【众智】【西安电子科技大学】【ID1232】【Deep_sort】NPU执行通过，提交功能验收-modelzoo-AtomGit

Adeepsort/LICENSE+674-0

@@ -0,0 +1,674 @@
	1	+ GNU GENERAL PUBLIC LICENSE
	2	+ Version 3, 29 June 2007
	3	+
	4	+ Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
	5	+ Everyone is permitted to copy and distribute verbatim copies
	6	+ of this license document, but changing it is not allowed.
	7	+
	8	+ Preamble
	9	+
	10	+ The GNU General Public License is a free, copyleft license for
	11	+software and other kinds of works.
	12	+
	13	+ The licenses for most software and other practical works are designed
	14	+to take away your freedom to share and change the works. By contrast,
	15	+the GNU General Public License is intended to guarantee your freedom to
	16	+share and change all versions of a program--to make sure it remains free
	17	+software for all its users. We, the Free Software Foundation, use the
	18	+GNU General Public License for most of our software; it applies also to
	19	+any other work released this way by its authors. You can apply it to
	20	+your programs, too.
	21	+
	22	+ When we speak of free software, we are referring to freedom, not
	23	+price. Our General Public Licenses are designed to make sure that you
	24	+have the freedom to distribute copies of free software (and charge for
	25	+them if you wish), that you receive source code or can get it if you
	26	+want it, that you can change the software or use pieces of it in new
	27	+free programs, and that you know you can do these things.
	28	+
	29	+ To protect your rights, we need to prevent others from denying you
	30	+these rights or asking you to surrender the rights. Therefore, you have
	31	+certain responsibilities if you distribute copies of the software, or if
	32	+you modify it: responsibilities to respect the freedom of others.
	33	+
	34	+ For example, if you distribute copies of such a program, whether
	35	+gratis or for a fee, you must pass on to the recipients the same
	36	+freedoms that you received. You must make sure that they, too, receive
	37	+or can get the source code. And you must show them these terms so they
	38	+know their rights.
	39	+
	40	+ Developers that use the GNU GPL protect your rights with two steps:
	41	+(1) assert copyright on the software, and (2) offer you this License
	42	+giving you legal permission to copy, distribute and/or modify it.
	43	+
	44	+ For the developers' and authors' protection, the GPL clearly explains
	45	+that there is no warranty for this free software. For both users' and
	46	+authors' sake, the GPL requires that modified versions be marked as
	47	+changed, so that their problems will not be attributed erroneously to
	48	+authors of previous versions.
	49	+
	50	+ Some devices are designed to deny users access to install or run
	51	+modified versions of the software inside them, although the manufacturer
	52	+can do so. This is fundamentally incompatible with the aim of
	53	+protecting users' freedom to change the software. The systematic
	54	+pattern of such abuse occurs in the area of products for individuals to
	55	+use, which is precisely where it is most unacceptable. Therefore, we
	56	+have designed this version of the GPL to prohibit the practice for those
	57	+products. If such problems arise substantially in other domains, we
	58	+stand ready to extend this provision to those domains in future versions
	59	+of the GPL, as needed to protect the freedom of users.
	60	+
	61	+ Finally, every program is threatened constantly by software patents.
	62	+States should not allow patents to restrict development and use of
	63	+software on general-purpose computers, but in those that do, we wish to
	64	+avoid the special danger that patents applied to a free program could
	65	+make it effectively proprietary. To prevent this, the GPL assures that
	66	+patents cannot be used to render the program non-free.
	67	+
	68	+ The precise terms and conditions for copying, distribution and
	69	+modification follow.
	70	+
	71	+ TERMS AND CONDITIONS
	72	+
	73	+ 0. Definitions.
	74	+
	75	+ "This License" refers to version 3 of the GNU General Public License.
	76	+
	77	+ "Copyright" also means copyright-like laws that apply to other kinds of
	78	+works, such as semiconductor masks.
	79	+
	80	+ "The Program" refers to any copyrightable work licensed under this
	81	+License. Each licensee is addressed as "you". "Licensees" and
	82	+"recipients" may be individuals or organizations.
	83	+
	84	+ To "modify" a work means to copy from or adapt all or part of the work
	85	+in a fashion requiring copyright permission, other than the making of an
	86	+exact copy. The resulting work is called a "modified version" of the
	87	+earlier work or a work "based on" the earlier work.
	88	+
	89	+ A "covered work" means either the unmodified Program or a work based
	90	+on the Program.
	91	+
	92	+ To "propagate" a work means to do anything with it that, without
	93	+permission, would make you directly or secondarily liable for
	94	+infringement under applicable copyright law, except executing it on a
	95	+computer or modifying a private copy. Propagation includes copying,
	96	+distribution (with or without modification), making available to the
	97	+public, and in some countries other activities as well.
	98	+
	99	+ To "convey" a work means any kind of propagation that enables other
	100	+parties to make or receive copies. Mere interaction with a user through
	101	+a computer network, with no transfer of a copy, is not conveying.
	102	+
	103	+ An interactive user interface displays "Appropriate Legal Notices"
	104	+to the extent that it includes a convenient and prominently visible
	105	+feature that (1) displays an appropriate copyright notice, and (2)
	106	+tells the user that there is no warranty for the work (except to the
	107	+extent that warranties are provided), that licensees may convey the
	108	+work under this License, and how to view a copy of this License. If
	109	+the interface presents a list of user commands or options, such as a
	110	+menu, a prominent item in the list meets this criterion.
	111	+
	112	+ 1. Source Code.
	113	+
	114	+ The "source code" for a work means the preferred form of the work
	115	+for making modifications to it. "Object code" means any non-source
	116	+form of a work.
	117	+
	118	+ A "Standard Interface" means an interface that either is an official
	119	+standard defined by a recognized standards body, or, in the case of
	120	+interfaces specified for a particular programming language, one that
	121	+is widely used among developers working in that language.
	122	+
	123	+ The "System Libraries" of an executable work include anything, other
	124	+than the work as a whole, that (a) is included in the normal form of
	125	+packaging a Major Component, but which is not part of that Major
	126	+Component, and (b) serves only to enable use of the work with that
	127	+Major Component, or to implement a Standard Interface for which an
	128	+implementation is available to the public in source code form. A
	129	+"Major Component", in this context, means a major essential component
	130	+(kernel, window system, and so on) of the specific operating system
	131	+(if any) on which the executable work runs, or a compiler used to
	132	+produce the work, or an object code interpreter used to run it.
	133	+
	134	+ The "Corresponding Source" for a work in object code form means all
	135	+the source code needed to generate, install, and (for an executable
	136	+work) run the object code and to modify the work, including scripts to
	137	+control those activities. However, it does not include the work's
	138	+System Libraries, or general-purpose tools or generally available free
	139	+programs which are used unmodified in performing those activities but
	140	+which are not part of the work. For example, Corresponding Source
	141	+includes interface definition files associated with source files for
	142	+the work, and the source code for shared libraries and dynamically
	143	+linked subprograms that the work is specifically designed to require,
	144	+such as by intimate data communication or control flow between those
	145	+subprograms and other parts of the work.
	146	+
	147	+ The Corresponding Source need not include anything that users
	148	+can regenerate automatically from other parts of the Corresponding
	149	+Source.
	150	+
	151	+ The Corresponding Source for a work in source code form is that
	152	+same work.
	153	+
	154	+ 2. Basic Permissions.
	155	+
	156	+ All rights granted under this License are granted for the term of
	157	+copyright on the Program, and are irrevocable provided the stated
	158	+conditions are met. This License explicitly affirms your unlimited
	159	+permission to run the unmodified Program. The output from running a
	160	+covered work is covered by this License only if the output, given its
	161	+content, constitutes a covered work. This License acknowledges your
	162	+rights of fair use or other equivalent, as provided by copyright law.
	163	+
	164	+ You may make, run and propagate covered works that you do not
	165	+convey, without conditions so long as your license otherwise remains
	166	+in force. You may convey covered works to others for the sole purpose
	167	+of having them make modifications exclusively for you, or provide you
	168	+with facilities for running those works, provided that you comply with
	169	+the terms of this License in conveying all material for which you do
	170	+not control copyright. Those thus making or running the covered works
	171	+for you must do so exclusively on your behalf, under your direction
	172	+and control, on terms that prohibit them from making any copies of
	173	+your copyrighted material outside their relationship with you.
	174	+
	175	+ Conveying under any other circumstances is permitted solely under
	176	+the conditions stated below. Sublicensing is not allowed; section 10
	177	+makes it unnecessary.
	178	+
	179	+ 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
	180	+
	181	+ No covered work shall be deemed part of an effective technological
	182	+measure under any applicable law fulfilling obligations under article
	183	+11 of the WIPO copyright treaty adopted on 20 December 1996, or
	184	+similar laws prohibiting or restricting circumvention of such
	185	+measures.
	186	+
	187	+ When you convey a covered work, you waive any legal power to forbid
	188	+circumvention of technological measures to the extent such circumvention
	189	+is effected by exercising rights under this License with respect to
	190	+the covered work, and you disclaim any intention to limit operation or
	191	+modification of the work as a means of enforcing, against the work's
	192	+users, your or third parties' legal rights to forbid circumvention of
	193	+technological measures.
	194	+
	195	+ 4. Conveying Verbatim Copies.
	196	+
	197	+ You may convey verbatim copies of the Program's source code as you
	198	+receive it, in any medium, provided that you conspicuously and
	199	+appropriately publish on each copy an appropriate copyright notice;
	200	+keep intact all notices stating that this License and any
	201	+non-permissive terms added in accord with section 7 apply to the code;
	202	+keep intact all notices of the absence of any warranty; and give all
	203	+recipients a copy of this License along with the Program.
	204	+
	205	+ You may charge any price or no price for each copy that you convey,
	206	+and you may offer support or warranty protection for a fee.
	207	+
	208	+ 5. Conveying Modified Source Versions.
	209	+
	210	+ You may convey a work based on the Program, or the modifications to
	211	+produce it from the Program, in the form of source code under the
	212	+terms of section 4, provided that you also meet all of these conditions:
	213	+
	214	+ a) The work must carry prominent notices stating that you modified
	215	+ it, and giving a relevant date.
	216	+
	217	+ b) The work must carry prominent notices stating that it is
	218	+ released under this License and any conditions added under section
	219	+ 7. This requirement modifies the requirement in section 4 to
	220	+ "keep intact all notices".
	221	+
	222	+ c) You must license the entire work, as a whole, under this
	223	+ License to anyone who comes into possession of a copy. This
	224	+ License will therefore apply, along with any applicable section 7
	225	+ additional terms, to the whole of the work, and all its parts,
	226	+ regardless of how they are packaged. This License gives no
	227	+ permission to license the work in any other way, but it does not
	228	+ invalidate such permission if you have separately received it.
	229	+
	230	+ d) If the work has interactive user interfaces, each must display
	231	+ Appropriate Legal Notices; however, if the Program has interactive
	232	+ interfaces that do not display Appropriate Legal Notices, your
	233	+ work need not make them do so.
	234	+
	235	+ A compilation of a covered work with other separate and independent
	236	+works, which are not by their nature extensions of the covered work,
	237	+and which are not combined with it such as to form a larger program,
	238	+in or on a volume of a storage or distribution medium, is called an
	239	+"aggregate" if the compilation and its resulting copyright are not
	240	+used to limit the access or legal rights of the compilation's users
	241	+beyond what the individual works permit. Inclusion of a covered work
	242	+in an aggregate does not cause this License to apply to the other
	243	+parts of the aggregate.
	244	+
	245	+ 6. Conveying Non-Source Forms.
	246	+
	247	+ You may convey a covered work in object code form under the terms
	248	+of sections 4 and 5, provided that you also convey the
	249	+machine-readable Corresponding Source under the terms of this License,
	250	+in one of these ways:
	251	+
	252	+ a) Convey the object code in, or embodied in, a physical product
	253	+ (including a physical distribution medium), accompanied by the
	254	+ Corresponding Source fixed on a durable physical medium
	255	+ customarily used for software interchange.
	256	+
	257	+ b) Convey the object code in, or embodied in, a physical product
	258	+ (including a physical distribution medium), accompanied by a
	259	+ written offer, valid for at least three years and valid for as
	260	+ long as you offer spare parts or customer support for that product
	261	+ model, to give anyone who possesses the object code either (1) a
	262	+ copy of the Corresponding Source for all the software in the
	263	+ product that is covered by this License, on a durable physical
	264	+ medium customarily used for software interchange, for a price no
	265	+ more than your reasonable cost of physically performing this
	266	+ conveying of source, or (2) access to copy the
	267	+ Corresponding Source from a network server at no charge.
	268	+
	269	+ c) Convey individual copies of the object code with a copy of the
	270	+ written offer to provide the Corresponding Source. This
	271	+ alternative is allowed only occasionally and noncommercially, and
	272	+ only if you received the object code with such an offer, in accord
	273	+ with subsection 6b.
	274	+
	275	+ d) Convey the object code by offering access from a designated
	276	+ place (gratis or for a charge), and offer equivalent access to the
	277	+ Corresponding Source in the same way through the same place at no
	278	+ further charge. You need not require recipients to copy the
	279	+ Corresponding Source along with the object code. If the place to
	280	+ copy the object code is a network server, the Corresponding Source
	281	+ may be on a different server (operated by you or a third party)
	282	+ that supports equivalent copying facilities, provided you maintain
	283	+ clear directions next to the object code saying where to find the
	284	+ Corresponding Source. Regardless of what server hosts the
	285	+ Corresponding Source, you remain obligated to ensure that it is
	286	+ available for as long as needed to satisfy these requirements.
	287	+
	288	+ e) Convey the object code using peer-to-peer transmission, provided
	289	+ you inform other peers where the object code and Corresponding
	290	+ Source of the work are being offered to the general public at no
	291	+ charge under subsection 6d.
	292	+
	293	+ A separable portion of the object code, whose source code is excluded
	294	+from the Corresponding Source as a System Library, need not be
	295	+included in conveying the object code work.
	296	+
	297	+ A "User Product" is either (1) a "consumer product", which means any
	298	+tangible personal property which is normally used for personal, family,
	299	+or household purposes, or (2) anything designed or sold for incorporation
	300	+into a dwelling. In determining whether a product is a consumer product,
	301	+doubtful cases shall be resolved in favor of coverage. For a particular
	302	+product received by a particular user, "normally used" refers to a
	303	+typical or common use of that class of product, regardless of the status
	304	+of the particular user or of the way in which the particular user
	305	+actually uses, or expects or is expected to use, the product. A product
	306	+is a consumer product regardless of whether the product has substantial
	307	+commercial, industrial or non-consumer uses, unless such uses represent
	308	+the only significant mode of use of the product.
	309	+
	310	+ "Installation Information" for a User Product means any methods,
	311	+procedures, authorization keys, or other information required to install
	312	+and execute modified versions of a covered work in that User Product from
	313	+a modified version of its Corresponding Source. The information must
	314	+suffice to ensure that the continued functioning of the modified object
	315	+code is in no case prevented or interfered with solely because
	316	+modification has been made.
	317	+
	318	+ If you convey an object code work under this section in, or with, or
	319	+specifically for use in, a User Product, and the conveying occurs as
	320	+part of a transaction in which the right of possession and use of the
	321	+User Product is transferred to the recipient in perpetuity or for a
	322	+fixed term (regardless of how the transaction is characterized), the
	323	+Corresponding Source conveyed under this section must be accompanied
	324	+by the Installation Information. But this requirement does not apply
	325	+if neither you nor any third party retains the ability to install
	326	+modified object code on the User Product (for example, the work has
	327	+been installed in ROM).
	328	+
	329	+ The requirement to provide Installation Information does not include a
	330	+requirement to continue to provide support service, warranty, or updates
	331	+for a work that has been modified or installed by the recipient, or for
	332	+the User Product in which it has been modified or installed. Access to a
	333	+network may be denied when the modification itself materially and
	334	+adversely affects the operation of the network or violates the rules and
	335	+protocols for communication across the network.
	336	+
