Upload 16 files

.gitattributes

@@ -1,35 +1,3 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
 *.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.caffemodel filter=lfs diff=lfs merge=lfs -text
+models/*.caffemodel filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,9 @@
+.idea/
+__pycache__/
+
+build
+caches
+support.egg-info
+pycocotools
+
+*.so

LICENSE

@@ -0,0 +1,674 @@
+                    GNU GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU General Public License is a free, copyleft license for
+software and other kinds of works.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+the GNU General Public License is intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.  We, the Free Software Foundation, use the
+GNU General Public License for most of our software; it applies also to
+any other work released this way by its authors.  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  To protect your rights, we need to prevent others from denying you
+these rights or asking you to surrender the rights.  Therefore, you have
+certain responsibilities if you distribute copies of the software, or if
+you modify it: responsibilities to respect the freedom of others.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must pass on to the recipients the same
+freedoms that you received.  You must make sure that they, too, receive
+or can get the source code.  And you must show them these terms so they
+know their rights.
+
+  Developers that use the GNU GPL protect your rights with two steps:
+(1) assert copyright on the software, and (2) offer you this License
+giving you legal permission to copy, distribute and/or modify it.
+
+  For the developers' and authors' protection, the GPL clearly explains
+that there is no warranty for this free software.  For both users' and
+authors' sake, the GPL requires that modified versions be marked as
+changed, so that their problems will not be attributed erroneously to
+authors of previous versions.
+
+  Some devices are designed to deny users access to install or run
+modified versions of the software inside them, although the manufacturer
+can do so.  This is fundamentally incompatible with the aim of
+protecting users' freedom to change the software.  The systematic
+pattern of such abuse occurs in the area of products for individuals to
+use, which is precisely where it is most unacceptable.  Therefore, we
+have designed this version of the GPL to prohibit the practice for those
+products.  If such problems arise substantially in other domains, we
+stand ready to extend this provision to those domains in future versions
+of the GPL, as needed to protect the freedom of users.
+
+  Finally, every program is threatened constantly by software patents.
+States should not allow patents to restrict development and use of
+software on general-purpose computers, but in those that do, we wish to
+avoid the special danger that patents applied to a free program could
+make it effectively proprietary.  To prevent this, the GPL assures that
+patents cannot be used to render the program non-free.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Use with the GNU Affero General Public License.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU Affero General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the special requirements of the GNU Affero General Public License,
+section 13, concerning interaction through a network will apply to the
+combination as such.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If the program does terminal interaction, make it output a short
+notice like this when it starts in an interactive mode:
+
+    <program>  Copyright (C) <year>  <name of author>
+    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, your program's commands
+might be different; for a GUI interface, you would use an "about box".
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU GPL, see
+<https://www.gnu.org/licenses/>.
+
+  The GNU General Public License does not permit incorporating your program
+into proprietary programs.  If your program is a subroutine library, you
+may consider it more useful to permit linking proprietary applications with
+the library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.  But first, please read
+<https://www.gnu.org/licenses/why-not-lgpl.html>.

README.md

@@ -1,8 +1,8 @@
 ---
-title: Fasterrcnn Project Demo
-emoji: 🚀
+title: Fasterrcnn
+emoji: 📷
 colorFrom: red
-colorTo: indigo
+colorTo: purple
 sdk: streamlit
 sdk_version: 1.26.0
 app_file: app.py
@@ -10,4 +10,4 @@ pinned: false
 license: lgpl-3.0
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Faster RCNN Demo with Web App using Hugging Face Spaces

app.py

@@ -0,0 +1,166 @@
+"""Object detection demo with MobileNet SSD.
+This model and code are based on
+https://github.com/robmarkcole/object-detection-app
+"""
+
+import logging
+import queue
+from pathlib import Path
+from typing import List, NamedTuple
+
+import av
+import cv2
+import numpy as np
+import streamlit as st
+from streamlit_webrtc import WebRtcMode, webrtc_streamer
+
+from sample_utils.download import download_file
+from sample_utils.turn import get_ice_servers
+
+HERE = Path(__file__).parent
+ROOT = HERE
+
+logger = logging.getLogger(__name__)
+
+
+MODEL_URL = "https://github.com/robmarkcole/object-detection-app/raw/master/model/MobileNetSSD_deploy.caffemodel"  # noqa: E501
+MODEL_LOCAL_PATH = ROOT / "./models/MobileNetSSD_deploy.caffemodel"
+PROTOTXT_URL = "https://github.com/robmarkcole/object-detection-app/raw/master/model/MobileNetSSD_deploy.prototxt.txt"  # noqa: E501
+PROTOTXT_LOCAL_PATH = ROOT / "./models/MobileNetSSD_deploy.prototxt.txt"
+
+CLASSES = [
+    "background",
+    "aeroplane",
+    "bicycle",
+    "bird",
+    "boat",
+    "bottle",
+    "bus",
+    "car",
+    "cat",
+    "chair",
+    "cow",
+    "diningtable",
+    "dog",
+    "horse",
+    "motorbike",
+    "person",
+    "pottedplant",
+    "sheep",
+    "sofa",
+    "train",
+    "tvmonitor",
+    ]
+
+
+class Detection(NamedTuple):
+    class_id: int
+    label: str
+    score: float
+    box: np.ndarray
+
+
+@st.cache_resource  # type: ignore
+def generate_label_colors():
+    return np.random.uniform(0, 255, size=(len(CLASSES), 3))
+
+
+COLORS = generate_label_colors()
+
+download_file(MODEL_URL, MODEL_LOCAL_PATH, expected_size=23147564)
+download_file(PROTOTXT_URL, PROTOTXT_LOCAL_PATH, expected_size=29353)
+
+
+# Session-specific caching
+cache_key = "object_detection_dnn"
+if cache_key in st.session_state:
+    net = st.session_state[cache_key]
+else:
+    net = cv2.dnn.readNetFromCaffe(str(PROTOTXT_LOCAL_PATH), str(MODEL_LOCAL_PATH))
+    st.session_state[cache_key] = net
+
+score_threshold = st.slider("Score threshold", 0.0, 1.0, 0.5, 0.05)
+
+# NOTE: The callback will be called in another thread,
+#       so use a queue here for thread-safety to pass the data
+#       from inside to outside the callback.
+# TODO: A general-purpose shared state object may be more useful.
+result_queue: "queue.Queue[List[Detection]]" = queue.Queue()
+
+
+def video_frame_callback(frame: av.VideoFrame) -> av.VideoFrame:
+    image = frame.to_ndarray(format="bgr24")
+
+    # Run inference
+    blob = cv2.dnn.blobFromImage(
+        cv2.resize(image, (300, 300)), 0.007843, (300, 300), 127.5
+    )
+    net.setInput(blob)
+    output = net.forward()
+
+    h, w = image.shape[:2]
+
+    # Convert the output array into a structured form.
+    output = output.squeeze()  # (1, 1, N, 7) -> (N, 7)
+    output = output[output[:, 2] >= score_threshold]
+    detections = [
+        Detection(
+            class_id=int(detection[1]),
+            label=CLASSES[int(detection[1])],
+            score=float(detection[2]),
+            box=(detection[3:7] * np.array([w, h, w, h])),
+        )
+        for detection in output
+    ]
+
+    # Render bounding boxes and captions
+    for detection in detections:
+        caption = f"{detection.label}: {round(detection.score * 100, 2)}%"
+        color = COLORS[detection.class_id]
+        xmin, ymin, xmax, ymax = detection.box.astype("int")
+
+        cv2.rectangle(image, (xmin, ymin), (xmax, ymax), color, 2)
+        cv2.putText(
+            image,
+            caption,
+            (xmin, ymin - 15 if ymin - 15 > 15 else ymin + 15),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.5,
+            color,
+            2,
+        )
+
+    result_queue.put(detections)
+
+    return av.VideoFrame.from_ndarray(image, format="bgr24")
+
+
+webrtc_ctx = webrtc_streamer(
+    key="object-detection",
+    mode=WebRtcMode.SENDRECV,
+    rtc_configuration={
+        "iceServers": get_ice_servers(),
+        "iceTransportPolicy": "relay",
+    },
+    video_frame_callback=video_frame_callback,
+    media_stream_constraints={"video": True, "audio": False},
+    async_processing=True,
+)
+
+if st.checkbox("Show the detected labels", value=True):
+    if webrtc_ctx.state.playing:
+        labels_placeholder = st.empty()
+        # NOTE: The video transformation with object detection and
+        # this loop displaying the result labels are running
+        # in different threads asynchronously.
+        # Then the rendered video frames and the labels displayed here
+        # are not strictly synchronized.
+        while True:
+            result = result_queue.get()
+            labels_placeholder.table(result)
+
+st.markdown(
+    "This demo uses a model and code from "
+    "https://github.com/robmarkcole/object-detection-app. "
+    "Many thanks to the project."
+)

