base, ext = os.path.splitext(filename_base) + if ext != ".jpg": + return None + person_id, cam_seq, frame_idx, detection_idx = filename_base.split('_') + return int(person_id), int(cam_seq[1]), filename_base, ext + + +def read_train_split_to_str(dataset_dir): + """Read training data to list of filenames. + + Parameters + ---------- + dataset_dir : str + Path to the Market 1501 dataset directory. + + Returns + ------- + (List[str], List[int], List[int]) + Returns a tuple with the following values: + + * List of image filenames (full path to image files). + * List of unique IDs for the individuals in the images. + * List of camera indices. + + """ + filenames, ids, camera_indices = [], [], [] + + image_dir = os.path.join(dataset_dir, "bounding_box_train") + for filename in sorted(os.listdir(image_dir)): + meta_data = _parse_filename(filename) + if meta_data is None: + # This is not a valid filename (e.g., Thumbs.db). + continue + + filenames.append(os.path.join(image_dir, filename)) + ids.append(meta_data[0]) + camera_indices.append(meta_data[1]) + + return filenames, ids, camera_indices + + +def read_train_split_to_image(dataset_dir): + """Read training images to memory. This consumes a lot of memory. + + Parameters + ---------- + dataset_dir : str + Path to the Market 1501 dataset directory. + + Returns + ------- + (ndarray, ndarray, ndarray) + Returns a tuple with the following values: + + * Tensor of images in BGR color space of shape 128x64x3. + * One dimensional array of unique IDs for the individuals in the images. + * One dimensional array of camera indices. + + """ + filenames, ids, camera_indices = read_train_split_to_str(dataset_dir) + + images = np.zeros((len(filenames), 128, 64, 3), np.uint8) + for i, filename in enumerate(filenames): + images[i] = cv2.imread(filename, cv2.IMREAD_COLOR) + + ids = np.asarray(ids, np.int64) + camera_indices = np.asarray(camera_indices, np.int64) + return images, ids, camera_indices + + +def read_test_split_to_str(dataset_dir): + """Read query and gallery data to list of filenames. + + Parameters + ---------- + dataset_dir : str + Path to the Market 1501 dataset directory. + + Returns + ------- + (List[str], List[int], List[str], List[int], ndarray) + Returns a tuple with the following values: + + * List of N gallery filenames (full path to image files). + * List of N unique IDs for the individuals in the gallery. + * List of M query filenames (full path to image files). + * List of M unique IDs for the individuals in the queries. + * Matrix of shape MxN such that element (i, j) evaluates to 0 if + gallery image j should be excluded from metrics computation of + query i and 1 otherwise. + + """ + # Read gallery. + gallery_filenames, gallery_ids = [], [] + + image_dir = os.path.join(dataset_dir, "bounding_box_test") + for filename in sorted(os.listdir(image_dir)): + meta_data = _parse_filename(filename) + if meta_data is None: + # This is not a valid filename (e.g., Thumbs.db). + continue + + gallery_filenames.append(os.path.join(image_dir, filename)) + gallery_ids.append(meta_data[0]) + + # Read queries. + query_filenames, query_ids, query_junk_indices = [], [], [] + + image_dir = os.path.join(dataset_dir, "query") + for filename in sorted(os.listdir(image_dir)): + meta_data = _parse_filename(filename) + if meta_data is None: + # This is not a valid filename (e.g., Thumbs.db). + continue + + filename_base = meta_data[2] + junk_matfile = filename_base + "_junk.mat" + mat = sio.loadmat(os.path.join(dataset_dir, "gt_query", junk_matfile)) + if np.any(mat["junk_index"] < 1): + indices = [] + else: + # MATLAB to Python index. + indices = list(mat["junk_index"].astype(np.int64).ravel() - 1) + + query_junk_indices.append(indices) + query_filenames.append(os.path.join(image_dir, filename)) + query_ids.append(meta_data[0]) + + # The following matrix maps from query (row) to gallery image (column) such + # that element (i, j) evaluates to 0 if query i and gallery image j should + # be excluded from computation of the evaluation metrics and 1 otherwise. + good_mask = np.ones( + (len(query_filenames), len(gallery_filenames)), np.float32) + for i, junk_indices in enumerate(query_junk_indices): + good_mask[i, junk_indices] = 0. + + return gallery_filenames, gallery_ids, query_filenames, query_ids, good_mask + + +def read_test_split_to_image(dataset_dir): + """Read query and gallery data to memory. This consumes a lot of memory. + + Parameters + ---------- + dataset_dir : str + Path to the Market 1501 dataset directory. + + Returns + ------- + (ndarray, ndarray, ndarray, ndarray, ndarray) + Returns a tuple with the following values: + + * Tensor of shape Nx128x64x3 of N gallery images in BGR color space. + * One dimensional array of N unique gallery IDs. + * Tensor of shape Mx128x64x3 of M query images in BGR color space. + * One dimensional array of M unique query IDs. + * Matrix of shape MxN such that element (i, j) evaluates to 0 if + gallery image j should be excluded from metrics computation of + query i and 1 otherwise. + + """ + gallery_filenames, gallery_ids, query_filenames, query_ids, good_mask = ( + read_test_split_to_str(dataset_dir)) + + gallery_images = np.zeros((len(gallery_filenames), 128, 64, 3), np.uint8) + for i, filename in enumerate(gallery_filenames): + gallery_images[i] = cv2.imread(filename, cv2.IMREAD_COLOR) + + query_images = np.zeros((len(query_filenames), 128, 64, 3), np.uint8) + for i, filename in enumerate(query_filenames): + query_images[i] = cv2.imread(filename, cv2.IMREAD_COLOR) + + gallery_ids = np.asarray(gallery_ids, np.int64) + query_ids = np.asarray(query_ids, np.int64) + return gallery_images, gallery_ids, query_images, query_ids, good_mask + diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/mars.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/mars.py new file mode 100644 index 000000000..2f7d117ed --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/mars.py @@ -0,0 +1,190 @@ +# vim: expandtab:ts=4:sw=4 +from npu_bridge.npu_init import * +import os +import numpy as np +import cv2 + + +# The maximum person ID in the dataset. +MAX_LABEL = 1500 + +IMAGE_SHAPE = 256, 128, 3 + + +def read_train_test_directory_to_str(directory): + """Read bbox_train/bbox_test directory. + + Parameters + ---------- + directory : str + Path to bbox_train/bbox_test directory. + + Returns + ------- + (List[str], List[int], List[int], List[int]) + Returns a tuple with the following entries: + + * List of image filenames. + * List of corresponding unique IDs for the individuals in the images. + * List of camera indices. + * List of tracklet indices. + + """ + + def to_label(x): + return int(x) if x.isdigit() else -1 + + dirnames = os.listdir(directory) + image_filenames, ids, camera_indices, tracklet_indices = [], [], [], [] + for dirname in dirnames: + filenames = os.listdir(os.path.join(directory, dirname)) + filenames = [ + f for f in filenames if os.path.splitext(f)[1] == ".jpg"] + image_filenames += [ + os.path.join(directory, dirname, f) for f in filenames] + ids += [to_label(dirname) for _ in filenames] + camera_indices += [int(f[5]) for f in filenames] + tracklet_indices += [int(f[7:11]) for f in filenames] + + return image_filenames, ids, camera_indices, tracklet_indices + + +def read_train_test_directory_to_image(directory, image_shape=(128, 64)): + """Read images in bbox_train/bbox_test directory. + + Parameters + ---------- + directory : str + Path to bbox_train/bbox_test directory. + image_shape : Tuple[int, int] + A tuple (height, width) of the desired image size. + + Returns + ------- + (np.ndarray, np.ndarray, np.ndarray, np.ndarray) + Returns a tuple with the following entries: + + * Tensor of images in BGR color space. + * One dimensional array of unique IDs for the individuals in the images. + * One dimensional array of camera indices. + * One dimensional array of tracklet indices. + + """ + reshape_fn = ( + (lambda x: x) if image_shape == IMAGE_SHAPE[:2] + else (lambda x: cv2.resize(x, image_shape[::-1]))) + + filenames, ids, camera_indices, tracklet_indices = ( + read_train_test_directory_to_str(directory)) + + images = np.zeros((len(filenames), ) + image_shape + (3, ), np.uint8) + for i, filename in enumerate(filenames): + if i % 1000 == 0: + print("Reading %s, %d / %d" % (directory, i, len(filenames))) + image = cv2.imread(filename, cv2.IMREAD_COLOR) + images[i] = reshape_fn(image) + ids = np.asarray(ids, dtype=np.int64) + camera_indices = np.asarray(camera_indices, dtype=np.int64) + tracklet_indices = np.asarray(tracklet_indices, dtype=np.int64) + return images, ids, camera_indices, tracklet_indices + + +def read_train_split_to_str(dataset_dir): + """Read training data to list of filenames. + + Parameters + ---------- + dataset_dir : str + Path to the MARS dataset directory; ``bbox_train`` should be a + subdirectory of this folder. + + Returns + ------- + (List[str], List[int], List[int], List[int]) + Returns a tuple with the following entries: + + * List of image filenames. + * List of corresponding unique IDs for the individuals in the images. + * List of camera indices. + * List of tracklet indices. + + """ + train_dir = os.path.join(dataset_dir, "bbox_train") + return read_train_test_directory_to_str(train_dir) + + +def read_train_split_to_image(dataset_dir, image_shape=(128, 64)): + """Read training images to memory. This consumes a lot of memory. + + Parameters + ---------- + dataset_dir : str + Path to the MARS dataset directory; ``bbox_train`` should be a + subdirectory of this folder. + image_shape : Tuple[int, int] + A tuple (height, width) of the desired image size. + + Returns + ------- + (np.ndarray, np.ndarray, np.ndarray, np.ndarray) + Returns a tuple with the following entries: + + * Tensor of images in BGR color space. + * One dimensional array of unique IDs for the individuals in the images. + * One dimensional array of camera indices. + * One dimensional array of tracklet indices. + + """ + train_dir = os.path.join(dataset_dir, "bbox_train") + return read_train_test_directory_to_image(train_dir, image_shape) + + +def read_test_split_to_str(dataset_dir): + """Read training data to list of filenames. + + Parameters + ---------- + dataset_dir : str + Path to the MARS dataset directory; ``bbox_test`` should be a + subdirectory of this folder. + + Returns + ------- + (List[str], List[int], List[int], List[int]) + Returns a tuple with the following entries: + + * List of image filenames. + * List of corresponding unique IDs for the individuals in the images. + * List of camera indices. + * List of tracklet indices. + + """ + test_dir = os.path.join(dataset_dir, "bbox_test") + return read_train_test_directory_to_str(test_dir) + + +def read_test_split_to_image(dataset_dir, image_shape=(128, 64)): + """Read test images to memory. This consumes a lot of memory. + + Parameters + ---------- + dataset_dir : str + Path to the MARS dataset directory; ``bbox_test`` should be a + subdirectory of this folder. + image_shape : Tuple[int, int] + A tuple (height, width) of the desired image size. + + Returns + ------- + (np.ndarray, np.ndarray, np.ndarray, np.ndarray) + Returns a tuple with the following entries: + + * Tensor of images in BGR color space. + * One dimensional array of unique IDs for the individuals in the images. + * One dimensional array of camera indices. + * One dimensional array of tracklet indices. + + """ + test_dir = os.path.join(dataset_dir, "bbox_test") + return read_train_test_directory_to_image(test_dir, image_shape) + diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/util.