	337	+ Corresponding Source conveyed, and Installation Information provided,
	338	+in accord with this section must be in a format that is publicly
	339	+documented (and with an implementation available to the public in
	340	+source code form), and must require no special password or key for
	341	+unpacking, reading or copying.
	342	+
	343	+ 7. Additional Terms.
	344	+
	345	+ "Additional permissions" are terms that supplement the terms of this
	346	+License by making exceptions from one or more of its conditions.
	347	+Additional permissions that are applicable to the entire Program shall
	348	+be treated as though they were included in this License, to the extent
	349	+that they are valid under applicable law. If additional permissions
	350	+apply only to part of the Program, that part may be used separately
	351	+under those permissions, but the entire Program remains governed by
	352	+this License without regard to the additional permissions.
	353	+
	354	+ When you convey a copy of a covered work, you may at your option
	355	+remove any additional permissions from that copy, or from any part of
	356	+it. (Additional permissions may be written to require their own
	357	+removal in certain cases when you modify the work.) You may place
	358	+additional permissions on material, added by you to a covered work,
	359	+for which you have or can give appropriate copyright permission.
	360	+
	361	+ Notwithstanding any other provision of this License, for material you
	362	+add to a covered work, you may (if authorized by the copyright holders of
	363	+that material) supplement the terms of this License with terms:
	364	+
	365	+ a) Disclaiming warranty or limiting liability differently from the
	366	+ terms of sections 15 and 16 of this License; or
	367	+
	368	+ b) Requiring preservation of specified reasonable legal notices or
	369	+ author attributions in that material or in the Appropriate Legal
	370	+ Notices displayed by works containing it; or
	371	+
	372	+ c) Prohibiting misrepresentation of the origin of that material, or
	373	+ requiring that modified versions of such material be marked in
	374	+ reasonable ways as different from the original version; or
	375	+
	376	+ d) Limiting the use for publicity purposes of names of licensors or
	377	+ authors of the material; or
	378	+
	379	+ e) Declining to grant rights under trademark law for use of some
	380	+ trade names, trademarks, or service marks; or
	381	+
	382	+ f) Requiring indemnification of licensors and authors of that
	383	+ material by anyone who conveys the material (or modified versions of
	384	+ it) with contractual assumptions of liability to the recipient, for
	385	+ any liability that these contractual assumptions directly impose on
	386	+ those licensors and authors.
	387	+
	388	+ All other non-permissive additional terms are considered "further
	389	+restrictions" within the meaning of section 10. If the Program as you
	390	+received it, or any part of it, contains a notice stating that it is
	391	+governed by this License along with a term that is a further
	392	+restriction, you may remove that term. If a license document contains
	393	+a further restriction but permits relicensing or conveying under this
	394	+License, you may add to a covered work material governed by the terms
	395	+of that license document, provided that the further restriction does
	396	+not survive such relicensing or conveying.
	397	+
	398	+ If you add terms to a covered work in accord with this section, you
	399	+must place, in the relevant source files, a statement of the
	400	+additional terms that apply to those files, or a notice indicating
	401	+where to find the applicable terms.
	402	+
	403	+ Additional terms, permissive or non-permissive, may be stated in the
	404	+form of a separately written license, or stated as exceptions;
	405	+the above requirements apply either way.
	406	+
	407	+ 8. Termination.
	408	+
	409	+ You may not propagate or modify a covered work except as expressly
	410	+provided under this License. Any attempt otherwise to propagate or
	411	+modify it is void, and will automatically terminate your rights under
	412	+this License (including any patent licenses granted under the third
	413	+paragraph of section 11).
	414	+
	415	+ However, if you cease all violation of this License, then your
	416	+license from a particular copyright holder is reinstated (a)
	417	+provisionally, unless and until the copyright holder explicitly and
	418	+finally terminates your license, and (b) permanently, if the copyright
	419	+holder fails to notify you of the violation by some reasonable means
	420	+prior to 60 days after the cessation.
	421	+
	422	+ Moreover, your license from a particular copyright holder is
	423	+reinstated permanently if the copyright holder notifies you of the
	424	+violation by some reasonable means, this is the first time you have
	425	+received notice of violation of this License (for any work) from that
	426	+copyright holder, and you cure the violation prior to 30 days after
	427	+your receipt of the notice.
	428	+
	429	+ Termination of your rights under this section does not terminate the
	430	+licenses of parties who have received copies or rights from you under
	431	+this License. If your rights have been terminated and not permanently
	432	+reinstated, you do not qualify to receive new licenses for the same
	433	+material under section 10.
	434	+
	435	+ 9. Acceptance Not Required for Having Copies.
	436	+
	437	+ You are not required to accept this License in order to receive or
	438	+run a copy of the Program. Ancillary propagation of a covered work
	439	+occurring solely as a consequence of using peer-to-peer transmission
	440	+to receive a copy likewise does not require acceptance. However,
	441	+nothing other than this License grants you permission to propagate or
	442	+modify any covered work. These actions infringe copyright if you do
	443	+not accept this License. Therefore, by modifying or propagating a
	444	+covered work, you indicate your acceptance of this License to do so.
	445	+
	446	+ 10. Automatic Licensing of Downstream Recipients.
	447	+
	448	+ Each time you convey a covered work, the recipient automatically
	449	+receives a license from the original licensors, to run, modify and
	450	+propagate that work, subject to this License. You are not responsible
	451	+for enforcing compliance by third parties with this License.
	452	+
	453	+ An "entity transaction" is a transaction transferring control of an
	454	+organization, or substantially all assets of one, or subdividing an
	455	+organization, or merging organizations. If propagation of a covered
	456	+work results from an entity transaction, each party to that
	457	+transaction who receives a copy of the work also receives whatever
	458	+licenses to the work the party's predecessor in interest had or could
	459	+give under the previous paragraph, plus a right to possession of the
	460	+Corresponding Source of the work from the predecessor in interest, if
	461	+the predecessor has it or can get it with reasonable efforts.
	462	+
	463	+ You may not impose any further restrictions on the exercise of the
	464	+rights granted or affirmed under this License. For example, you may
	465	+not impose a license fee, royalty, or other charge for exercise of
	466	+rights granted under this License, and you may not initiate litigation
	467	+(including a cross-claim or counterclaim in a lawsuit) alleging that
	468	+any patent claim is infringed by making, using, selling, offering for
	469	+sale, or importing the Program or any portion of it.
	470	+
	471	+ 11. Patents.
	472	+
	473	+ A "contributor" is a copyright holder who authorizes use under this
	474	+License of the Program or a work on which the Program is based. The
	475	+work thus licensed is called the contributor's "contributor version".
	476	+
	477	+ A contributor's "essential patent claims" are all patent claims
	478	+owned or controlled by the contributor, whether already acquired or
	479	+hereafter acquired, that would be infringed by some manner, permitted
	480	+by this License, of making, using, or selling its contributor version,
	481	+but do not include claims that would be infringed only as a
	482	+consequence of further modification of the contributor version. For
	483	+purposes of this definition, "control" includes the right to grant
	484	+patent sublicenses in a manner consistent with the requirements of
	485	+this License.
	486	+
	487	+ Each contributor grants you a non-exclusive, worldwide, royalty-free
	488	+patent license under the contributor's essential patent claims, to
	489	+make, use, sell, offer for sale, import and otherwise run, modify and
	490	+propagate the contents of its contributor version.
	491	+
	492	+ In the following three paragraphs, a "patent license" is any express
	493	+agreement or commitment, however denominated, not to enforce a patent
	494	+(such as an express permission to practice a patent or covenant not to
	495	+sue for patent infringement). To "grant" such a patent license to a
	496	+party means to make such an agreement or commitment not to enforce a
	497	+patent against the party.
	498	+
	499	+ If you convey a covered work, knowingly relying on a patent license,
	500	+and the Corresponding Source of the work is not available for anyone
	501	+to copy, free of charge and under the terms of this License, through a
	502	+publicly available network server or other readily accessible means,
	503	+then you must either (1) cause the Corresponding Source to be so
	504	+available, or (2) arrange to deprive yourself of the benefit of the
	505	+patent license for this particular work, or (3) arrange, in a manner
	506	+consistent with the requirements of this License, to extend the patent
	507	+license to downstream recipients. "Knowingly relying" means you have
	508	+actual knowledge that, but for the patent license, your conveying the
	509	+covered work in a country, or your recipient's use of the covered work
	510	+in a country, would infringe one or more identifiable patents in that
	511	+country that you have reason to believe are valid.
	512	+
	513	+ If, pursuant to or in connection with a single transaction or
	514	+arrangement, you convey, or propagate by procuring conveyance of, a
	515	+covered work, and grant a patent license to some of the parties
	516	+receiving the covered work authorizing them to use, propagate, modify
	517	+or convey a specific copy of the covered work, then the patent license
	518	+you grant is automatically extended to all recipients of the covered
	519	+work and works based on it.
	520	+
	521	+ A patent license is "discriminatory" if it does not include within
	522	+the scope of its coverage, prohibits the exercise of, or is
	523	+conditioned on the non-exercise of one or more of the rights that are
	524	+specifically granted under this License. You may not convey a covered
	525	+work if you are a party to an arrangement with a third party that is
	526	+in the business of distributing software, under which you make payment
	527	+to the third party based on the extent of your activity of conveying
	528	+the work, and under which the third party grants, to any of the
	529	+parties who would receive the covered work from you, a discriminatory
	530	+patent license (a) in connection with copies of the covered work
	531	+conveyed by you (or copies made from those copies), or (b) primarily
	532	+for and in connection with specific products or compilations that
	533	+contain the covered work, unless you entered into that arrangement,
	534	+or that patent license was granted, prior to 28 March 2007.
	535	+
	536	+ Nothing in this License shall be construed as excluding or limiting
	537	+any implied license or other defenses to infringement that may
	538	+otherwise be available to you under applicable patent law.
	539	+
	540	+ 12. No Surrender of Others' Freedom.
	541	+
	542	+ If conditions are imposed on you (whether by court order, agreement or
	543	+otherwise) that contradict the conditions of this License, they do not
	544	+excuse you from the conditions of this License. If you cannot convey a
	545	+covered work so as to satisfy simultaneously your obligations under this
	546	+License and any other pertinent obligations, then as a consequence you may
	547	+not convey it at all. For example, if you agree to terms that obligate you
	548	+to collect a royalty for further conveying from those to whom you convey
	549	+the Program, the only way you could satisfy both those terms and this
	550	+License would be to refrain entirely from conveying the Program.
	551	+
	552	+ 13. Use with the GNU Affero General Public License.
	553	+
	554	+ Notwithstanding any other provision of this License, you have
	555	+permission to link or combine any covered work with a work licensed
	556	+under version 3 of the GNU Affero General Public License into a single
	557	+combined work, and to convey the resulting work. The terms of this
	558	+License will continue to apply to the part which is the covered work,
	559	+but the special requirements of the GNU Affero General Public License,
	560	+section 13, concerning interaction through a network will apply to the
	561	+combination as such.
	562	+
	563	+ 14. Revised Versions of this License.
	564	+
	565	+ The Free Software Foundation may publish revised and/or new versions of
	566	+the GNU General Public License from time to time. Such new versions will
	567	+be similar in spirit to the present version, but may differ in detail to
	568	+address new problems or concerns.
	569	+
	570	+ Each version is given a distinguishing version number. If the
	571	+Program specifies that a certain numbered version of the GNU General
	572	+Public License "or any later version" applies to it, you have the
	573	+option of following the terms and conditions either of that numbered
	574	+version or of any later version published by the Free Software
	575	+Foundation. If the Program does not specify a version number of the
	576	+GNU General Public License, you may choose any version ever published
	577	+by the Free Software Foundation.
	578	+
	579	+ If the Program specifies that a proxy can decide which future
	580	+versions of the GNU General Public License can be used, that proxy's
	581	+public statement of acceptance of a version permanently authorizes you
	582	+to choose that version for the Program.
	583	+
	584	+ Later license versions may give you additional or different
	585	+permissions. However, no additional obligations are imposed on any
	586	+author or copyright holder as a result of your choosing to follow a
	587	+later version.
	588	+
	589	+ 15. Disclaimer of Warranty.
	590	+
	591	+ THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
	592	+APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
	593	+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
	594	+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
	595	+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
	596	+PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
	597	+IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
	598	+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
	599	+
	600	+ 16. Limitation of Liability.
	601	+
	602	+ IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
	603	+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
	604	+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
	605	+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
	606	+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
	607	+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
	608	+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
	609	+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
	610	+SUCH DAMAGES.
	611	+
	612	+ 17. Interpretation of Sections 15 and 16.
	613	+
	614	+ If the disclaimer of warranty and limitation of liability provided
	615	+above cannot be given local legal effect according to their terms,
	616	+reviewing courts shall apply local law that most closely approximates
	617	+an absolute waiver of all civil liability in connection with the
	618	+Program, unless a warranty or assumption of liability accompanies a
	619	+copy of the Program in return for a fee.
	620	+
	621	+ END OF TERMS AND CONDITIONS
	622	+
	623	+ How to Apply These Terms to Your New Programs
	624	+
	625	+ If you develop a new program, and you want it to be of the greatest
	626	+possible use to the public, the best way to achieve this is to make it
	627	+free software which everyone can redistribute and change under these terms.
	628	+
	629	+ To do so, attach the following notices to the program. It is safest
	630	+to attach them to the start of each source file to most effectively
	631	+state the exclusion of warranty; and each file should have at least
	632	+the "copyright" line and a pointer to where the full notice is found.
	633	+
	634	+ <one line to give the program's name and a brief idea of what it does.>
	635	+ Copyright (C) <year> <name of author>
	636	+
	637	+ This program is free software: you can redistribute it and/or modify
	638	+ it under the terms of the GNU General Public License as published by
	639	+ the Free Software Foundation, either version 3 of the License, or
	640	+ (at your option) any later version.
	641	+
	642	+ This program is distributed in the hope that it will be useful,
	643	+ but WITHOUT ANY WARRANTY; without even the implied warranty of
	644	+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
	645	+ GNU General Public License for more details.
	646	+
	647	+ You should have received a copy of the GNU General Public License
	648	+ along with this program. If not, see <http://www.gnu.org/licenses/>.
	649	+
	650	+Also add information on how to contact you by electronic and paper mail.
	651	+
	652	+ If the program does terminal interaction, make it output a short
	653	+notice like this when it starts in an interactive mode:
	654	+
	655	+ <program> Copyright (C) <year> <name of author>
	656	+ This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
	657	+ This is free software, and you are welcome to redistribute it
	658	+ under certain conditions; type `show c' for details.
	659	+
	660	+The hypothetical commands `show w' and `show c' should show the appropriate
	661	+parts of the General Public License. Of course, your program's commands
	662	+might be different; for a GUI interface, you would use an "about box".
	663	+
	664	+ You should also get your employer (if you work as a programmer) or school,
	665	+if any, to sign a "copyright disclaimer" for the program, if necessary.
	666	+For more information on this, and how to apply and follow the GNU GPL, see
	667	+<http://www.gnu.org/licenses/>.
	668	+
	669	+ The GNU General Public License does not permit incorporating your program
	670	+into proprietary programs. If your program is a subroutine library, you
	671	+may consider it more useful to permit linking proprietary applications with
	672	+the library. If this is what you want to do, use the GNU Lesser General
	673	+Public License instead of this License. But first, please read
	674	+<http://www.gnu.org/philosophy/why-not-lgpl.html>.