bbox.py

@@ -0,0 +1,92 @@
+from typing import List
+
+import torch
+from torch import Tensor
+
+
+class BBox(object):
+
+    def __init__(self, left: float, top: float, right: float, bottom: float):
+        super().__init__()
+        self.left = left
+        self.top = top
+        self.right = right
+        self.bottom = bottom
+
+    def __repr__(self) -> str:
+        return 'BBox[l={:.1f}, t={:.1f}, r={:.1f}, b={:.1f}]'.format(
+            self.left, self.top, self.right, self.bottom)
+
+    def tolist(self) -> List[float]:
+        return [self.left, self.top, self.right, self.bottom]
+
+    @staticmethod
+    def to_center_base(bboxes: Tensor) -> Tensor:
+        return torch.stack([
+            (bboxes[..., 0] + bboxes[..., 2]) / 2,
+            (bboxes[..., 1] + bboxes[..., 3]) / 2,
+            bboxes[..., 2] - bboxes[..., 0],
+            bboxes[..., 3] - bboxes[..., 1]
+        ], dim=-1)
+
+    @staticmethod
+    def from_center_base(center_based_bboxes: Tensor) -> Tensor:
+        return torch.stack([
+            center_based_bboxes[..., 0] - center_based_bboxes[..., 2] / 2,
+            center_based_bboxes[..., 1] - center_based_bboxes[..., 3] / 2,
+            center_based_bboxes[..., 0] + center_based_bboxes[..., 2] / 2,
+            center_based_bboxes[..., 1] + center_based_bboxes[..., 3] / 2
+        ], dim=-1)
+
+    @staticmethod
+    def calc_transformer(src_bboxes: Tensor, dst_bboxes: Tensor) -> Tensor:
+        center_based_src_bboxes = BBox.to_center_base(src_bboxes)
+        center_based_dst_bboxes = BBox.to_center_base(dst_bboxes)
+        transformers = torch.stack([
+            (center_based_dst_bboxes[..., 0] - center_based_src_bboxes[..., 0]) / center_based_src_bboxes[..., 2],
+            (center_based_dst_bboxes[..., 1] - center_based_src_bboxes[..., 1]) / center_based_src_bboxes[..., 3],
+            torch.log(center_based_dst_bboxes[..., 2] / center_based_src_bboxes[..., 2]),
+            torch.log(center_based_dst_bboxes[..., 3] / center_based_src_bboxes[..., 3])
+        ], dim=-1)
+        return transformers
+
+    @staticmethod
+    def apply_transformer(src_bboxes: Tensor, transformers: Tensor) -> Tensor:
+        center_based_src_bboxes = BBox.to_center_base(src_bboxes)
+        center_based_dst_bboxes = torch.stack([
+            transformers[..., 0] * center_based_src_bboxes[..., 2] + center_based_src_bboxes[..., 0],
+            transformers[..., 1] * center_based_src_bboxes[..., 3] + center_based_src_bboxes[..., 1],
+            torch.exp(transformers[..., 2]) * center_based_src_bboxes[..., 2],
+            torch.exp(transformers[..., 3]) * center_based_src_bboxes[..., 3]
+        ], dim=-1)
+        dst_bboxes = BBox.from_center_base(center_based_dst_bboxes)
+        return dst_bboxes
+
+    @staticmethod
+    def iou(source: Tensor, other: Tensor) -> Tensor:
+        source, other = source.unsqueeze(dim=-2).repeat(1, 1, other.shape[-2], 1), \
+                        other.unsqueeze(dim=-3).repeat(1, source.shape[-2], 1, 1)
+
+        source_area = (source[..., 2] - source[..., 0]) * (source[..., 3] - source[..., 1])
+        other_area = (other[..., 2] - other[..., 0]) * (other[..., 3] - other[..., 1])
+
+        intersection_left = torch.max(source[..., 0], other[..., 0])
+        intersection_top = torch.max(source[..., 1], other[..., 1])
+        intersection_right = torch.min(source[..., 2], other[..., 2])
+        intersection_bottom = torch.min(source[..., 3], other[..., 3])
+        intersection_width = torch.clamp(intersection_right - intersection_left, min=0)
+        intersection_height = torch.clamp(intersection_bottom - intersection_top, min=0)
+        intersection_area = intersection_width * intersection_height
+
+        return intersection_area / (source_area + other_area - intersection_area)
+
+    @staticmethod
+    def inside(bboxes: Tensor, left: float, top: float, right: float, bottom: float) -> Tensor:
+        return ((bboxes[..., 0] >= left) * (bboxes[..., 1] >= top) *
+                (bboxes[..., 2] <= right) * (bboxes[..., 3] <= bottom))
+
+    @staticmethod
+    def clip(bboxes: Tensor, left: float, top: float, right: float, bottom: float) -> Tensor:
+        bboxes[..., [0, 2]] = bboxes[..., [0, 2]].clamp(min=left, max=right)
+        bboxes[..., [1, 3]] = bboxes[..., [1, 3]].clamp(min=top, max=bottom)
+        return bboxes

