added lfs

app.py

@@ -0,0 +1,33 @@
+import gradio as gr
+import numpy as np
+import uuid
+from PIL import Image
+import os
+
+
+def inference(img):
+    name_tag = str(uuid.uuid4())
+    script_command = f"python yolov9/detect_dual.py --source {img} --img 640 --device cpu --weights yolov9/runs/train/exp2/weights/best.pt --name {name_tag}"
+    os.system(script_command)
+    output_file = f"yolov9/runs/detect/{name_tag}/{img.split('/')[-1]}"
+    return output_file
+
+title = "Sketch GUI Element Detection"
+
+description = "This is a demo for detecting GUI elements in a sketch image. Upload a sketch image and the model will detect the GUI elements in the image."
+
+img_input = gr.Image(type="filepath", label="Upload a sketch image", width=300, height=300)
+prediction_output = gr.Image(label="Output Image", width=640, height=640)
+
+example_lst = [
+    ["test_images/Shipping-1.png"],
+]
+
+demo = gr.Interface(fn=inference, 
+                    inputs=img_input, 
+                    outputs=prediction_output, 
+                    title=title, 
+                    description=description,
+                    examples=example_lst)
+
+demo.launch(debug=True)

yolov9/LICENSE.md

@@ -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>.

yolov9/README.md

@@ -0,0 +1,335 @@
+# YOLOv9
+
+Implementation of paper - [YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information](https://arxiv.org/abs/2402.13616)
+
+[![arxiv.org](http://img.shields.io/badge/cs.CV-arXiv%3A2402.13616-B31B1B.svg)](https://arxiv.org/abs/2402.13616)
+[![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/kadirnar/Yolov9)
+[![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/merve/yolov9)
+[![Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/roboflow-ai/notebooks/blob/main/notebooks/train-yolov9-object-detection-on-custom-dataset.ipynb)
+[![OpenCV](https://img.shields.io/badge/OpenCV-BlogPost-black?logo=opencv&labelColor=blue&color=black)](https://learnopencv.com/yolov9-advancing-the-yolo-legacy/)
+
+<div align="center">
+    <a href="./">
+        <img src="./figure/performance.png" width="79%"/>
+    </a>
+</div>
+
+
+## Performance 
+
+MS COCO
+
+| Model | Test Size | AP<sup>val</sup> | AP<sub>50</sub><sup>val</sup> | AP<sub>75</sub><sup>val</sup> | Param. | FLOPs |
+| :-- | :-: | :-: | :-: | :-: | :-: | :-: |
+| [**YOLOv9-T**](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-t-converted.pt) | 640 | **38.3%** | **53.1%** | **41.3%** | **2.0M** | **7.7G** |
+| [**YOLOv9-S**](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-s-converted.pt) | 640 | **46.8%** | **63.4%** | **50.7%** | **7.1M** | **26.4G** |
+| [**YOLOv9-M**](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-m-converted.pt) | 640 | **51.4%** | **68.1%** | **56.1%** | **20.0M** | **76.3G** |
+| [**YOLOv9-C**](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-c-converted.pt) | 640 | **53.0%** | **70.2%** | **57.8%** | **25.3M** | **102.1G** |
+| [**YOLOv9-E**](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-e-converted.pt) | 640 | **55.6%** | **72.8%** | **60.6%** | **57.3M** | **189.0G** |
+<!-- | [**YOLOv9 (ReLU)**]() | 640 | **51.9%** | **69.1%** | **56.5%** | **25.3M** | **102.1G** | -->
+
+<!-- tiny, small, and medium models will be released after the paper be accepted and published. -->
+
+## Useful Links
+
+<details><summary> <b>Expand</b> </summary>
+
+Custom training: https://github.com/WongKinYiu/yolov9/issues/30#issuecomment-1960955297
+    
+ONNX export: https://github.com/WongKinYiu/yolov9/issues/2#issuecomment-1960519506 https://github.com/WongKinYiu/yolov9/issues/40#issue-2150697688 https://github.com/WongKinYiu/yolov9/issues/130#issue-2162045461
+
+ONNX export for segmentation: https://github.com/WongKinYiu/yolov9/issues/260#issue-2191162150
+
+TensorRT inference: https://github.com/WongKinYiu/yolov9/issues/143#issuecomment-1975049660 https://github.com/WongKinYiu/yolov9/issues/34#issue-2150393690 https://github.com/WongKinYiu/yolov9/issues/79#issue-2153547004 https://github.com/WongKinYiu/yolov9/issues/143#issue-2164002309
+
+QAT TensorRT: https://github.com/WongKinYiu/yolov9/issues/327#issue-2229284136 https://github.com/WongKinYiu/yolov9/issues/253#issue-2189520073