Adeepsort/README.md+165-0

@@ -0,0 +1,165 @@
	1	+# cosine_metric_learning
	2	+
	3	+## Introduction
	4	+
	5	+This repository contains code for training a metric feature representation to be
	6	+used with the [deep_sort tracker](https://github.com/nwojke/deep_sort). The
	7	+approach is described in
	8	+
	9	+ @inproceedings{Wojke2018deep,
	10	+ title={Deep Cosine Metric Learning for Person Re-identification},
	11	+ author={Wojke, Nicolai and Bewley, Alex},
	12	+ booktitle={2018 IEEE Winter Conference on Applications of Computer Vision (WACV)},
	13	+ year={2018},
	14	+ pages={748--756},
	15	+ organization={IEEE},
	16	+ doi={10.1109/WACV.2018.00087}
	17	+ }
	18	+
	19	+Pre-trained models used in the paper can be found
	20	+[here](https://drive.google.com/open?id=13HtkxD6ggcrGJLWaUcqgXl2UO6-p4PK0).
	21	+A preprint of the paper is available [here](http://elib.dlr.de/116408/).
	22	+The repository comes with code to train a model on the
	23	+[Market1501](http://www.liangzheng.org/Project/project_reid.html)
	24	+and [MARS](http://www.liangzheng.com.cn/Project/project_mars.html) datasets.
	25	+
	26	+## Training on Market1501
	27	+
	28	+The following description assumes you have downloaded the Market1501 dataset to
	29	+``./Market-1501-v15.09.15``. The following command starts training
	30	+using the cosine-softmax classifier described in the above paper:
	31	+```
	32	+python train_market1501.py \
	33	+ --dataset_dir=./Market-1501-v15.09.15/ \
	34	+ --loss_mode=cosine-softmax \
	35	+ --log_dir=./output/market1501/ \
	36	+ --run_id=cosine-softmax
	37	+```
	38	+This will create a directory `./output/market1501/cosine-softmax` where
	39	+TensorFlow checkpoints are stored and which can be monitored using
	40	+``tensorboard``:
	41	+```
	42	+tensorboard --logdir ./output/market1501/cosine-softmax --port 6006
	43	+```
	44	+The code splits off 10% of the training data for validation.
	45	+Concurrently to training, run the following command to run CMC evaluation
	46	+metrics on the validation set:
	47	+```
	48	+CUDA_VISIBLE_DEVICES="" python train_market1501.py \
	49	+ --mode=eval \
	50	+ --dataset_dir=./Market-1501-v15.09.15/ \
	51	+ --loss_mode=cosine-softmax \
	52	+ --log_dir=./output/market1501/ \
	53	+ --run_id=cosine-softmax \
	54	+ --eval_log_dir=./eval_output/market1501
	55	+```
	56	+The command will block indefinitely to monitor the training directory for saved
	57	+checkpoints and each stored checkpoint in the training directory is evaluated on
	58	+the validation set. The results of this evaluation are stored in
	59	+``./eval_output/market1501/cosine-softmax`` to be monitored using
	60	+``tensorboard``:
	61	+```
	62	+tensorboard --logdir ./eval_output/market1501/cosine-softmax --port 6007
	63	+```
	64	+
	65	+## Training on MARS
	66	+
	67	+To train on MARS, download the
	68	+[evaluation software](https://github.com/liangzheng06/MARS-evaluation) and
	69	+extract ``bbox_train.zip`` and ``bbox_test.zip`` from the
	70	+[dataset website](http://www.liangzheng.com.cn/Project/project_mars.html)
	71	+into the evaluation software directory. The following description assumes they
	72	+are stored in ``./MARS-evaluation-master/bbox_train`` and
	73	+``./MARS-evaluation-master/bbox_test``. Training can be started with the following
	74	+command:
	75	+```
	76	+python train_mars.py \
	77	+ --dataset_dir=./MARS-evaluation-master \
	78	+ --loss_mode=cosine-softmax \
	79	+ --log_dir=./output/mars/ \
	80	+ --run_id=cosine-softmax
	81	+```
	82	+Again, this will create a directory `./output/mars/cosine-softmax` where
	83	+TensorFlow checkpoints are stored and which can be monitored using
	84	+``tensorboard``:
	85	+```
	86	+tensorboard --logdir ./output/mars/cosine-softmax --port 7006
	87	+```
	88	+As for Market1501, 10% of the training data are split off for validation.
	89	+Concurrently to training, run the following command to run CMC evaluation
	90	+metrics on the validation set:
	91	+```
	92	+CUDA_VISIBLE_DEVICES="" python train_mars.py \
	93	+ --mode=eval \
	94	+ --dataset_dir=./MARS-evaluation-master/ \
	95	+ --loss_mode=cosine-softmax \
	96	+ --log_dir=./output/mars/ \
	97	+ --run_id=cosine-softmax \
	98	+ --eval_log_dir=./eval_output/mars
	99	+```
	100	+Evaluation metrics on the validation set can be monitored with ``tensorboard``
	101	+```
	102	+tensorboard --logdir ./eval_output/mars/cosine-softmax
	103	+```
	104	+
	105	+## Testing
	106	+
	107	+Final model testing has been carried out using evaluation software provided by
	108	+the dataset authors. The training scripts can be used to write features of the
	109	+test split. The following command exports MARS test features to
	110	+``./MARS-evaluation-master/feat_test.mat``
	111	+```
	112	+python train_mars.py \
	113	+ --mode=export \
	114	+ --dataset_dir=./MARS-evaluation-master \
	115	+ --loss_mode=cosine-softmax .\
	116	+ --restore_path=PATH_TO_CHECKPOINT
	117	+```
	118	+where ``PATH_TO_CHECKPOINT`` the checkpoint file to evaluate. Note that the
	119	+evaluation script needs minor adjustments to apply the cosine similarity metric.
	120	+More precisely, change the feature computation in
	121	+``utils/process_box_features.m`` to average pooling (line 8) and apply
	122	+a re-normalization at the end of the file. The modified file should look like
	123	+this:
	124	+```
	125	+function video_feat = process_box_feat(box_feat, video_info)
	126	+
	127	+nVideo = size(video_info, 1);
	128	+video_feat = zeros(size(box_feat, 1), nVideo);
	129	+for n = 1:nVideo
	130	+ feature_set = box_feat(:, video_info(n, 1):video_info(n, 2));
	131	+% video_feat(:, n) = max(feature_set, [], 2); % max pooling
	132	+ video_feat(:, n) = mean(feature_set, 2); % avg pooling
	133	+end
	134	+
	135	+%%% normalize train and test features
	136	+sum_val = sqrt(sum(video_feat.^2));
	137	+for n = 1:size(video_feat, 1)
	138	+ video_feat(n, :) = video_feat(n, :)./sum_val;
	139	+end
	140	+```
	141	+The Market1501 script contains a similar export functionality which can be
	142	+applied in the same way as described for MARS:
	143	+```
	144	+python train_market1501.py \
	145	+ --mode=export \
	146	+ --dataset_dir=./Market-1501-v15.09.15/
	147	+ --sdk_dir=./Market-1501_baseline-v16.01.14/
	148	+ --loss_mode=cosine-softmax \
	149	+ --restore_path=PATH_TO_CHECKPOINT
	150	+```
	151	+This command creates ``./Market-1501_baseline-v16.01.14/feat_query.mat`` and
	152	+``./Market-1501_baseline-v16.01.14/feat_test.mat`` to be used with the
	153	+Market1501 evaluation code.
	154	+
	155	+## Model export
	156	+
	157	+To export your trained model for use with the
	158	+[deep_sort tracker](https://github.com/nwojke/deep_sort), run the following
	159	+command:
	160	+```
	161	+python train_mars.py --mode=freeze --restore_path=PATH_TO_CHECKPOINT
	162	+```
	163	+This will create a ``mars.pb`` file which can be supplied to Deep SORT. Again,
	164	+the Market1501 script contains a similar function.
	165	+##kernel_meta&ckpt&resources （three missed doc）

Adeepsort/boot_modelarts.py+72-0

@@ -0,0 +1,72 @@
	1	+# Copyright 2021 Huawei Technologies Co., Ltd
	2	+#
	3	+# Licensed under the Apache License, Version 2.0 (the "License");
	4	+# you may not use this file except in compliance with the License.
	5	+# You may obtain a copy of the License at
	6	+#
	7	+# http://www.apache.org/licenses/LICENSE-2.0
	8	+#
	9	+# Unless required by applicable law or agreed to in writing, software
	10	+# distributed under the License is distributed on an "AS IS" BASIS,
	11	+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	12	+# See the License for the specific language governing permissions and
	13	+# limitations under the License.
	14	+"""
	15	+This is the boot file for ModelArts platform.
	16	+Firstly, the train datasets are copyed from obs to ModelArts.
	17	+Then, the string of train shell command is concated and using 'os.system()' to execute
	18	+"""
	19	+import os
	20	+import numpy as np
	21	+import argparse
	22	+from help_modelarts import obs_data2modelarts
	23	+from cfg import make_config
	24	+print(os.system('env'))
	25	+
	26	+
	27	+import argparse
	28	+import moxing as mox
	29	+# 解析输入参数data_url
	30	+
	31	+
	32	+
	33	+if __name__ == '__main__':
	34	+ ## Note: the code dir is not the same as work dir on ModelArts Platform!!!
	35	+ code_dir = os.path.dirname(__file__)
	36	+ work_dir = os.getcwd()
	37	+ print("===>>>code_dir:{}, work_dir:{}".format(code_dir, work_dir))
	38	+
	39	+ parser = argparse.ArgumentParser()
	40	+ parser.add_argument("--data_url", type=str, default="./dataset")
	41	+ parser.add_argument("--train_url", type=str, default="./output")
	42	+ config = parser.parse_args()
	43	+ # 在ModelArts容器创建数据存放目录
	44	+ data_dir = "/cache/dataset"
	45	+ model_dir = "/cache/result"
	46	+ os.makedirs(data_dir)
	47	+ os.makedirs(model_dir)
	48	+ # OBS数据拷贝到ModelArts容器内
	49	+ mox.file.copy_parallel(config.data_url, data_dir)
	50	+ mox.file.copy_parallel(model_dir, config.train_url)
	51	+ #parser.add_argument("--train_url", type=str, default="s3://boom/Market-1501-v15.09.15/")
	52	+ #parser.add_argument("--data_url", type=str, default="s3://boom/Market-1501-v15.09.15/")
	53	+ #parser.add_argument("--modelarts_data_dir", type=str, default="/cache/Market-1501-v15.09.15")
	54	+ #parser.add_argument("--modelarts_result_dir", type=str, default="./output")
	55	+
	56	+
	57	+ print("--------config----------")
	58	+ for k in list(vars(config).keys()):
	59	+ print("key:{}: value:{}".format(k, vars(config)[k]))
	60	+ print("--------config----------")
	61	+
	62	+ ## copy dataset from obs to modelarts
	63	+ obs_data2modelarts(config)
	64	+
	65	+ ## start to train on Modelarts platform
	66	+ if not os.path.exists(config.modelarts_result_dir):
	67	+ os.makedirs(config.modelarts_result_dir)
	68	+ bash_header = os.path.join(code_dir, 'scripts/run_1p.sh')
	69	+ arg_url = '%s %s %s %s' % (code_dir, config.modelarts_data_dir, config.modelarts_result_dir)
	70	+ bash_command = 'bash %s %s' % (bash_header, arg_url)
	71	+ print("bash command:", bash_command)
	72	+ os.system(bash_command)

Adeepsort/cfg.py+46-0

@@ -0,0 +1,46 @@
	1	+# Copyright 2021 Huawei Technologies Co., Ltd
	2	+#
	3	+# Licensed under the Apache License, Version 2.0 (the "License");
	4	+# you may not use this file except in compliance with the License.
	5	+# You may obtain a copy of the License at
	6	+#
	7	+# http://www.apache.org/licenses/LICENSE-2.0
	8	+#
	9	+# Unless required by applicable law or agreed to in writing, software
	10	+# distributed under the License is distributed on an "AS IS" BASIS,
	11	+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	12	+# See the License for the specific language governing permissions and
	13	+# limitations under the License.
	14	+"""
	15	+Here, define the configuration of tensorflow session
	16	+For different chips, the config is not the same.
	17	+"""
	18	+import tensorflow as tf
	19	+import os
	20	+
	21	+
	22	+def make_config(FLAGS):
	23	+ chip = FLAGS.chip.lower()
	24	+ tf.logging.info("chip is [%s]", chip)
	25	+
	26	+ if chip == 'cpu':
	27	+ config = tf.ConfigProto()
	28	+ elif chip == 'gpu':
	29	+ config = tf.ConfigProto(allow_soft_placement=True)
	30	+ config.gpu_options.allow_growth = True
	31	+ elif chip == 'npu':
	32	+ from tensorflow.core.protobuf.rewriter_config_pb2 import RewriterConfig
	33	+
	34	+ config = tf.ConfigProto()
	35	+ custom_op = config.graph_options.rewrite_options.custom_optimizers.add()
	36	+ custom_op.name = "NpuOptimizer"
	37	+ config.graph_options.rewrite_options.remapping = RewriterConfig.OFF
	38	+ config.graph_options.rewrite_options.memory_optimization = RewriterConfig.OFF # 必须显式关闭
	39	+ ## Performance Profiling
	40	+ ## refer to link:https://support.huaweicloud.com/Development-tg-cann202training1/atlasprofilingtrain_16_0003.html
	41	+
	42	+
	43	+ else:
	44	+ raise RuntimeError('chip [%s] has not supported' % chip)
	45	+
	46	+ return config

Adeepsort/datasets/__init__.py+1-0

	@@ -0,0 +1 @@
		1	+# vim: expandtab:ts=4:sw=4

Adeepsort/datasets/market1501.py+214-0

@@ -0,0 +1,214 @@
	1	+# vim: expandtab:ts=4:sw=4
	2	+from npu_bridge.npu_init import *
	3	+import os
	4	+import numpy as np
	5	+import cv2
	6	+import scipy.io as sio
	7	+
	8	+
	9	+# The maximum person ID in the dataset.
	10	+MAX_LABEL = 1501
	11	+
	12	+IMAGE_SHAPE = 128, 64, 3
	13	+
	14	+
	15	+def _parse_filename(filename):
	16	+ """Parse meta-information from given filename.
	17	+
	18	+ Parameters
	19	+ ----------
	20	+ filename : str
	21	+ A Market 1501 image filename.
	22	+
	23	+ Returns
	24	+ -------
	25	+ (int, int, str, str) \| NoneType
	26	+ Returns a tuple with the following entries:
	27	+
	28	+ * Unique ID of the individual in the image
	29	+ * Index of the camera which has observed the individual
	30	+ * Filename without extension
	31	+ * File extension
	32	+
	33	+ Returns None if the given filename is not a valid filename.
	34	+
	35	+ """
	36	+ filename_base, ext = os.path.splitext(filename)
	37	+ if '.' in filename_base:
	38	+ # Some images have double filename extensions.
	39	+ filename_base, ext = os.path.splitext(filename_base)
	40	+ if ext != ".jpg":
	41	+ return None
	42	+ person_id, cam_seq, frame_idx, detection_idx = filename_base.split('_')
	43	+ return int(person_id), int(cam_seq[1]), filename_base, ext
	44	+
	45	+
	46	+def read_train_split_to_str(dataset_dir):
	47	+ """Read training data to list of filenames.
	48	+
	49	+ Parameters
	50	+ ----------
	51	+ dataset_dir : str
	52	+ Path to the Market 1501 dataset directory.
	53	+
	54	+ Returns
	55	+ -------
	56	+ (List[str], List[int], List[int])
	57	+ Returns a tuple with the following values:
	58	+
	59	+ * List of image filenames (full path to image files).
	60	+ * List of unique IDs for the individuals in the images.
	61	+ * List of camera indices.
	62	+
	63	+ """
	64	+ filenames, ids, camera_indices = [], [], []
	65	+
	66	+ image_dir = os.path.join(dataset_dir, "bounding_box_train")
	67	+ for filename in sorted(os.listdir(image_dir)):
	68	+ meta_data = _parse_filename(filename)
	69	+ if meta_data is None:
	70	+ # This is not a valid filename (e.g., Thumbs.db).
	71	+ continue
	72	+
	73	+ filenames.append(os.path.join(image_dir, filename))
	74	+ ids.append(meta_data[0])
	75	+ camera_indices.append(meta_data[1])
	76	+
	77	+ return filenames, ids, camera_indices
	78	+
	79	+
	80	+def read_train_split_to_image(dataset_dir):
	81	+ """Read training images to memory. This consumes a lot of memory.
	82	+
	83	+ Parameters
	84	+ ----------
	85	+ dataset_dir : str
	86	+ Path to the Market 1501 dataset directory.
	87	+
	88	+ Returns
	89	+ -------
	90	+ (ndarray, ndarray, ndarray)
	91	+ Returns a tuple with the following values:
	92	+
	93	+ * Tensor of images in BGR color space of shape 128x64x3.
	94	+ * One dimensional array of unique IDs for the individuals in the images.
	95	+ * One dimensional array of camera indices.
	96	+
	97	+ """
	98	+ filenames, ids, camera_indices = read_train_split_to_str(dataset_dir)
	99	+
	100	+ images = np.zeros((len(filenames), 128, 64, 3), np.uint8)
	101	+ for i, filename in enumerate(filenames):
	102	+ images[i] = cv2.imread(filename, cv2.IMREAD_COLOR)
	103	+
	104	+ ids = np.asarray(ids, np.int64)
	105	+ camera_indices = np.asarray(camera_indices, np.int64)
	106	+ return images, ids, camera_indices
	107	+
	108	+
	109	+def read_test_split_to_str(dataset_dir):
	110	+ """Read query and gallery data to list of filenames.
	111	+
	112	+ Parameters
	113	+ ----------
	114	+ dataset_dir : str
	115	+ Path to the Market 1501 dataset directory.
	116	+
	117	+ Returns
	118	+ -------
	119	+ (List[str], List[int], List[str], List[int], ndarray)
	120	+ Returns a tuple with the following values:
	121	+
	122	+ * List of N gallery filenames (full path to image files).
	123	+ * List of N unique IDs for the individuals in the gallery.
	124	+ * List of M query filenames (full path to image files).
	125	+ * List of M unique IDs for the individuals in the queries.
	126	+ * Matrix of shape MxN such that element (i, j) evaluates to 0 if
	127	+ gallery image j should be excluded from metrics computation of
	128	+ query i and 1 otherwise.
	129	+
	130	+ """
	131	+ # Read gallery.
	132	+ gallery_filenames, gallery_ids = [], []
	133	+
	134	+ image_dir = os.path.join(dataset_dir, "bounding_box_test")
	135	+ for filename in sorted(os.listdir(image_dir)):
	136	+ meta_data = _parse_filename(filename)
	137	+ if meta_data is None:
	138	+ # This is not a valid filename (e.g., Thumbs.db).
	139	+ continue
	140	+
	141	+ gallery_filenames.append(os.path.join(image_dir, filename))
	142	+ gallery_ids.append(meta_data[0])
	143	+
	144	+ # Read queries.
	145	+ query_filenames, query_ids, query_junk_indices = [], [], []
	146	+
	147	+ image_dir = os.path.join(dataset_dir, "query")
	148	+ for filename in sorted(os.listdir(image_dir)):
	149	+ meta_data = _parse_filename(filename)
	150	+ if meta_data is None:
	151	+ # This is not a valid filename (e.g., Thumbs.db).
	152	+ continue
	153	+
	154	+ filename_base = meta_data[2]
	155	+ junk_matfile = filename_base + "_junk.mat"
	156	+ mat = sio.loadmat(os.path.join(dataset_dir, "gt_query", junk_matfile))
	157	+ if np.any(mat["junk_index"] < 1):
	158	+ indices = []
	159	+ else:
	160	+ # MATLAB to Python index.
	161	+ indices = list(mat["junk_index"].astype(np.int64).ravel() - 1)
	162	+
	163	+ query_junk_indices.append(indices)
	164	+ query_filenames.append(os.path.join(image_dir, filename))
	165	+ query_ids.append(meta_data[0])
	166	+
	167	+ # The following matrix maps from query (row) to gallery image (column) such
	168	+ # that element (i, j) evaluates to 0 if query i and gallery image j should
	169	+ # be excluded from computation of the evaluation metrics and 1 otherwise.
	170	+ good_mask = np.ones(
	171	+ (len(query_filenames), len(gallery_filenames)), np.float32)
	172	+ for i, junk_indices in enumerate(query_junk_indices):
	173	+ good_mask[i, junk_indices] = 0.
	174	+
	175	+ return gallery_filenames, gallery_ids, query_filenames, query_ids, good_mask
	176	+
	177	+
	178	+def read_test_split_to_image(dataset_dir):
	179	+ """Read query and gallery data to memory. This consumes a lot of memory.
	180	+
	181	+ Parameters
	182	+ ----------
	183	+ dataset_dir : str
	184	+ Path to the Market 1501 dataset directory.
	185	+
	186	+ Returns
	187	+ -------
	188	+ (ndarray, ndarray, ndarray, ndarray, ndarray)
	189	+ Returns a tuple with the following values:
	190	+
	191	+ * Tensor of shape Nx128x64x3 of N gallery images in BGR color space.
	192	+ * One dimensional array of N unique gallery IDs.
	193	+ * Tensor of shape Mx128x64x3 of M query images in BGR color space.
	194	+ * One dimensional array of M unique query IDs.
	195	+ * Matrix of shape MxN such that element (i, j) evaluates to 0 if
	196	+ gallery image j should be excluded from metrics computation of
	197	+ query i and 1 otherwise.
	198	+
	199	+ """
	200	+ gallery_filenames, gallery_ids, query_filenames, query_ids, good_mask = (
	201	+ read_test_split_to_str(dataset_dir))
	202	+
	203	+ gallery_images = np.zeros((len(gallery_filenames), 128, 64, 3), np.uint8)
	204	+ for i, filename in enumerate(gallery_filenames):
	205	+ gallery_images[i] = cv2.imread(filename, cv2.IMREAD_COLOR)
	206	+
	207	+ query_images = np.zeros((len(query_filenames), 128, 64, 3), np.uint8)
	208	+ for i, filename in enumerate(query_filenames):
	209	+ query_images[i] = cv2.imread(filename, cv2.IMREAD_COLOR)
	210	+
	211	+ gallery_ids = np.asarray(gallery_ids, np.int64)
	212	+ query_ids = np.asarray(query_ids, np.int64)
	213	+ return gallery_images, gallery_ids, query_images, query_ids, good_mask
	214	+