eval.py

@@ -0,0 +1,74 @@
+import argparse
+import os
+import time
+
+import uuid
+
+from backbone.base import Base as BackboneBase
+from config.eval_config import EvalConfig as Config
+from dataset.base import Base as DatasetBase
+from evaluator import Evaluator
+from logger import Logger as Log
+from model import Model
+from roi.pooler import Pooler
+
+
+def _eval(path_to_checkpoint: str, dataset_name: str, backbone_name: str, path_to_data_dir: str, path_to_results_dir: str):
+    dataset = DatasetBase.from_name(dataset_name)(path_to_data_dir, DatasetBase.Mode.EVAL, Config.IMAGE_MIN_SIDE, Config.IMAGE_MAX_SIDE)
+    evaluator = Evaluator(dataset, path_to_data_dir, path_to_results_dir)
+
+    Log.i('Found {:d} samples'.format(len(dataset)))
+
+    backbone = BackboneBase.from_name(backbone_name)(pretrained=False)
+    model = Model(backbone, dataset.num_classes(), pooler_mode=Config.POOLER_MODE,
+                  anchor_ratios=Config.ANCHOR_RATIOS, anchor_sizes=Config.ANCHOR_SIZES,
+                  rpn_pre_nms_top_n=Config.RPN_PRE_NMS_TOP_N, rpn_post_nms_top_n=Config.RPN_POST_NMS_TOP_N).cuda()
+    model.load(path_to_checkpoint)
+
+    Log.i('Start evaluating with 1 GPU (1 batch per GPU)')
+    mean_ap, detail = evaluator.evaluate(model)
+    Log.i('Done')
+
+    Log.i('mean AP = {:.4f}'.format(mean_ap))
+    Log.i('\n' + detail)
+
+
+if __name__ == '__main__':
+    def main():
+        parser = argparse.ArgumentParser()
+        parser.add_argument('-s', '--dataset', type=str, choices=DatasetBase.OPTIONS, required=True, help='name of dataset')
+        parser.add_argument('-b', '--backbone', type=str, choices=BackboneBase.OPTIONS, required=True, help='name of backbone model')
+        parser.add_argument('-d', '--data_dir', type=str, default='./data', help='path to data directory')
+        parser.add_argument('--image_min_side', type=float, help='default: {:g}'.format(Config.IMAGE_MIN_SIDE))
+        parser.add_argument('--image_max_side', type=float, help='default: {:g}'.format(Config.IMAGE_MAX_SIDE))
+        parser.add_argument('--anchor_ratios', type=str, help='default: "{!s}"'.format(Config.ANCHOR_RATIOS))
+        parser.add_argument('--anchor_sizes', type=str, help='default: "{!s}"'.format(Config.ANCHOR_SIZES))
+        parser.add_argument('--pooler_mode', type=str, choices=Pooler.OPTIONS, help='default: {.value:s}'.format(Config.POOLER_MODE))
+        parser.add_argument('--rpn_pre_nms_top_n', type=int, help='default: {:d}'.format(Config.RPN_PRE_NMS_TOP_N))
+        parser.add_argument('--rpn_post_nms_top_n', type=int, help='default: {:d}'.format(Config.RPN_POST_NMS_TOP_N))
+        parser.add_argument('--checkpoint', type=str, help='path to evaluating checkpoint')
+        args = parser.parse_args()
+
+        path_to_checkpoint = args.checkpoint
+        dataset_name = args.dataset
+        backbone_name = args.backbone
+        path_to_data_dir = args.data_dir
+
+        path_to_results_dir = os.path.join(os.path.dirname(path_to_checkpoint), 'results-{:s}-{:s}-{:s}'.format(
+            time.strftime('%Y%m%d%H%M%S'), path_to_checkpoint.split(os.path.sep)[-1].split(os.path.curdir)[0],
+            str(uuid.uuid4()).split('-')[0]))
+        os.makedirs(path_to_results_dir)
+
+        Config.setup(image_min_side=args.image_min_side, image_max_side=args.image_max_side,
+                     anchor_ratios=args.anchor_ratios, anchor_sizes=args.anchor_sizes, pooler_mode=args.pooler_mode,
+                     rpn_pre_nms_top_n=args.rpn_pre_nms_top_n, rpn_post_nms_top_n=args.rpn_post_nms_top_n)
+
+        Log.initialize(os.path.join(path_to_results_dir, 'eval.log'))
+        Log.i('Arguments:')
+        for k, v in vars(args).items():
+            Log.i(f'\t{k} = {v}')
+        Log.i(Config.describe())
+
+        _eval(path_to_checkpoint, dataset_name, backbone_name, path_to_data_dir, path_to_results_dir)
+
+    main()

evaluator.py

@@ -0,0 +1,46 @@
+from typing import Tuple
+
+import torch
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+
+from dataset.base import Base as DatasetBase
+from model import Model
+
+
+class Evaluator(object):
+    def __init__(self, dataset: DatasetBase, path_to_data_dir: str, path_to_results_dir: str):
+        super().__init__()
+        self._dataset = dataset
+        self._dataloader = DataLoader(dataset, batch_size=1, shuffle=False, num_workers=8, pin_memory=True)
+        self._path_to_data_dir = path_to_data_dir
+        self._path_to_results_dir = path_to_results_dir
+
+    def evaluate(self, model: Model) -> Tuple[float, str]:
+        all_image_ids, all_detection_bboxes, all_detection_classes, all_detection_probs = [], [], [], []
+
+        with torch.no_grad():
+            for _, (image_id_batch, image_batch, scale_batch, _, _) in enumerate(tqdm(self._dataloader)):
+                image_batch = image_batch.cuda()
+                assert image_batch.shape[0] == 1, 'do not use batch size more than 1 on evaluation'
+
+                detection_bboxes, detection_classes, detection_probs, detection_batch_indices = model.eval().forward(image_batch)
+
+                scale_batch = scale_batch[detection_batch_indices].unsqueeze(dim=-1).expand_as(detection_bboxes).to(device=detection_bboxes.device)
+                detection_bboxes = detection_bboxes / scale_batch
+
+                kept_indices = (detection_probs > 0.05).nonzero().view(-1)
+                detection_bboxes = detection_bboxes[kept_indices]
+                detection_classes = detection_classes.to(device=detection_bboxes.device)
+                detection_classes = detection_classes[kept_indices]
+                detection_probs = detection_probs[kept_indices]
+                detection_batch_indices = detection_batch_indices.to(device=detection_bboxes.device)
+                detection_batch_indices = detection_batch_indices[kept_indices]
+
+                all_detection_bboxes.extend(detection_bboxes.tolist())
+                all_detection_classes.extend(detection_classes.tolist())
+                all_detection_probs.extend(detection_probs.tolist())
+                all_image_ids.extend([image_id_batch[i] for i in detection_batch_indices])
+
+        mean_ap, detail = self._dataset.evaluate(self._path_to_results_dir, all_image_ids, all_detection_bboxes, all_detection_classes, all_detection_probs)
+        return mean_ap, detail