+
+TensorRT inference for segmentation: https://github.com/WongKinYiu/yolov9/issues/446
+
+TFLite: https://github.com/WongKinYiu/yolov9/issues/374#issuecomment-2065751706
+
+OpenVINO: https://github.com/WongKinYiu/yolov9/issues/164#issue-2168540003
+
+C# ONNX inference: https://github.com/WongKinYiu/yolov9/issues/95#issue-2155974619
+
+C# OpenVINO inference: https://github.com/WongKinYiu/yolov9/issues/95#issuecomment-1968131244
+
+OpenCV: https://github.com/WongKinYiu/yolov9/issues/113#issuecomment-1971327672
+
+Hugging Face demo: https://github.com/WongKinYiu/yolov9/issues/45#issuecomment-1961496943
+
+CoLab demo: https://github.com/WongKinYiu/yolov9/pull/18
+
+ONNXSlim export: https://github.com/WongKinYiu/yolov9/pull/37
+
+YOLOv9 ROS: https://github.com/WongKinYiu/yolov9/issues/144#issue-2164210644
+
+YOLOv9 ROS TensorRT: https://github.com/WongKinYiu/yolov9/issues/145#issue-2164218595
+
+YOLOv9 Julia: https://github.com/WongKinYiu/yolov9/issues/141#issuecomment-1973710107
+
+YOLOv9 MLX: https://github.com/WongKinYiu/yolov9/issues/258#issue-2190586540
+
+YOLOv9 StrongSORT with OSNet: https://github.com/WongKinYiu/yolov9/issues/299#issue-2212093340
+
+YOLOv9 ByteTrack: https://github.com/WongKinYiu/yolov9/issues/78#issue-2153512879
+
+YOLOv9 DeepSORT: https://github.com/WongKinYiu/yolov9/issues/98#issue-2156172319
+
+YOLOv9 counting: https://github.com/WongKinYiu/yolov9/issues/84#issue-2153904804
+
+YOLOv9 speed estimation: https://github.com/WongKinYiu/yolov9/issues/456
+
+YOLOv9 face detection: https://github.com/WongKinYiu/yolov9/issues/121#issue-2160218766
+
+YOLOv9 segmentation onnxruntime: https://github.com/WongKinYiu/yolov9/issues/151#issue-2165667350
+
+Comet logging: https://github.com/WongKinYiu/yolov9/pull/110
+
+MLflow logging: https://github.com/WongKinYiu/yolov9/pull/87
+
+AnyLabeling tool: https://github.com/WongKinYiu/yolov9/issues/48#issue-2152139662
+
+AX650N deploy: https://github.com/WongKinYiu/yolov9/issues/96#issue-2156115760
+
+Conda environment: https://github.com/WongKinYiu/yolov9/pull/93
+
+AutoDL docker environment: https://github.com/WongKinYiu/yolov9/issues/112#issue-2158203480
+
+</details>
+
+
+## Installation
+
+Docker environment (recommended)
+<details><summary> <b>Expand</b> </summary>
+
+``` shell
+# create the docker container, you can change the share memory size if you have more.
+nvidia-docker run --name yolov9 -it -v your_coco_path/:/coco/ -v your_code_path/:/yolov9 --shm-size=64g nvcr.io/nvidia/pytorch:21.11-py3
+
+# apt install required packages
+apt update
+apt install -y zip htop screen libgl1-mesa-glx
+
+# pip install required packages
+pip install seaborn thop
+
+# go to code folder
+cd /yolov9
+```
+
+</details>
+
+
+## Evaluation
+
+[`yolov9-s-converted.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-s-converted.pt) [`yolov9-m-converted.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-m-converted.pt) [`yolov9-c-converted.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-c-converted.pt) [`yolov9-e-converted.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-e-converted.pt) 
+[`yolov9-s.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-s.pt) [`yolov9-m.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-m.pt) [`yolov9-c.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-c.pt) [`yolov9-e.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-e.pt) 
+[`gelan-s.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/gelan-s.pt) [`gelan-m.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/gelan-m.pt) [`gelan-c.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/gelan-c.pt) [`gelan-e.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/gelan-e.pt)
+
+``` shell
+# evaluate converted yolov9 models
+python val.py --data data/coco.yaml --img 640 --batch 32 --conf 0.001 --iou 0.7 --device 0 --weights './yolov9-c-converted.pt' --save-json --name yolov9_c_c_640_val
+
+# evaluate yolov9 models
+# python val_dual.py --data data/coco.yaml --img 640 --batch 32 --conf 0.001 --iou 0.7 --device 0 --weights './yolov9-c.pt' --save-json --name yolov9_c_640_val
+
+# evaluate gelan models