Adeepsort/datasets/mars.py+190-0

@@ -0,0 +1,190 @@
	1	+# vim: expandtab:ts=4:sw=4
	2	+from npu_bridge.npu_init import *
	3	+import os
	4	+import numpy as np
	5	+import cv2
	6	+
	7	+
	8	+# The maximum person ID in the dataset.
	9	+MAX_LABEL = 1500
	10	+
	11	+IMAGE_SHAPE = 256, 128, 3
	12	+
	13	+
	14	+def read_train_test_directory_to_str(directory):
	15	+ """Read bbox_train/bbox_test directory.
	16	+
	17	+ Parameters
	18	+ ----------
	19	+ directory : str
	20	+ Path to bbox_train/bbox_test directory.
	21	+
	22	+ Returns
	23	+ -------
	24	+ (List[str], List[int], List[int], List[int])
	25	+ Returns a tuple with the following entries:
	26	+
	27	+ * List of image filenames.
	28	+ * List of corresponding unique IDs for the individuals in the images.
	29	+ * List of camera indices.
	30	+ * List of tracklet indices.
	31	+
	32	+ """
	33	+
	34	+ def to_label(x):
	35	+ return int(x) if x.isdigit() else -1
	36	+
	37	+ dirnames = os.listdir(directory)
	38	+ image_filenames, ids, camera_indices, tracklet_indices = [], [], [], []
	39	+ for dirname in dirnames:
	40	+ filenames = os.listdir(os.path.join(directory, dirname))
	41	+ filenames = [
	42	+ f for f in filenames if os.path.splitext(f)[1] == ".jpg"]
	43	+ image_filenames += [
	44	+ os.path.join(directory, dirname, f) for f in filenames]
	45	+ ids += [to_label(dirname) for _ in filenames]
	46	+ camera_indices += [int(f[5]) for f in filenames]
	47	+ tracklet_indices += [int(f[7:11]) for f in filenames]
	48	+
	49	+ return image_filenames, ids, camera_indices, tracklet_indices
	50	+
	51	+
	52	+def read_train_test_directory_to_image(directory, image_shape=(128, 64)):
	53	+ """Read images in bbox_train/bbox_test directory.
	54	+
	55	+ Parameters
	56	+ ----------
	57	+ directory : str
	58	+ Path to bbox_train/bbox_test directory.
	59	+ image_shape : Tuple[int, int]
	60	+ A tuple (height, width) of the desired image size.
	61	+
	62	+ Returns
	63	+ -------
	64	+ (np.ndarray, np.ndarray, np.ndarray, np.ndarray)
	65	+ Returns a tuple with the following entries:
	66	+
	67	+ * Tensor of images in BGR color space.
	68	+ * One dimensional array of unique IDs for the individuals in the images.
	69	+ * One dimensional array of camera indices.
	70	+ * One dimensional array of tracklet indices.
	71	+
	72	+ """
	73	+ reshape_fn = (
	74	+ (lambda x: x) if image_shape == IMAGE_SHAPE[:2]
	75	+ else (lambda x: cv2.resize(x, image_shape[::-1])))
	76	+
	77	+ filenames, ids, camera_indices, tracklet_indices = (
	78	+ read_train_test_directory_to_str(directory))
	79	+
	80	+ images = np.zeros((len(filenames), ) + image_shape + (3, ), np.uint8)
	81	+ for i, filename in enumerate(filenames):
	82	+ if i % 1000 == 0:
	83	+ print("Reading %s, %d / %d" % (directory, i, len(filenames)))
	84	+ image = cv2.imread(filename, cv2.IMREAD_COLOR)
	85	+ images[i] = reshape_fn(image)
	86	+ ids = np.asarray(ids, dtype=np.int64)
	87	+ camera_indices = np.asarray(camera_indices, dtype=np.int64)
	88	+ tracklet_indices = np.asarray(tracklet_indices, dtype=np.int64)
	89	+ return images, ids, camera_indices, tracklet_indices
	90	+
	91	+
	92	+def read_train_split_to_str(dataset_dir):
	93	+ """Read training data to list of filenames.
	94	+
	95	+ Parameters
	96	+ ----------
	97	+ dataset_dir : str
	98	+ Path to the MARS dataset directory; ``bbox_train`` should be a
	99	+ subdirectory of this folder.
	100	+
	101	+ Returns
	102	+ -------
	103	+ (List[str], List[int], List[int], List[int])
	104	+ Returns a tuple with the following entries:
	105	+
	106	+ * List of image filenames.
	107	+ * List of corresponding unique IDs for the individuals in the images.
	108	+ * List of camera indices.
	109	+ * List of tracklet indices.
	110	+
	111	+ """
	112	+ train_dir = os.path.join(dataset_dir, "bbox_train")
	113	+ return read_train_test_directory_to_str(train_dir)
	114	+
	115	+
	116	+def read_train_split_to_image(dataset_dir, image_shape=(128, 64)):
	117	+ """Read training images to memory. This consumes a lot of memory.
	118	+
	119	+ Parameters
	120	+ ----------
	121	+ dataset_dir : str
	122	+ Path to the MARS dataset directory; ``bbox_train`` should be a
	123	+ subdirectory of this folder.
	124	+ image_shape : Tuple[int, int]
	125	+ A tuple (height, width) of the desired image size.
	126	+
	127	+ Returns
	128	+ -------
	129	+ (np.ndarray, np.ndarray, np.ndarray, np.ndarray)
	130	+ Returns a tuple with the following entries:
	131	+
	132	+ * Tensor of images in BGR color space.
	133	+ * One dimensional array of unique IDs for the individuals in the images.
	134	+ * One dimensional array of camera indices.
	135	+ * One dimensional array of tracklet indices.
	136	+
	137	+ """
	138	+ train_dir = os.path.join(dataset_dir, "bbox_train")
	139	+ return read_train_test_directory_to_image(train_dir, image_shape)
	140	+
	141	+
	142	+def read_test_split_to_str(dataset_dir):
	143	+ """Read training data to list of filenames.
	144	+
	145	+ Parameters
	146	+ ----------
	147	+ dataset_dir : str
	148	+ Path to the MARS dataset directory; ``bbox_test`` should be a
	149	+ subdirectory of this folder.
	150	+
	151	+ Returns
	152	+ -------
	153	+ (List[str], List[int], List[int], List[int])
	154	+ Returns a tuple with the following entries:
	155	+
	156	+ * List of image filenames.
	157	+ * List of corresponding unique IDs for the individuals in the images.
	158	+ * List of camera indices.
	159	+ * List of tracklet indices.
	160	+
	161	+ """
	162	+ test_dir = os.path.join(dataset_dir, "bbox_test")
	163	+ return read_train_test_directory_to_str(test_dir)
	164	+
	165	+
	166	+def read_test_split_to_image(dataset_dir, image_shape=(128, 64)):
	167	+ """Read test images to memory. This consumes a lot of memory.
	168	+
	169	+ Parameters
	170	+ ----------
	171	+ dataset_dir : str
	172	+ Path to the MARS dataset directory; ``bbox_test`` should be a
	173	+ subdirectory of this folder.
	174	+ image_shape : Tuple[int, int]
	175	+ A tuple (height, width) of the desired image size.
	176	+
	177	+ Returns
	178	+ -------
	179	+ (np.ndarray, np.ndarray, np.ndarray, np.ndarray)
	180	+ Returns a tuple with the following entries:
	181	+
	182	+ * Tensor of images in BGR color space.
	183	+ * One dimensional array of unique IDs for the individuals in the images.
	184	+ * One dimensional array of camera indices.
	185	+ * One dimensional array of tracklet indices.
	186	+
	187	+ """
	188	+ test_dir = os.path.join(dataset_dir, "bbox_test")
	189	+ return read_train_test_directory_to_image(test_dir, image_shape)
	190	+

Adeepsort/datasets/util.py+173-0

@@ -0,0 +1,173 @@
	1	+# vim: expandtab:ts=4:sw=4
	2	+from npu_bridge.npu_init import *
	3	+import numpy as np
	4	+import cv2
	5	+
	6	+
	7	+def crop_to_shape(images, patch_shape):
	8	+ """Crop images to desired shape, respecting the target aspect ratio.
	9	+
	10	+ Parameters
	11	+ ----------
	12	+ images : List[ndarray]
	13	+ A list of images in BGR format (dtype np.uint8)
	14	+ patch_shape : (int, int)
	15	+ Target image patch shape (height, width).
	16	+
	17	+ Returns
	18	+ -------
	19	+ ndarray
	20	+ A tensor of output images.
	21	+
	22	+ """
	23	+ assert len(images) > 0, "Empty image list is not allowed."
	24	+ channels = () if len(images[0].shape) == 0 else (images[0].shape[-1], )
	25	+ output_images = np.zeros(
	26	+ (len(images), ) + patch_shape + channels, dtype=np.uint8)
	27	+
	28	+ target_aspect_ratio = float(patch_shape[1]) / patch_shape[0]
	29	+ for i, image in enumerate(images):
	30	+ image_aspect_ratio = float(image.shape[1]) / image.shape[0]
	31	+ if target_aspect_ratio > image_aspect_ratio:
	32	+ # Fix width, modify height.
	33	+ crop_height = image.shape[1] / target_aspect_ratio
	34	+ crop_width = image.shape[1]
	35	+ else:
	36	+ # Fix height, modify width.
	37	+ crop_width = target_aspect_ratio * image.shape[0]
	38	+ crop_height = image.shape[0]
	39	+
	40	+ sx = int((image.shape[1] - crop_width) / 2)
	41	+ sy = int((image.shape[0] - crop_height) / 2)
	42	+ ex = int(min(sx + crop_width, image.shape[1]))
	43	+ ey = int(min(sy + crop_height, image.shape[0]))
	44	+ output_images[i, ...] = cv2.resize(
	45	+ image[sy:ey, sx:ex], patch_shape[::-1],
	46	+ interpolation=cv2.INTER_CUBIC)
	47	+
	48	+ return output_images
	49	+
	50	+
	51	+def create_validation_split(data_y, num_validation_y, seed=None):
	52	+ """Split dataset into training and validation set with disjoint classes.
	53	+
	54	+ Parameters
	55	+ ----------
	56	+ data_y : ndarray
	57	+ A label vector.
	58	+ num_validation_y : int \| float
	59	+ The number of identities to split off for validation. If an integer
	60	+ is given, this value should be at least 1 and is interpreted as absolute
	61	+ number of validation identities. If a float is given, this value should
	62	+ be in [0, 1[ and is interpreted as fraction of validation identities.
	63	+ seed : Optional[int]
	64	+ A random generator seed used to select the validation idenities.
	65	+
	66	+ Returns
	67	+ -------
	68	+ (ndarray, ndarray)
	69	+ Returns indices of training and validation set.
	70	+
	71	+ """
	72	+ unique_y = np.unique(data_y)
	73	+ if isinstance(num_validation_y, float):
	74	+ num_validation_y = int(num_validation_y * len(unique_y))
	75	+
	76	+ random_generator = np.random.RandomState(seed=seed)
	77	+ validation_y = random_generator.choice(
	78	+ unique_y, num_validation_y, replace=False)
	79	+
	80	+ validation_mask = np.full((len(data_y), ), False, bool)
	81	+ for y in validation_y:
	82	+ validation_mask = np.logical_or(validation_mask, data_y == y)
	83	+ training_mask = np.logical_not(validation_mask)
	84	+ return np.where(training_mask)[0], np.where(validation_mask)[0]
	85	+
	86	+
	87	+def limit_num_elements_per_identity(data_y, max_num_images_per_id, seed=None):
	88	+ """Limit the number of elements per identity to `max_num_images_per_id`.
	89	+
	90	+ Parameters
	91	+ ----------
	92	+ data_y : ndarray
	93	+ A label vector.
	94	+ max_num_images_per_id : int
	95	+ The maximum number of elements per identity that should remain in
	96	+ the data set.
	97	+ seed : Optional[int]
	98	+ Random generator seed.
	99	+
	100	+ Returns
	101	+ -------
	102	+ ndarray
	103	+ A boolean mask that evaluates to True if the corresponding
	104	+ should remain in the data set.
	105	+
	106	+ """
	107	+ random_generator = np.random.RandomState(seed=seed)
	108	+ valid_mask = np.full((len(data_y), ), False, bool)
	109	+ for y in np.unique(data_y):
	110	+ indices = np.where(data_y == y)[0]
	111	+ num_select = min(len(indices), max_num_images_per_id)
	112	+ indices = random_generator.choice(indices, num_select, replace=False)
	113	+ valid_mask[indices] = True
	114	+ return valid_mask
	115	+
	116	+
	117	+def create_cmc_probe_and_gallery(data_y, camera_indices=None, seed=None):
	118	+ """Create probe and gallery images for evaluation of CMC top-k statistics.
	119	+
	120	+ For every identity, this function selects one image as probe and one image
	121	+ for the gallery. Cross-view validation is performed when multiple cameras
	122	+ are given.
	123	+
	124	+ Parameters
	125	+ ----------
	126	+ data_y : ndarray
	127	+ Vector of data labels.
	128	+ camera_indices : Optional[ndarray]
	129	+ Optional array of camera indices. If possible, probe and gallery images
	130	+ are selected from different cameras (i.e., cross-view validation).
	131	+ If None given, assumes all images are taken from the same camera.
	132	+ seed : Optional[int]
	133	+ The random seed used to select probe and gallery images.
	134	+
	135	+ Returns
	136	+ -------
	137	+ (ndarray, ndarray)
	138	+ Returns a tuple of indices to probe and gallery images.
	139	+
	140	+ """
	141	+ data_y = np.asarray(data_y)
	142	+ if camera_indices is None:
	143	+ camera_indices = np.zeros_like(data_y, dtype=np.int)
	144	+ camera_indices = np.asarray(camera_indices)
	145	+
	146	+ random_generator = np.random.RandomState(seed=seed)
	147	+ unique_y = np.unique(data_y)
	148	+ probe_indices, gallery_indices = [], []
	149	+ for y in unique_y:
	150	+ mask_y = data_y == y
	151	+
	152	+ unique_cameras = np.unique(camera_indices[mask_y])
	153	+ if len(unique_cameras) == 1:
	154	+ # If we have only one camera, take any two images from this device.
	155	+ c = unique_cameras[0]
	156	+ indices = np.where(np.logical_and(mask_y, camera_indices == c))[0]
	157	+ if len(indices) < 2:
	158	+ continue # Cannot generate a pair for this identity.
	159	+ i1, i2 = random_generator.choice(indices, 2, replace=False)
	160	+ else:
	161	+ # If we have multiple cameras, take images of two (randomly chosen)
	162	+ # different devices.
	163	+ c1, c2 = random_generator.choice(unique_cameras, 2, replace=False)
	164	+ indices1 = np.where(np.logical_and(mask_y, camera_indices == c1))[0]
	165	+ indices2 = np.where(np.logical_and(mask_y, camera_indices == c2))[0]
	166	+ i1 = random_generator.choice(indices1)
	167	+ i2 = random_generator.choice(indices2)
	168	+
	169	+ probe_indices.append(i1)
	170	+ gallery_indices.append(i2)
	171	+
	172	+ return np.asarray(probe_indices), np.asarray(gallery_indices)
	173	+