infer.py

@@ -0,0 +1,89 @@
+import argparse
+import os
+import random
+import torch
+
+from PIL import ImageDraw
+from torchvision.transforms import transforms
+from dataset.base import Base as DatasetBase
+from backbone.base import Base as BackboneBase
+from bbox import BBox
+from model import Model
+from roi.pooler import Pooler
+from config.eval_config import EvalConfig as Config
+
+
+def _infer(path_to_input_image: str, path_to_output_image: str, path_to_checkpoint: str, dataset_name: str, backbone_name: str, prob_thresh: float):
+    dataset_class = DatasetBase.from_name(dataset_name)
+    backbone = BackboneBase.from_name(backbone_name)(pretrained=False)
+    model = Model(backbone, dataset_class.num_classes(), pooler_mode=Config.POOLER_MODE,
+                  anchor_ratios=Config.ANCHOR_RATIOS, anchor_sizes=Config.ANCHOR_SIZES,
+                  rpn_pre_nms_top_n=Config.RPN_PRE_NMS_TOP_N, rpn_post_nms_top_n=Config.RPN_POST_NMS_TOP_N).cuda()
+    model.load(path_to_checkpoint)
+
+    with torch.no_grad():
+        image = transforms.Image.open(path_to_input_image)
+        image_tensor, scale = dataset_class.preprocess(image, Config.IMAGE_MIN_SIDE, Config.IMAGE_MAX_SIDE)
+
+        detection_bboxes, detection_classes, detection_probs, _ = \
+            model.eval().forward(image_tensor.unsqueeze(dim=0).cuda())
+        detection_bboxes /= scale
+
+        kept_indices = detection_probs > prob_thresh
+        detection_bboxes = detection_bboxes[kept_indices]
+        detection_classes = detection_classes[kept_indices]
+        detection_probs = detection_probs[kept_indices]
+
+        draw = ImageDraw.Draw(image)
+
+        for bbox, cls, prob in zip(detection_bboxes.tolist(), detection_classes.tolist(), detection_probs.tolist()):
+            color = random.choice(['red', 'green', 'blue', 'yellow', 'purple', 'white'])
+            bbox = BBox(left=bbox[0], top=bbox[1], right=bbox[2], bottom=bbox[3])
+            category = dataset_class.LABEL_TO_CATEGORY_DICT[cls]
+
+            draw.rectangle(((bbox.left, bbox.top), (bbox.right, bbox.bottom)), outline=color)
+            draw.text((bbox.left, bbox.top), text=f'{category:s} {prob:.3f}', fill=color)
+
+        image.save(path_to_output_image)
+        print(f'Output image is saved to {path_to_output_image}')
+
+
+if __name__ == '__main__':
+    def main():
+        parser = argparse.ArgumentParser()
+        parser.add_argument('-s', '--dataset', type=str, choices=DatasetBase.OPTIONS, required=True, help='name of dataset')
+        parser.add_argument('-b', '--backbone', type=str, choices=BackboneBase.OPTIONS, required=True, help='name of backbone model')
+        parser.add_argument('-c', '--checkpoint', type=str, required=True, help='path to checkpoint')
+        parser.add_argument('-p', '--probability_threshold', type=float, default=0.6, help='threshold of detection probability')
+        parser.add_argument('--image_min_side', type=float, help='default: {:g}'.format(Config.IMAGE_MIN_SIDE))
+        parser.add_argument('--image_max_side', type=float, help='default: {:g}'.format(Config.IMAGE_MAX_SIDE))
+        parser.add_argument('--anchor_ratios', type=str, help='default: "{!s}"'.format(Config.ANCHOR_RATIOS))
+        parser.add_argument('--anchor_sizes', type=str, help='default: "{!s}"'.format(Config.ANCHOR_SIZES))
+        parser.add_argument('--pooler_mode', type=str, choices=Pooler.OPTIONS, help='default: {.value:s}'.format(Config.POOLER_MODE))
+        parser.add_argument('--rpn_pre_nms_top_n', type=int, help='default: {:d}'.format(Config.RPN_PRE_NMS_TOP_N))
+        parser.add_argument('--rpn_post_nms_top_n', type=int, help='default: {:d}'.format(Config.RPN_POST_NMS_TOP_N))
+        parser.add_argument('input', type=str, help='path to input image')
+        parser.add_argument('output', type=str, help='path to output result image')
+        args = parser.parse_args()
+
+        path_to_input_image = args.input
+        path_to_output_image = args.output
+        dataset_name = args.dataset
+        backbone_name = args.backbone
+        path_to_checkpoint = args.checkpoint
+        prob_thresh = args.probability_threshold
+
+        os.makedirs(os.path.join(os.path.curdir, os.path.dirname(path_to_output_image)), exist_ok=True)
+
+        Config.setup(image_min_side=args.image_min_side, image_max_side=args.image_max_side,
+                     anchor_ratios=args.anchor_ratios, anchor_sizes=args.anchor_sizes, pooler_mode=args.pooler_mode,
+                     rpn_pre_nms_top_n=args.rpn_pre_nms_top_n, rpn_post_nms_top_n=args.rpn_post_nms_top_n)
+
+        print('Arguments:')
+        for k, v in vars(args).items():
+            print(f'\t{k} = {v}')
+        print(Config.describe())
+
+        _infer(path_to_input_image, path_to_output_image, path_to_checkpoint, dataset_name, backbone_name, prob_thresh)
+
+    main()

infer_stream.py

@@ -0,0 +1,117 @@
+import argparse
+import itertools
+import random
+import time
+import torch
+
+import cv2
+import numpy as np
+from PIL import ImageDraw, Image
+
+from backbone.base import Base as BackboneBase
+from config.eval_config import EvalConfig as Config
+from dataset.base import Base as DatasetBase
+from bbox import BBox
+from model import Model
+from roi.pooler import Pooler
+
+
+def _infer_stream(path_to_input_stream_endpoint: str, period_of_inference: int, path_to_checkpoint: str, dataset_name: str, backbone_name: str, prob_thresh: float):
+    dataset_class = DatasetBase.from_name(dataset_name)
+    backbone = BackboneBase.from_name(backbone_name)(pretrained=False)
+    model = Model(backbone, dataset_class.num_classes(), pooler_mode=Config.POOLER_MODE,
+                  anchor_ratios=Config.ANCHOR_RATIOS, anchor_sizes=Config.ANCHOR_SIZES,
+                  rpn_pre_nms_top_n=Config.RPN_PRE_NMS_TOP_N, rpn_post_nms_top_n=Config.RPN_POST_NMS_TOP_N).cuda()
+    model.load(path_to_checkpoint)
+
+    if path_to_input_stream_endpoint.isdigit():
+        path_to_input_stream_endpoint = int(path_to_input_stream_endpoint)
+    video_capture = cv2.VideoCapture(path_to_input_stream_endpoint)
+
+    with torch.no_grad():
+        for sn in itertools.count(start=1):
+            success, frame = video_capture.read()
+            
+            if not success:
+                break
+
+            if sn % period_of_inference != 0:
+                continue
+
+            timestamp = time.time()
+
+            image = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            image = Image.fromarray(image)
+            image_tensor, scale = dataset_class.preprocess(image, Config.IMAGE_MIN_SIDE, Config.IMAGE_MAX_SIDE)
+
+            detection_bboxes, detection_classes, detection_probs, _ = \
+                model.eval().forward(image_tensor.unsqueeze(dim=0).cuda())
+            detection_bboxes /= scale
+
+            kept_indices = detection_probs > prob_thresh
+            detection_bboxes = detection_bboxes[kept_indices]
+            detection_classes = detection_classes[kept_indices]
+            detection_probs = detection_probs[kept_indices]
+
+            draw = ImageDraw.Draw(image)
+
+            for bbox, cls, prob in zip(detection_bboxes.tolist(), detection_classes.tolist(), detection_probs.tolist()):
+                color = random.choice(['red', 'green', 'blue', 'yellow', 'purple', 'white'])
+                bbox = BBox(left=bbox[0], top=bbox[1], right=bbox[2], bottom=bbox[3])
+                category = dataset_class.LABEL_TO_CATEGORY_DICT[cls]
+
+                draw.rectangle(((bbox.left, bbox.top), (bbox.right, bbox.bottom)), outline=color)
+                draw.text((bbox.left, bbox.top), text=f'{category:s} {prob:.3f}', fill=color)
+
+            image = np.array(image)
+            frame = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
+
+            elapse = time.time() - timestamp
+            fps = 1 / elapse
+            cv2.putText(frame, f'FPS = {fps:.1f}', (20, 20), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 1, cv2.LINE_AA)
+
+            cv2.imshow('easy-faster-rcnn-pytorch', frame)
+            if cv2.waitKey(10) == 27:
+                break
+
+    video_capture.release()
+    cv2.destroyAllWindows()
+
+
+if __name__ == '__main__':
+    def main():
+        parser = argparse.ArgumentParser()
+        parser.add_argument('-s', '--dataset', type=str, choices=DatasetBase.OPTIONS, required=True, help='name of dataset')
+        parser.add_argument('-b', '--backbone', type=str, choices=BackboneBase.OPTIONS, required=True, help='name of backbone model')
+        parser.add_argument('-c', '--checkpoint', type=str, required=True, help='path to checkpoint')
+        parser.add_argument('-p', '--probability_threshold', type=float, default=0.6, help='threshold of detection probability')
+        parser.add_argument('--image_min_side', type=float, help='default: {:g}'.format(Config.IMAGE_MIN_SIDE))
+        parser.add_argument('--image_max_side', type=float, help='default: {:g}'.format(Config.IMAGE_MAX_SIDE))
+        parser.add_argument('--anchor_ratios', type=str, help='default: "{!s}"'.format(Config.ANCHOR_RATIOS))
+        parser.add_argument('--anchor_sizes', type=str, help='default: "{!s}"'.format(Config.ANCHOR_SIZES))
+        parser.add_argument('--pooler_mode', type=str, choices=Pooler.OPTIONS, help='default: {.value:s}'.format(Config.POOLER_MODE))
+        parser.add_argument('--rpn_pre_nms_top_n', type=int, help='default: {:d}'.format(Config.RPN_PRE_NMS_TOP_N))
+        parser.add_argument('--rpn_post_nms_top_n', type=int, help='default: {:d}'.format(Config.RPN_POST_NMS_TOP_N))
+        parser.add_argument('--stream_input', type=str, help='path to input stream endpoint')
+        parser.add_argument('--period', type=int, help='period of inference')
+        args = parser.parse_args()
+
+        path_to_input_stream_endpoint = args.stream_input
+        period_of_inference = args.period
+        dataset_name = args.dataset
+        backbone_name = args.backbone
+        path_to_checkpoint = args.checkpoint
+        prob_thresh = args.probability_threshold
+
+        Config.setup(image_min_side=args.image_min_side, image_max_side=args.image_max_side,
+                     anchor_ratios=args.anchor_ratios, anchor_sizes=args.anchor_sizes, pooler_mode=args.pooler_mode,
+                     rpn_pre_nms_top_n=args.rpn_pre_nms_top_n, rpn_post_nms_top_n=args.rpn_post_nms_top_n)
+
+        print('Arguments:')
+        for k, v in vars(args).items():
+            print(f'\t{k} = {v}')
+        print(Config.describe())
+
+        _infer_stream(path_to_input_stream_endpoint, period_of_inference, path_to_checkpoint, dataset_name, backbone_name, prob_thresh)
+
+    main()