+# python val.py --data data/coco.yaml --img 640 --batch 32 --conf 0.001 --iou 0.7 --device 0 --weights './gelan-c.pt' --save-json --name gelan_c_640_val
+```
+
+You will get the results:
+
+```
+ Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.530
+ Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.702
+ Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.578
+ Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.362
+ Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.585
+ Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.693
+ Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.392
+ Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.652
+ Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.702
+ Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.541
+ Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.760
+ Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.844
+```
+
+
+## Training
+
+Data preparation
+
+``` shell
+bash scripts/get_coco.sh
+```
+
+* Download MS COCO dataset images ([train](http://images.cocodataset.org/zips/train2017.zip), [val](http://images.cocodataset.org/zips/val2017.zip), [test](http://images.cocodataset.org/zips/test2017.zip)) and [labels](https://github.com/WongKinYiu/yolov7/releases/download/v0.1/coco2017labels-segments.zip). If you have previously used a different version of YOLO, we strongly recommend that you delete `train2017.cache` and `val2017.cache` files, and redownload [labels](https://github.com/WongKinYiu/yolov7/releases/download/v0.1/coco2017labels-segments.zip) 
+
+Single GPU training
+
+``` shell
+# train yolov9 models
+python train_dual.py --workers 8 --device 0 --batch 16 --data data/coco.yaml --img 640 --cfg models/detect/yolov9-c.yaml --weights '' --name yolov9-c --hyp hyp.scratch-high.yaml --min-items 0 --epochs 500 --close-mosaic 15
+
+# train gelan models
+# python train.py --workers 8 --device 0 --batch 32 --data data/coco.yaml --img 640 --cfg models/detect/gelan-c.yaml --weights '' --name gelan-c --hyp hyp.scratch-high.yaml --min-items 0 --epochs 500 --close-mosaic 15
+```
+
+Multiple GPU training
+
+``` shell
+# train yolov9 models
+python -m torch.distributed.launch --nproc_per_node 8 --master_port 9527 train_dual.py --workers 8 --device 0,1,2,3,4,5,6,7 --sync-bn --batch 128 --data data/coco.yaml --img 640 --cfg models/detect/yolov9-c.yaml --weights '' --name yolov9-c --hyp hyp.scratch-high.yaml --min-items 0 --epochs 500 --close-mosaic 15
+
+# train gelan models
+# python -m torch.distributed.launch --nproc_per_node 4 --master_port 9527 train.py --workers 8 --device 0,1,2,3 --sync-bn --batch 128 --data data/coco.yaml --img 640 --cfg models/detect/gelan-c.yaml --weights '' --name gelan-c --hyp hyp.scratch-high.yaml --min-items 0 --epochs 500 --close-mosaic 15
+```
+
+
+## Re-parameterization
+
+See [reparameterization.ipynb](https://github.com/WongKinYiu/yolov9/blob/main/tools/reparameterization.ipynb).
+
+
+## Inference
+
+<div align="center">
+    <a href="./">
+        <img src="./figure/horses_prediction.jpg" width="49%"/>
+    </a>
+</div>
+
+``` shell
+# inference converted yolov9 models
+python detect.py --source './data/images/horses.jpg' --img 640 --device 0 --weights './yolov9-c-converted.pt' --name yolov9_c_c_640_detect
+
+# inference yolov9 models
+# python detect_dual.py --source './data/images/horses.jpg' --img 640 --device 0 --weights './yolov9-c.pt' --name yolov9_c_640_detect
+
+# inference gelan models
+# python detect.py --source './data/images/horses.jpg' --img 640 --device 0 --weights './gelan-c.pt' --name gelan_c_c_640_detect
+```
+
+
+## Citation
+
+```
+@article{wang2024yolov9,
+  title={{YOLOv9}: Learning What You Want to Learn Using Programmable Gradient Information},
+  author={Wang, Chien-Yao  and Liao, Hong-Yuan Mark},
+  booktitle={arXiv preprint arXiv:2402.13616},
+  year={2024}
+}
+```
+
+```
+@article{chang2023yolor,
+  title={{YOLOR}-Based Multi-Task Learning},
+  author={Chang, Hung-Shuo and Wang, Chien-Yao and Wang, Richard Robert and Chou, Gene and Liao, Hong-Yuan Mark},
+  journal={arXiv preprint arXiv:2309.16921},
+  year={2023}
+}
+```
+
+
+## Teaser
+
+Parts of code of [YOLOR-Based Multi-Task Learning](https://arxiv.org/abs/2309.16921) are released in the repository.