Adeepsort/help_modelarts.py+80-0

@@ -0,0 +1,80 @@
	1	+# Copyright 2021 Huawei Technologies Co., Ltd
	2	+#
	3	+# Licensed under the Apache License, Version 2.0 (the "License");
	4	+# you may not use this file except in compliance with the License.
	5	+# You may obtain a copy of the License at
	6	+#
	7	+# http://www.apache.org/licenses/LICENSE-2.0
	8	+#
	9	+# Unless required by applicable law or agreed to in writing, software
	10	+# distributed under the License is distributed on an "AS IS" BASIS,
	11	+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	12	+# See the License for the specific language governing permissions and
	13	+# limitations under the License.
	14	+import os
	15	+import datetime
	16	+import moxing as mox
	17	+
	18	+def obs_data2modelarts(config):
	19	+ """
	20	+ Copy train data from obs to modelarts by using moxing api.
	21	+ """
	22	+ start = datetime.datetime.now()
	23	+ print("===>>>Copy files from obs:{} to modelarts dir:{}".format(config.data_url, config.modelarts_data_dir))
	24	+ mox.file.copy_parallel(src_url=config.data_url, dst_url=config.modelarts_data_dir)
	25	+ end = datetime.datetime.now()
	26	+ print("===>>>Copy from obs to modelarts, time use:{}(s)".format((end - start).seconds))
	27	+ files = os.listdir(config.modelarts_data_dir)
	28	+ print("===>>>Files:", files)
	29	+
	30	+
	31	+def modelarts_result2obs(FLAGS):
	32	+ """
	33	+ Copy debug data from modelarts to obs.
	34	+ According to the swich flags, the debug data may contains auto tune repository,
	35	+ dump data for precision comparision, even the computation graph and profiling data.
	36	+ """
	37	+ work_dir = os.getcwd()
	38	+
	39	+ ## copy result from modelarts to obs
	40	+ obs_result_dir = os.path.join(FLAGS.obs_dir, 'result')
	41	+ if not mox.file.exists(obs_result_dir):
	42	+ mox.file.make_dirs(obs_result_dir)
	43	+ mox.file.copy_parallel(src_url=FLAGS.result, dst_url=obs_result_dir)
	44	+ print("===>>>Copy Event or Checkpoint from modelarts dir:{} to obs:{}".format(FLAGS.result, obs_result_dir))
	45	+
	46	+ ## Copy auto tune repository. Comment this snippets if npu_auto_tune is off.
	47	+ if FLAGS.npu_auto_tune:
	48	+ modelarts_auto_tune_dir = os.path.join(work_dir, "npu_auto_tune")
	49	+ obs_auto_tune_dir = os.path.join(FLAGS.obs_dir, 'npu_auto_tune')
	50	+ if not mox.file.exists(obs_auto_tune_dir):
	51	+ mox.file.make_dirs(obs_auto_tune_dir)
	52	+ mox.file.copy_parallel(modelarts_auto_tune_dir, obs_auto_tune_dir)
	53	+ print("===>>>Auto tune:{} on OBS dir:{}".format(mox.file.list_directory(obs_auto_tune_dir), obs_auto_tune_dir))
	54	+
	55	+ ## Copy dump data. Comment this snippets if npu_dump_data is off.
	56	+ if FLAGS.npu_dump_data:
	57	+ modelarts_dump_data_dir = os.path.join(work_dir, "npu_dump_data")
	58	+ obs_dump_data_dir = os.path.join(FLAGS.obs_dir, 'npu_dump_data')
	59	+ if not mox.file.exists(obs_dump_data_dir):
	60	+ mox.file.make_dirs(obs_dump_data_dir)
	61	+ mox.file.copy_parallel(modelarts_dump_data_dir, obs_dump_data_dir)
	62	+ print("===>>>Dumped graph:{} on OBS dir:{}".format(mox.file.list_directory(obs_dump_data_dir), obs_dump_data_dir))
	63	+
	64	+ ## Copy compute graph. Comment this snippets if npu_dump_graph is off.
	65	+ if FLAGS.npu_dump_graph:
	66	+ modelarts_dump_graph_dir = os.path.join(work_dir, "npu_dump_graph")
	67	+ obs_dump_graph_dir = os.path.join(FLAGS.obs_dir, 'npu_dump_graph')
	68	+ if not mox.file.exists(obs_dump_graph_dir):
	69	+ mox.file.make_dirs(obs_dump_graph_dir)
	70	+ mox.file.copy_parallel(modelarts_dump_graph_dir, obs_dump_graph_dir)
	71	+ print("===>>>Dumped data:{} on OBS dir:{}".format(mox.file.list_directory(obs_dump_graph_dir), obs_dump_graph_dir))
	72	+
	73	+ ## Copy profiling data. Comment this snippets if npu_profiling is off.
	74	+ if FLAGS.npu_profiling:
	75	+ modelarts_profiling_dir = os.path.join(work_dir, "npu_profiling")
	76	+ obs_profiling_dir = os.path.join(FLAGS.obs_dir, 'npu_profiling')
	77	+ if not mox.file.exists(obs_profiling_dir):
	78	+ mox.file.make_dirs(obs_profiling_dir)
	79	+ mox.file.copy_parallel(modelarts_profiling_dir, obs_profiling_dir)
	80	+ print("===>>>Profiling data:{} on OBS dir:{}".format(mox.file.list_directory(obs_profiling_dir), obs_profiling_dir))

Adeepsort/losses.py+145-0

@@ -0,0 +1,145 @@
	1	+# vim: expandtab:ts=4:sw=4
	2	+from npu_bridge.npu_init import *
	3	+import tensorflow as tf
	4	+
	5	+
	6	+def _pdist(a, b=None):
	7	+ sq_sum_a = tf.reduce_sum(tf.square(a), reduction_indices=[1])
	8	+ if b is None:
	9	+ return -2 * tf.matmul(a, tf.transpose(a)) + \
	10	+ tf.reshape(sq_sum_a, (-1, 1)) + tf.reshape(sq_sum_a, (1, -1))
	11	+ sq_sum_b = tf.reduce_sum(tf.square(b), reduction_indices=[1])
	12	+ return -2 * tf.matmul(a, tf.transpose(b)) + \
	13	+ tf.reshape(sq_sum_a, (-1, 1)) + tf.reshape(sq_sum_b, (1, -1))
	14	+
	15	+
	16	+def softmargin_triplet_loss(features, labels, create_summaries=True):
	17	+ """Softmargin triplet loss.
	18	+
	19	+ See::
	20	+
	21	+ Hermans, Beyer, Leibe: In Defense of the Triplet Loss for Person
	22	+ Re-Identification. arXiv, 2017.
	23	+
	24	+ Parameters
	25	+ ----------
	26	+ features : tf.Tensor
	27	+ A matrix of shape NxM that contains the M-dimensional feature vectors
	28	+ of N objects (floating type).
	29	+ labels : tf.Tensor
	30	+ The one-dimensional array of length N that contains for each feature
	31	+ the associated class label (integer type).
	32	+ create_summaries : Optional[bool]
	33	+ If True, creates summaries to monitor training behavior.
	34	+
	35	+ Returns
	36	+ -------
	37	+ tf.Tensor
	38	+ A scalar loss tensor.
	39	+
	40	+ """
	41	+ eps = tf.constant(1e-5, tf.float32)
	42	+ nil = tf.constant(0., tf.float32)
	43	+ almost_inf = tf.constant(1e+10, tf.float32)
	44	+
	45	+ squared_distance_mat = _pdist(features)
	46	+ distance_mat = tf.sqrt(tf.maximum(nil, eps + squared_distance_mat))
	47	+ label_mat = tf.cast(tf.equal(
	48	+ tf.reshape(labels, (-1, 1)), tf.reshape(labels, (1, -1))), tf.float32)
	49	+
	50	+ positive_distance = tf.reduce_max(label_mat * distance_mat, axis=1)
	51	+ negative_distance = tf.reduce_min(
	52	+ (label_mat * almost_inf) + distance_mat, axis=1)
	53	+ loss = tf.nn.softplus(positive_distance - negative_distance)
	54	+ if create_summaries:
	55	+ fraction_invalid_pdist = tf.reduce_mean(
	56	+ tf.cast(tf.less_equal(squared_distance_mat, -eps), tf.float32))
	57	+ tf.summary.scalar("fraction_invalid_pdist", fraction_invalid_pdist)
	58	+
	59	+ fraction_active_triplets = tf.reduce_mean(
	60	+ tf.cast(tf.greater_equal(loss, 1e-5), tf.float32))
	61	+ tf.summary.scalar("fraction_active_triplets", fraction_active_triplets)
	62	+
	63	+ embedding_squared_norm = tf.reduce_mean(
	64	+ tf.reduce_sum(tf.square(features), axis=1))
	65	+ tf.summary.scalar("mean squared feature norm", embedding_squared_norm)
	66	+
	67	+ mean_distance = tf.reduce_mean(distance_mat)
	68	+ tf.summary.scalar("mean feature distance", mean_distance)
	69	+
	70	+ mean_positive_distance = tf.reduce_mean(positive_distance)
	71	+ tf.summary.scalar("mean positive distance", mean_positive_distance)
	72	+
	73	+ mean_negative_distance = tf.reduce_mean(negative_distance)
	74	+ tf.summary.scalar("mean negative distance", mean_negative_distance)
	75	+
	76	+ return tf.reduce_mean(loss)
	77	+
	78	+
	79	+def magnet_loss(features, labels, margin=1.0, unique_labels=None):
	80	+ """Simple unimodal magnet loss.
	81	+
	82	+ See::
	83	+
	84	+ Rippel, Paluri, Dollar, Bourdev: Metric Learning With Adaptive
	85	+ Density Discrimination. ICLR, 2016.
	86	+
	87	+ Parameters
	88	+ ----------
	89	+ features : tf.Tensor
	90	+ A matrix of shape NxM that contains the M-dimensional feature vectors
	91	+ of N objects (floating type).
	92	+ labels : tf.Tensor
	93	+ The one-dimensional array of length N that contains for each feature
	94	+ the associated class label (integer type).
	95	+ margin : float
	96	+ A scalar margin hyperparameter.
	97	+ unique_labels : Optional[tf.Tensor]
	98	+ Optional tensor of unique values in `labels`. If None given, computed
	99	+ from data.
	100	+
	101	+ Returns
	102	+ -------
	103	+ tf.Tensor
	104	+ A scalar loss tensor.
	105	+
	106	+ """
	107	+ nil = tf.constant(0., tf.float32)
	108	+ one = tf.constant(1., tf.float32)
	109	+ minus_two = tf.constant(-2., tf.float32)
	110	+ eps = tf.constant(1e-4, tf.float32)
	111	+ margin = tf.constant(margin, tf.float32)
	112	+
	113	+ num_per_class = None
	114	+ if unique_labels is None:
	115	+ unique_labels, sample_to_unique_y, num_per_class = tf.unique_with_counts(labels)
	116	+ num_per_class = tf.cast(num_per_class, tf.float32)
	117	+
	118	+ y_mat = tf.cast(tf.equal(
	119	+ tf.reshape(labels, (-1, 1)), tf.reshape(unique_labels, (1, -1))),
	120	+ dtype=tf.float32)
	121	+
	122	+ # If class_means is None, compute from batch data.
	123	+ if num_per_class is None:
	124	+ num_per_class = tf.reduce_sum(y_mat, reduction_indices=[0])
	125	+ class_means = tf.reduce_sum(
	126	+ tf.expand_dims(tf.transpose(y_mat), -1) * tf.expand_dims(features, 0),
	127	+ reduction_indices=[1]) / tf.expand_dims(num_per_class, -1)
	128	+
	129	+ squared_distance = _pdist(features, class_means)
	130	+
	131	+ num_samples = tf.cast(tf.shape(labels)[0], tf.float32)
	132	+ variance = tf.reduce_sum(
	133	+ y_mat * squared_distance) / (num_samples - one)
	134	+
	135	+ const = one / (minus_two * (variance + eps))
	136	+ linear = const * squared_distance - y_mat * margin
	137	+
	138	+ maxi = tf.reduce_max(linear, reduction_indices=[1], keepdims=True)
	139	+ loss_mat = tf.exp(linear - maxi)
	140	+
	141	+ a = tf.reduce_sum(y_mat * loss_mat, reduction_indices=[1])
	142	+ b = tf.reduce_sum((one - y_mat) * loss_mat, reduction_indices=[1])
	143	+ loss = tf.maximum(nil, -tf.log(eps + a / (eps + b)))
	144	+ return tf.reduce_mean(loss), class_means, variance
	145	+