logger.py

@@ -0,0 +1,35 @@
+import logging
+
+
+class Logger(object):
+    Initialized = False
+
+    @staticmethod
+    def initialize(path_to_log_file):
+        logging.basicConfig(level=logging.INFO,
+                            format='%(asctime)s %(levelname)-8s %(message)s',
+                            datefmt='%Y-%m-%d %H:%M:%S',
+                            handlers=[logging.FileHandler(path_to_log_file),
+                                      logging.StreamHandler()])
+        Logger.Initialized = True
+
+    @staticmethod
+    def log(level, message):
+        assert Logger.Initialized, 'Logger has not been initialized'
+        logging.log(level, message)
+
+    @staticmethod
+    def d(message):
+        Logger.log(logging.DEBUG, message)
+
+    @staticmethod
+    def i(message):
+        Logger.log(logging.INFO, message)
+
+    @staticmethod
+    def w(message):
+        Logger.log(logging.WARNING, message)
+
+    @staticmethod
+    def e(message):
+        Logger.log(logging.ERROR, message)

model.py

@@ -0,0 +1,218 @@
+import os
+from typing import Union, Tuple, List, Optional
+
+import torch
+from torch import nn, Tensor
+from torch.nn import functional as F
+from torch.optim import Optimizer
+from torch.optim.lr_scheduler import _LRScheduler
+
+from backbone.base import Base as BackboneBase
+from bbox import BBox
+from extension.functional import beta_smooth_l1_loss
+from roi.pooler import Pooler
+from rpn.region_proposal_network import RegionProposalNetwork
+#from support.layer.nms import nms
+from torchvision.ops import nms
+
+
+class Model(nn.Module):
+
+    def __init__(self, backbone: BackboneBase, num_classes: int, pooler_mode: Pooler.Mode,
+                 anchor_ratios: List[Tuple[int, int]], anchor_sizes: List[int],
+                 rpn_pre_nms_top_n: int, rpn_post_nms_top_n: int,
+                 anchor_smooth_l1_loss_beta: Optional[float] = None, proposal_smooth_l1_loss_beta: Optional[float] = None):
+        super().__init__()
+
+        self.features, hidden, num_features_out, num_hidden_out = backbone.features()
+        self._bn_modules = nn.ModuleList([it for it in self.features.modules() if isinstance(it, nn.BatchNorm2d)] +
+                                         [it for it in hidden.modules() if isinstance(it, nn.BatchNorm2d)])
+
+        # NOTE: It's crucial to freeze batch normalization modules for few batches training, which can be done by following processes
+        #       (1) Change mode to `eval`
+        #       (2) Disable gradient (we move this process into `forward`)
+        for bn_module in self._bn_modules:
+            for parameter in bn_module.parameters():
+                parameter.requires_grad = False
+
+        self.rpn = RegionProposalNetwork(num_features_out, anchor_ratios, anchor_sizes, rpn_pre_nms_top_n, rpn_post_nms_top_n, anchor_smooth_l1_loss_beta)
+        self.detection = Model.Detection(pooler_mode, hidden, num_hidden_out, num_classes, proposal_smooth_l1_loss_beta)
+
+    def forward(self, image_batch: Tensor,
+                gt_bboxes_batch: Tensor = None, gt_classes_batch: Tensor = None) -> Union[Tuple[Tensor, Tensor, Tensor, Tensor],
+                                                                                          Tuple[Tensor, Tensor, Tensor, Tensor]]:
+        # disable gradient for each forwarding process just in case model was switched to `train` mode at any time
+        for bn_module in self._bn_modules:
+            bn_module.eval()
+
+        features = self.features(image_batch)
+
+        batch_size, _, image_height, image_width = image_batch.shape
+        _, _, features_height, features_width = features.shape
+
+        anchor_bboxes = self.rpn.generate_anchors(image_width, image_height, num_x_anchors=features_width, num_y_anchors=features_height).to(features).repeat(batch_size, 1, 1)
+
+        if self.training:
+            anchor_objectnesses, anchor_transformers, anchor_objectness_losses, anchor_transformer_losses = self.rpn.forward(features, anchor_bboxes, gt_bboxes_batch, image_width, image_height)
+            proposal_bboxes = self.rpn.generate_proposals(anchor_bboxes, anchor_objectnesses, anchor_transformers, image_width, image_height).detach()  # it's necessary to detach `proposal_bboxes` here
+            proposal_classes, proposal_transformers, proposal_class_losses, proposal_transformer_losses = self.detection.forward(features, proposal_bboxes, gt_classes_batch, gt_bboxes_batch)
+            return anchor_objectness_losses, anchor_transformer_losses, proposal_class_losses, proposal_transformer_losses
+        else:
+            anchor_objectnesses, anchor_transformers = self.rpn.forward(features)
+            proposal_bboxes = self.rpn.generate_proposals(anchor_bboxes, anchor_objectnesses, anchor_transformers, image_width, image_height)
+            proposal_classes, proposal_transformers = self.detection.forward(features, proposal_bboxes)
+            detection_bboxes, detection_classes, detection_probs, detection_batch_indices = self.detection.generate_detections(proposal_bboxes, proposal_classes, proposal_transformers, image_width, image_height)
+            return detection_bboxes, detection_classes, detection_probs, detection_batch_indices
+
+    def save(self, path_to_checkpoints_dir: str, step: int, optimizer: Optimizer, scheduler: _LRScheduler) -> str:
+        path_to_checkpoint = os.path.join(path_to_checkpoints_dir, f'model-{step}.pth')
+        checkpoint = {
+            'state_dict': self.state_dict(),
+            'step': step,
+            'optimizer_state_dict': optimizer.state_dict(),
+            'scheduler_state_dict': scheduler.state_dict()
+        }
+        torch.save(checkpoint, path_to_checkpoint)
+        return path_to_checkpoint
+
+    def load(self, path_to_checkpoint: str, optimizer: Optimizer = None, scheduler: _LRScheduler = None) -> 'Model':
+        checkpoint = torch.load(path_to_checkpoint)
+        self.load_state_dict(checkpoint['state_dict'])
+        step = checkpoint['step']
+        if optimizer is not None:
+            optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
+        if scheduler is not None:
+            scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
+        return step
+
+    class Detection(nn.Module):
+
+        def __init__(self, pooler_mode: Pooler.Mode, hidden: nn.Module, num_hidden_out: int, num_classes: int, proposal_smooth_l1_loss_beta: float):
+            super().__init__()
+            self._pooler_mode = pooler_mode
+            self.hidden = hidden
+            self.num_classes = num_classes
+            self._proposal_class = nn.Linear(num_hidden_out, num_classes)
+            self._proposal_transformer = nn.Linear(num_hidden_out, num_classes * 4)
+            self._proposal_smooth_l1_loss_beta = proposal_smooth_l1_loss_beta
+            self._transformer_normalize_mean = torch.tensor([0., 0., 0., 0.], dtype=torch.float)
+            self._transformer_normalize_std = torch.tensor([.1, .1, .2, .2], dtype=torch.float)
+
+        def forward(self, features: Tensor, proposal_bboxes: Tensor,
+                    gt_classes_batch: Optional[Tensor] = None, gt_bboxes_batch: Optional[Tensor] = None) -> Union[Tuple[Tensor, Tensor], Tuple[Tensor, Tensor, Tensor, Tensor]]:
+            batch_size = features.shape[0]
+
+            if not self.training:
+                proposal_batch_indices = torch.arange(end=batch_size, dtype=torch.long, device=proposal_bboxes.device).view(-1, 1).repeat(1, proposal_bboxes.shape[1])
+                pool = Pooler.apply(features, proposal_bboxes.view(-1, 4), proposal_batch_indices.view(-1), mode=self._pooler_mode)