+
+<div align="center">
+    <a href="./">
+        <img src="./figure/multitask.png" width="99%"/>
+    </a>
+</div>
+
+#### Object Detection
+
+[`gelan-c-det.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/gelan-c-det.pt)
+
+`object detection`
+
+``` shell
+# coco/labels/{split}/*.txt
+# bbox or polygon (1 instance 1 line)
+python train.py --workers 8 --device 0 --batch 32 --data data/coco.yaml --img 640 --cfg models/detect/gelan-c.yaml --weights '' --name gelan-c-det --hyp hyp.scratch-high.yaml --min-items 0 --epochs 300 --close-mosaic 10
+```
+
+| Model | Test Size | Param. | FLOPs | AP<sup>box</sup> |
+| :-- | :-: | :-: | :-: | :-: |
+| [**GELAN-C-DET**](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/gelan-c-det.pt) | 640 | 25.3M | 102.1G |**52.3%** |
+| [**YOLOv9-C-DET**]() | 640 | 25.3M | 102.1G | **53.0%** |
+
+#### Instance Segmentation
+
+[`gelan-c-seg.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/gelan-c-seg.pt)
+
+`object detection` `instance segmentation`
+
+``` shell
+# coco/labels/{split}/*.txt
+# polygon (1 instance 1 line)
+python segment/train.py --workers 8 --device 0 --batch 32  --data coco.yaml --img 640 --cfg models/segment/gelan-c-seg.yaml --weights '' --name gelan-c-seg --hyp hyp.scratch-high.yaml --no-overlap --epochs 300 --close-mosaic 10
+```
+
+| Model | Test Size | Param. | FLOPs | AP<sup>box</sup> | AP<sup>mask</sup>  |
+| :-- | :-: | :-: | :-: | :-: | :-: |
+| [**GELAN-C-SEG**](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/gelan-c-seg.pt) | 640 | 27.4M | 144.6G | **52.3%** | **42.4%** |
+| [**YOLOv9-C-SEG**]() | 640 | 27.4M | 145.5G | **53.3%** | **43.5%** |
+
+#### Panoptic Segmentation
+
+[`gelan-c-pan.pt`](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/gelan-c-pan.pt)
+
+`object detection` `instance segmentation` `semantic segmentation` `stuff segmentation` `panoptic segmentation`
+
+``` shell
+# coco/labels/{split}/*.txt
+# polygon (1 instance 1 line)
+# coco/stuff/{split}/*.txt
+# polygon (1 semantic 1 line)
+python panoptic/train.py --workers 8 --device 0 --batch 32  --data coco.yaml --img 640 --cfg models/panoptic/gelan-c-pan.yaml --weights '' --name gelan-c-pan --hyp hyp.scratch-high.yaml --no-overlap --epochs 300 --close-mosaic 10
+```
+
+| Model | Test Size | Param. | FLOPs | AP<sup>box</sup> | AP<sup>mask</sup>  | mIoU<sub>164k/10k</sub><sup>semantic</sup> | mIoU<sup>stuff</sup> | PQ<sup>panoptic</sup> |
+| :-- | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: |
+| [**GELAN-C-PAN**](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/gelan-c-pan.pt) | 640 | 27.6M | 146.7G | **52.6%** | **42.5%** | **39.0%/48.3%** | **52.7%** | **39.4%** |
+| [**YOLOv9-C-PAN**]() | 640 | 28.8M | 187.0G | **52.7%** | **43.0%** | **39.8%/-** | **52.2%** | **40.5%** |
+
+#### Image Captioning (not yet released)
+
+<!--[`gelan-c-cap.pt`]()-->
+
+`object detection` `instance segmentation` `semantic segmentation` `stuff segmentation` `panoptic segmentation` `image captioning`
+
+``` shell
+# coco/labels/{split}/*.txt
+# polygon (1 instance 1 line)
+# coco/stuff/{split}/*.txt
+# polygon (1 semantic 1 line)
+# coco/annotations/*.json
+# json (1 split 1 file)
+python caption/train.py --workers 8 --device 0 --batch 32  --data coco.yaml --img 640 --cfg models/caption/gelan-c-cap.yaml --weights '' --name gelan-c-cap --hyp hyp.scratch-high.yaml --no-overlap --epochs 300 --close-mosaic 10
+```
+
+| Model | Test Size | Param. | FLOPs |  AP<sup>box</sup> | AP<sup>mask</sup>  | mIoU<sub>164k/10k</sub><sup>semantic</sup>  | mIoU<sup>stuff</sup> | PQ<sup>panoptic</sup> | BLEU@4<sup>caption</sup> | CIDEr<sup>caption</sup> |
+| :-- | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: |
+| [**GELAN-C-CAP**]() | 640 | 47.5M | - | **51.9%** | **42.6%** | **42.5%/-** | **56.5%** | **41.7%** | **38.8** | **122.3** |
+| [**YOLOv9-C-CAP**]() | 640 | 47.5M | - | **52.1%** | **42.6%** | **43.0%/-** | **56.4%** | **42.1%** | **39.1** | **122.0** |
+<!--| [**YOLOR-MT**]() | 640 | 79.3M | - | **51.0%** | **41.7%** | **-/49.6%** | **55.9%** | **40.5%** | **35.7** | **112.7** |-->
+
+
+## Acknowledgements
+
+<details><summary> <b>Expand</b> </summary>
+
+* [https://github.com/AlexeyAB/darknet](https://github.com/AlexeyAB/darknet)
+* [https://github.com/WongKinYiu/yolor](https://github.com/WongKinYiu/yolor)
+* [https://github.com/WongKinYiu/yolov7](https://github.com/WongKinYiu/yolov7)
+* [https://github.com/VDIGPKU/DynamicDet](https://github.com/VDIGPKU/DynamicDet)
+* [https://github.com/DingXiaoH/RepVGG](https://github.com/DingXiaoH/RepVGG)
+* [https://github.com/ultralytics/yolov5](https://github.com/ultralytics/yolov5)
+* [https://github.com/meituan/YOLOv6](https://github.com/meituan/YOLOv6)
+
+</details>

yolov9/benchmarks.py

@@ -0,0 +1,142 @@
+import argparse
+import platform
+import sys
+import time
+from pathlib import Path
+
+import pandas as pd
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+# ROOT = ROOT.relative_to(Path.cwd())  # relative
+
+import export
+from models.experimental import attempt_load
+from models.yolo import SegmentationModel
+from segment.val import run as val_seg
+from utils import notebook_init
+from utils.general import LOGGER, check_yaml, file_size, print_args
+from utils.torch_utils import select_device
+from val import run as val_det
+
+
+def run(
+        weights=ROOT / 'yolo.pt',  # weights path
+        imgsz=640,  # inference size (pixels)
+        batch_size=1,  # batch size
+        data=ROOT / 'data/coco.yaml',  # dataset.yaml path
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        half=False,  # use FP16 half-precision inference
+        test=False,  # test exports only
+        pt_only=False,  # test PyTorch only
+        hard_fail=False,  # throw error on benchmark failure
+):
+    y, t = [], time.time()
+    device = select_device(device)
+    model_type = type(attempt_load(weights, fuse=False))  # DetectionModel, SegmentationModel, etc.