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/util.py new file mode 100644 index 000000000..5762ff7b2 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/datasets/util.py @@ -0,0 +1,173 @@ +# vim: expandtab:ts=4:sw=4 +from npu_bridge.npu_init import * +import numpy as np +import cv2 + + +def crop_to_shape(images, patch_shape): + """Crop images to desired shape, respecting the target aspect ratio. + + Parameters + ---------- + images : List[ndarray] + A list of images in BGR format (dtype np.uint8) + patch_shape : (int, int) + Target image patch shape (height, width). + + Returns + ------- + ndarray + A tensor of output images. + + """ + assert len(images) > 0, "Empty image list is not allowed." + channels = () if len(images[0].shape) == 0 else (images[0].shape[-1], ) + output_images = np.zeros( + (len(images), ) + patch_shape + channels, dtype=np.uint8) + + target_aspect_ratio = float(patch_shape[1]) / patch_shape[0] + for i, image in enumerate(images): + image_aspect_ratio = float(image.shape[1]) / image.shape[0] + if target_aspect_ratio > image_aspect_ratio: + # Fix width, modify height. + crop_height = image.shape[1] / target_aspect_ratio + crop_width = image.shape[1] + else: + # Fix height, modify width. + crop_width = target_aspect_ratio * image.shape[0] + crop_height = image.shape[0] + + sx = int((image.shape[1] - crop_width) / 2) + sy = int((image.shape[0] - crop_height) / 2) + ex = int(min(sx + crop_width, image.shape[1])) + ey = int(min(sy + crop_height, image.shape[0])) + output_images[i, ...] = cv2.resize( + image[sy:ey, sx:ex], patch_shape[::-1], + interpolation=cv2.INTER_CUBIC) + + return output_images + + +def create_validation_split(data_y, num_validation_y, seed=None): + """Split dataset into training and validation set with disjoint classes. + + Parameters + ---------- + data_y : ndarray + A label vector. + num_validation_y : int | float + The number of identities to split off for validation. If an integer + is given, this value should be at least 1 and is interpreted as absolute + number of validation identities. If a float is given, this value should + be in [0, 1[ and is interpreted as fraction of validation identities. + seed : Optional[int] + A random generator seed used to select the validation idenities. + + Returns + ------- + (ndarray, ndarray) + Returns indices of training and validation set. + + """ + unique_y = np.unique(data_y) + if isinstance(num_validation_y, float): + num_validation_y = int(num_validation_y * len(unique_y)) + + random_generator = np.random.RandomState(seed=seed) + validation_y = random_generator.choice( + unique_y, num_validation_y, replace=False) + + validation_mask = np.full((len(data_y), ), False, bool) + for y in validation_y: + validation_mask = np.logical_or(validation_mask, data_y == y) + training_mask = np.logical_not(validation_mask) + return np.where(training_mask)[0], np.where(validation_mask)[0] + + +def limit_num_elements_per_identity(data_y, max_num_images_per_id, seed=None): + """Limit the number of elements per identity to `max_num_images_per_id`. + + Parameters + ---------- + data_y : ndarray + A label vector. + max_num_images_per_id : int + The maximum number of elements per identity that should remain in + the data set. + seed : Optional[int] + Random generator seed. + + Returns + ------- + ndarray + A boolean mask that evaluates to True if the corresponding + should remain in the data set. + + """ + random_generator = np.random.RandomState(seed=seed) + valid_mask = np.full((len(data_y), ), False, bool) + for y in np.unique(data_y): + indices = np.where(data_y == y)[0] + num_select = min(len(indices), max_num_images_per_id) + indices = random_generator.choice(indices, num_select, replace=False) + valid_mask[indices] = True + return valid_mask + + +def create_cmc_probe_and_gallery(data_y, camera_indices=None, seed=None): + """Create probe and gallery images for evaluation of CMC top-k statistics. + + For every identity, this function selects one image as probe and one image + for the gallery. Cross-view validation is performed when multiple cameras + are given. + + Parameters + ---------- + data_y : ndarray + Vector of data labels. + camera_indices : Optional[ndarray] + Optional array of camera indices. If possible, probe and gallery images + are selected from different cameras (i.e., cross-view validation). + If None given, assumes all images are taken from the same camera. + seed : Optional[int] + The random seed used to select probe and gallery images. + + Returns + ------- + (ndarray, ndarray) + Returns a tuple of indices to probe and gallery images. + + """ + data_y = np.asarray(data_y) + if camera_indices is None: + camera_indices = np.zeros_like(data_y, dtype=np.int) + camera_indices = np.asarray(camera_indices) + + random_generator = np.random.RandomState(seed=seed) + unique_y = np.unique(data_y) + probe_indices, gallery_indices = [], [] + for y in unique_y: + mask_y = data_y == y + + unique_cameras = np.unique(camera_indices[mask_y]) + if len(unique_cameras) == 1: + # If we have only one camera, take any two images from this device. + c = unique_cameras[0] + indices = np.where(np.logical_and(mask_y, camera_indices == c))[0] + if len(indices) < 2: + continue # Cannot generate a pair for this identity. + i1, i2 = random_generator.choice(indices, 2, replace=False) + else: + # If we have multiple cameras, take images of two (randomly chosen) + # different devices. + c1, c2 = random_generator.choice(unique_cameras, 2, replace=False) + indices1 = np.where(np.logical_and(mask_y, camera_indices == c1))[0] + indices2 = np.where(np.logical_and(mask_y, camera_indices == c2))[0] + i1 = random_generator.choice(indices1) + i2 = random_generator.choice(indices2) + + probe_indices.append(i1) + gallery_indices.append(i2) + + return np.asarray(probe_indices), np.asarray(gallery_indices) + diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/help_modelarts.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/help_modelarts.py new file mode 100644 index 000000000..edd3b80fc --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/help_modelarts.py @@ -0,0 +1,80 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import os +import datetime +import moxing as mox + +def obs_data2modelarts(config): + """ + Copy train data from obs to modelarts by using moxing api. + """ + start = datetime.datetime.now() + print("===>>>Copy files from obs:{} to modelarts dir:{}".format(config.data_url, config.modelarts_data_dir)) + mox.file.copy_parallel(src_url=config.data_url, dst_url=config.modelarts_data_dir) + end = datetime.datetime.now() + print("===>>>Copy from obs to modelarts, time use:{}(s)".format((end - start).seconds)) + files = os.listdir(config.modelarts_data_dir) + print("===>>>Files:", files) + + +def modelarts_result2obs(FLAGS): + """ + Copy debug data from modelarts to obs. + According to the swich flags, the debug data may contains auto tune repository, + dump data for precision comparision, even the computation graph and profiling data. + """ + work_dir = os.getcwd() + + ## copy result from modelarts to obs + obs_result_dir = os.path.join(FLAGS.obs_dir, 'result') + if not mox.file.exists(obs_result_dir): + mox.file.make_dirs(obs_result_dir) + mox.file.copy_parallel(src_url=FLAGS.result, dst_url=obs_result_dir) + print("===>>>Copy Event or Checkpoint from modelarts dir:{} to obs:{}".format(FLAGS.result, obs_result_dir)) + + ## Copy auto tune repository. Comment this snippets if npu_auto_tune is off. + if FLAGS.npu_auto_tune: + modelarts_auto_tune_dir = os.path.join(work_dir, "npu_auto_tune") + obs_auto_tune_dir = os.path.join(FLAGS.obs_dir, 'npu_auto_tune') + if not mox.file.exists(obs_auto_tune_dir): + mox.file.make_dirs(obs_auto_tune_dir) + mox.file.copy_parallel(modelarts_auto_tune_dir, obs_auto_tune_dir) + print("===>>>Auto tune:{} on OBS dir:{}".format(mox.file.list_directory(obs_auto_tune_dir), obs_auto_tune_dir)) + + ## Copy dump data. Comment this snippets if npu_dump_data is off. + if FLAGS.npu_dump_data: + modelarts_dump_data_dir = os.path.join(work_dir, "npu_dump_data") + obs_dump_data_dir = os.path.join(FLAGS.obs_dir, 'npu_dump_data') + if not mox.file.exists(obs_dump_data_dir): + mox.file.make_dirs(obs_dump_data_dir) + mox.file.copy_parallel(modelarts_dump_data_dir, obs_dump_data_dir) + print("===>>>Dumped graph:{} on OBS dir:{}".format(mox.file.list_directory(obs_dump_data_dir), obs_dump_data_dir)) + + ## Copy compute graph. Comment this snippets if npu_dump_graph is off. + if FLAGS.npu_dump_graph: + modelarts_dump_graph_dir = os.path.join(work_dir, "npu_dump_graph") + obs_dump_graph_dir = os.path.join(FLAGS.obs_dir, 'npu_dump_graph') + if not mox.file.exists(obs_dump_graph_dir): + mox.file.make_dirs(obs_dump_graph_dir) + mox.file.copy_parallel(modelarts_dump_graph_dir, obs_dump_graph_dir) + print("===>>>Dumped data:{} on OBS dir:{}".format(mox.file.list_directory(obs_dump_graph_dir), obs_dump_graph_dir)) + + ## Copy profiling data. Comment this snippets if npu_profiling is off. + if FLAGS.npu_profiling: + modelarts_profiling_dir = os.path.join(work_dir, "npu_profiling") + obs_profiling_dir = os.path.join(FLAGS.obs_dir, 'npu_profiling') + if not mox.file.exists(obs_profiling_dir): + mox.file.make_dirs(obs_profiling_dir) + mox.file.copy_parallel(modelarts_profiling_dir, obs_profiling_dir) + print("===>>>Profiling data:{} on OBS dir:{}".format(mox.file.list_directory(obs_profiling_dir), obs_profiling_dir)) \ No newline at end of file diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/losses.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/losses.py new file mode 100644 index 000000000..fda1c33e2 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/losses.py @@ -0,0 +1,145 @@ +# vim: expandtab:ts=4:sw=4 +from npu_bridge.npu_init import * +import tensorflow as tf + + +def _pdist(a, b=None): + sq_sum_a = tf.reduce_sum(tf.square(a), reduction_indices=[1]) + if b is None: + return -2 * tf.matmul(a, tf.transpose(a)) + \ + tf.reshape(sq_sum_a, (-1, 1)) + tf.reshape(sq_sum_a, (1, -1)) + sq_sum_b = tf.reduce_sum(tf.square(b), reduction_indices=[1]) + return -2 * tf.matmul(a, tf.transpose(b)) + \ + tf.reshape(sq_sum_a, (-1, 1)) + tf.reshape(sq_sum_b, (1, -1)) + + +def softmargin_triplet_loss(features, labels, create_summaries=True): + """Softmargin triplet loss. + + See:: + + Hermans, Beyer, Leibe: In Defense of the Triplet Loss for Person + Re-Identification. arXiv, 2017. + + Parameters + ---------- + features : tf.Tensor + A matrix of shape NxM that contains the M-dimensional feature vectors + of N objects (floating type). + labels : tf.Tensor + The one-dimensional array of length N that contains for each feature + the associated class label (integer type). + create_summaries : Optional[bool] + If True, creates summaries to monitor training behavior. + + Returns + ------- + tf.Tensor + A scalar loss tensor. + + """ + eps = tf.constant(1e-5, tf.float32) + nil = tf.constant(0., tf.float32) + almost_inf = tf.constant(1e+10, tf.float32) + + squared_distance_mat = _pdist(features) + distance_mat = tf.sqrt(tf.maximum(nil, eps + squared_distance_mat)) + label_mat = tf.cast(tf.equal( + tf.reshape(labels, (-1, 1)), tf.reshape(labels, (1, -1))), tf.float32) + + positive_distance = tf.reduce_max(label_mat * distance_mat, axis=1) + negative_distance = tf.reduce_min( + (label_mat * almost_inf) + distance_mat, axis=1) + loss = tf.nn.softplus(positive_distance - negative_distance) + if create_summaries: + fraction_invalid_pdist = tf.reduce_mean( + tf.cast(tf.less_equal(squared_distance_mat, -eps), tf.float32)) + tf.summary.scalar("fraction_invalid_pdist", fraction_invalid_pdist) + + fraction_active_triplets = tf.reduce_mean( + tf.cast(tf.greater_equal(loss, 1e-5), tf.float32)) + tf.summary.scalar("fraction_active_triplets", fraction_active_triplets) + + embedding_squared_norm = tf.reduce_mean( + tf.reduce_sum(tf.square(features), axis=1)) + tf.summary.scalar("mean squared feature norm", embedding_squared_norm) + + mean_distance = tf.reduce_mean(distance_mat) + tf.summary.scalar("mean feature distance", mean_distance) + + mean_positive_distance = tf.reduce_mean(positive_distance) + tf.summary.scalar("mean positive distance", mean_positive_distance) + + mean_negative_distance = tf.reduce_mean(negative_distance) + tf.summary.scalar("mean negative distance", mean_negative_distance) + + return tf.reduce_mean(loss) + + +def magnet_loss(features, labels, margin=1.0, unique_labels=None): + """Simple unimodal magnet loss. + + See:: + + Rippel, Paluri, Dollar, Bourdev: Metric Learning With Adaptive + Density Discrimination. ICLR, 2016. + + Parameters + ---------- + features : tf.Tensor + A matrix of shape NxM that contains the M-dimensional feature vectors + of N objects (floating type). + labels : tf.Tensor + The one-dimensional array of length N that contains for each feature + the associated class label (integer type). + margin : float + A scalar margin hyperparameter. + unique_labels : Optional[tf.Tensor] + Optional tensor of unique values in `labels`. If None given, computed + from data. + + Returns + ------- + tf.Tensor + A scalar loss tensor. + + """ + nil = tf.constant(0., tf.float32) + one = tf.constant(1., tf.float32) + minus_two = tf.constant(-2., tf.float32) + eps = tf.constant(1e-4, tf.float32) + margin = tf.constant(margin, tf.float32) + + num_per_class = None + if unique_labels is None: + unique_labels, sample_to_unique_y, num_per_class = tf.unique_with_counts(labels) + num_per_class = tf.cast(num_per_class, tf.float32) + + y_mat = tf.cast(tf.equal( + tf.reshape(labels, (-1, 1)), tf.reshape(unique_labels, (1, -1))), + dtype=tf.float32) + + # If class_means is None, compute from batch data. + if num_per_class is None: + num_per_class = tf.reduce_sum(y_mat, reduction_indices=[0]) + class_means = tf.reduce_sum( + tf.expand_dims(tf.transpose(y_mat), -1) * tf.expand_dims(features, 0), + reduction_indices=[1]) / tf.expand_dims(num_per_class, -1) + + squared_distance = _pdist(features, class_means) + + num_samples = tf.cast(tf.shape(labels)[0], tf.float32) + variance = tf.reduce_sum( + y_mat * squared_distance) / (num_samples - one) + + const = one / (minus_two * (variance + eps)) + linear = const * squared_distance - y_mat * margin + + maxi = tf.reduce_max(linear, reduction_indices=[1], keepdims=True) + loss_mat = tf.exp(linear - maxi) + + a = tf.reduce_sum(y_mat * loss_mat, reduction_indices=[1]) + b = tf.reduce_sum((one - y_mat) * loss_mat, reduction_indices=[1]) + loss = tf.maximum(nil, -tf.log(eps + a / (eps + b))) + return tf.reduce_mean(loss), class_means, variance + diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/metrics.