+                hidden = self.hidden(pool)
+                hidden = F.adaptive_max_pool2d(input=hidden, output_size=1)
+                hidden = hidden.view(hidden.shape[0], -1)
+
+                proposal_classes = self._proposal_class(hidden)
+                proposal_transformers = self._proposal_transformer(hidden)
+
+                proposal_classes = proposal_classes.view(batch_size, -1, proposal_classes.shape[-1])
+                proposal_transformers = proposal_transformers.view(batch_size, -1, proposal_transformers.shape[-1])
+                return proposal_classes, proposal_transformers
+            else:
+                # find labels for each `proposal_bboxes`
+                labels = torch.full((batch_size, proposal_bboxes.shape[1]), -1, dtype=torch.long, device=proposal_bboxes.device)
+                ious = BBox.iou(proposal_bboxes, gt_bboxes_batch)
+                proposal_max_ious, proposal_assignments = ious.max(dim=2)
+                labels[proposal_max_ious < 0.5] = 0
+                fg_masks = proposal_max_ious >= 0.5
+                if len(fg_masks.nonzero()) > 0:
+                    labels[fg_masks] = gt_classes_batch[fg_masks.nonzero()[:, 0], proposal_assignments[fg_masks]]
+
+                # select 128 x `batch_size` samples
+                fg_indices = (labels > 0).nonzero()
+                bg_indices = (labels == 0).nonzero()
+                fg_indices = fg_indices[torch.randperm(len(fg_indices))[:min(len(fg_indices), 32 * batch_size)]]
+                bg_indices = bg_indices[torch.randperm(len(bg_indices))[:128 * batch_size - len(fg_indices)]]
+                selected_indices = torch.cat([fg_indices, bg_indices], dim=0)
+                selected_indices = selected_indices[torch.randperm(len(selected_indices))].unbind(dim=1)
+
+                proposal_bboxes = proposal_bboxes[selected_indices]
+                gt_bboxes = gt_bboxes_batch[selected_indices[0], proposal_assignments[selected_indices]]
+                gt_proposal_classes = labels[selected_indices]
+                gt_proposal_transformers = BBox.calc_transformer(proposal_bboxes, gt_bboxes)
+                batch_indices = selected_indices[0]
+
+                pool = Pooler.apply(features, proposal_bboxes, proposal_batch_indices=batch_indices, mode=self._pooler_mode)
+                hidden = self.hidden(pool)
+                hidden = F.adaptive_max_pool2d(input=hidden, output_size=1)
+                hidden = hidden.view(hidden.shape[0], -1)
+
+                proposal_classes = self._proposal_class(hidden)
+                proposal_transformers = self._proposal_transformer(hidden)
+                proposal_class_losses, proposal_transformer_losses = self.loss(proposal_classes, proposal_transformers,
+                                                                               gt_proposal_classes, gt_proposal_transformers,
+                                                                               batch_size, batch_indices)
+
+                return proposal_classes, proposal_transformers, proposal_class_losses, proposal_transformer_losses
+
+        def loss(self, proposal_classes: Tensor, proposal_transformers: Tensor,
+                 gt_proposal_classes: Tensor, gt_proposal_transformers: Tensor,
+                 batch_size, batch_indices) -> Tuple[Tensor, Tensor]:
+            proposal_transformers = proposal_transformers.view(-1, self.num_classes, 4)[torch.arange(end=len(proposal_transformers), dtype=torch.long), gt_proposal_classes]
+            transformer_normalize_mean = self._transformer_normalize_mean.to(device=gt_proposal_transformers.device)
+            transformer_normalize_std = self._transformer_normalize_std.to(device=gt_proposal_transformers.device)
+            gt_proposal_transformers = (gt_proposal_transformers - transformer_normalize_mean) / transformer_normalize_std  # scale up target to make regressor easier to learn
+
+            cross_entropies = torch.empty(batch_size, dtype=torch.float, device=proposal_classes.device)
+            smooth_l1_losses = torch.empty(batch_size, dtype=torch.float, device=proposal_transformers.device)
+
+            for batch_index in range(batch_size):
+                selected_indices = (batch_indices == batch_index).nonzero().view(-1)
+
+                cross_entropy = F.cross_entropy(input=proposal_classes[selected_indices],
+                                                target=gt_proposal_classes[selected_indices])
+
+                fg_indices = gt_proposal_classes[selected_indices].nonzero().view(-1)
+                smooth_l1_loss = beta_smooth_l1_loss(input=proposal_transformers[selected_indices][fg_indices],
+                                                     target=gt_proposal_transformers[selected_indices][fg_indices],
+                                                     beta=self._proposal_smooth_l1_loss_beta)
+
+                cross_entropies[batch_index] = cross_entropy
+                smooth_l1_losses[batch_index] = smooth_l1_loss
+
+            return cross_entropies, smooth_l1_losses
+
+        def generate_detections(self, proposal_bboxes: Tensor, proposal_classes: Tensor, proposal_transformers: Tensor, image_width: int, image_height: int) -> Tuple[Tensor, Tensor, Tensor, Tensor]:
+            batch_size = proposal_bboxes.shape[0]
+
+            proposal_transformers = proposal_transformers.view(batch_size, -1, self.num_classes, 4)
+            transformer_normalize_std = self._transformer_normalize_std.to(device=proposal_transformers.device)
+            transformer_normalize_mean = self._transformer_normalize_mean.to(device=proposal_transformers.device)
+            proposal_transformers = proposal_transformers * transformer_normalize_std + transformer_normalize_mean
+
+            proposal_bboxes = proposal_bboxes.unsqueeze(dim=2).repeat(1, 1, self.num_classes, 1)
+            detection_bboxes = BBox.apply_transformer(proposal_bboxes, proposal_transformers)
+            detection_bboxes = BBox.clip(detection_bboxes, left=0, top=0, right=image_width, bottom=image_height)
+            detection_probs = F.softmax(proposal_classes, dim=-1)
+
+            all_detection_bboxes = []
+            all_detection_classes = []
+            all_detection_probs = []
+            all_detection_batch_indices = []
+
+            for batch_index in range(batch_size):
+                for c in range(1, self.num_classes):
+                    class_bboxes = detection_bboxes[batch_index, :, c, :]
+                    class_probs = detection_probs[batch_index, :, c]
+                    threshold = 0.3
+                    kept_indices = nms(class_bboxes, class_probs, threshold)
+                    class_bboxes = class_bboxes[kept_indices]
+                    class_probs = class_probs[kept_indices]
+
+                    all_detection_bboxes.append(class_bboxes)
+                    all_detection_classes.append(torch.full((len(kept_indices),), c, dtype=torch.int))
+                    all_detection_probs.append(class_probs)
+                    all_detection_batch_indices.append(torch.full((len(kept_indices),), batch_index, dtype=torch.long))
+
+            all_detection_bboxes = torch.cat(all_detection_bboxes, dim=0)
+            all_detection_classes = torch.cat(all_detection_classes, dim=0)
+            all_detection_probs = torch.cat(all_detection_probs, dim=0)
+            all_detection_batch_indices = torch.cat(all_detection_batch_indices, dim=0)
+            return all_detection_bboxes, all_detection_classes, all_detection_probs, all_detection_batch_indices