+    for i, (name, f, suffix, cpu, gpu) in export.export_formats().iterrows():  # index, (name, file, suffix, CPU, GPU)
+        try:
+            assert i not in (9, 10), 'inference not supported'  # Edge TPU and TF.js are unsupported
+            assert i != 5 or platform.system() == 'Darwin', 'inference only supported on macOS>=10.13'  # CoreML
+            if 'cpu' in device.type:
+                assert cpu, 'inference not supported on CPU'
+            if 'cuda' in device.type:
+                assert gpu, 'inference not supported on GPU'
+
+            # Export
+            if f == '-':
+                w = weights  # PyTorch format
+            else:
+                w = export.run(weights=weights, imgsz=[imgsz], include=[f], device=device, half=half)[-1]  # all others
+            assert suffix in str(w), 'export failed'
+
+            # Validate
+            if model_type == SegmentationModel:
+                result = val_seg(data, w, batch_size, imgsz, plots=False, device=device, task='speed', half=half)
+                metric = result[0][7]  # (box(p, r, map50, map), mask(p, r, map50, map), *loss(box, obj, cls))
+            else:  # DetectionModel:
+                result = val_det(data, w, batch_size, imgsz, plots=False, device=device, task='speed', half=half)
+                metric = result[0][3]  # (p, r, map50, map, *loss(box, obj, cls))
+            speed = result[2][1]  # times (preprocess, inference, postprocess)
+            y.append([name, round(file_size(w), 1), round(metric, 4), round(speed, 2)])  # MB, mAP, t_inference
+        except Exception as e:
+            if hard_fail:
+                assert type(e) is AssertionError, f'Benchmark --hard-fail for {name}: {e}'
+            LOGGER.warning(f'WARNING ⚠️ Benchmark failure for {name}: {e}')
+            y.append([name, None, None, None])  # mAP, t_inference
+        if pt_only and i == 0:
+            break  # break after PyTorch
+
+    # Print results
+    LOGGER.info('\n')
+    parse_opt()
+    notebook_init()  # print system info
+    c = ['Format', 'Size (MB)', 'mAP50-95', 'Inference time (ms)'] if map else ['Format', 'Export', '', '']
+    py = pd.DataFrame(y, columns=c)
+    LOGGER.info(f'\nBenchmarks complete ({time.time() - t:.2f}s)')
+    LOGGER.info(str(py if map else py.iloc[:, :2]))
+    if hard_fail and isinstance(hard_fail, str):
+        metrics = py['mAP50-95'].array  # values to compare to floor
+        floor = eval(hard_fail)  # minimum metric floor to pass
+        assert all(x > floor for x in metrics if pd.notna(x)), f'HARD FAIL: mAP50-95 < floor {floor}'
+    return py
+
+
+def test(
+        weights=ROOT / 'yolo.pt',  # weights path
+        imgsz=640,  # inference size (pixels)
+        batch_size=1,  # batch size
+        data=ROOT / 'data/coco128.yaml',  # dataset.yaml path
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        half=False,  # use FP16 half-precision inference
+        test=False,  # test exports only
+        pt_only=False,  # test PyTorch only
+        hard_fail=False,  # throw error on benchmark failure
+):
+    y, t = [], time.time()
+    device = select_device(device)
+    for i, (name, f, suffix, gpu) in export.export_formats().iterrows():  # index, (name, file, suffix, gpu-capable)
+        try:
+            w = weights if f == '-' else \
+                export.run(weights=weights, imgsz=[imgsz], include=[f], device=device, half=half)[-1]  # weights
+            assert suffix in str(w), 'export failed'
+            y.append([name, True])
+        except Exception:
+            y.append([name, False])  # mAP, t_inference
+
+    # Print results
+    LOGGER.info('\n')
+    parse_opt()
+    notebook_init()  # print system info
+    py = pd.DataFrame(y, columns=['Format', 'Export'])
+    LOGGER.info(f'\nExports complete ({time.time() - t:.2f}s)')
+    LOGGER.info(str(py))
+    return py
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', type=str, default=ROOT / 'yolo.pt', help='weights path')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)')
+    parser.add_argument('--batch-size', type=int, default=1, help='batch size')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--test', action='store_true', help='test exports only')
+    parser.add_argument('--pt-only', action='store_true', help='test PyTorch only')
+    parser.add_argument('--hard-fail', nargs='?', const=True, default=False, help='Exception on error or < min metric')
+    opt = parser.parse_args()
+    opt.data = check_yaml(opt.data)  # check YAML
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    test(**vars(opt)) if opt.test else run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

yolov9/classify/predict.py

@@ -0,0 +1,224 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Run YOLOv5 classification inference on images, videos, directories, globs, YouTube, webcam, streams, etc.