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/metrics.py new file mode 100644 index 000000000..568d378b6 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/metrics.py @@ -0,0 +1,223 @@ +# vim: expandtab:ts=4:sw=4 +from npu_bridge.npu_init import * +import tensorflow as tf +import tensorflow.contrib.slim as slim + + +def pdist(a, b=None): + """Compute element-wise squared distance between `a` and `b`. + + Parameters + ---------- + a : tf.Tensor + A matrix of shape NxL with N row-vectors of dimensionality L. + b : tf.Tensor + A matrix of shape MxL with M row-vectors of dimensionality L. + + Returns + ------- + tf.Tensor + A matrix of shape NxM where element (i, j) contains the squared + distance between elements `a[i]` and `b[j]`. + + """ + sq_sum_a = tf.reduce_sum(tf.square(a), reduction_indices=[1]) + if b is None: + return -2 * tf.matmul(a, tf.transpose(a)) + \ + tf.reshape(sq_sum_a, (-1, 1)) + tf.reshape(sq_sum_a, (1, -1)) + sq_sum_b = tf.reduce_sum(tf.square(b), reduction_indices=[1]) + return -2 * tf.matmul(a, tf.transpose(b)) + \ + tf.reshape(sq_sum_a, (-1, 1)) + tf.reshape(sq_sum_b, (1, -1)) + + +def cosine_distance(a, b=None): + """Compute element-wise cosine distance between `a` and `b`. + + Parameters + ---------- + a : tf.Tensor + A matrix of shape NxL with N row-vectors of dimensionality L. + b : tf.Tensor + A matrix of shape NxL with N row-vectors of dimensionality L. + + Returns + ------- + tf.Tensor + A matrix of shape NxM where element (i, j) contains the cosine distance + between elements `a[i]` and `b[j]`. + + """ + a_normed = tf.nn.l2_normalize(a, dim=1) + b_normed = a_normed if b is None else tf.nn.l2_normalize(b, dim=1) + return ( + tf.constant(1.0, tf.float32) - + tf.matmul(a_normed, tf.transpose(b_normed))) + + +def recognition_rate_at_k(probe_x, probe_y, gallery_x, gallery_y, k, + measure=pdist): + """Compute the recognition rate at a given level `k`. + + For a given probe and ranked gallery that is sorted according to a distance + measure `measure` in descending order, the recognition rate at `k` is:: + + recognition_rate_at_k = num_correct / min(k, num_relevant) + + where num_correct refers to the fraction of images in the top k entries of + the ranked gallery that have the same label as the probe and `num_relevant` + refers to the total number of elements in the gallery that have the same + label. + + Parameters + ---------- + probe_x: tf.Tensor + A tensor of probe images. + probe_y: tf.Tensor + A tensor of probe labels. + gallery_x: tf.Tensor + A tensor of gallery images. + gallery_y: tf.Tensor + A tensor of gallery labels. + k: int + See description above. + measure: Callable[tf.Tensor, tf.Tensor] -> tf.Tensor + A callable that computes for two matrices of row-vectors a matrix of + element-wise distances. See `pdist` for an example. + + Returns + ------- + tf.Tensor + Returns a scalar tensor which represents the computed metric. + + """ + # Build a matrix of shape (num_probes, num_gallery_images) where element + # (i, j) is 1 if probe image i and the gallery image j have the same + # identity, otherwise 0. + label_eq_mat = tf.cast(tf.equal(tf.reshape( + probe_y, (-1, 1)), tf.reshape(gallery_y, (1, -1))), + tf.float32) + + # For each probe image, compute the number of relevant images in the + # gallery (same identity). This should always be one for CMC evaluation + # because we always have exactly one probe and one gallery image for each + # identity. + num_relevant = tf.minimum(tf.cast(k, tf.float32), tf.reduce_sum( + label_eq_mat, reduction_indices=[1])) + + # Rank gallery images by the similarity measure to build a matrix of + # shape (num_probes, k) where element (i, j) contains the label of the + # j-th ranked gallery image for probe i. + predictions = tf.exp(-measure(probe_x, gallery_x)) # Compute similarity. + _, prediction_indices = tf.nn.top_k(predictions, k=k) + label_mat = tf.gather(gallery_y, prediction_indices) + + # Just as we have done before, build a matrix where element (i, j) is + # one if probe i and gallery image j share the same label (same identity). + # This time, the matrix is ranked by the similarity measure and we only + # keep the top-k predictions. + label_eq_mat = tf.cast(tf.equal( + label_mat, tf.reshape(probe_y, (-1, 1))), tf.float32) + + # Compute the number of true positives in [0, k[, i.e., check if we find + # the correct gallery image within the top-k ranked results. Then, compute + # the recognition rate, which in our case is either 0 or 1 since we have + # only one gallery image that shares the same identity with the probe. + # + # This is the final output of our CMC metric. + true_positives_at_k = tf.reduce_sum(label_eq_mat, reduction_indices=[1]) + return true_positives_at_k / num_relevant + + +def streaming_mean_cmc_at_k(probe_x, probe_y, gallery_x, gallery_y, k, + measure=pdist): + """Compute cumulated matching characteristics (CMC) at level `k` over + a stream of data (i.e., multiple batches). + + The function is compatible with TensorFlow-Slim's streaming metrics + interface, e.g., `slim.metrics.aggregate_metric_map`. + + Parameters + ---------- + probe_x: tf.Tensor + A tensor of probe images. + probe_y: tf.Tensor + A tensor of probe labels. + gallery_x: tf.Tensor + A tensor of gallery images. + gallery_y: tf.Tensor + A tensor of gallery labels. + k: int + See description above. + measure: Callable[tf.Tensor, tf.Tensor] -> tf.Tensor + A callable that computes for two matrices of row-vectors a matrix of + element-wise distances. See `pdist` for an example. + + Returns + ------- + Tuple[tf.Tensor, tf.Tensor] + The first element in the tuple is the current result. The second element + is an operation that updates the computed metric based on new data. + + """ + recognition_rate = recognition_rate_at_k( + probe_x, probe_y, gallery_x, gallery_y, k, measure) + return slim.metrics.streaming_mean(recognition_rate) + + +def streaming_mean_averge_precision(probe_x, probe_y, gallery_x, gallery_y, + good_mask, measure=pdist): + """Compute mean average precision (mAP) over a stream of data. + + Parameters + ---------- + probe_x: tf.Tensor + A tensor of N probe images. + probe_y: tf.Tensor + A tensor of N probe labels. + gallery_x: tf.Tensor + A tensor of M gallery images. + gallery_y: tf.Tensor + A tensor of M gallery labels. + good_mask: Optional[tf.Tensor] + A matrix of shape NxM where element (i, j) evaluates to 0.0 if the pair + of i-th probe and j-th gallery image should be excluded from metric + computation. All other elements should evaluate to 1.0. + measure: Callable[tf.Tensor, tf.Tensor] -> tf.Tensor + A callable that computes for two matrices of row-vectors a matrix of + element-wise distances. See `pdist` for an example. + + Returns + ------- + Tuple[tf.Tensor, tf.Tensor] + The first element in the tuple is the current result. The second element + is an operation that updates the computed metric based on new data. + + """ + # See Wikipedia: + # https://en.wikipedia.org/wiki/Information_retrieval#Average_precision + if good_mask.dtype != tf.float32: + good_mask = tf.cast(good_mask, tf.float32) + + # Compute similarity measure and mask out diagonal (similarity to self). + predictions = good_mask * tf.exp(-measure(probe_x, gallery_x)) + + # Compute matrix of predicted labels. + k = tf.shape(gallery_y)[0] + _, prediction_indices = tf.nn.top_k(predictions, k=k) + predicted_label_mat = tf.gather(gallery_y, prediction_indices) + label_eq_mat = tf.cast(tf.equal( + predicted_label_mat, tf.reshape(probe_y, (-1, 1))), tf.float32) + + # Compute statistics. + num_relevant = tf.reduce_sum( + good_mask * label_eq_mat, reduction_indices=[1], keep_dims=True) + true_positives_at_k = tf.cumsum(label_eq_mat, axis=1) + retrieved_at_k = tf.cumsum(tf.ones_like(label_eq_mat), axis=1) + precision_at_k = true_positives_at_k / retrieved_at_k + relevant_at_k = label_eq_mat + average_precision = ( + tf.reduce_sum(precision_at_k * relevant_at_k, reduction_indices=[1]) / + tf.cast(tf.squeeze(num_relevant), tf.float32)) + + return slim.metrics.streaming_mean(average_precision) + diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/__init__.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/queued_trainer.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/queued_trainer.py new file mode 100644 index 000000000..8939fbe28 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/queued_trainer.py @@ -0,0 +1,616 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# vim: expandtab:ts=4:sw=4 +from npu_bridge.npu_init import * +import string +import os +import threading +import numpy as np + +import tensorflow as tf +import tensorflow.contrib.slim as slim +import time + +config = tf.ConfigProto() +custom_op = config.graph_options.rewrite_options.custom_optimizers.add() +custom_op.name = "NpuOptimizer" +custom_op.parameter_map["use_off_line"].b = True +custom_op.parameter_map["mix_compile_mode"].b = True +custom_op.parameter_map["precision_mode"].s = tf.compat.as_bytes("force_fp32")#allow_mix_precision +#custom_op.parameter_map["modify_mixlist"].s = tf.compat.as_bytes("/home/test/ops_info.json") +config.graph_options.rewrite_options.remapping = RewriterConfig.OFF +config.graph_options.rewrite_options.memory_optimization = RewriterConfig.OFF + +def run_in_batches(f, data_dict, out, batch_size): + """Process data in batches. + + Parameters + ---------- + f : Callable[Dict[tf.Tensor, np.ndarray] -> np.ndarray + A function that maps a given input (one or multiple inpu arrays) to a + single output array. + data_dict : Dict[tf.Tensor, np.ndarray] + Maps from symbolic input tensor to numpy data array. + out : np.ndarray + The computed function output will be stored in this array; must be have + compatible shape and length to the output computed by `f`. + batch_size : int + The number of samples to compute in each call to `f`. If the length of + the input array is not divisible by the batch size, the final call to + `f` contains fewer examples. + + """ + data_len = len(out) + num_batches = int(data_len / batch_size) + + def pad(x): + x = np.asarray(x) + y = np.full((batch_size, ) + x.shape[1:], x[0], dtype=x.dtype) + y[:x.shape[0]] = x + return y + + s, e = 0, batch_size + for i in range(num_batches): + s, e = i * batch_size, (i + 1) * batch_size + batch_data_dict = {k: v[s:e] for k, v in data_dict.items()} + out[s:e] = f(batch_data_dict) + if e < len(out): + remaining_len = len(out) - e + batch_data_dict = {k: pad(v[e:]) for k, v in data_dict.items()} + out[e:] = f(batch_data_dict)[:remaining_len] + + +def iterate_forever(batch_size, data, *other_data): + """Iterate over dataset indefenitely. + + Parameters + ---------- + batch_size : int + The batch size. + data : ndarray + The first input array. + other_data + Additional input arrays; must be of type np.ndarray. + + Returns + ------- + List[np.ndarray] + A dataset batch. The length of each entry in the list is `batch_size`. + + """ + data_len = len(data) + num_batches = int(data_len / batch_size) + + while True: + data_list = [data] + list(other_data) + s, e = 0, 0 + for i in range(num_batches): + s, e = i * batch_size, (i + 1) * batch_size + batch = [x[s:e] for x in data_list] + yield batch[0] if len(batch) == 1 else batch + if e < data_len: + batch = [x[e:] for x in data_list] + yield batch[0] if len(batch) == 1 else batch + + +def random_shuffle_forever(batch_size, data, *other_data): + """A generator that randomly selects `batch_size` entries from the data. + + Parameters + ---------- + batch_size : int + The batch size. + data : np.ndarray + The first input array. + other_data + Additional input arrays; must be of type np.ndarray + + Returns + ------- + List[np.ndarray] + A batch of randomly selected entries. The length of each entry in the + list is `batch_size`. + + """ + data_list = [data] + list(other_data) + indices = np.arange(len(data)) + while True: + batch_indices = np.random.choice(indices, batch_size, replace=False) + batch = [x[batch_indices] for x in data_list] + yield batch[0] if len(batch) == 1 else batch + + +def random_sample_identities_forever(batch_size, num_samples_per_id, data_x, + data_y, num_fa_images=0): + """A generator that randomly selects a fixed number of entries per label. + + If false alarms are passed into this function, they should have a negative + label, i.e., `data_y[i] < 0` if the i-th example corresponds to a false + alarm. + + Parameters + ---------- + batch_size : int + The batch size. + num_samples_per_id : int + Number of examples per label in each batch. If the `batch_size` is not + divisible by `num_samples_per_id` then the batch is filled with false + alarms. A warning is printed if no false alarms are available to fill + up the batch. + data_x : List[string] | np.ndarray + The data array; either a list of filenames or a tensor of input images. + data_y : List[int] | np.ndarray + The label array (either as list of one-dimensional numpy array). + num_fa_images : Optional[int] + Number of false alarm images to include in each batch; defaults to zero. + + Returns + ------- + List[np.ndarray] + Returns a list of length two where the first entry is the data array + corresponding to `data_x` and the second entry is the label array + corresponding to `data_y`. The elements in the list are of length + `batch_size`. + + """ + assert (batch_size - num_fa_images) % num_samples_per_id == 0 + num_ids_per_batch = int((batch_size - num_fa_images) / num_samples_per_id) + + data_x = np.asarray(data_x) + data_y = np.asarray(data_y) + + unique_y = np.unique(data_y[data_y >= 0]) + y_to_idx = {y: np.where(data_y == y)[0] for y in unique_y} + fa_indices = np.where(data_y < 0)[0] + + while True: + # Draw the desired number of identities. + indices = np.random.choice( + len(unique_y), num_ids_per_batch, replace=False) + batch_unique_y = unique_y[indices] + + batch_x = np.zeros((batch_size, ) + data_x.shape[1:], data_x.dtype) + batch_y = np.zeros((batch_size, ), data_y.dtype) + e = 0 + for i, y in enumerate(batch_unique_y): + num_samples = min(num_samples_per_id, len(y_to_idx[y])) + indices = np.random.choice(y_to_idx[y], num_samples, replace=False) + s, e = e, e + num_samples + batch_x[s:e] = data_x[indices] + batch_y[s:e] = y + + # Fill up remaining space with false alarms. + num_samples = len(batch_x) - e + if num_fa_images > 0: + num_batch_fa_samples = min(num_samples, len(fa_indices)) + indices = np.random.choice( + fa_indices, num_batch_fa_samples, replace=False) + s, e = e, e + num_batch_fa_samples + batch_x[s:e] = data_x[indices] + batch_y[s:e] = data_y[indices] + + # If we need to add more data, random sample ids until we have reached + # the batch size. + num_samples = len(batch_x) - e + num_tries = 0 + while num_samples > 0 and num_tries < 100: + y = np.random.choice(unique_y) + if y in batch_unique_y: + # Find a target that we have not yet in this batch. + num_tries += 1 + continue + + num_samples = min(num_samples, len(y_to_idx[y])) + indices = np.random.choice(y_to_idx[y], num_samples, replace=False) + s, e = e, e + num_samples + batch_x[s:e] = data_x[indices] + batch_y[s:e] = y + num_samples = len(batch_x) - e + + if e < batch_size: + print("ERROR: Failed to sample a full batch. Adding corrupt data.") + yield [batch_x, batch_y] + + +def _truncate_dataset_to_batch_size(batch_size, data, *other_data): + """Truncate given input data to a multiple of the batch size. + + Parameters + ---------- + batch_size : int + The batch size. The length of the truncated data is a multiple of this + value. + data : np.ndarray + The first input array. + *other_data + Additional input arrays; must be of type np.ndarray. + + Returns + ------- + List[np.ndarray] + The truncated data. The length of each entry in the list is a multiple + of the batch size. + + """ + num_batches = int(len(data) / batch_size) + new_len = num_batches * batch_size + dataset = [data] + list(other_data) + if new_len < len(data): + print( + "WARNING dataset length is not a multiple of batch size. " + "Truncating from %d to %d." % (len(data), new_len)) + dataset = [x[:new_len] for x in dataset] + return num_batches, dataset[0] if len(dataset) == 1 else dataset + + +def _generate_run_id(size=6, chars=None): + """Generate a random ID of length `size`. + + Parameters + ---------- + size : int + chars : Optional[str] + Optional list of characters to use for generating the ID. + + Returns + ------- + str + Returns a random identifier of length `size`. + + """ + if chars is None: + chars = string.ascii_uppercase + string.digits + import random + return ''.join(random.choice(chars) for _ in range(size)) + + +class ThreadSafeIterator(object): + """ + This class wraps an iterator (or generator) such that only one thread at a + time is granted access. + + Parameters + ---------- + iterator_or_generator + An iterator or generator to be wrapped. + + """ + + def __init__(self, iterator_or_generator): + self._iterator_or_generator = iterator_or_generator + self._lock = threading.Lock() + + def __iter__(self): + return self + + def __next__(self): + with self._lock: + return next(self._iterator_or_generator) + + def next(self): + with self._lock: + return self._iterator_or_generator.next() + + +class QueuedTrainer(object): + """ + This class implements code to train and evaluate TensorFlow models based on + TensorFlow-Slim. Image loading and preprocessing is de-coupled from the + training steps using a tf.FIFOQueue. + + Parameters + ---------- + enqueue_vars : List[tf.Tensor] + A list of tensors to be enqueued; usually the labels and preprocessed + images. + input_vars : Optional[List[tf.Tensor]] + An optional list of input tensors; usually the labels and raw (not + preprocessed) images or filenames to the images. The list must be of the + same length as the `enqueue_vars` and there must be a one-to-one + correspondence, i.e., the i-th element in `enqueue_vars` is i-th + preprocessed element in `input_vars`. If None, the input_vars are set to + `enqueue_vars`. + num_enqueue_threads : Optional[int] + Number of threads used to preprocess data in parallel. + queue_capacity : Optional[int] + Maximum number of elements in the queue; defaults to 512. + + """ + + def __init__(self, enqueue_vars, input_vars=None, num_enqueue_threads=4, + queue_capacity=512): + if input_vars is None: + input_vars = enqueue_vars + self._input_vars = input_vars + self._enqueue_vars = enqueue_vars + + shapes = [var.get_shape().as_list()[1:] for var in enqueue_vars] + dtypes = [var.dtype for var in enqueue_vars] + self._queue = tf.FIFOQueue(queue_capacity, dtypes, shapes) + + self._num_enqueue_threads = num_enqueue_threads + self._enqueue_threads = [] + self._enqueue_op = self._queue.enqueue_many(self._enqueue_vars) + self._stop_op = self._queue.close(cancel_pending_enqueues=True) + self._coordinator = None + + self._feed_generator = None + self._batch_size = None + self._init_fns = [] + + def get_input_vars(self, batch_size): + """Get the top `batch_size` elements from the queue. The tensors + returned by this functions should be passed on the the TensorFlow model. + + Parameters + ---------- + batch_size : int + The batch size. + + Returns + ------- + List[tf.Tensor] + Returns the top `batch_size` elements from the queue. There is a + one-to-one correspondence between the `enqueue_vars` passed in to + the constructor of this class and the tensors in the list returned + by this function. + + """ + self._batch_size = batch_size + return self._queue.dequeue_many(batch_size) + + def run(self, feed_generator, train_op, log_dir="/tmp/slim_trainer/", + restore_path=None, variables_to_restore=None, run_id=None, + max_checkpoints_to_keep=0, **kwargs): + """ Run training. + + Parameters + ---------- + feed_generator : Iterator[ndarray, ...] + An iterator or generator that returns batches of training data; must + return a one-to-one correspondence with the `enqueue_vars` passed + to the constructor of this class. + train_op : tf.Tensor + The training operation created with `slim.learning.create_train_op`. + log_dir : Optional[str] + Path to TensorFlow log directory. This value is used in conjunction + with `run_id` to generate the checkpoint and summary directory; + defaults to '/tmp/slim_trainer'. + restore_path : Optional[str] + An optional checkpoint path. If not None, resumes training from the + given checkpoint. + variables_to_restore : Optional[List[str]] + An optional list of variable scopes. If not None, only restores + variables under the given scope. This value is ignored if + `restore_path` is None. + run_id : Optional[str] + A string that identifies this training run. The checkpoints and + TensorFlow summaries are stored in `log_dir/run_id`. If None, a + random ID will be generated. Point tensorboard to this directory to + monitor training progress. + max_checkpoints_to_keep : int + Keep only the `max_checkpoints_to_keep` newest checkpoints. If 0, + keep all checkpoints. + kwargs: + Additional named arguments passed on to tf.slim.learning.train, + e.g., `number_of_steps=100` to run 100 iterations of training. + + """ + if restore_path is not None: + if variables_to_restore is None: + variables_to_restore = slim.get_variables_to_restore() + init_assign_op, init_feed_dict = slim.assign_from_checkpoint( + restore_path, variables_to_restore) + self._init_fns.append(lambda sess: sess.run( + init_assign_op, init_feed_dict)) + self._feed_generator = ThreadSafeIterator(feed_generator) + self._coordinator = tf.train.Coordinator() + + if run_id is None: + run_id = _generate_run_id(6) + log_dir = os.path.join(log_dir, run_id) + print("---------------------------------------") + print("Run ID: ", run_id) + print("Log directory: ", log_dir) + print("---------------------------------------") + + saver = tf.train.Saver(max_to_keep=max_checkpoints_to_keep) + try: + slim.learning.train( + train_op, log_dir, self._train_step_fn, session_config=config, saver=saver, + **kwargs) + except UnboundLocalError: + # NOTE(nwojke): Due to a bug in slim, a local variable 'total_loss' + # is referenced when an exception is raised during training. We + # catch the exception here because it occurs whenever we close the + # queue with self._stop_all_threads(). + pass + self._wait_for_threads() + + def evaluate(self, dataset, checkpoint_dir, log_dir, run_id=None, + init_op=None, eval_op=None, final_op=None, + summary_op=None, variables_to_restore=None, + eval_interval_secs=60, max_num_evaluations=None): + """Run evaluation. Monitors files in the log directory and computes + evaluation metrics. This function must be called concurrently to + training (in a separate process). + + WARNING: The dataset is truncated to the batch size. Thus, the computed + metrics are only accurate if the dataset length is divisible by the + batch size. + + Parameters + ---------- + dataset : List[T] + The dataset is a list (or tuple) of data arrays. The length of the + list must be the same as the `input_vars` passed to the constructor + of this class and there must be a one-to-one correspondence such + that `dataset[i]` corresponds to the numeric data of its symbolic + equivalent in `input_vars[i]`. + checkpoint_dir : str + The directory where checkpoints are stored. Should be set to + `log_dir` of the training process. + log_dir : str + Path to TensorFlow log directory where evaluation logs will be + stored. This directory should be different from the `log_dir` + passed to `run`. + run_id : Optional[str] + A string that identifies the training runrun. Should be set to + `run_id` passed to `run`. + init_op : Optional[tf.Tensor] + Optional operation to execute prior to processing the `dataset`. + eval_op : Optional[tf.Tensor] + Evaluation operation; will be executed for each batch in the + `dataset`. + final_op : Optional[tf.Tensor] + Optional operation to execute after processing the `dataset`. + summary_op : Optional[tf.Tensor] + Summary operation; defaults to `tf.summary.merge_all()`. + variables_to_restore : Optional[List[tf.Tensor]] + List of variables to restore; defaults to + `slim.get_variables_to_restore()`. + eval_interval_secs : Optional[int] + Poll the `checkpoint_dir` every `eval_interval_secs` seconds for + new checkpoints. + max_num_evaluations : Optional[int] + Evaluate at most `max_num_evaluations` checkpoints. + + Returns + ------- + T + Returns the value of the last call to `final_op` or None. + + """ + if run_id is None: + print("---------------------------------------") + print("Checkpoint directory: ", checkpoint_dir) + print("Log directory: ", log_dir) + print("---------------------------------------") + else: + checkpoint_dir = os.path.join(checkpoint_dir, run_id) + log_dir = os.path.join(log_dir, run_id) + print("---------------------------------------") + print("Run ID: ", run_id) + print("Checkpoint directory: ", checkpoint_dir) + print("Log directory: ", log_dir) + print("---------------------------------------") + + if summary_op is None: + summary_op = tf.summary.merge_all() + + global_step = tf.train.get_or_create_global_step() + + if variables_to_restore is None: + variables_to_restore = slim.get_variables_to_restore() + saver = tf.train.Saver(variables_to_restore) + summary_writer = tf.summary.FileWriter(log_dir) + sv = tf.train.Supervisor( + graph=tf.get_default_graph(), logdir=log_dir, summary_op=None, + summary_writer=None, global_step=None, saver=saver) + + print("Entering evaluation loop. Waiting for checkpoints.") + num_batches, dataset = _truncate_dataset_to_batch_size( + self._batch_size, *dataset) + + final_op_value = None + num_evaluations = 0 + for checkpoint_path in slim.evaluation.checkpoints_iterator( + checkpoint_dir, eval_interval_secs): + with sv.managed_session(start_standard_services=False) as session: + sv.saver.restore(session, checkpoint_path) + sv.start_queue_runners(session) + + print("Starting evaluation of '%s'" % checkpoint_path) + self._feed_generator = iterate_forever( + self._batch_size, *dataset) + self._coordinator = tf.train.Coordinator() + for fn in self._init_fns: + fn(session) + self._start_enqueue(session, num_threads=1) + + if init_op is not None: + session.run(init_op) + + if eval_op is not None: + for i in range(num_batches): + session.run(eval_op) + + if final_op is not None: + final_op_value = session.run(final_op) + else: + final_op_value = None + + summary_str = session.run(summary_op) + global_step_value = session.run(global_step) + summary_writer.add_summary(summary_str, global_step_value) + summary_writer.flush() + + self._stop_all_threads(session) + print("Finished evaluation of '%s'" % checkpoint_path) + + num_evaluations += 1 + if max_num_evaluations is not None \ + and num_evaluations >= max_num_evaluations: + break + return final_op_value + + def _train_step_fn(self, session, train_op, global_step, + train_step_kwargs): + if len(self._enqueue_threads) == 0: + for fn in self._init_fns: + fn(session) + self._start_enqueue(session) + train_start = time.time() + total_loss, should_stop = slim.learning.train_step( + session, train_op, global_step, train_step_kwargs) + print("========each step time: {:.2f}".format((time.time() - train_start)*1000)) + if should_stop or self._coordinator.should_stop(): + self._stop_all_threads(session) + print("========loss: ", total_loss) + return total_loss, should_stop + + def _stop_all_threads(self, session): + self._coordinator.request_stop() + session.run(self._stop_op) # Close the queue. + + def _wait_for_threads(self): + self._coordinator.join(self._enqueue_threads) + self._enqueue_threads = [] + + def _start_enqueue(self, session, num_threads=None): + if num_threads is None: + num_threads = self._num_enqueue_threads + for _ in range(num_threads): + thread = threading.Thread( + target=self._run_enqueue_thread, args=(session, )) + thread.start() + self._enqueue_threads.append(thread) + + def _run_enqueue_thread(self, session): + try: + for data in self._feed_generator: + if self._coordinator.should_stop(): + break + try: + feed_dict = { + var: value for var, value in + zip(self._input_vars, data)} + session.run(self._enqueue_op, feed_dict=feed_dict) + except (tf.errors.CancelledError, tf.errors.AbortedError): + # We have been requested to stop enqueuing data. + break + except Exception as e: + print("EnqueueError:", e) + self._stop_all_threads(session) diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/run_1p.sh b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/run_1p.sh new file mode 100644 index 000000000..260e3c525 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/run_1p.sh @@ -0,0 +1,25 @@ +#!