packages.txt

@@ -0,0 +1 @@
+ffmpeg

requirements.txt

@@ -0,0 +1,11 @@
+tqdm
+numpy
+torch
+torchvision
+Pillow
+opencv-python-headless==4.8.0.76
+streamlit_webrtc==0.47.0
+pyOpenSSL==23.1.0
+pydub==0.25.1
+twilio~=8.5.0
+matplotlib

test_streamlit.py

@@ -0,0 +1,170 @@
+"""Object detection demo with MobileNet SSD.
+This model and code are based on
+https://github.com/robmarkcole/object-detection-app
+"""
+
+import logging
+import queue
+from pathlib import Path
+from typing import List, NamedTuple
+
+import av
+import cv2
+import numpy as np
+import streamlit as st
+from streamlit_webrtc import WebRtcMode, webrtc_streamer
+
+from sample_utils.download import download_file
+from sample_utils.turn import get_ice_servers
+
+HERE = Path(__file__).parent
+ROOT = HERE
+
+logger = logging.getLogger(__name__)
+
+
+MODEL_URL = "https://github.com/robmarkcole/object-detection-app/raw/master/model/MobileNetSSD_deploy.caffemodel"  # noqa: E501
+MODEL_LOCAL_PATH = ROOT / "./models/MobileNetSSD_deploy.caffemodel"
+PROTOTXT_URL = "https://github.com/robmarkcole/object-detection-app/raw/master/model/MobileNetSSD_deploy.prototxt.txt"  # noqa: E501
+PROTOTXT_LOCAL_PATH = ROOT / "./models/MobileNetSSD_deploy.prototxt.txt"
+
+CLASSES = [
+    "background",
+    "aeroplane",
+    "bicycle",
+    "bird",
+    "boat",
+    "bottle",
+    "bus",
+    "car",
+    "cat",
+    "chair",
+    "cow",
+    "diningtable",
+    "dog",
+    "horse",
+    "motorbike",
+    "person",
+    "pottedplant",
+    "sheep",
+    "sofa",
+    "train",
+    "tvmonitor",
+]
+
+
+class Detection(NamedTuple):
+    class_id: int
+    label: str
+    score: float
+    box: np.ndarray
+
+
+@st.cache_resource  # type: ignore
+def generate_label_colors():
+    return np.random.uniform(0, 255, size=(len(CLASSES), 3))
+
+
+COLORS = generate_label_colors()
+
+download_file(MODEL_URL, MODEL_LOCAL_PATH, expected_size=23147564)
+download_file(PROTOTXT_URL, PROTOTXT_LOCAL_PATH, expected_size=29353)
+
+
+# Session-specific caching
+cache_key = "object_detection_dnn"
+if cache_key in st.session_state:
+    net = st.session_state[cache_key]
+    net.setPreferableBackend(cv2.dnn.DNN_BACKEND_CUDA)
+    net.setPreferableTarget(cv2.dnn.DNN_TARGET_CUDA)
+else:
+    net = cv2.dnn.readNetFromCaffe(str(PROTOTXT_LOCAL_PATH), str(MODEL_LOCAL_PATH))
+    net.setPreferableBackend(cv2.dnn.DNN_BACKEND_CUDA)
+    net.setPreferableTarget(cv2.dnn.DNN_TARGET_CUDA)
+    st.session_state[cache_key] = net
+
+score_threshold = st.slider("Score threshold", 0.0, 1.0, 0.5, 0.05)
+
+# NOTE: The callback will be called in another thread,
+#       so use a queue here for thread-safety to pass the data
+#       from inside to outside the callback.
+# TODO: A general-purpose shared state object may be more useful.
+result_queue: "queue.Queue[List[Detection]]" = queue.Queue()
+
+
+def video_frame_callback(frame: av.VideoFrame) -> av.VideoFrame:
+    image = frame.to_ndarray(format="bgr24")
+
+    # Run inference
+    blob = cv2.dnn.blobFromImage(
+        cv2.resize(image, (300, 300)), 0.007843, (300, 300), 127.5
+    )
+    net.setInput(blob)
+    output = net.forward()
+
+    h, w = image.shape[:2]
+
+    # Convert the output array into a structured form.
+    output = output.squeeze()  # (1, 1, N, 7) -> (N, 7)
+    output = output[output[:, 2] >= score_threshold]
+    detections = [
+        Detection(
+            class_id=int(detection[1]),
+            label=CLASSES[int(detection[1])],
+            score=float(detection[2]),
+            box=(detection[3:7] * np.array([w, h, w, h])),
+        )
+        for detection in output
+    ]
+
+    # Render bounding boxes and captions
+    for detection in detections:
+        caption = f"{detection.label}: {round(detection.score * 100, 2)}%"
+        color = COLORS[detection.class_id]
+        xmin, ymin, xmax, ymax = detection.box.astype("int")
+
+        cv2.rectangle(image, (xmin, ymin), (xmax, ymax), color, 2)
+        cv2.putText(
+            image,
+            caption,
+            (xmin, ymin - 15 if ymin - 15 > 15 else ymin + 15),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.5,
+            color,
+            2,
+        )
+
+    result_queue.put(detections)
+
+    return av.VideoFrame.from_ndarray(image, format="bgr24")
+
+
+webrtc_ctx = webrtc_streamer(
+    key="object-detection",
+    mode=WebRtcMode.SENDRECV,
+    rtc_configuration={
+        "iceServers": get_ice_servers(),
+        "iceTransportPolicy": "relay",
+    },
+    video_frame_callback=video_frame_callback,
+    media_stream_constraints={"video": True, "audio": False},
+    async_processing=True,
+)
+
+if st.checkbox("Show the detected labels", value=True):
+    if webrtc_ctx.state.playing:
+        labels_placeholder = st.empty()
+        # NOTE: The video transformation with object detection and
+        # this loop displaying the result labels are running
+        # in different threads asynchronously.
+        # Then the rendered video frames and the labels displayed here
+        # are not strictly synchronized.
+        while True:
+            result = result_queue.get()
+            labels_placeholder.table(result)
+
+st.markdown(
+    "This demo uses a model and code from "
+    "https://github.com/robmarkcole/object-detection-app. "
+    "Many thanks to the project."
+)