+
+Usage - sources:
+    $ python classify/predict.py --weights yolov5s-cls.pt --source 0                               # webcam
+                                                                   img.jpg                         # image
+                                                                   vid.mp4                         # video
+                                                                   screen                          # screenshot
+                                                                   path/                           # directory
+                                                                   'path/*.jpg'                    # glob
+                                                                   'https://youtu.be/Zgi9g1ksQHc'  # YouTube
+                                                                   'rtsp://example.com/media.mp4'  # RTSP, RTMP, HTTP stream
+
+Usage - formats:
+    $ python classify/predict.py --weights yolov5s-cls.pt                 # PyTorch
+                                           yolov5s-cls.torchscript        # TorchScript
+                                           yolov5s-cls.onnx               # ONNX Runtime or OpenCV DNN with --dnn
+                                           yolov5s-cls_openvino_model     # OpenVINO
+                                           yolov5s-cls.engine             # TensorRT
+                                           yolov5s-cls.mlmodel            # CoreML (macOS-only)
+                                           yolov5s-cls_saved_model        # TensorFlow SavedModel
+                                           yolov5s-cls.pb                 # TensorFlow GraphDef
+                                           yolov5s-cls.tflite             # TensorFlow Lite
+                                           yolov5s-cls_edgetpu.tflite     # TensorFlow Edge TPU
+                                           yolov5s-cls_paddle_model       # PaddlePaddle
+"""
+
+import argparse
+import os
+import platform
+import sys
+from pathlib import Path
+
+import torch
+import torch.nn.functional as F
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLOv5 root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.common import DetectMultiBackend
+from utils.augmentations import classify_transforms
+from utils.dataloaders import IMG_FORMATS, VID_FORMATS, LoadImages, LoadScreenshots, LoadStreams
+from utils.general import (LOGGER, Profile, check_file, check_img_size, check_imshow, check_requirements, colorstr, cv2,
+                           increment_path, print_args, strip_optimizer)
+from utils.plots import Annotator
+from utils.torch_utils import select_device, smart_inference_mode
+
+
+@smart_inference_mode()
+def run(
+        weights=ROOT / 'yolov5s-cls.pt',  # model.pt path(s)
+        source=ROOT / 'data/images',  # file/dir/URL/glob/screen/0(webcam)
+        data=ROOT / 'data/coco128.yaml',  # dataset.yaml path
+        imgsz=(224, 224),  # inference size (height, width)
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        view_img=False,  # show results
+        save_txt=False,  # save results to *.txt
+        nosave=False,  # do not save images/videos
+        augment=False,  # augmented inference
+        visualize=False,  # visualize features
+        update=False,  # update all models
+        project=ROOT / 'runs/predict-cls',  # save results to project/name
+        name='exp',  # save results to project/name
+        exist_ok=False,  # existing project/name ok, do not increment
+        half=False,  # use FP16 half-precision inference
+        dnn=False,  # use OpenCV DNN for ONNX inference
+        vid_stride=1,  # video frame-rate stride
+):
+    source = str(source)
+    save_img = not nosave and not source.endswith('.txt')  # save inference images
+    is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS)
+    is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://'))
+    webcam = source.isnumeric() or source.endswith('.txt') or (is_url and not is_file)
+    screenshot = source.lower().startswith('screen')
+    if is_url and is_file:
+        source = check_file(source)  # download
+
+    # Directories
+    save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+    (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+    # Load model
+    device = select_device(device)
+    model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+    stride, names, pt = model.stride, model.names, model.pt
+    imgsz = check_img_size(imgsz, s=stride)  # check image size
+
+    # Dataloader
+    bs = 1  # batch_size
+    if webcam:
+        view_img = check_imshow(warn=True)
+        dataset = LoadStreams(source, img_size=imgsz, transforms=classify_transforms(imgsz[0]), vid_stride=vid_stride)
+        bs = len(dataset)
+    elif screenshot:
+        dataset = LoadScreenshots(source, img_size=imgsz, stride=stride, auto=pt)
+    else:
+        dataset = LoadImages(source, img_size=imgsz, transforms=classify_transforms(imgsz[0]), vid_stride=vid_stride)
+    vid_path, vid_writer = [None] * bs, [None] * bs
+
+    # Run inference
+    model.warmup(imgsz=(1 if pt else bs, 3, *imgsz))  # warmup
+    seen, windows, dt = 0, [], (Profile(), Profile(), Profile())
+    for path, im, im0s, vid_cap, s in dataset:
+        with dt[0]:
+            im = torch.Tensor(im).to(model.device)
+            im = im.half() if model.fp16 else im.float()  # uint8 to fp16/32