/bin/bash +### Do not need to Configure CANN Environment on Modelarts Platform, because it has been set already. +### Modelarts Platform command for train +export TF_CPP_MIN_LOG_LEVEL=2 ## Tensorflow api print Log Config +export ASCEND_SLOG_PRINT_TO_STDOUT=0 ## Print log on terminal on(1), off(0) + +code_dir=${1} +data_dir=${2} +result_dir=${3} +obs_url=${4} + +current_time=`date "+%Y-%m-%d-%H-%M-%S"` + +python3.7 ${code_dir}/train_market1501.py \ + --dataset_dir=${data_dir} \ + #--result=${result_dir} \ + #--obs_dir=${obs_url} \ + #--chip='npu' \ + #--loss_mode=cosine-softmax \ + --run_id=cosine-softmax \ + #--platform='modelarts' \ + #--npu_dump_data=False \ + #--npu_dump_graph=False \ + #--npu_profiling=False \ + #--npu_auto_tune=False 2>&1 | tee ${result_dir}/${current_time}_train_npu.log diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/scripts/run_1p.sh b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/scripts/run_1p.sh new file mode 100644 index 000000000..80fd6a6c6 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/scripts/run_1p.sh @@ -0,0 +1,25 @@ +#!/bin/bash +### Do not need to Configure CANN Environment on Modelarts Platform, because it has been set already. +### Modelarts Platform command for train +export TF_CPP_MIN_LOG_LEVEL=2 ## Tensorflow api print Log Config +export ASCEND_SLOG_PRINT_TO_STDOUT=0 ## Print log on terminal on(1), off(0) + +code_dir=${1} +data_dir=${2} +result_dir=${3} +obs_url=${4} + +current_time=`date "+%Y-%m-%d-%H-%M-%S"` + +python3.7 ${code_dir}/train_market1501.py + #--dataset_dir=${data_dir} \ + #--result=${result_dir} \ + #--obs_dir=${obs_url} \ + #--chip='npu' \ + #--loss_mode=cosine-softmax \ + #--run_id=cosine-softmax \ + #--platform='modelarts' \ + #--npu_dump_data=False \ + #--npu_dump_graph=False \ + #--npu_profiling=False \ + #--npu_auto_tune=False 2>&1 | tee ${result_dir}/${current_time}_train_npu.log diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/scripts/run_apulis.sh b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/scripts/run_apulis.sh new file mode 100644 index 000000000..d3c278f85 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/scripts/run_apulis.sh @@ -0,0 +1,30 @@ +#!/bin/bash +### Apulis Platform command for train (CANN Version-20.2) +## Set Ascend Log Level, if u wanna to print on terminal, you should open 'ASCEND_SLOG_PRINT_TO_STDOUT'. +## refer to link: https://support.huaweicloud.com/Graphdevg-cann202training1/atlasag_25_0077.html +export ASCEND_SLOG_PRINT_TO_STDOUT=0 ## Print log on terminal on(1), off(0) +export ASCEND_GLOBAL_LOG_LEVEL=3 ## Ascend log level. debug(0), info(1), warning(2), error(3) +export TF_CPP_MIN_LOG_LEVEL=2 ## Tensorflow api print Log Config + +## Configure Environment for Auto Tune +## refer to link:https://support.huaweicloud.com/developmenttg-cann330alphaXtraining/atlasautotune_16_0014.html +export install_path=/home/HwHiAiUser/Ascend/ascend-toolkit/latest +export PATH=${install_path}/fwkacllib/bin:$PATH +export LD_LIBRARY_PATH=${install_path}/fwkacllib/lib64:$LD_LIBRARY_PATH +export PYTHONPATH=${install_path}/fwkacllib/python/site-packages:$PYTHONPATH + +code_dir=$(cd "$(dirname "$0")"; cd ..; pwd) +echo "===>>>Python boot file dir: ${code_dir}" + +current_time=`date "+%Y-%m-%d-%H-%M-%S"` + +python3.7 ${code_dir}/train.py \ + --chip='npu' \ + --dataset=/data/dataset/storage/flowers/images \ + --result=./log \ + --num_classes=5 \ + --train_step=2 \ + --npu_dump_data=False \ + --npu_dump_graph=False \ + --npu_profiling=False \ + --npu_auto_tune=False 2>&1 | tee ${code_dir}/${current_time}_train_npu.log \ No newline at end of file diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/train_app.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/train_app.py new file mode 100644 index 000000000..f0da1c216 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/train_app.py @@ -0,0 +1,690 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# vim: expandtab:ts=4:sw=4 +from npu_bridge.npu_init import * +import argparse +import numpy as np +import tensorflow as tf +import tensorflow.contrib.slim as slim + +from datasets import util +import queued_trainer +import metrics +import losses + + +def create_default_argument_parser(dataset_name): + """Create an argument parser with default arguments. + + Parameters + ---------- + dataset_name : str + Name of the dataset. This value is used to set default directories. + + Returns + ------- + argparse.ArgumentParser + Returns an argument parser with default arguments. + + """ + parser = argparse.ArgumentParser( + description="Metric trainer (%s)" % dataset_name) + parser.add_argument( + "--batch_size", help="Training batch size", default=128, type=int) + parser.add_argument( + "--learning_rate", help="Learning rate", default=1e-3, type=float) + parser.add_argument( + "--eval_log_dir", + help="Evaluation log directory (only used in mode 'evaluation').", + default="/tmp/%s_evaldir" % dataset_name) + parser.add_argument( + "--number_of_steps", help="Number of train/eval steps. If None given, " + "runs indefenitely", default=None, type=int) + parser.add_argument( + "--log_dir", help="Log and checkpoints directory.", + default="/tmp/%s_logdir" % dataset_name) + parser.add_argument( + "--loss_mode", help="One of 'cosine-softmax', 'magnet', 'triplet'", + type=str, default="cosine-softmax") + parser.add_argument( + "--mode", help="One of 'train', 'eval', 'finalize', 'freeze'.", + type=str, default="train") + parser.add_argument( + "--restore_path", help="If not None, resume training of a given " + "checkpoint (mode 'train').", default=None) + parser.add_argument( + "--run_id", help="An optional run-id. If None given, a new one is " + "created", type=str, default=None) + return parser + + +def to_train_kwargs(args): + """Parse command-line training arguments. + + Parameters + ---------- + args : argparse.Namespace + Namespace of an argument parser that was created with + create_default_argument_parser. + + Returns + ------- + Dict[str, T] + Returns a dictionary of named arguments to be passed on to + train_loop. + + """ + kwargs_dict = { + "batch_size": args.batch_size, + "learning_rate": args.learning_rate, + "log_dir": args.log_dir, + "loss_mode": args.loss_mode, + "number_of_steps": args.number_of_steps, + "restore_path": args.restore_path, + "run_id": args.run_id, + } + return kwargs_dict + + +def to_eval_kwargs(args): + """Parse command-line evaluation arguments. + + Parameters + ---------- + args : argparse.Namespace + Namespace of an argument parser that was created with + create_default_argument_parser. + + Returns + ------- + Dict[str, T] + Returns a dictionary of named arguments to be passed on to + eval_loop. + + """ + kwargs_dict = { + "eval_log_dir": args.eval_log_dir, + "log_dir": args.log_dir, + "loss_mode": args.loss_mode, + "run_id": args.run_id, + } + return kwargs_dict + + +def train_loop(preprocess_fn, network_factory, train_x, train_y, + num_images_per_id, batch_size, log_dir, image_shape=None, + restore_path=None, exclude_from_restore=None, run_id=None, + number_of_steps=None, loss_mode="cosine-softmax", + learning_rate=1e-3, trainable_scopes=None, + save_summaries_secs=60, save_interval_secs=300): + """Start training. + + Parameters + ---------- + preprocess_fn : Callable[tf.Tensor] -> tf.Tensor + A callable that applies preprocessing to a given input image tensor of + dtype tf.uint8 and returns a floating point representation (tf.float32). + network_factory : Callable[tf.Tensor] -> (tf.Tensor, tf.Tensor) + A callable that takes as argument a preprocessed input image of dtype + tf.float32 and returns the feature representation as well as a logits + tensors. The logits may be set to None if not required by the loss. + train_x : List[str] | np.ndarray + A list of image filenames or a tensor of images. + train_y : List[int] | np.ndarray + A list or one-dimensional array of labels for the images in `train_x`. + num_images_per_id : int + Sample `num_images_per_id` images for each label at each training + iteration. The number of identities sampled at each iteration is + computed as `batch_size / num_images_per_id`. The `batch_size` must be + divisible by this number. + batch_size : int + The number of images at each training iteration. + log_dir : str + Used to construct the log and checkpoint directory. They are stored in + `log_dir/run_id`. + image_shape : Tuple[int, int, int] | NoneType + Image shape (height, width, channels) or None. If None, `train_x` must + be an array of images such that the shape can be queries from this + variable. + restore_path : Optional[str] + If not None, resumes training from the given checkpoint file. + exclude_from_restore : Optional[List[str]] + An optional list of variable scopes to be used in conjunction with + `restore_path`. If not None, variables in the given scopes are not + restored from the checkpoint file. + run_id : Optional[str] + A string that identifies the training run; used to construct the + log and checkpoint directory `log_dir/run_id`. If None, a random + string is created. + number_of_steps : Optional[int] + The total number of training iterations. If None, training runs + indefenitely. + loss_mode : Optional[str] + A string that identifies the loss function used for training; must be + one of 'cosine-softmax', 'magnet', 'triplet'. This value defaults to + 'cosine-softmax'. + learning_rate : Optional[float] + Adam learning rate; defaults to 1e-3. + trainable_scopes : Optional[List[str]] + Optional list of variable scopes. If not None, only variables within the + given scopes are trained. Otherwise all variables are trained. + save_summaries_secs : Optional[int] + Save training summaries every `save_summaries_secs` seconds to the + log directory. + save_interval_secs : Optional[int] + Save checkpoints every `save_interval_secs` seconds to the log + directory. + + """ + if image_shape is None: + # If image_shape is not set, train_x must be an image array. Here we + # query the image shape from the array of images. + assert type(train_x) == np.ndarray + image_shape = train_x.shape[1:] + elif type(train_x) == np.ndarray: + assert train_x.shape[1:] == image_shape + read_from_file = type(train_x) != np.ndarray + + trainer, train_op = create_trainer( + preprocess_fn, network_factory, read_from_file, image_shape, batch_size, + loss_mode, learning_rate=learning_rate, + trainable_scopes=trainable_scopes) + + feed_generator = queued_trainer.random_sample_identities_forever( + batch_size, num_images_per_id, train_x, train_y) + + variables_to_restore = slim.get_variables_to_restore( + exclude=exclude_from_restore) + trainer.run( + feed_generator, train_op, log_dir, restore_path=restore_path, + variables_to_restore=variables_to_restore, + run_id=run_id, save_summaries_secs=save_summaries_secs, + save_interval_secs=save_interval_secs, number_of_steps=number_of_steps) + + +def create_trainer(preprocess_fn, network_factory, read_from_file, image_shape, + batch_size, loss_mode, learning_rate=1e-3, + trainable_scopes=None): + """Create trainer. + + Parameters + ---------- + preprocess_fn : Callable[tf.Tensor] -> tf.Tensor + A callable that applies preprocessing to a given input image tensor of + dtype tf.uint8 and returns a floating point representation (tf.float32). + network_factory : Callable[tf.Tensor] -> (tf.Tensor, tf.Tensor) + A callable that takes as argument a preprocessed input