voc_eval.py

@@ -0,0 +1,198 @@
+# --------------------------------------------------------
+# Fast/er R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Bharath Hariharan
+# --------------------------------------------------------
+
+import xml.etree.ElementTree as ET
+import os
+import _pickle as cPickle
+import numpy as np
+
+def parse_rec(filename):
+    """ Parse a PASCAL VOC xml file """
+    tree = ET.parse(filename)
+    objects = []
+    for obj in tree.findall('object'):
+        obj_struct = {}
+        obj_struct['name'] = obj.find('name').text
+        obj_struct['pose'] = obj.find('pose').text
+        obj_struct['truncated'] = int(obj.find('truncated').text)
+        obj_struct['difficult'] = int(obj.find('difficult').text)
+        bbox = obj.find('bndbox')
+        obj_struct['bbox'] = [int(bbox.find('xmin').text),
+                              int(bbox.find('ymin').text),
+                              int(bbox.find('xmax').text),
+                              int(bbox.find('ymax').text)]
+        objects.append(obj_struct)
+
+    return objects
+
+def voc_ap(rec, prec, use_07_metric=False):
+    """ ap = voc_ap(rec, prec, [use_07_metric])
+    Compute VOC AP given precision and recall.
+    If use_07_metric is true, uses the
+    VOC 07 11 point method (default:False).
+    """
+    if use_07_metric:
+        # 11 point metric
+        ap = 0.
+        for t in np.arange(0., 1.1, 0.1):
+            if np.sum(rec >= t) == 0:
+                p = 0
+            else:
+                p = np.max(prec[rec >= t])
+            ap = ap + p / 11.
+    else:
+        # correct AP calculation
+        # first append sentinel values at the end
+        mrec = np.concatenate(([0.], rec, [1.]))
+        mpre = np.concatenate(([0.], prec, [0.]))
+
+        # compute the precision envelope
+        for i in range(mpre.size - 1, 0, -1):
+            mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i])
+
+        # to calculate area under PR curve, look for points
+        # where X axis (recall) changes value
+        i = np.where(mrec[1:] != mrec[:-1])[0]
+
+        # and sum (\Delta recall) * prec
+        ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])
+    return ap
+
+def voc_eval(detpath,
+             annopath,
+             imagesetfile,
+             classname,
+             cachedir,
+             ovthresh=0.5,
+             use_07_metric=False):
+    """rec, prec, ap = voc_eval(detpath,
+                                annopath,
+                                imagesetfile,
+                                classname,
+                                [ovthresh],
+                                [use_07_metric])
+    Top level function that does the PASCAL VOC evaluation.
+    detpath: Path to detections
+        detpath.format(classname) should produce the detection results file.
+    annopath: Path to annotations
+        annopath.format(imagename) should be the xml annotations file.
+    imagesetfile: Text file containing the list of images, one image per line.
+    classname: Category name (duh)
+    cachedir: Directory for caching the annotations
+    [ovthresh]: Overlap threshold (default = 0.5)
+    [use_07_metric]: Whether to use VOC07's 11 point AP computation
+        (default False)
+    """
+    # assumes detections are in detpath.format(classname)
+    # assumes annotations are in annopath.format(imagename)
+    # assumes imagesetfile is a text file with each line an image name
+    # cachedir caches the annotations in a pickle file
+
+    # first load gt
+    if not os.path.isdir(cachedir):
+        os.mkdir(cachedir)
+    cachefile = os.path.join(cachedir, 'annots.pkl')
+    # read list of images
+    with open(imagesetfile, 'r') as f:
+        lines = f.readlines()
+    imagenames = [x.strip() for x in lines]
+
+    if not os.path.isfile(cachefile):
+        # load annots
+        recs = {}
+        for i, imagename in enumerate(imagenames):
+            recs[imagename] = parse_rec(annopath.format(imagename))
+            if i % 100 == 0:
+                print('Reading annotation for {:d}/{:d}'.format(
+                    i + 1, len(imagenames)))
+        # save
+        print('Saving cached annotations to {:s}'.format(cachefile))
+        with open(cachefile, 'wb') as f:
+            cPickle.dump(recs, f)
+    else:
+        # load
+        with open(cachefile, 'rb') as f:
+            recs = cPickle.load(f)
+
+    # extract gt objects for this class
+    class_recs = {}
+    npos = 0
+    for imagename in imagenames:
+        R = [obj for obj in recs[imagename] if obj['name'] == classname]
+        bbox = np.array([x['bbox'] for x in R])
+        difficult = np.array([x['difficult'] for x in R]).astype(np.bool_)
+        det = [False] * len(R)
+        npos = npos + sum(~difficult)
+        class_recs[imagename] = {'bbox': bbox,
+                                 'difficult': difficult,
+                                 'det': det}
+
+    # read dets
+    detfile = detpath.format(classname)
+    with open(detfile, 'r') as f:
+        lines = f.readlines()
+
+    splitlines = [x.strip().split(' ') for x in lines]
+    image_ids = [x[0] for x in splitlines]
+    confidence = np.array([float(x[1]) for x in splitlines])
+    BB = np.array([[float(z) for z in x[2:]] for x in splitlines])
+
+    # sort by confidence
+    sorted_ind = np.argsort(-confidence)
+    sorted_scores = np.sort(-confidence)
+    BB = BB[sorted_ind, :]
+    image_ids = [image_ids[x] for x in sorted_ind]
+
+    # go down dets and mark TPs and FPs
+    nd = len(image_ids)
+    tp = np.zeros(nd)
+    fp = np.zeros(nd)
+    for d in range(nd):
+        R = class_recs[image_ids[d]]
+        bb = BB[d, :].astype(float)
+        ovmax = -np.inf
+        BBGT = R['bbox'].astype(float)
+
+        if BBGT.size > 0:
+            # compute overlaps
+            # intersection
+            ixmin = np.maximum(BBGT[:, 0], bb[0])
+            iymin = np.maximum(BBGT[:, 1], bb[1])
+            ixmax = np.minimum(BBGT[:, 2], bb[2])
+            iymax = np.minimum(BBGT[:, 3], bb[3])
+            iw = np.maximum(ixmax - ixmin + 1., 0.)
+            ih = np.maximum(iymax - iymin + 1., 0.)
+            inters = iw * ih
+
+            # union
+            uni = ((bb[2] - bb[0] + 1.) * (bb[3] - bb[1] + 1.) +
+                   (BBGT[:, 2] - BBGT[:, 0] + 1.) *
+                   (BBGT[:, 3] - BBGT[:, 1] + 1.) - inters)
+
+            overlaps = inters / uni
+            ovmax = np.max(overlaps)
+            jmax = np.argmax(overlaps)
+
+        if ovmax > ovthresh:
+            if not R['difficult'][jmax]:
+                if not R['det'][jmax]:
+                    tp[d] = 1.
+                    R['det'][jmax] = 1
+                else:
+                    fp[d] = 1.
+        else:
+            fp[d] = 1.
+
+    # compute precision recall
+    fp = np.cumsum(fp)
+    tp = np.cumsum(tp)
+    rec = tp / float(npos)
+    # avoid divide by zero in case the first detection matches a difficult
+    # ground truth
+    prec = tp / np.maximum(tp + fp, np.finfo(np.float64).eps)
+    ap = voc_ap(rec, prec, use_07_metric)
+
+    return rec, prec, ap