+            if len(im.shape) == 3:
+                im = im[None]  # expand for batch dim
+
+        # Inference
+        with dt[1]:
+            results = model(im)
+
+        # Post-process
+        with dt[2]:
+            pred = F.softmax(results, dim=1)  # probabilities
+
+        # Process predictions
+        for i, prob in enumerate(pred):  # per image
+            seen += 1
+            if webcam:  # batch_size >= 1
+                p, im0, frame = path[i], im0s[i].copy(), dataset.count
+                s += f'{i}: '
+            else:
+                p, im0, frame = path, im0s.copy(), getattr(dataset, 'frame', 0)
+
+            p = Path(p)  # to Path
+            save_path = str(save_dir / p.name)  # im.jpg
+            txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}')  # im.txt
+
+            s += '%gx%g ' % im.shape[2:]  # print string
+            annotator = Annotator(im0, example=str(names), pil=True)
+
+            # Print results
+            top5i = prob.argsort(0, descending=True)[:5].tolist()  # top 5 indices
+            s += f"{', '.join(f'{names[j]} {prob[j]:.2f}' for j in top5i)}, "
+
+            # Write results
+            text = '\n'.join(f'{prob[j]:.2f} {names[j]}' for j in top5i)
+            if save_img or view_img:  # Add bbox to image
+                annotator.text((32, 32), text, txt_color=(255, 255, 255))
+            if save_txt:  # Write to file
+                with open(f'{txt_path}.txt', 'a') as f:
+                    f.write(text + '\n')
+
+            # Stream results
+            im0 = annotator.result()
+            if view_img:
+                if platform.system() == 'Linux' and p not in windows:
+                    windows.append(p)
+                    cv2.namedWindow(str(p), cv2.WINDOW_NORMAL | cv2.WINDOW_KEEPRATIO)  # allow window resize (Linux)
+                    cv2.resizeWindow(str(p), im0.shape[1], im0.shape[0])
+                cv2.imshow(str(p), im0)
+                cv2.waitKey(1)  # 1 millisecond
+
+            # Save results (image with detections)
+            if save_img:
+                if dataset.mode == 'image':
+                    cv2.imwrite(save_path, im0)
+                else:  # 'video' or 'stream'
+                    if vid_path[i] != save_path:  # new video
+                        vid_path[i] = save_path
+                        if isinstance(vid_writer[i], cv2.VideoWriter):
+                            vid_writer[i].release()  # release previous video writer
+                        if vid_cap:  # video
+                            fps = vid_cap.get(cv2.CAP_PROP_FPS)
+                            w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+                            h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                        else:  # stream
+                            fps, w, h = 30, im0.shape[1], im0.shape[0]
+                        save_path = str(Path(save_path).with_suffix('.mp4'))  # force *.mp4 suffix on results videos
+                        vid_writer[i] = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
+                    vid_writer[i].write(im0)
+
+        # Print time (inference-only)
+        LOGGER.info(f"{s}{dt[1].dt * 1E3:.1f}ms")
+
+    # Print results
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {(1, 3, *imgsz)}' % t)
+    if save_txt or save_img:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    if update:
+        strip_optimizer(weights[0])  # update model (to fix SourceChangeWarning)
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolov5s-cls.pt', help='model path(s)')
+    parser.add_argument('--source', type=str, default=ROOT / 'data/images', help='file/dir/URL/glob/screen/0(webcam)')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='(optional) dataset.yaml path')
+    parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[224], help='inference size h,w')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--view-img', action='store_true', help='show results')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--nosave', action='store_true', help='do not save images/videos')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--visualize', action='store_true', help='visualize features')
+    parser.add_argument('--update', action='store_true', help='update all models')
+    parser.add_argument('--project', default=ROOT / 'runs/predict-cls', help='save results to project/name')
+    parser.add_argument('--name', default='exp', help='save results to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    parser.add_argument('--vid-stride', type=int, default=1, help='video frame-rate stride')
+    opt = parser.parse_args()
+    opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1  # expand
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    check_requirements(exclude=('tensorboard', 'thop'))
+    run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

yolov9/classify/train.py

@@ -0,0 +1,333 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Train a YOLOv5 classifier model on a classification dataset
+
+Usage - Single-GPU training:
+    $ python classify/train.py --model yolov5s-cls.pt --data imagenette160 --epochs 5 --img 224
+
+Usage - Multi-GPU DDP training:
+    $ python -m torch.distributed.run --nproc_per_node 4 --master_port 1 classify/train.py --model yolov5s-cls.pt --data imagenet --epochs 5 --img 224 --device 0,1,2,3