Adeepsort/metrics.py+223-0

@@ -0,0 +1,223 @@
	1	+# vim: expandtab:ts=4:sw=4
	2	+from npu_bridge.npu_init import *
	3	+import tensorflow as tf
	4	+import tensorflow.contrib.slim as slim
	5	+
	6	+
	7	+def pdist(a, b=None):
	8	+ """Compute element-wise squared distance between `a` and `b`.
	9	+
	10	+ Parameters
	11	+ ----------
	12	+ a : tf.Tensor
	13	+ A matrix of shape NxL with N row-vectors of dimensionality L.
	14	+ b : tf.Tensor
	15	+ A matrix of shape MxL with M row-vectors of dimensionality L.
	16	+
	17	+ Returns
	18	+ -------
	19	+ tf.Tensor
	20	+ A matrix of shape NxM where element (i, j) contains the squared
	21	+ distance between elements `a[i]` and `b[j]`.
	22	+
	23	+ """
	24	+ sq_sum_a = tf.reduce_sum(tf.square(a), reduction_indices=[1])
	25	+ if b is None:
	26	+ return -2 * tf.matmul(a, tf.transpose(a)) + \
	27	+ tf.reshape(sq_sum_a, (-1, 1)) + tf.reshape(sq_sum_a, (1, -1))
	28	+ sq_sum_b = tf.reduce_sum(tf.square(b), reduction_indices=[1])
	29	+ return -2 * tf.matmul(a, tf.transpose(b)) + \
	30	+ tf.reshape(sq_sum_a, (-1, 1)) + tf.reshape(sq_sum_b, (1, -1))
	31	+
	32	+
	33	+def cosine_distance(a, b=None):
	34	+ """Compute element-wise cosine distance between `a` and `b`.
	35	+
	36	+ Parameters
	37	+ ----------
	38	+ a : tf.Tensor
	39	+ A matrix of shape NxL with N row-vectors of dimensionality L.
	40	+ b : tf.Tensor
	41	+ A matrix of shape NxL with N row-vectors of dimensionality L.
	42	+
	43	+ Returns
	44	+ -------
	45	+ tf.Tensor
	46	+ A matrix of shape NxM where element (i, j) contains the cosine distance
	47	+ between elements `a[i]` and `b[j]`.
	48	+
	49	+ """
	50	+ a_normed = tf.nn.l2_normalize(a, dim=1)
	51	+ b_normed = a_normed if b is None else tf.nn.l2_normalize(b, dim=1)
	52	+ return (
	53	+ tf.constant(1.0, tf.float32) -
	54	+ tf.matmul(a_normed, tf.transpose(b_normed)))
	55	+
	56	+
	57	+def recognition_rate_at_k(probe_x, probe_y, gallery_x, gallery_y, k,
	58	+ measure=pdist):
	59	+ """Compute the recognition rate at a given level `k`.
	60	+
	61	+ For a given probe and ranked gallery that is sorted according to a distance
	62	+ measure `measure` in descending order, the recognition rate at `k` is::
	63	+
	64	+ recognition_rate_at_k = num_correct / min(k, num_relevant)
	65	+
	66	+ where num_correct refers to the fraction of images in the top k entries of
	67	+ the ranked gallery that have the same label as the probe and `num_relevant`
	68	+ refers to the total number of elements in the gallery that have the same
	69	+ label.
	70	+
	71	+ Parameters
	72	+ ----------
	73	+ probe_x: tf.Tensor
	74	+ A tensor of probe images.
	75	+ probe_y: tf.Tensor
	76	+ A tensor of probe labels.
	77	+ gallery_x: tf.Tensor
	78	+ A tensor of gallery images.
	79	+ gallery_y: tf.Tensor
	80	+ A tensor of gallery labels.
	81	+ k: int
	82	+ See description above.
	83	+ measure: Callable[tf.Tensor, tf.Tensor] -> tf.Tensor
	84	+ A callable that computes for two matrices of row-vectors a matrix of
	85	+ element-wise distances. See `pdist` for an example.
	86	+
	87	+ Returns
	88	+ -------
	89	+ tf.Tensor
	90	+ Returns a scalar tensor which represents the computed metric.
	91	+
	92	+ """
	93	+ # Build a matrix of shape (num_probes, num_gallery_images) where element
	94	+ # (i, j) is 1 if probe image i and the gallery image j have the same
	95	+ # identity, otherwise 0.
	96	+ label_eq_mat = tf.cast(tf.equal(tf.reshape(
	97	+ probe_y, (-1, 1)), tf.reshape(gallery_y, (1, -1))),
	98	+ tf.float32)
	99	+
	100	+ # For each probe image, compute the number of relevant images in the
	101	+ # gallery (same identity). This should always be one for CMC evaluation
	102	+ # because we always have exactly one probe and one gallery image for each
	103	+ # identity.
	104	+ num_relevant = tf.minimum(tf.cast(k, tf.float32), tf.reduce_sum(
	105	+ label_eq_mat, reduction_indices=[1]))
	106	+
	107	+ # Rank gallery images by the similarity measure to build a matrix of
	108	+ # shape (num_probes, k) where element (i, j) contains the label of the
	109	+ # j-th ranked gallery image for probe i.
	110	+ predictions = tf.exp(-measure(probe_x, gallery_x)) # Compute similarity.
	111	+ _, prediction_indices = tf.nn.top_k(predictions, k=k)
	112	+ label_mat = tf.gather(gallery_y, prediction_indices)
	113	+
	114	+ # Just as we have done before, build a matrix where element (i, j) is
	115	+ # one if probe i and gallery image j share the same label (same identity).
	116	+ # This time, the matrix is ranked by the similarity measure and we only
	117	+ # keep the top-k predictions.
	118	+ label_eq_mat = tf.cast(tf.equal(
	119	+ label_mat, tf.reshape(probe_y, (-1, 1))), tf.float32)
	120	+
	121	+ # Compute the number of true positives in [0, k[, i.e., check if we find
	122	+ # the correct gallery image within the top-k ranked results. Then, compute
	123	+ # the recognition rate, which in our case is either 0 or 1 since we have
	124	+ # only one gallery image that shares the same identity with the probe.
	125	+ #
	126	+ # This is the final output of our CMC metric.
	127	+ true_positives_at_k = tf.reduce_sum(label_eq_mat, reduction_indices=[1])
	128	+ return true_positives_at_k / num_relevant
	129	+
	130	+
	131	+def streaming_mean_cmc_at_k(probe_x, probe_y, gallery_x, gallery_y, k,
	132	+ measure=pdist):
	133	+ """Compute cumulated matching characteristics (CMC) at level `k` over
	134	+ a stream of data (i.e., multiple batches).
	135	+
	136	+ The function is compatible with TensorFlow-Slim's streaming metrics
	137	+ interface, e.g., `slim.metrics.aggregate_metric_map`.
	138	+
	139	+ Parameters
	140	+ ----------
	141	+ probe_x: tf.Tensor
	142	+ A tensor of probe images.
	143	+ probe_y: tf.Tensor
	144	+ A tensor of probe labels.
	145	+ gallery_x: tf.Tensor
	146	+ A tensor of gallery images.
	147	+ gallery_y: tf.Tensor
	148	+ A tensor of gallery labels.
	149	+ k: int
	150	+ See description above.
	151	+ measure: Callable[tf.Tensor, tf.Tensor] -> tf.Tensor
	152	+ A callable that computes for two matrices of row-vectors a matrix of
	153	+ element-wise distances. See `pdist` for an example.
	154	+
	155	+ Returns
	156	+ -------
	157	+ Tuple[tf.Tensor, tf.Tensor]
	158	+ The first element in the tuple is the current result. The second element
	159	+ is an operation that updates the computed metric based on new data.
	160	+
	161	+ """
	162	+ recognition_rate = recognition_rate_at_k(
	163	+ probe_x, probe_y, gallery_x, gallery_y, k, measure)
	164	+ return slim.metrics.streaming_mean(recognition_rate)
	165	+
	166	+
	167	+def streaming_mean_averge_precision(probe_x, probe_y, gallery_x, gallery_y,
	168	+ good_mask, measure=pdist):
	169	+ """Compute mean average precision (mAP) over a stream of data.
	170	+
	171	+ Parameters
	172	+ ----------
	173	+ probe_x: tf.Tensor
	174	+ A tensor of N probe images.
	175	+ probe_y: tf.Tensor
	176	+ A tensor of N probe labels.
	177	+ gallery_x: tf.Tensor
	178	+ A tensor of M gallery images.
	179	+ gallery_y: tf.Tensor
	180	+ A tensor of M gallery labels.
	181	+ good_mask: Optional[tf.Tensor]
	182	+ A matrix of shape NxM where element (i, j) evaluates to 0.0 if the pair
	183	+ of i-th probe and j-th gallery image should be excluded from metric
	184	+ computation. All other elements should evaluate to 1.0.
	185	+ measure: Callable[tf.Tensor, tf.Tensor] -> tf.Tensor
	186	+ A callable that computes for two matrices of row-vectors a matrix of
	187	+ element-wise distances. See `pdist` for an example.
	188	+
	189	+ Returns
	190	+ -------
	191	+ Tuple[tf.Tensor, tf.Tensor]
	192	+ The first element in the tuple is the current result. The second element
	193	+ is an operation that updates the computed metric based on new data.
	194	+
	195	+ """
	196	+ # See Wikipedia:
	197	+ # https://en.wikipedia.org/wiki/Information_retrieval#Average_precision
	198	+ if good_mask.dtype != tf.float32:
	199	+ good_mask = tf.cast(good_mask, tf.float32)
	200	+
	201	+ # Compute similarity measure and mask out diagonal (similarity to self).
	202	+ predictions = good_mask * tf.exp(-measure(probe_x, gallery_x))
	203	+
	204	+ # Compute matrix of predicted labels.
	205	+ k = tf.shape(gallery_y)[0]
	206	+ _, prediction_indices = tf.nn.top_k(predictions, k=k)
	207	+ predicted_label_mat = tf.gather(gallery_y, prediction_indices)
	208	+ label_eq_mat = tf.cast(tf.equal(
	209	+ predicted_label_mat, tf.reshape(probe_y, (-1, 1))), tf.float32)
	210	+
	211	+ # Compute statistics.
	212	+ num_relevant = tf.reduce_sum(
	213	+ good_mask * label_eq_mat, reduction_indices=[1], keep_dims=True)
	214	+ true_positives_at_k = tf.cumsum(label_eq_mat, axis=1)
	215	+ retrieved_at_k = tf.cumsum(tf.ones_like(label_eq_mat), axis=1)
	216	+ precision_at_k = true_positives_at_k / retrieved_at_k
	217	+ relevant_at_k = label_eq_mat
	218	+ average_precision = (
	219	+ tf.reduce_sum(precision_at_k * relevant_at_k, reduction_indices=[1]) /
	220	+ tf.cast(tf.squeeze(num_relevant), tf.float32))
	221	+
	222	+ return slim.metrics.streaming_mean(average_precision)
	223	+

Adeepsort/nets/__init__.py+0-0

The file is empty

Adeepsort/nets/deep_sort/.keep+0-0

The file is empty

Adeepsort/nets/deep_sort/__init__.py+0-0

The file is empty

Adeepsort/nets/deep_sort/network_definition.py+134-0

@@ -0,0 +1,134 @@
	1	+# vim: expandtab:ts=4:sw=4
	2	+from npu_bridge.npu_init import *
	3	+import tensorflow as tf
	4	+import tensorflow.contrib.slim as slim
	5	+
	6	+from . import residual_net
	7	+
	8	+
	9	+def create_network(images, num_classes=None, add_logits=True, reuse=None,
	10	+ create_summaries=True, weight_decay=1e-8):
	11	+ nonlinearity = tf.nn.elu
	12	+ conv_weight_init = tf.truncated_normal_initializer(stddev=1e-3)
	13	+ conv_bias_init = tf.zeros_initializer()
	14	+ conv_regularizer = slim.l2_regularizer(weight_decay)
	15	+ fc_weight_init = tf.truncated_normal_initializer(stddev=1e-3)
	16	+ fc_bias_init = tf.zeros_initializer()
	17	+ fc_regularizer = slim.l2_regularizer(weight_decay)
	18	+
	19	+ def batch_norm_fn(x):
	20	+ return slim.batch_norm(x, scope=tf.get_variable_scope().name + "/bn")
	21	+
	22	+ network = images
	23	+ network = slim.conv2d(
	24	+ network, 32, [3, 3], stride=1, activation_fn=nonlinearity,
	25	+ padding="SAME", normalizer_fn=batch_norm_fn, scope="conv1_1",
	26	+ weights_initializer=conv_weight_init, biases_initializer=conv_bias_init,
	27	+ weights_regularizer=conv_regularizer)
	28	+ if create_summaries:
	29	+ tf.summary.histogram(network.name + "/activations", network)
	30	+ tf.summary.image("conv1_1/weights", tf.transpose(
	31	+ slim.get_variables("conv1_1/weights:0")[0], [3, 0, 1, 2]),
	32	+ max_outputs=128)
	33	+ network = slim.conv2d(
	34	+ network, 32, [3, 3], stride=1, activation_fn=nonlinearity,
	35	+ padding="SAME", normalizer_fn=batch_norm_fn, scope="conv1_2",
	36	+ weights_initializer=conv_weight_init, biases_initializer=conv_bias_init,
	37	+ weights_regularizer=conv_regularizer)
	38	+ if create_summaries:
	39	+ tf.summary.histogram(network.name + "/activations", network)
	40	+
	41	+ network = slim.max_pool2d(
	42	+ network, [3, 3], [2, 2], scope="pool1", padding="SAME")
	43	+
	44	+ network = residual_net.residual_block(
	45	+ network, "conv2_1", nonlinearity, conv_weight_init, conv_bias_init,
	46	+ conv_regularizer, increase_dim=False, is_first=True,
	47	+ summarize_activations=create_summaries)
	48	+ network = residual_net.residual_block(
	49	+ network, "conv2_3", nonlinearity, conv_weight_init, conv_bias_init,
	50	+ conv_regularizer, increase_dim=False,
	51	+ summarize_activations=create_summaries)
	52	+
	53	+ network = residual_net.residual_block(
	54	+ network, "conv3_1", nonlinearity, conv_weight_init, conv_bias_init,
	55	+ conv_regularizer, increase_dim=True,
	56	+ summarize_activations=create_summaries)
	57	+ network = residual_net.residual_block(
	58	+ network, "conv3_3", nonlinearity, conv_weight_init, conv_bias_init,
	59	+ conv_regularizer, increase_dim=False,
	60	+ summarize_activations=create_summaries)
	61	+
	62	+ network = residual_net.residual_block(
	63	+ network, "conv4_1", nonlinearity, conv_weight_init, conv_bias_init,
	64	+ conv_regularizer, increase_dim=True,
	65	+ summarize_activations=create_summaries)
	66	+ network = residual_net.residual_block(
	67	+ network, "conv4_3", nonlinearity, conv_weight_init, conv_bias_init,
	68	+ conv_regularizer, increase_dim=False,
	69	+ summarize_activations=create_summaries)
	70	+
	71	+ feature_dim = network.get_shape().as_list()[-1]
	72	+ print("feature dimensionality: ", feature_dim)
	73	+ network = slim.flatten(network)
	74	+
	75	+ network = slim.dropout(network, keep_prob=0.6)
	76	+ network = slim.fully_connected(
	77	+ network, feature_dim, activation_fn=nonlinearity,
	78	+ normalizer_fn=batch_norm_fn, weights_regularizer=fc_regularizer,
	79	+ scope="fc1", weights_initializer=fc_weight_init,
	80	+ biases_initializer=fc_bias_init)
	81	+
	82	+ features = network
	83	+
	84	+ # Features in rows, normalize axis 1.
	85	+ features = tf.nn.l2_normalize(features, dim=1)
	86	+
	87	+ if add_logits:
	88	+ with slim.variable_scope.variable_scope("ball", reuse=reuse):
	89	+ weights = slim.model_variable(
	90	+ "mean_vectors", (feature_dim, int(num_classes)),
	91	+ initializer=tf.truncated_normal_initializer(stddev=1e-3),
	92	+ regularizer=None)
	93	+ scale = slim.model_variable(
	94	+ "scale", (), tf.float32,
	95	+ initializer=tf.constant_initializer(0., tf.float32),
	96	+ regularizer=slim.l2_regularizer(1e-1))
	97	+ if create_summaries:
	98	+ tf.summary.scalar("scale", scale)
	99	+ scale = tf.nn.softplus(scale)
	100	+
	101	+ # Mean vectors in colums, normalize axis 0.
	102	+ weights_normed = tf.nn.l2_normalize(weights, dim=0)
	103	+ logits = scale * tf.matmul(features, weights_normed)
	104	+ else:
	105	+ logits = None
	106	+ return features, logits
	107	+
	108	+
	109	+def create_network_factory(is_training, num_classes, add_logits,
	110	+ weight_decay=1e-8, reuse=None):
	111	+
	112	+ def factory_fn(image):
	113	+ with slim.arg_scope([slim.batch_norm, slim.dropout],
	114	+ is_training=is_training):
	115	+ with slim.arg_scope([slim.conv2d, slim.fully_connected,
	116	+ slim.batch_norm, slim.layer_norm],
	117	+ reuse=reuse):
	118	+ features, logits = create_network(
	119	+ image, num_classes=num_classes, add_logits=add_logits,
	120	+ reuse=reuse, create_summaries=is_training,
	121	+ weight_decay=weight_decay)
	122	+ return features, logits
	123	+
	124	+ return factory_fn
	125	+
	126	+
	127	+def preprocess(image, is_training=False, input_is_bgr=False):
	128	+ if input_is_bgr:
	129	+ image = image[:, :, ::-1] # BGR to RGB
	130	+ image = tf.divide(tf.cast(image, tf.float32), 255.0)
	131	+ if is_training:
	132	+ image = tf.image.random_flip_left_right(image)
	133	+ return image
	134	+

Adeepsort/nets/deep_sort/residual_net.py+86-0

@@ -0,0 +1,86 @@
	1	+# vim: expandtab:ts=4:sw=4
	2	+from npu_bridge.npu_init import *
	3	+import tensorflow as tf
	4	+from tensorflow.contrib import slim
	5	+
	6	+
	7	+def _batch_norm_fn(x, scope=None):
	8	+ if scope is None:
	9	+ scope = tf.get_variable_scope().name + "/bn"
	10	+ return slim.batch_norm(x, scope=scope)
	11	+
	12	+
	13	+def create_link(
	14	+ incoming, network_builder, scope, nonlinearity=tf.nn.elu,
	15	+ weights_initializer=tf.truncated_normal_initializer(stddev=1e-3),
	16	+ regularizer=None, is_first=False, summarize_activations=True):
	17	+ if is_first:
	18	+ network = incoming
	19	+ else:
	20	+ network = _batch_norm_fn(incoming, scope=scope + "/bn")
	21	+ network = nonlinearity(network)
	22	+ if summarize_activations:
	23	+ tf.summary.histogram(scope+"/activations", network)
	24	+
	25	+ pre_block_network = incoming
	26	+ post_block_network = network_builder(network, scope)
	27	+
	28	+ incoming_dim = pre_block_network.get_shape().as_list()[-1]
	29	+ outgoing_dim = post_block_network.get_shape().as_list()[-1]
	30	+ if incoming_dim != outgoing_dim:
	31	+ assert outgoing_dim == 2 * incoming_dim, \
	32	+ "%d != %d" % (outgoing_dim, 2 * incoming)
	33	+ projection = slim.conv2d(
	34	+ incoming, outgoing_dim, 1, 2, padding="SAME", activation_fn=None,
	35	+ scope=scope+"/projection", weights_initializer=weights_initializer,
	36	+ biases_initializer=None, weights_regularizer=regularizer)
	37	+ network = projection + post_block_network
	38	+ else:
	39	+ network = incoming + post_block_network
	40	+ return network
	41	+
	42	+
	43	+def create_inner_block(
	44	+ incoming, scope, nonlinearity=tf.nn.elu,
	45	+ weights_initializer=tf.truncated_normal_initializer(1e-3),
	46	+ bias_initializer=tf.zeros_initializer(), regularizer=None,
	47	+ increase_dim=False, summarize_activations=True):
	48	+ n = incoming.get_shape().as_list()[-1]
	49	+ stride = 1
	50	+ if increase_dim:
	51	+ n *= 2
	52	+ stride = 2
	53	+
	54	+ incoming = slim.conv2d(
	55	+ incoming, n, [3, 3], stride, activation_fn=nonlinearity, padding="SAME",
	56	+ normalizer_fn=_batch_norm_fn, weights_initializer=weights_initializer,
	57	+ biases_initializer=bias_initializer, weights_regularizer=regularizer,
	58	+ scope=scope + "/1")
	59	+ if summarize_activations:
	60	+ tf.summary.histogram(incoming.name + "/activations", incoming)
	61	+
	62	+ incoming = slim.dropout(incoming, keep_prob=0.6)
	63	+
	64	+ incoming = slim.conv2d(
	65	+ incoming, n, [3, 3], 1, activation_fn=None, padding="SAME",
	66	+ normalizer_fn=None, weights_initializer=weights_initializer,
	67	+ biases_initializer=bias_initializer, weights_regularizer=regularizer,
	68	+ scope=scope + "/2")
	69	+ return incoming
	70	+
	71	+
	72	+def residual_block(incoming, scope, nonlinearity=tf.nn.elu,
	73	+ weights_initializer=tf.truncated_normal_initializer(1e3),
	74	+ bias_initializer=tf.zeros_initializer(), regularizer=None,
	75	+ increase_dim=False, is_first=False,
	76	+ summarize_activations=True):
	77	+
	78	+ def network_builder(x, s):
	79	+ return create_inner_block(
	80	+ x, s, nonlinearity, weights_initializer, bias_initializer,
	81	+ regularizer, increase_dim, summarize_activations)
	82	+
	83	+ return create_link(
	84	+ incoming, network_builder, scope, nonlinearity, weights_initializer,
	85	+ regularizer, is_first, summarize_activations)
	86	+