image of dtype + tf.float32 and returns the feature representation as well as a logits + tensors. The logits may be set to None if not required by the loss. + read_from_file: + Set to True if images are read from file. If False, the trainer expects + input images as numpy arrays (i.e., data loading must be handled outside + of the trainer). + image_shape: Tuple[int, int, int] + Image shape (height, width, channels). + batch_size: + Number of images per batch. + loss_mode : str + One of 'cosine-softmax', 'magnet', 'triplet'. If 'cosine-softmax', the + logits tensor returned by the `network_factory` must not be None. + learning_rate: float + Adam learning rate; defauls to 1e-3. + trainable_scopes: Optional[List[str]] + Optional list of variable scopes. If not None, only variables within the + given scopes are trained. Otherwise all variables are trained. + + Returns + ------- + QueuedTrainer + Returns a trainer object to be used for training and evaluating the + given TensorFlow model. + + """ + num_channels = image_shape[-1] if len(image_shape) == 3 else 1 + + with tf.device('/cpu:0'): + label_var = tf.placeholder(tf.int64, (None,)) + + if read_from_file: + # NOTE(nwojke): tf.image.decode_jpg handles various image types. + filename_var = tf.placeholder(tf.string, (None, )) + image_var = tf.map_fn( + lambda x: tf.image.decode_jpeg( + tf.read_file(x), channels=num_channels), + filename_var, back_prop=False, dtype=tf.uint8) + image_var = tf.image.resize_images(image_var, image_shape[:2]) + input_vars = [filename_var, label_var] + else: + image_var = tf.placeholder(tf.uint8, (None,) + image_shape) + input_vars = [image_var, label_var] + + enqueue_vars = [ + tf.map_fn( + lambda x: preprocess_fn(x, is_training=True), + image_var, back_prop=False, dtype=tf.float32), + label_var] + + trainer = queued_trainer.QueuedTrainer(enqueue_vars, input_vars) + image_var, label_var = trainer.get_input_vars(batch_size) + tf.summary.image("images", image_var) + + feature_var, logit_var = network_factory(image_var) + _create_loss(feature_var, logit_var, label_var, mode=loss_mode) + + if trainable_scopes is None: + variables_to_train = tf.trainable_variables() + else: + variables_to_train = [] + for scope in trainable_scopes: + variables = tf.get_collection( + tf.GraphKeys.TRAINABLE_VARIABLES, scope) + variables_to_train.extend(variables) + + global_step = tf.train.get_or_create_global_step() + + loss_var = tf.losses.get_total_loss() + + train_optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate) + + opt_tmp = train_optimizer + + loss_scale_manager = ExponentialUpdateLossScaleManager(init_loss_scale=2**32, incr_every_n_steps=1000, decr_every_n_nan_or_inf=2, decr_ratio=0.5) + opt_tmp = NPULossScaleOptimizer(opt_tmp, loss_scale_manager) + + #train_op = slim.learning.create_train_op( + # loss_var, tf.train.AdamOptimizer(opt_tmp), + # global_step, summarize_gradients=False, + # variables_to_train=variables_to_train) + + train_op = slim.learning.create_train_op( + loss_var, opt_tmp, + global_step, summarize_gradients=False, + variables_to_train=variables_to_train) + + #train_op = slim.learning.create_train_op( + # loss_var, npu_tf_optimizer(tf.train.AdamOptimizer(learning_rate=learning_rate)), + # global_step, summarize_gradients=False, + # variables_to_train=variables_to_train) + tf.summary.scalar("total_loss", loss_var) + tf.summary.scalar("learning_rate", learning_rate) + + + regularization_var = tf.reduce_sum(tf.losses.get_regularization_loss()) + tf.summary.scalar("weight_loss", regularization_var) + return trainer, train_op + + +def eval_loop(preprocess_fn, network_factory, data_x, data_y, camera_indices, + log_dir, eval_log_dir, image_shape=None, run_id=None, + loss_mode="cosine-softmax", num_galleries=10, random_seed=4321): + """Evaluate a running training session using CMC metric averaged over + `num_galleries` galleries where each gallery contains for every identity a + randomly selected image-pair. + + A call to this function will block indefinitely, monitoring the + `log_dir/run_id` for saved checkpoints. Then, creates summaries in + `eval_log_dir/run_id` that can be monitored with tensorboard. + + Parameters + ---------- + preprocess_fn : Callable[tf.Tensor] -> tf.Tensor + A callable that applies preprocessing to a given input image tensor of + dtype tf.uint8 and returns a floating point representation (tf.float32). + network_factory : Callable[tf.Tensor] -> (tf.Tensor, tf.Tensor) + A callable that takes as argument a preprocessed input image of dtype + tf.float32 and returns the feature representation as well as a logits + tensors. The logits may be set to None if not required by the loss. + data_x : List[str] | np.ndarray + A list of image filenames or a tensor of images. + data_y : List[int] | np.ndarray + A list or one-dimensional array of labels for the images in `data_x`. + camera_indices: Optional[List[int] | np.ndarray] + A list or one-dimensional array of camera indices for the images in + `data_x`. If not None, CMC galleries are created such that image pairs + are collected from different cameras. + log_dir: str + Should be equivalent to the `log_dir` passed into `train_loop` of the + training run to monitor. + eval_log_dir: + Used to construct the tensorboard log directory where metrics are + summarized. + image_shape : Tuple[int, int, int] | NoneType + Image shape (height, width, channels) or None. If None, `train_x` must + be an array of images such that the shape can be queries from this + variable. + run_id : str + A string that identifies the training run; must be set to the same + `run_id` passed into `train_loop`. + loss_mode : Optional[str] + A string that identifies the loss function used for training; must be + one of 'cosine-softmax', 'magnet', 'triplet'. This value defaults to + 'cosine-softmax'. + num_galleries: int + The number of galleries to be constructed for evaluation of CMC + metrics. + random_seed: Optional[int] + If not None, the NumPy random seed is fixed to this number; can be used + to produce the same galleries over multiple runs. + + """ + if image_shape is None: + # If image_shape is not set, train_x must be an image array. Here we + # query the image shape from the array of images. + assert type(data_x) == np.ndarray + image_shape = data_x.shape[1:] + elif type(data_x) == np.ndarray: + assert data_x.shape[1:] == image_shape + read_from_file = type(data_x) != np.ndarray + + # Create num_galleries random CMC galleries to average CMC top-k over. + probes, galleries = [], [] + for i in range(num_galleries): + probe_indices, gallery_indices = util.create_cmc_probe_and_gallery( + data_y, camera_indices, seed=random_seed + i) + probes.append(probe_indices) + galleries.append(gallery_indices) + probes, galleries = np.asarray(probes), np.asarray(galleries) + + # Set up the data feed. + with tf.device('/cpu:0'): + # Feed probe and gallery indices to the trainer. + num_probes, num_gallery_images = probes.shape[1], galleries.shape[1] + probe_idx_var = tf.placeholder(tf.int64, (None, num_probes)) + gallery_idx_var = tf.placeholder(tf.int64, (None, num_gallery_images)) + trainer = queued_trainer.QueuedTrainer( + [probe_idx_var, gallery_idx_var]) + + # Retrieve indices from trainer and gather data from constant memory. + data_x_var = tf.constant(data_x) + data_y_var = tf.constant(data_y) + + probe_idx_var, gallery_idx_var = trainer.get_input_vars(batch_size=1) + probe_idx_var = tf.squeeze(probe_idx_var) + gallery_idx_var = tf.squeeze(gallery_idx_var) + + # Apply preprocessing. + probe_x_var = tf.gather(data_x_var, probe_idx_var) + if read_from_file: + # NOTE(nwojke): tf.image.decode_jpg handles various image types. + num_channels = image_shape[-1] if len(image_shape) == 3 else 1 + probe_x_var = tf.map_fn( + lambda x: tf.image.decode_jpeg( + tf.read_file(x), channels=num_channels), + probe_x_var, dtype=tf.uint8) + probe_x_var = tf.image.resize_images(probe_x_var, image_shape[:2]) + probe_x_var = tf.map_fn( + lambda x: preprocess_fn(x, is_training=False), + probe_x_var, back_prop=False, dtype=tf.float32) + probe_y_var = tf.gather(data_y_var, probe_idx_var) + + gallery_x_var = tf.gather(data_x_var, gallery_idx_var) + if read_from_file: + # NOTE(nwojke): tf.image.decode_jpg handles various image types. + num_channels = image_shape[-1] if len(image_shape) == 3 else 1 + gallery_x_var = tf.map_fn( + lambda x: tf.image.decode_jpeg( + tf.read_file(x), channels=num_channels), + gallery_x_var, dtype=tf.uint8) + gallery_x_var = tf.image.resize_images( + gallery_x_var, image_shape[:2]) + gallery_x_var = tf.map_fn( + lambda x: preprocess_fn(x, is_training=False), + gallery_x_var, back_prop=False, dtype=tf.float32) + gallery_y_var = tf.gather(data_y_var, gallery_idx_var) + + # Construct the network and compute features. + probe_and_gallery_x_var = tf.concat( + axis=0, values=[probe_x_var, gallery_x_var]) + probe_and_gallery_x_var, _ = network_factory(probe_and_gallery_x_var) + + num_probe = tf.shape(probe_x_var)[0] + probe_x_var = tf.slice( + probe_and_gallery_x_var, [0, 0], [num_probe, -1]) + gallery_x_var = tf.slice( + probe_and_gallery_x_var, [num_probe, 0], [-1, -1]) + + # Set up the metrics. + distance_measure = ( + metrics.cosine_distance if loss_mode == "cosine-softmax" + else metrics.pdist) + + def cmc_metric_at_k(k): + return metrics.streaming_mean_cmc_at_k( + probe_x_var, probe_y_var, gallery_x_var, gallery_y_var, + k=k, measure=distance_measure) + + names_to_values, names_to_updates = slim.metrics.aggregate_metric_map({ + "Precision@%d" % k: cmc_metric_at_k(k) for k in [1, 5, 10, 20]}) + for metric_name, metric_value in names_to_values.items(): + tf.summary.scalar(metric_name, metric_value) + + # Start evaluation loop. + trainer.evaluate( + (probes, galleries), log_dir, eval_log_dir, run_id=run_id, + eval_op=list(names_to_updates.values()), eval_interval_secs=60) + + +def finalize(preprocess_fn, network_factory, checkpoint_path, image_shape, + output_filename): + """Finalize model, i.e., strip off training variables and only save model + variables to checkpoint file. + + Parameters + ---------- + preprocess_fn : Callable[tf.Tensor] -> tf.Tensor + A callable that applies preprocessing to a given input image tensor of + dtype tf.uint8 and returns a floating point representation (tf.float32). + network_factory : Callable[tf.Tensor] -> (tf.Tensor, tf.Tensor) + A callable that takes as argument a preprocessed input image of dtype + tf.float32 and returns the feature representation as well as a logits + tensors. The logits may be set to None if not required by the loss. + checkpoint_path : str + The checkpoint file to load. + image_shape : Tuple[int, int, int] + Image shape (height, width, channels). + output_filename : str + The checkpoint file to write. + + """ + with tf.Session(config=npu_config_proto(), graph=tf.Graph()) as session: + input_var = tf.placeholder(tf.uint8, (None, ) + image_shape) + image_var = tf.map_fn( + lambda x: preprocess_fn(x, is_training=False), + input_var, back_prop=False, dtype=tf.float32) + network_factory(image_var) + + loader = tf.train.Saver(slim.get_variables_to_restore()) + loader.restore(session, checkpoint_path) + + saver = tf.train.Saver(slim.get_model_variables()) + saver.save(session, output_filename, global_step=None) + + +def freeze(preprocess_fn, network_factory, checkpoint_path, image_shape, + output_filename, input_name="images", feature_name="features"): + """Write frozen inference graph that takes as input a list of images and + returns their feature representation. + + Parameters + ---------- + preprocess_fn : Callable[tf.Tensor] -> tf.Tensor + A callable that applies preprocessing to a given input image tensor of + dtype tf.uint8 and returns a floating point representation (tf.float32). + network_factory : Callable[tf.Tensor] -> (tf.Tensor, tf.Tensor) + A callable that takes as argument a preprocessed input image of dtype + tf.float32 and returns the feature representation as well as a logits + tensors. The logits may be set to None if not required by the loss. + checkpoint_path : str + The checkpoint file to load. + image_shape : Tuple[int, int, int] + Image shape (height, width, channels). + output_filename : str + Path to the file to write to. + input_name : Optional[str] + The input (image) placeholder will be given this name; defaults + to `images`. + feature_name : Optional[str] + The output (feature) tensor will be given this name; defaults to + `features`. + + """ + with tf.Session(config=npu_config_proto(), graph=tf.Graph()) as session: + input_var = tf.placeholder( + tf.uint8, (None, ) + image_shape, name=input_name) + image_var = tf.map_fn( + lambda x: preprocess_fn(x, is_training=False), + input_var, back_prop=False, dtype=tf.float32) + features, _ = network_factory(image_var) + features = tf.identity(features, name=feature_name) + + saver = tf.train.Saver(slim.get_variables_to_restore()) + saver.restore(session, checkpoint_path) + + output_graph_def = tf.graph_util.convert_variables_to_constants( + session, tf.get_default_graph().as_graph_def(), + [features.name.split(":")[0]]) + with tf.gfile.GFile(output_filename, "wb") as file_handle: + file_handle.write(output_graph_def.SerializeToString()) + + +def