Adeepsort/queued_trainer.py+616-0

@@ -0,0 +1,616 @@
	1	+# Copyright 2021 Huawei Technologies Co., Ltd
	2	+#
	3	+# Licensed under the Apache License, Version 2.0 (the "License");
	4	+# you may not use this file except in compliance with the License.
	5	+# You may obtain a copy of the License at
	6	+#
	7	+# http://www.apache.org/licenses/LICENSE-2.0
	8	+#
	9	+# Unless required by applicable law or agreed to in writing, software
	10	+# distributed under the License is distributed on an "AS IS" BASIS,
	11	+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	12	+# See the License for the specific language governing permissions and
	13	+# limitations under the License.
	14	+
	15	+# vim: expandtab:ts=4:sw=4
	16	+from npu_bridge.npu_init import *
	17	+import string
	18	+import os
	19	+import threading
	20	+import numpy as np
	21	+
	22	+import tensorflow as tf
	23	+import tensorflow.contrib.slim as slim
	24	+import time
	25	+
	26	+config = tf.ConfigProto()
	27	+custom_op = config.graph_options.rewrite_options.custom_optimizers.add()
	28	+custom_op.name = "NpuOptimizer"
	29	+custom_op.parameter_map["use_off_line"].b = True
	30	+custom_op.parameter_map["mix_compile_mode"].b = True
	31	+custom_op.parameter_map["precision_mode"].s = tf.compat.as_bytes("force_fp32")#allow_mix_precision
	32	+#custom_op.parameter_map["modify_mixlist"].s = tf.compat.as_bytes("/home/test/ops_info.json")
	33	+config.graph_options.rewrite_options.remapping = RewriterConfig.OFF
	34	+config.graph_options.rewrite_options.memory_optimization = RewriterConfig.OFF
	35	+
	36	+def run_in_batches(f, data_dict, out, batch_size):
	37	+ """Process data in batches.
	38	+
	39	+ Parameters
	40	+ ----------
	41	+ f : Callable[Dict[tf.Tensor, np.ndarray] -> np.ndarray
	42	+ A function that maps a given input (one or multiple inpu arrays) to a
	43	+ single output array.
	44	+ data_dict : Dict[tf.Tensor, np.ndarray]
	45	+ Maps from symbolic input tensor to numpy data array.
	46	+ out : np.ndarray
	47	+ The computed function output will be stored in this array; must be have
	48	+ compatible shape and length to the output computed by `f`.
	49	+ batch_size : int
	50	+ The number of samples to compute in each call to `f`. If the length of
	51	+ the input array is not divisible by the batch size, the final call to
	52	+ `f` contains fewer examples.
	53	+
	54	+ """
	55	+ data_len = len(out)
	56	+ num_batches = int(data_len / batch_size)
	57	+
	58	+ def pad(x):
	59	+ x = np.asarray(x)
	60	+ y = np.full((batch_size, ) + x.shape[1:], x[0], dtype=x.dtype)
	61	+ y[:x.shape[0]] = x
	62	+ return y
	63	+
	64	+ s, e = 0, batch_size
	65	+ for i in range(num_batches):
	66	+ s, e = i * batch_size, (i + 1) * batch_size
	67	+ batch_data_dict = {k: v[s:e] for k, v in data_dict.items()}
	68	+ out[s:e] = f(batch_data_dict)
	69	+ if e < len(out):
	70	+ remaining_len = len(out) - e
	71	+ batch_data_dict = {k: pad(v[e:]) for k, v in data_dict.items()}
	72	+ out[e:] = f(batch_data_dict)[:remaining_len]
	73	+
	74	+
	75	+def iterate_forever(batch_size, data, *other_data):
	76	+ """Iterate over dataset indefenitely.
	77	+
	78	+ Parameters
	79	+ ----------
	80	+ batch_size : int
	81	+ The batch size.
	82	+ data : ndarray
	83	+ The first input array.
	84	+ other_data
	85	+ Additional input arrays; must be of type np.ndarray.
	86	+
	87	+ Returns
	88	+ -------
	89	+ List[np.ndarray]
	90	+ A dataset batch. The length of each entry in the list is `batch_size`.
	91	+
	92	+ """
	93	+ data_len = len(data)
	94	+ num_batches = int(data_len / batch_size)
	95	+
	96	+ while True:
	97	+ data_list = [data] + list(other_data)
	98	+ s, e = 0, 0
	99	+ for i in range(num_batches):
	100	+ s, e = i * batch_size, (i + 1) * batch_size
	101	+ batch = [x[s:e] for x in data_list]
	102	+ yield batch[0] if len(batch) == 1 else batch
	103	+ if e < data_len:
	104	+ batch = [x[e:] for x in data_list]
	105	+ yield batch[0] if len(batch) == 1 else batch
	106	+
	107	+
	108	+def random_shuffle_forever(batch_size, data, *other_data):
	109	+ """A generator that randomly selects `batch_size` entries from the data.
	110	+
	111	+ Parameters
	112	+ ----------
	113	+ batch_size : int
	114	+ The batch size.
	115	+ data : np.ndarray
	116	+ The first input array.
	117	+ other_data
	118	+ Additional input arrays; must be of type np.ndarray
	119	+
	120	+ Returns
	121	+ -------
	122	+ List[np.ndarray]
	123	+ A batch of randomly selected entries. The length of each entry in the
	124	+ list is `batch_size`.
	125	+
	126	+ """
	127	+ data_list = [data] + list(other_data)
	128	+ indices = np.arange(len(data))
	129	+ while True:
	130	+ batch_indices = np.random.choice(indices, batch_size, replace=False)
	131	+ batch = [x[batch_indices] for x in data_list]
	132	+ yield batch[0] if len(batch) == 1 else batch
	133	+
	134	+
	135	+def random_sample_identities_forever(batch_size, num_samples_per_id, data_x,
	136	+ data_y, num_fa_images=0):
	137	+ """A generator that randomly selects a fixed number of entries per label.
	138	+
	139	+ If false alarms are passed into this function, they should have a negative
	140	+ label, i.e., `data_y[i] < 0` if the i-th example corresponds to a false
	141	+ alarm.
	142	+
	143	+ Parameters
	144	+ ----------
	145	+ batch_size : int
	146	+ The batch size.
	147	+ num_samples_per_id : int
	148	+ Number of examples per label in each batch. If the `batch_size` is not
	149	+ divisible by `num_samples_per_id` then the batch is filled with false
	150	+ alarms. A warning is printed if no false alarms are available to fill
	151	+ up the batch.
	152	+ data_x : List[string] \| np.ndarray
	153	+ The data array; either a list of filenames or a tensor of input images.
	154	+ data_y : List[int] \| np.ndarray
	155	+ The label array (either as list of one-dimensional numpy array).
	156	+ num_fa_images : Optional[int]
	157	+ Number of false alarm images to include in each batch; defaults to zero.
	158	+
	159	+ Returns
	160	+ -------
	161	+ List[np.ndarray]
	162	+ Returns a list of length two where the first entry is the data array
	163	+ corresponding to `data_x` and the second entry is the label array
	164	+ corresponding to `data_y`. The elements in the list are of length
	165	+ `batch_size`.
	166	+
	167	+ """
	168	+ assert (batch_size - num_fa_images) % num_samples_per_id == 0
	169	+ num_ids_per_batch = int((batch_size - num_fa_images) / num_samples_per_id)
	170	+
	171	+ data_x = np.asarray(data_x)
	172	+ data_y = np.asarray(data_y)
	173	+
	174	+ unique_y = np.unique(data_y[data_y >= 0])
	175	+ y_to_idx = {y: np.where(data_y == y)[0] for y in unique_y}
	176	+ fa_indices = np.where(data_y < 0)[0]
	177	+
	178	+ while True:
	179	+ # Draw the desired number of identities.
	180	+ indices = np.random.choice(
	181	+ len(unique_y), num_ids_per_batch, replace=False)
	182	+ batch_unique_y = unique_y[indices]
	183	+
	184	+ batch_x = np.zeros((batch_size, ) + data_x.shape[1:], data_x.dtype)
	185	+ batch_y = np.zeros((batch_size, ), data_y.dtype)
	186	+ e = 0
	187	+ for i, y in enumerate(batch_unique_y):
	188	+ num_samples = min(num_samples_per_id, len(y_to_idx[y]))
	189	+ indices = np.random.choice(y_to_idx[y], num_samples, replace=False)
	190	+ s, e = e, e + num_samples
	191	+ batch_x[s:e] = data_x[indices]
	192	+ batch_y[s:e] = y
	193	+
	194	+ # Fill up remaining space with false alarms.
	195	+ num_samples = len(batch_x) - e
	196	+ if num_fa_images > 0:
	197	+ num_batch_fa_samples = min(num_samples, len(fa_indices))
	198	+ indices = np.random.choice(
	199	+ fa_indices, num_batch_fa_samples, replace=False)
	200	+ s, e = e, e + num_batch_fa_samples
	201	+ batch_x[s:e] = data_x[indices]
	202	+ batch_y[s:e] = data_y[indices]
	203	+
	204	+ # If we need to add more data, random sample ids until we have reached
	205	+ # the batch size.
	206	+ num_samples = len(batch_x) - e
	207	+ num_tries = 0
	208	+ while num_samples > 0 and num_tries < 100:
	209	+ y = np.random.choice(unique_y)
	210	+ if y in batch_unique_y:
	211	+ # Find a target that we have not yet in this batch.
	212	+ num_tries += 1
	213	+ continue
	214	+
	215	+ num_samples = min(num_samples, len(y_to_idx[y]))
	216	+ indices = np.random.choice(y_to_idx[y], num_samples, replace=False)
	217	+ s, e = e, e + num_samples
	218	+ batch_x[s:e] = data_x[indices]
	219	+ batch_y[s:e] = y
	220	+ num_samples = len(batch_x) - e
	221	+
	222	+ if e < batch_size:
	223	+ print("ERROR: Failed to sample a full batch. Adding corrupt data.")
	224	+ yield [batch_x, batch_y]
	225	+
	226	+
	227	+def _truncate_dataset_to_batch_size(batch_size, data, *other_data):
	228	+ """Truncate given input data to a multiple of the batch size.
	229	+
	230	+ Parameters
	231	+ ----------
	232	+ batch_size : int
	233	+ The batch size. The length of the truncated data is a multiple of this
	234	+ value.
	235	+ data : np.ndarray
	236	+ The first input array.
	237	+ *other_data
	238	+ Additional input arrays; must be of type np.ndarray.
	239	+
	240	+ Returns
	241	+ -------
	242	+ List[np.ndarray]
	243	+ The truncated data. The length of each entry in the list is a multiple
	244	+ of the batch size.
	245	+
	246	+ """
	247	+ num_batches = int(len(data) / batch_size)
	248	+ new_len = num_batches * batch_size
	249	+ dataset = [data] + list(other_data)
	250	+ if new_len < len(data):
	251	+ print(
	252	+ "WARNING dataset length is not a multiple of batch size. "
	253	+ "Truncating from %d to %d." % (len(data), new_len))
	254	+ dataset = [x[:new_len] for x in dataset]
	255	+ return num_batches, dataset[0] if len(dataset) == 1 else dataset
	256	+
	257	+
	258	+def _generate_run_id(size=6, chars=None):
	259	+ """Generate a random ID of length `size`.
	260	+
	261	+ Parameters
	262	+ ----------
	263	+ size : int
	264	+ chars : Optional[str]
	265	+ Optional list of characters to use for generating the ID.
	266	+
	267	+ Returns
	268	+ -------
	269	+ str
	270	+ Returns a random identifier of length `size`.
	271	+
	272	+ """
	273	+ if chars is None:
	274	+ chars = string.ascii_uppercase + string.digits
	275	+ import random
	276	+ return ''.join(random.choice(chars) for _ in range(size))
	277	+
	278	+
	279	+class ThreadSafeIterator(object):
	280	+ """
	281	+ This class wraps an iterator (or generator) such that only one thread at a
	282	+ time is granted access.
	283	+
	284	+ Parameters
	285	+ ----------
	286	+ iterator_or_generator
	287	+ An iterator or generator to be wrapped.
	288	+
	289	+ """
	290	+
	291	+ def __init__(self, iterator_or_generator):
	292	+ self._iterator_or_generator = iterator_or_generator
	293	+ self._lock = threading.Lock()
	294	+
	295	+ def __iter__(self):
	296	+ return self
	297	+
	298	+ def __next__(self):
	299	+ with self._lock:
	300	+ return next(self._iterator_or_generator)
	301	+
	302	+ def next(self):
	303	+ with self._lock:
	304	+ return self._iterator_or_generator.next()
	305	+
	306	+
	307	+class QueuedTrainer(object):
	308	+ """
	309	+ This class implements code to train and evaluate TensorFlow models based on
	310	+ TensorFlow-Slim. Image loading and preprocessing is de-coupled from the
	311	+ training steps using a tf.FIFOQueue.
	312	+
	313	+ Parameters
	314	+ ----------
	315	+ enqueue_vars : List[tf.Tensor]
	316	+ A list of tensors to be enqueued; usually the labels and preprocessed
	317	+ images.
	318	+ input_vars : Optional[List[tf.Tensor]]
	319	+ An optional list of input tensors; usually the labels and raw (not
	320	+ preprocessed) images or filenames to the images. The list must be of the
	321	+ same length as the `enqueue_vars` and there must be a one-to-one
	322	+ correspondence, i.e., the i-th element in `enqueue_vars` is i-th
	323	+ preprocessed element in `input_vars`. If None, the input_vars are set to
	324	+ `enqueue_vars`.
	325	+ num_enqueue_threads : Optional[int]
	326	+ Number of threads used to preprocess data in parallel.
	327	+ queue_capacity : Optional[int]
	328	+ Maximum number of elements in the queue; defaults to 512.
	329	+
	330	+ """
	331	+
	332	+ def __init__(self, enqueue_vars, input_vars=None, num_enqueue_threads=4,
	333	+ queue_capacity=512):
	334	+ if input_vars is None:
	335	+ input_vars = enqueue_vars
	336	+ self._input_vars = input_vars
	337	+ self._enqueue_vars = enqueue_vars
	338	+
	339	+ shapes = [var.get_shape().as_list()[1:] for var in enqueue_vars]
	340	+ dtypes = [var.dtype for var in enqueue_vars]
	341	+ self._queue = tf.FIFOQueue(queue_capacity, dtypes, shapes)
	342	+
	343	+ self._num_enqueue_threads = num_enqueue_threads
	344	+ self._enqueue_threads = []
	345	+ self._enqueue_op = self._queue.enqueue_many(self._enqueue_vars)
	346	+ self._stop_op = self._queue.close(cancel_pending_enqueues=True)
	347	+ self._coordinator = None
	348	+
	349	+ self._feed_generator = None
	350	+ self._batch_size = None
	351	+ self._init_fns = []
	352	+
	353	+ def get_input_vars(self, batch_size):
	354	+ """Get the top `batch_size` elements from the queue. The tensors
	355	+ returned by this functions should be passed on the the TensorFlow model.
	356	+
	357	+ Parameters
	358	+ ----------
	359	+ batch_size : int
	360	+ The batch size.
	361	+
	362	+ Returns
	363	+ -------
	364	+ List[tf.Tensor]
	365	+ Returns the top `batch_size` elements from the queue. There is a
	366	+ one-to-one correspondence between the `enqueue_vars` passed in to
	367	+ the constructor of this class and the tensors in the list returned
	368	+ by this function.
	369	+
	370	+ """
	371	+ self._batch_size = batch_size
	372	+ return self._queue.dequeue_many(batch_size)
	373	+
	374	+ def run(self, feed_generator, train_op, log_dir="/tmp/slim_trainer/",
	375	+ restore_path=None, variables_to_restore=None, run_id=None,
	376	+ max_checkpoints_to_keep=0, **kwargs):
	377	+ """ Run training.
	378	+
	379	+ Parameters
	380	+ ----------
	381	+ feed_generator : Iterator[ndarray, ...]
	382	+ An iterator or generator that returns batches of training data; must
	383	+ return a one-to-one correspondence with the `enqueue_vars` passed
	384	+ to the constructor of this class.
	385	+ train_op : tf.Tensor
	386	+ The training operation created with `slim.learning.create_train_op`.
	387	+ log_dir : Optional[str]
	388	+ Path to TensorFlow log directory. This value is used in conjunction
	389	+ with `run_id` to generate the checkpoint and summary directory;
	390	+ defaults to '/tmp/slim_trainer'.
	391	+ restore_path : Optional[str]
	392	+ An optional checkpoint path. If not None, resumes training from the
	393	+ given checkpoint.
	394	+ variables_to_restore : Optional[List[str]]
	395	+ An optional list of variable scopes. If not None, only restores
	396	+ variables under the given scope. This value is ignored if
	397	+ `restore_path` is None.
	398	+ run_id : Optional[str]
	399	+ A string that identifies this training run. The checkpoints and
	400	+ TensorFlow summaries are stored in `log_dir/run_id`. If None, a
	401	+ random ID will be generated. Point tensorboard to this directory to
	402	+ monitor training progress.
	403	+ max_checkpoints_to_keep : int
	404	+ Keep only the `max_checkpoints_to_keep` newest checkpoints. If 0,
	405	+ keep all checkpoints.
	406	+ kwargs:
	407	+ Additional named arguments passed on to tf.slim.learning.train,
	408	+ e.g., `number_of_steps=100` to run 100 iterations of training.
	409	+
	410	+ """
	411	+ if restore_path is not None:
	412	+ if variables_to_restore is None:
	413	+ variables_to_restore = slim.get_variables_to_restore()
	414	+ init_assign_op, init_feed_dict = slim.assign_from_checkpoint(
	415	+ restore_path, variables_to_restore)
	416	+ self._init_fns.append(lambda sess: sess.run(
	417	+ init_assign_op, init_feed_dict))
	418	+ self._feed_generator = ThreadSafeIterator(feed_generator)
	419	+ self._coordinator = tf.train.Coordinator()
	420	+
	421	+ if run_id is None:
	422	+ run_id = _generate_run_id(6)
	423	+ log_dir = os.path.join(log_dir, run_id)
	424	+ print("---------------------------------------")
	425	+ print("Run ID: ", run_id)
	426	+ print("Log directory: ", log_dir)
	427	+ print("---------------------------------------")
	428	+
	429	+ saver = tf.train.Saver(max_to_keep=max_checkpoints_to_keep)
	430	+ try:
	431	+ slim.learning.train(
	432	+ train_op, log_dir, self._train_step_fn, session_config=config, saver=saver,
	433	+ **kwargs)
	434	+ except UnboundLocalError:
	435	+ # NOTE(nwojke): Due to a bug in slim, a local variable 'total_loss'
	436	+ # is referenced when an exception is raised during training. We
	437	+ # catch the exception here because it occurs whenever we close the
	438	+ # queue with self._stop_all_threads().
	439	+ pass
	440	+ self._wait_for_threads()
	441	+
	442	+ def evaluate(self, dataset, checkpoint_dir, log_dir, run_id=None,
	443	+ init_op=None, eval_op=None, final_op=None,
	444	+ summary_op=None, variables_to_restore=None,
	445	+ eval_interval_secs=60, max_num_evaluations=None):
	446	+ """Run evaluation. Monitors files in the log directory and computes
	447	+ evaluation metrics. This function must be called concurrently to
	448	+ training (in a separate process).
	449	+
	450	+ WARNING: The dataset is truncated to the batch size. Thus, the computed
	451	+ metrics are only accurate if the dataset length is divisible by the
	452	+ batch size.
	453	+
	454	+ Parameters
	455	+ ----------
	456	+ dataset : List[T]
	457	+ The dataset is a list (or tuple) of data arrays. The length of the
	458	+ list must be the same as the `input_vars` passed to the constructor
	459	+ of this class and there must be a one-to-one correspondence such
	460	+ that `dataset[i]` corresponds to the numeric data of its symbolic
	461	+ equivalent in `input_vars[i]`.
	462	+ checkpoint_dir : str
	463	+ The directory where checkpoints are stored. Should be set to
	464	+ `log_dir` of the training process.
	465	+ log_dir : str
	466	+ Path to TensorFlow log directory where evaluation logs will be
	467	+ stored. This directory should be different from the `log_dir`
	468	+ passed to `run`.
	469	+ run_id : Optional[str]
	470	+ A string that identifies the training runrun. Should be set to
	471	+ `run_id` passed to `run`.
	472	+ init_op : Optional[tf.Tensor]
	473	+ Optional operation to execute prior to processing the `dataset`.
	474	+ eval_op : Optional[tf.Tensor]
	475	+ Evaluation operation; will be executed for each batch in the
	476	+ `dataset`.
	477	+ final_op : Optional[tf.Tensor]
	478	+ Optional operation to execute after processing the `dataset`.
	479	+ summary_op : Optional[tf.Tensor]
	480	+ Summary operation; defaults to `tf.summary.merge_all()`.
	481	+ variables_to_restore : Optional[List[tf.Tensor]]
	482	+ List of variables to restore; defaults to
	483	+ `slim.get_variables_to_restore()`.
	484	+ eval_interval_secs : Optional[int]
	485	+ Poll the `checkpoint_dir` every `eval_interval_secs` seconds for
	486	+ new checkpoints.
	487	+ max_num_evaluations : Optional[int]
	488	+ Evaluate at most `max_num_evaluations` checkpoints.
	489	+
	490	+ Returns
	491	+ -------
	492	+ T
	493	+ Returns the value of the last call to `final_op` or None.
	494	+
	495	+ """
	496	+ if run_id is None:
	497	+ print("---------------------------------------")
	498	+ print("Checkpoint directory: ", checkpoint_dir)
	499	+ print("Log directory: ", log_dir)
	500	+ print("---------------------------------------")
	501	+ else:
	502	+ checkpoint_dir = os.path.join(checkpoint_dir, run_id)
	503	+ log_dir = os.path.join(log_dir, run_id)
	504	+ print("---------------------------------------")
	505	+ print("Run ID: ", run_id)
	506	+ print("Checkpoint directory: ", checkpoint_dir)
	507	+ print("Log directory: ", log_dir)
	508	+ print("---------------------------------------")
	509	+
	510	+ if summary_op is None:
	511	+ summary_op = tf.summary.merge_all()
	512	+
	513	+ global_step = tf.train.get_or_create_global_step()
	514	+
	515	+ if variables_to_restore is None:
	516	+ variables_to_restore = slim.get_variables_to_restore()
	517	+ saver = tf.train.Saver(variables_to_restore)
	518	+ summary_writer = tf.summary.FileWriter(log_dir)
	519	+ sv = tf.train.Supervisor(
	520	+ graph=tf.get_default_graph(), logdir=log_dir, summary_op=None,
	521	+ summary_writer=None, global_step=None, saver=saver)
	522	+
	523	+ print("Entering evaluation loop. Waiting for checkpoints.")
	524	+ num_batches, dataset = _truncate_dataset_to_batch_size(
	525	+ self._batch_size, *dataset)
	526	+
	527	+ final_op_value = None
	528	+ num_evaluations = 0
	529	+ for checkpoint_path in slim.evaluation.checkpoints_iterator(
	530	+ checkpoint_dir, eval_interval_secs):
	531	+ with sv.managed_session(start_standard_services=False) as session:
	532	+ sv.saver.restore(session, checkpoint_path)
	533	+ sv.start_queue_runners(session)
	534	+
	535	+ print("Starting evaluation of '%s'" % checkpoint_path)
	536	+ self._feed_generator = iterate_forever(
	537	+ self._batch_size, *dataset)
	538	+ self._coordinator = tf.train.Coordinator()
	539	+ for fn in self._init_fns:
	540	+ fn(session)
	541	+ self._start_enqueue(session, num_threads=1)
	542	+
	543	+ if init_op is not None:
	544	+ session.run(init_op)
	545	+
	546	+ if eval_op is not None:
	547	+ for i in range(num_batches):
	548	+ session.run(eval_op)
	549	+
	550	+ if final_op is not None:
	551	+ final_op_value = session.run(final_op)
	552	+ else:
	553	+ final_op_value = None
	554	+
	555	+ summary_str = session.run(summary_op)
	556	+ global_step_value = session.run(global_step)
	557	+ summary_writer.add_summary(summary_str, global_step_value)
	558	+ summary_writer.flush()
	559	+
	560	+ self._stop_all_threads(session)
	561	+ print("Finished evaluation of '%s'" % checkpoint_path)
	562	+
	563	+ num_evaluations += 1
	564	+ if max_num_evaluations is not None \
	565	+ and num_evaluations >= max_num_evaluations:
	566	+ break
	567	+ return final_op_value
	568	+
	569	+ def _train_step_fn(self, session, train_op, global_step,
	570	+ train_step_kwargs):
	571	+ if len(self._enqueue_threads) == 0:
	572	+ for fn in self._init_fns:
	573	+ fn(session)
	574	+ self._start_enqueue(session)
	575	+ train_start = time.time()
	576	+ total_loss, should_stop = slim.learning.train_step(
	577	+ session, train_op, global_step, train_step_kwargs)
	578	+ print("========each step time: {:.2f}".format((time.time() - train_start)*1000))
	579	+ if should_stop or self._coordinator.should_stop():
	580	+ self._stop_all_threads(session)
	581	+ print("========loss: ", total_loss)
	582	+ return total_loss, should_stop
	583	+
	584	+ def _stop_all_threads(self, session):
	585	+ self._coordinator.request_stop()
	586	+ session.run(self._stop_op) # Close the queue.
	587	+
	588	+ def _wait_for_threads(self):
	589	+ self._coordinator.join(self._enqueue_threads)
	590	+ self._enqueue_threads = []
	591	+
	592	+ def _start_enqueue(self, session, num_threads=None):
	593	+ if num_threads is None:
	594	+ num_threads = self._num_enqueue_threads
	595	+ for _ in range(num_threads):
	596	+ thread = threading.Thread(
	597	+ target=self._run_enqueue_thread, args=(session, ))
	598	+ thread.start()
	599	+ self._enqueue_threads.append(thread)
	600	+
	601	+ def _run_enqueue_thread(self, session):
	602	+ try:
	603	+ for data in self._feed_generator:
	604	+ if self._coordinator.should_stop():
	605	+ break
	606	+ try:
	607	+ feed_dict = {
	608	+ var: value for var, value in
	609	+ zip(self._input_vars, data)}
	610	+ session.run(self._enqueue_op, feed_dict=feed_dict)
	611	+ except (tf.errors.CancelledError, tf.errors.AbortedError):
	612	+ # We have been requested to stop enqueuing data.
	613	+ break
	614	+ except Exception as e:
	615	+ print("EnqueueError:", e)
	616	+ self._stop_all_threads(session)