encode(preprocess_fn, network_factory, checkpoint_path, images_or_filenames, + batch_size=32, session=None, image_shape=None): + """ + + Parameters + ---------- + preprocess_fn : Callable[tf.Tensor] -> tf.Tensor + A callable that applies preprocessing to a given input image tensor of + dtype tf.uint8 and returns a floating point representation (tf.float32). + network_factory : Callable[tf.Tensor] -> (tf.Tensor, tf.Tensor) + A callable that takes as argument a preprocessed input image of dtype + tf.float32 and returns the feature representation as well as a logits + tensors. The logits may be set to None if not required by the loss. + checkpoint_path : str + Checkpoint file to load. + images_or_filenames : List[str] | np.ndarray + Either a list of filenames or an array of images. + batch_size : Optional[int] + Optional batch size; defaults to 32. + session : Optional[tf.Session] + Optional TensorFlow session. If None, a new session is created. + image_shape : Tuple[int, int, int] | NoneType + Image shape (height, width, channels) or None. If None, `train_x` must + be an array of images such that the shape can be queries from this + variable. + + Returns + ------- + np.ndarray + + """ + if image_shape is None: + assert type(images_or_filenames) == np.ndarray + image_shape = images_or_filenames.shape[1:] + elif type(images_or_filenames) == np.ndarray: + assert images_or_filenames.shape[1:] == image_shape + read_from_file = type(images_or_filenames) != np.ndarray + + encoder_fn = _create_encoder( + preprocess_fn, network_factory, image_shape, batch_size, session, + checkpoint_path, read_from_file) + features = encoder_fn(images_or_filenames) + return features + + +def _create_encoder(preprocess_fn, network_factory, image_shape, batch_size=32, + session=None, checkpoint_path=None, read_from_file=False): + if read_from_file: + num_channels = image_shape[-1] if len(image_shape) == 3 else 1 + input_var = tf.placeholder(tf.string, (None, )) + image_var = tf.map_fn( + lambda x: tf.image.decode_jpeg( + tf.read_file(x), channels=num_channels), + input_var, back_prop=False, dtype=tf.uint8) + image_var = tf.image.resize_images(image_var, image_shape[:2]) + else: + input_var = tf.placeholder(tf.uint8, (None, ) + image_shape) + image_var = input_var + + preprocessed_image_var = tf.map_fn( + lambda x: preprocess_fn(x, is_training=False), + image_var, back_prop=False, dtype=tf.float32) + + feature_var, _ = network_factory(preprocessed_image_var) + feature_dim = feature_var.get_shape().as_list()[-1] + + if session is None: + session = tf.Session(config=npu_config_proto()) + if checkpoint_path is not None: + tf.train.get_or_create_global_step() + init_assign_op, init_feed_dict = slim.assign_from_checkpoint( + checkpoint_path, slim.get_model_variables()) + session.run(init_assign_op, feed_dict=init_feed_dict) + + def encoder(data_x): + out = np.zeros((len(data_x), feature_dim), np.float32) + queued_trainer.run_in_batches( + lambda x: session.run(feature_var, feed_dict=x), + {input_var: data_x}, out, batch_size) + return out + + return encoder + + +def _create_softmax_loss(feature_var, logit_var, label_var): + del feature_var # Unused variable + cross_entropy_var = slim.losses.sparse_softmax_cross_entropy( + logit_var, tf.cast(label_var, tf.int64)) + tf.summary.scalar("cross_entropy_loss", cross_entropy_var) + + accuracy_var = slim.metrics.accuracy( + tf.cast(tf.argmax(logit_var, 1), tf.int64), label_var) + tf.summary.scalar("classification accuracy", accuracy_var) + + +def _create_magnet_loss(feature_var, logit_var, label_var, monitor_mode=False): + del logit_var # Unusued variable + magnet_loss, _, _ = losses.magnet_loss(feature_var, label_var) + tf.summary.scalar("magnet_loss", magnet_loss) + if not monitor_mode: + slim.losses.add_loss(magnet_loss) + + +def _create_triplet_loss(feature_var, logit_var, label_var, monitor_mode=False): + del logit_var # Unusued variables + triplet_loss = losses.softmargin_triplet_loss(feature_var, label_var) + tf.summary.scalar("triplet_loss", triplet_loss) + if not monitor_mode: + slim.losses.add_loss(triplet_loss) + + +def _create_loss( + feature_var, logit_var, label_var, mode, monitor_magnet=True, + monitor_triplet=True): + if mode == "cosine-softmax": + _create_softmax_loss(feature_var, logit_var, label_var) + elif mode == "magnet": + _create_magnet_loss(feature_var, logit_var, label_var) + elif mode == "triplet": + _create_triplet_loss(feature_var, logit_var, label_var) + else: + raise ValueError("Unknown loss mode: '%s'" % mode) + + if monitor_magnet and mode != "magnet": + _create_magnet_loss( + feature_var, logit_var, label_var, monitor_mode=monitor_magnet) + if monitor_triplet and mode != "triplet": + _create_triplet_loss( + feature_var, logit_var, label_var, monitor_mode=True) diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/train_market1501.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/train_market1501.py new file mode 100644 index 000000000..21667df89 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/train_market1501.py @@ -0,0 +1,139 @@ +# vim: expandtab:ts=4:sw=4 +from npu_bridge.npu_init import * +import functools +import os +import numpy as np +import scipy.io as sio +import train_app +from datasets import market1501 +from datasets import util +import nets.deep_sort.network_definition as net +import moxing as mox +from npu_bridge.estimator import npu_ops +from tensorflow.core.protobuf import rewriter_config_pb2 +import argparse + + + + +class Market1501(object): + + def __init__(self, data_url, num_validation_y=0.1, seed=1234): + self._data_url = data_url + self._num_validation_y = num_validation_y + self._seed = seed + + def read_train(self): + filenames, ids, camera_indices = market1501.read_train_split_to_str( + self._data_url) + train_indices, _ = util.create_validation_split( + np.asarray(ids, np.int64), self._num_validation_y, self._seed) + + filenames = [filenames[i] for i in train_indices] + ids = [ids[i] for i in train_indices] + camera_indices = [camera_indices[i] for i in train_indices] + return filenames, ids, camera_indices + + def read_validation(self): + filenames, ids, camera_indices = market1501.read_train_split_to_str( + self._data_url) + _, valid_indices = util.create_validation_split( + np.asarray(ids, np.int64), self._num_validation_y, self._seed) + + filenames = [filenames[i] for i in valid_indices] + ids = [ids[i] for i in valid_indices] + camera_indices = [camera_indices[i] for i in valid_indices] + return filenames, ids, camera_indices + + def read_test(self): + return market1501.read_test_split_to_str(self._data_url) + + +def main(): + + + arg_parser = train_app.create_default_argument_parser("market1501") + arg_parser.add_argument( + "--dataset_dir", help="Path to Market1501 dataset directory.", + default="resources/Market-1501-v15.09.15") + arg_parser.add_argument( + "--sdk_dir", help="Path to Market1501 baseline evaluation software.", + default="resources/Market-1501-v15.09.15-baseline") + arg_parser.add_argument( + "--device_id", type=int, default=4, help="Device id, default is 0.") + args = arg_parser.parse_args() + dataset = Market1501(args.dataset_dir, num_validation_y=0.1, seed=1234) + + + + if args.mode == "train": + train_x, train_y, _ = dataset.read_train() + print("Train set size: %d images, %d identities" % ( + len(train_x), len(np.unique(train_y)))) + + network_factory = net.create_network_factory( + is_training=True, num_classes=market1501.MAX_LABEL + 1, + add_logits=args.loss_mode == "cosine-softmax") + train_kwargs = train_app.to_train_kwargs(args) + train_app.train_loop( + net.preprocess, network_factory, train_x, train_y, + num_images_per_id=4, image_shape=market1501.IMAGE_SHAPE, + **train_kwargs) + + elif args.mode == "export": + # Export one specific model. + gallery_filenames, _, query_filenames, _, _ = dataset.read_test() + + network_factory = net.create_network_factory( + is_training=False, num_classes=market1501.MAX_LABEL + 1, + add_logits=False, reuse=None) + gallery_features = train_app.encode( + net.preprocess, network_factory, args.restore_path, + gallery_filenames, image_shape=market1501.IMAGE_SHAPE) + sio.savemat( + os.path.join(args.sdk_dir, "feat_test.mat"), + {"features": gallery_features}) + + network_factory = net.create_network_factory( + is_training=False, num_classes=market1501.MAX_LABEL + 1, + add_logits=False, reuse=True) + query_features = train_app.encode( + net.preprocess, network_factory, args.restore_path, + query_filenames, image_shape=market1501.IMAGE_SHAPE) + sio.savemat( + os.path.join(args.sdk_dir, "feat_query.mat"), + {"features": query_features}) + elif args.mode == "finalize": + network_factory = net.create_network_factory( + is_training=False, num_classes=market1501.MAX_LABEL + 1, + add_logits=False, reuse=None) + train_app.finalize( + functools.partial(net.preprocess, input_is_bgr=True), + network_factory, args.restore_path, + image_shape=market1501.IMAGE_SHAPE, + output_filename="./market1501.ckpt") + elif args.mode == "freeze": + network_factory = net.create_network_factory( + is_training=False, num_classes=market1501.MAX_LABEL + 1, + add_logits=False, reuse=None) + train_app.freeze( + functools.partial(net.preprocess, input_is_bgr=True), + network_factory, args.restore_path, + image_shape=market1501.IMAGE_SHAPE, + output_filename="./market1501.pb") + else: + raise ValueError("Invalid mode argument.") + # 在ModelArts容器创建训练输出目录 + model_dir = "/cache/result" + os.makedirs(model_dir) + # 训练结束后，将ModelArts容器内的训练输出拷贝到OBS + mox.file.copy_parallel(model_dir, arg.train_url) + + +if __name__ == "__main__": + #(npu_sess, npu_shutdown) = init_resource() + main() + + #shutdown_resource(npu_sess, npu_shutdown) + #close_session(npu_sess) + diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/train_mars.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/train_mars.py new file mode 100644 index 000000000..71d8356fb --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/train_mars.py @@ -0,0 +1,124 @@ +# vim: expandtab:ts=4:sw=4 +from npu_bridge.npu_init import * +import functools +import os +import numpy as np +import scipy.io as sio +import train_app +from datasets import mars +from datasets import util +import nets.deep_sort.network_definition as net + + +IMAGE_SHAPE = 128, 64, 3 + + +class Mars(object): + + def __init__(self, dataset_dir, num_validation_y=0.1, seed=1234): + self._dataset_dir = dataset_dir + self._num_validation_y = num_validation_y + self._seed = seed + + def read_train(self): + filenames, ids, camera_indices, _ = mars.read_train_split_to_str( + self._dataset_dir) + train_indices, _ = util.create_validation_split( + np.asarray(ids, np.int64), self._num_validation_y, self._seed) + + filenames = [filenames[i] for i in train_indices] + ids = [ids[i] for i in train_indices] + camera_indices = [camera_indices[i] for i in train_indices] + return filenames, ids, camera_indices + + def read_validation(self): + filenames, ids, camera_indices, _ = mars.read_train_split_to_str( + self._dataset_dir) + _, valid_indices = util.create_validation_split( + np.asarray(ids, np.int64), self._num_validation_y, self._seed) + + filenames = [filenames[i] for i in valid_indices] + ids = [ids[i] for i in valid_indices] + camera_indices = [camera_indices[i] for i in valid_indices] + return filenames, ids, camera_indices + + def read_test_filenames(self): + filename = os.path.join(self._dataset_dir, "info", "test_name.txt") + with open(filename, "r") as file_handle: + content = file_handle.read() + lines = content.splitlines() + + image_dir = os.path.join(self._dataset_dir, "bbox_test") + return [os.path.join(image_dir, f[:4], f) for f in lines] + + +def main(): + arg_parser = train_app.create_default_argument_parser("mars") + arg_parser.add_argument( + "--dataset_dir", help="Path to MARS dataset directory.", + default="resources/MARS-evaluation-master") + args = arg_parser.parse_args() + dataset = Mars(args.dataset_dir, num_validation_y=0.1, seed=1234) + + if args.mode == "train": + train_x, train_y, _ = dataset.read_train() + print("Train set size: %d images, %d identities" % ( + len(train_x), len(np.unique(train_y)))) + + network_factory = net.create_network_factory( + is_training=True, num_classes=mars.MAX_LABEL + 1, + add_logits=args.loss_mode == "cosine-softmax") + train_kwargs = train_app.to_train_kwargs(args) + train_app.train_loop( + net.preprocess, network_factory, train_x, train_y, + num_images_per_id=4, image_shape=IMAGE_SHAPE, **train_kwargs) + elif args.mode == "eval": + valid_x, valid_y, camera_indices = dataset.read_validation() + print("Validation set size: %d images, %d identities" % ( + len(valid_x), len(np.unique(valid_y)))) + + network_factory = net.create_network_factory( + is_training=False, num_classes=mars.MAX_LABEL + 1, + add_logits=args.loss_mode == "cosine-softmax") + eval_kwargs = train_app.to_eval_kwargs(args) + train_app.eval_loop( + net.preprocess, network_factory, valid_x, valid_y, camera_indices, + image_shape=IMAGE_SHAPE, num_galleries=20, **eval_kwargs) + elif args.mode == "export": + filenames = dataset.read_test_filenames() + + network_factory = net.create_network_factory( + is_training=False, num_classes=mars.MAX_LABEL + 1, + add_logits=False, reuse=None) + features = train_app.encode( + net.preprocess, network_factory, args.restore_path, + filenames, image_shape=IMAGE_SHAPE) + sio.savemat( + os.path.join(args.dataset_dir, "feat_test.mat"), + {"features": features}) + elif args.mode == "finalize": + network_factory = net.create_network_factory( + is_training=False, num_classes=mars.MAX_LABEL + 1, + add_logits=False, reuse=None) + train_app.finalize( + functools.partial(net.preprocess, input_is_bgr=True), + network_factory, args.restore_path, image_shape=IMAGE_SHAPE, + output_filename="./mars.ckpt") + elif args.mode == "freeze": + network_factory = net.create_network_factory( + is_training=False, num_classes=mars.MAX_LABEL + 1, + add_logits=False, reuse=None) + train_app.freeze( + functools.partial(net.preprocess, input_is_bgr=True), + network_factory, args.restore_path, image_shape=IMAGE_SHAPE, + output_filename="./mars.pb") + else: + raise ValueError("Invalid mode argument.") + + +if __name__ == "__main__": + (npu_sess, npu_shutdown) = init_resource() + main() + shutdown_resource(npu_sess, npu_shutdown) + close_session(npu_sess) + diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/vis_tools.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/vis_tools.py new file mode 100644 index 000000000..daae7b50d --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/vis_tools.py @@ -0,0 +1,82 @@ +# vim: expandtab:ts=4:sw=4 + +from npu_bridge.npu_init import * +from sklearn.manifold import TSNE +import numpy as np +import cv2 + + +def gray_to_color(img): + if len(img.shape) == 2: + img = np.dstack((img, img, img)) + return img + + +def min_resize(img, size): + """ + Resize an image so that it is size along the minimum spatial dimension. + """ + w, h = map(float, img.shape[:2]) + if min([w, h]) != size: + if w <= h: + img = cv2.resize(img, (int(round((h/w)*size)), int(size))) + else: + img = cv2.resize(img, (int(size), int(round((w/h)*size)))) + return img + + +def image_scatter(features, images, img_res, res=2000, cval=255): + """ + Embeds images via tsne into a scatter plot. + + Parameters + --------- + features: numpy array + Features to visualize + images: list or numpy array + Corresponding images to features. Expects float images from (0,1). + img_res: float or int + Resolution to embed images at + res: float or int + Size of embedding image in pixels + cval: float or numpy array + Background color value + + Returns + ------ + canvas: numpy array + Image of visualization + """ + features = np.copy(features).astype('float64') + images = [gray_to_color(image) for image in images] + images = [min_resize(image, img_res) for image in images] + max_width = max([image.shape[0] for image in images]) + max_height = max([image.shape[1] for image in images]) + + model = TSNE(n_components=2, random_state=0) + f2d = model.fit_transform(features) + + xx = f2d[:, 0] + yy = f2d[:, 1] + x_min, x_max = xx.min(), xx.max() + y_min, y_max = yy.min(), yy.max() + # Fix the ratios + sx = (x_max-x_min) + sy = (y_max-y_min) + if sx > sy: + res_x = sx/float(sy)*res + res_y = res + else: + res_x = res + res_y = sy/float(sx)*res + + canvas = np.ones((res_x+max_width, res_y+max_height, 3))*cval + x_coords = np.linspace(x_min, x_max, res_x) + y_coords = np.linspace(y_min, y_max, res_y) + for x, y, image in zip(xx, yy, images): + w, h = image.shape[:2] + x_idx = np.argmin((x - x_coords)**2) + y_idx = np.argmin((y - y_coords)**2) + canvas[x_idx:x_idx+w, y_idx:y_idx+h] = image + return canvas + -- Gitee From 8cf51f43d97de51b9a049ce16f2bc2ad38d7c510 Mon Sep 17 00:00:00 2001 From: Gently Date: Thu, 24 Mar 2022 14:02:07 +0000 Subject: [PATCH 3/4] =?UTF-8?q?=E6=96=B0=E5=BB=BA=20deep=5Fsort?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../DeepSort_ID0505_for_TensorFlow/nets/deep_sort/.keep | 0 1 file changed, 0 insertions(+), 0 deletions(-) create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/.keep diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/.keep b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/.keep new file mode 100644 index 000000000..e69de29bb -- Gitee From 711dde1ff4ddf75751d8b1c70d53365864185198 Mon Sep 17 00:00:00 2001 From: Gently Date: Thu, 24 Mar 2022 14:02:40 +0000 Subject: [PATCH 4/4] add deep_sort folder to nets folder --- .../nets/deep_sort/__init__.py | 0 .../nets/deep_sort/network_definition.py | 134 ++++++++++++++++++ .../nets/deep_sort/residual_net.py | 86 +++++++++++ 3 files changed, 220 insertions(+) create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/__init__.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/network_definition.py create mode 100644 TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/residual_net.py diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/__init__.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/network_definition.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/network_definition.py new file mode 100644 index 000000000..005da31b6 --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/network_definition.py @@ -0,0 +1,134 @@ +# vim: expandtab:ts=4:sw=4 +from npu_bridge.npu_init import * +import tensorflow as tf +import tensorflow.contrib.slim as slim + +from . import residual_net + + +def create_network(images, num_classes=None, add_logits=True, reuse=None, + create_summaries=True, weight_decay=1e-8): + nonlinearity = tf.nn.elu + conv_weight_init = tf.truncated_normal_initializer(stddev=1e-3) + conv_bias_init = tf.zeros_initializer() + conv_regularizer = slim.l2_regularizer(weight_decay) + fc_weight_init = tf.truncated_normal_initializer(stddev=1e-3) + fc_bias_init = tf.zeros_initializer() + fc_regularizer = slim.l2_regularizer(weight_decay) + + def batch_norm_fn(x): + return slim.batch_norm(x, scope=tf.get_variable_scope().name + "/bn") + + network = images + network = slim.conv2d( + network, 32, [3, 3], stride=1, activation_fn=nonlinearity, + padding="SAME", normalizer_fn=batch_norm_fn, scope="conv1_1", + weights_initializer=conv_weight_init, biases_initializer=conv_bias_init, + weights_regularizer=conv_regularizer) + if create_summaries: + tf.summary.histogram(network.name + "/activations", network) + tf.summary.image("conv1_1/weights", tf.transpose( + slim.get_variables("conv1_1/weights:0")[0], [3, 0, 1, 2]), + max_outputs=128) + network = slim.conv2d( + network, 32, [3, 3], stride=1, activation_fn=nonlinearity, + padding="SAME", normalizer_fn=batch_norm_fn, scope="conv1_2", + weights_initializer=conv_weight_init, biases_initializer=conv_bias_init, + weights_regularizer=conv_regularizer) + if create_summaries: + tf.summary.histogram(network.name + "/activations", network) + + network = slim.max_pool2d( + network, [3, 3], [2, 2], scope="pool1", padding="SAME") + + network = residual_net.residual_block( + network, "conv2_1", nonlinearity, conv_weight_init, conv_bias_init, + conv_regularizer, increase_dim=False, is_first=True, + summarize_activations=create_summaries) + network = residual_net.residual_block( + network, "conv2_3", nonlinearity, conv_weight_init, conv_bias_init, + conv_regularizer, increase_dim=False, + summarize_activations=create_summaries) + + network = residual_net.residual_block( + network, "conv3_1", nonlinearity, conv_weight_init, conv_bias_init, + conv_regularizer, increase_dim=True, + summarize_activations=create_summaries) + network = residual_net.residual_block( + network, "conv3_3", nonlinearity, conv_weight_init, conv_bias_init, + conv_regularizer, increase_dim=False, + summarize_activations=create_summaries) + + network = residual_net.residual_block( + network, "conv4_1", nonlinearity, conv_weight_init, conv_bias_init, + conv_regularizer, increase_dim=True, + summarize_activations=create_summaries) + network = residual_net.residual_block( + network, "conv4_3", nonlinearity, conv_weight_init, conv_bias_init, + conv_regularizer, increase_dim=False, + summarize_activations=create_summaries) + + feature_dim = network.get_shape().as_list()[-1] + print("feature dimensionality: ", feature_dim) + network = slim.flatten(network) + + network = slim.dropout(network, keep_prob=0.6) + network = slim.fully_connected( + network, feature_dim, activation_fn=nonlinearity, + normalizer_fn=batch_norm_fn, weights_regularizer=fc_regularizer, + scope="fc1", weights_initializer=fc_weight_init, + biases_initializer=fc_bias_init) + + features = network + + # Features in rows, normalize axis 1. + features = tf.nn.l2_normalize(features, dim=1) + + if add_logits: + with slim.variable_scope.variable_scope("ball", reuse=reuse): + weights = slim.model_variable( + "mean_vectors", (feature_dim, int(num_classes)), + initializer=tf.truncated_normal_initializer(stddev=1e-3), + regularizer=None) + scale = slim.model_variable( + "scale", (), tf.float32, + initializer=tf.constant_initializer(0., tf.float32), + regularizer=slim.l2_regularizer(1e-1)) + if create_summaries: + tf.summary.scalar("scale", scale) + scale = tf.nn.softplus(scale) + + # Mean vectors in colums, normalize axis 0. + weights_normed = tf.nn.l2_normalize(weights, dim=0) + logits = scale * tf.matmul(features, weights_normed) + else: + logits = None + return features, logits + + +def create_network_factory(is_training, num_classes, add_logits, + weight_decay=1e-8, reuse=None): + + def factory_fn(image): + with slim.arg_scope([slim.batch_norm, slim.dropout], + is_training=is_training): + with slim.arg_scope([slim.conv2d, slim.fully_connected, + slim.batch_norm, slim.layer_norm], + reuse=reuse): + features, logits = create_network( + image, num_classes=num_classes, add_logits=add_logits, + reuse=reuse, create_summaries=is_training, + weight_decay=weight_decay) + return features, logits + + return factory_fn + + +def preprocess(image, is_training=False, input_is_bgr=False): + if input_is_bgr: + image = image[:, :, ::-1] # BGR to RGB + image = tf.divide(tf.cast(image, tf.float32), 255.0) + if is_training: + image = tf.image.random_flip_left_right(image) + return image + diff --git a/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/residual_net.py b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/residual_net.py new file mode 100644 index 000000000..7c22fb5ed --- /dev/null +++ b/TensorFlow/contrib/DeepSort_ID0505_for_TensorFlow/DeepSort_ID0505_for_TensorFlow/nets/deep_sort/residual_net.py @@ -0,0 +1,86 @@ +# vim: expandtab:ts=4:sw=4 +from npu_bridge.npu_init import * +import tensorflow as tf +from tensorflow.contrib import slim + + +def _batch_norm_fn(x, scope=None): + if scope is None: + scope = tf.get_variable_scope().name + "/bn" + return slim.batch_norm(x, scope=scope) + + +def create_link( + incoming, network_builder, scope, nonlinearity=tf.nn.elu, + weights_initializer=tf.truncated_normal_initializer(stddev=1e-3), + regularizer=None, is_first=False, summarize_activations=True): + if is_first: + network = incoming + else: + network = _batch_norm_fn(incoming, scope=scope + "/bn") + network = nonlinearity(network) + if summarize_activations: + tf.summary.histogram(scope+"/activations", network) + + pre_block_network = incoming + post_block_network = network_builder(network, scope) + + incoming_dim = pre_block_network.get_shape().as_list()[-1] + outgoing_dim = post_block_network.get_shape().as_list()[-1] + if incoming_dim != outgoing_dim: + assert outgoing_dim == 2 * incoming_dim, \ + "%d != %d" % (outgoing_dim, 2 * incoming) + projection = slim.conv2d( + incoming, outgoing_dim, 1, 2, padding="SAME", activation_fn=None, + scope=scope+"/projection", weights_initializer=weights_initializer, + biases_initializer=None, weights_regularizer=regularizer) + network = projection + post_block_network + else: + network = incoming + post_block_network + return network + + +def create_inner_block( + incoming, scope, nonlinearity=tf.nn.elu, + weights_initializer=tf.truncated_normal_initializer(1e-3), + bias_initializer=tf.zeros_initializer(), regularizer=None, + increase_dim=False, summarize_activations=True): + n = incoming.get_shape().as_list()[-1] + stride = 1 + if increase_dim: + n *= 2 + stride = 2 + + incoming = slim.conv2d( + incoming, n, [3, 3], stride, activation_fn=nonlinearity, padding="SAME", + normalizer_fn=_batch_norm_fn, weights_initializer=weights_initializer, + biases_initializer=bias_initializer, weights_regularizer=regularizer, + scope=scope + "/1") + if summarize_activations: + tf.summary.histogram(incoming.name + "/activations", incoming) + + incoming = slim.dropout(incoming, keep_prob=0.6) + + incoming = slim.conv2d( + incoming, n, [3, 3], 1, activation_fn=None, padding="SAME", + normalizer_fn=None, weights_initializer=weights_initializer, + biases_initializer=bias_initializer, weights_regularizer=regularizer, + scope=scope + "/2") + return incoming + + +def residual_block(incoming, scope, nonlinearity=tf.nn.elu, + weights_initializer=tf.truncated_normal_initializer(1e3), + bias_initializer=tf.zeros_initializer(), regularizer=None, + increase_dim=False, is_first=False, + summarize_activations=True): + + def network_builder(x, s): + return create_inner_block( + x, s, nonlinearity, weights_initializer, bias_initializer, + regularizer, increase_dim, summarize_activations) + + return create_link( + incoming, network_builder, scope, nonlinearity, weights_initializer, + regularizer, is_first, summarize_activations) + -- Gitee