Adeepsort/run_1p.sh+25-0

@@ -0,0 +1,25 @@
	1	+#!/bin/bash
	2	+### Do not need to Configure CANN Environment on Modelarts Platform, because it has been set already.
	3	+### Modelarts Platform command for train
	4	+export TF_CPP_MIN_LOG_LEVEL=2 ## Tensorflow api print Log Config
	5	+export ASCEND_SLOG_PRINT_TO_STDOUT=0 ## Print log on terminal on(1), off(0)
	6	+
	7	+code_dir=${1}
	8	+data_dir=${2}
	9	+result_dir=${3}
	10	+obs_url=${4}
	11	+
	12	+current_time=`date "+%Y-%m-%d-%H-%M-%S"`
	13	+
	14	+python3.7 ${code_dir}/train_market1501.py \
	15	+ --dataset_dir=${data_dir} \
	16	+ #--result=${result_dir} \
	17	+ #--obs_dir=${obs_url} \
	18	+ #--chip='npu' \
	19	+ #--loss_mode=cosine-softmax \
	20	+ --run_id=cosine-softmax \
	21	+ #--platform='modelarts' \
	22	+ #--npu_dump_data=False \
	23	+ #--npu_dump_graph=False \
	24	+ #--npu_profiling=False \
	25	+ #--npu_auto_tune=False 2>&1 \| tee ${result_dir}/${current_time}_train_npu.log

Adeepsort/scripts/run_1p.sh+25-0

@@ -0,0 +1,25 @@
	1	+#!/bin/bash
	2	+### Do not need to Configure CANN Environment on Modelarts Platform, because it has been set already.
	3	+### Modelarts Platform command for train
	4	+export TF_CPP_MIN_LOG_LEVEL=2 ## Tensorflow api print Log Config
	5	+export ASCEND_SLOG_PRINT_TO_STDOUT=0 ## Print log on terminal on(1), off(0)
	6	+
	7	+code_dir=${1}
	8	+data_dir=${2}
	9	+result_dir=${3}
	10	+obs_url=${4}
	11	+
	12	+current_time=`date "+%Y-%m-%d-%H-%M-%S"`
	13	+
	14	+python3.7 ${code_dir}/train_market1501.py
	15	+ #--dataset_dir=${data_dir} \
	16	+ #--result=${result_dir} \
	17	+ #--obs_dir=${obs_url} \
	18	+ #--chip='npu' \
	19	+ #--loss_mode=cosine-softmax \
	20	+ #--run_id=cosine-softmax \
	21	+ #--platform='modelarts' \
	22	+ #--npu_dump_data=False \
	23	+ #--npu_dump_graph=False \
	24	+ #--npu_profiling=False \
	25	+ #--npu_auto_tune=False 2>&1 \| tee ${result_dir}/${current_time}_train_npu.log

Adeepsort/scripts/run_apulis.sh+30-0

@@ -0,0 +1,30 @@
	1	+#!/bin/bash
	2	+### Apulis Platform command for train (CANN Version-20.2)
	3	+## Set Ascend Log Level, if u wanna to print on terminal, you should open 'ASCEND_SLOG_PRINT_TO_STDOUT'.
	4	+## refer to link: https://support.huaweicloud.com/Graphdevg-cann202training1/atlasag_25_0077.html
	5	+export ASCEND_SLOG_PRINT_TO_STDOUT=0 ## Print log on terminal on(1), off(0)
	6	+export ASCEND_GLOBAL_LOG_LEVEL=3 ## Ascend log level. debug(0), info(1), warning(2), error(3)
	7	+export TF_CPP_MIN_LOG_LEVEL=2 ## Tensorflow api print Log Config
	8	+
	9	+## Configure Environment for Auto Tune
	10	+## refer to link:https://support.huaweicloud.com/developmenttg-cann330alphaXtraining/atlasautotune_16_0014.html
	11	+export install_path=/home/HwHiAiUser/Ascend/ascend-toolkit/latest
	12	+export PATH=${install_path}/fwkacllib/bin:$PATH
	13	+export LD_LIBRARY_PATH=${install_path}/fwkacllib/lib64:$LD_LIBRARY_PATH
	14	+export PYTHONPATH=${install_path}/fwkacllib/python/site-packages:$PYTHONPATH
	15	+
	16	+code_dir=$(cd "$(dirname "$0")"; cd ..; pwd)
	17	+echo "===>>>Python boot file dir: ${code_dir}"
	18	+
	19	+current_time=`date "+%Y-%m-%d-%H-%M-%S"`
	20	+
	21	+python3.7 ${code_dir}/train.py \
	22	+ --chip='npu' \
	23	+ --dataset=/data/dataset/storage/flowers/images \
	24	+ --result=./log \
	25	+ --num_classes=5 \
	26	+ --train_step=2 \
	27	+ --npu_dump_data=False \
	28	+ --npu_dump_graph=False \
	29	+ --npu_profiling=False \
	30	+ --npu_auto_tune=False 2>&1 \| tee ${code_dir}/${current_time}_train_npu.log

Adeepsort/train_app.py+690-0

Adeepsort/train_market1501.py+139-0

Adeepsort/train_mars.py+124-0

Adeepsort/vis_tools.py+82-0