+
+Datasets:           --data mnist, fashion-mnist, cifar10, cifar100, imagenette, imagewoof, imagenet, or 'path/to/data'
+YOLOv5-cls models:  --model yolov5n-cls.pt, yolov5s-cls.pt, yolov5m-cls.pt, yolov5l-cls.pt, yolov5x-cls.pt
+Torchvision models: --model resnet50, efficientnet_b0, etc. See https://pytorch.org/vision/stable/models.html
+"""
+
+import argparse
+import os
+import subprocess
+import sys
+import time
+from copy import deepcopy
+from datetime import datetime
+from pathlib import Path
+
+import torch
+import torch.distributed as dist
+import torch.hub as hub
+import torch.optim.lr_scheduler as lr_scheduler
+import torchvision
+from torch.cuda import amp
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLOv5 root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from classify import val as validate
+from models.experimental import attempt_load
+from models.yolo import ClassificationModel, DetectionModel
+from utils.dataloaders import create_classification_dataloader
+from utils.general import (DATASETS_DIR, LOGGER, TQDM_BAR_FORMAT, WorkingDirectory, check_git_info, check_git_status,
+                           check_requirements, colorstr, download, increment_path, init_seeds, print_args, yaml_save)
+from utils.loggers import GenericLogger
+from utils.plots import imshow_cls
+from utils.torch_utils import (ModelEMA, model_info, reshape_classifier_output, select_device, smart_DDP,
+                               smart_optimizer, smartCrossEntropyLoss, torch_distributed_zero_first)
+
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+GIT_INFO = check_git_info()
+
+
+def train(opt, device):
+    init_seeds(opt.seed + 1 + RANK, deterministic=True)
+    save_dir, data, bs, epochs, nw, imgsz, pretrained = \
+        opt.save_dir, Path(opt.data), opt.batch_size, opt.epochs, min(os.cpu_count() - 1, opt.workers), \
+        opt.imgsz, str(opt.pretrained).lower() == 'true'
+    cuda = device.type != 'cpu'
+
+    # Directories
+    wdir = save_dir / 'weights'
+    wdir.mkdir(parents=True, exist_ok=True)  # make dir
+    last, best = wdir / 'last.pt', wdir / 'best.pt'
+
+    # Save run settings
+    yaml_save(save_dir / 'opt.yaml', vars(opt))
+
+    # Logger
+    logger = GenericLogger(opt=opt, console_logger=LOGGER) if RANK in {-1, 0} else None
+
+    # Download Dataset
+    with torch_distributed_zero_first(LOCAL_RANK), WorkingDirectory(ROOT):
+        data_dir = data if data.is_dir() else (DATASETS_DIR / data)
+        if not data_dir.is_dir():
+            LOGGER.info(f'\nDataset not found ⚠️, missing path {data_dir}, attempting download...')
+            t = time.time()
+            if str(data) == 'imagenet':
+                subprocess.run(f"bash {ROOT / 'data/scripts/get_imagenet.sh'}", shell=True, check=True)
+            else:
+                url = f'https://github.com/ultralytics/yolov5/releases/download/v1.0/{data}.zip'
+                download(url, dir=data_dir.parent)
+            s = f"Dataset download success ✅ ({time.time() - t:.1f}s), saved to {colorstr('bold', data_dir)}\n"
+            LOGGER.info(s)
+
+    # Dataloaders
+    nc = len([x for x in (data_dir / 'train').glob('*') if x.is_dir()])  # number of classes
+    trainloader = create_classification_dataloader(path=data_dir / 'train',
+                                                   imgsz=imgsz,
+                                                   batch_size=bs // WORLD_SIZE,
+                                                   augment=True,
+                                                   cache=opt.cache,
+                                                   rank=LOCAL_RANK,
+                                                   workers=nw)
+
+    test_dir = data_dir / 'test' if (data_dir / 'test').exists() else data_dir / 'val'  # data/test or data/val
+    if RANK in {-1, 0}:
+        testloader = create_classification_dataloader(path=test_dir,
+                                                      imgsz=imgsz,
+                                                      batch_size=bs // WORLD_SIZE * 2,
+                                                      augment=False,
+                                                      cache=opt.cache,
+                                                      rank=-1,
+                                                      workers=nw)
+
+    # Model
+    with torch_distributed_zero_first(LOCAL_RANK), WorkingDirectory(ROOT):
+        if Path(opt.model).is_file() or opt.model.endswith('.pt'):
+            model = attempt_load(opt.model, device='cpu', fuse=False)
+        elif opt.model in torchvision.models.__dict__:  # TorchVision models i.e. resnet50, efficientnet_b0
+            model = torchvision.models.__dict__[opt.model](weights='IMAGENET1K_V1' if pretrained else None)
+        else:
+            m = hub.list('ultralytics/yolov5')  # + hub.list('pytorch/vision')  # models
+            raise ModuleNotFoundError(f'--model {opt.model} not found. Available models are: \n' + '\n'.join(m))
+        if isinstance(model, DetectionModel):
+            LOGGER.warning("WARNING ⚠️ pass YOLOv5 classifier model with '-cls' suffix, i.e. '--model yolov5s-cls.pt'")
+            model = ClassificationModel(model=model, nc=nc, cutoff=opt.cutoff or 10)  # convert to classification model
+        reshape_classifier_output(model, nc)  # update class count
+    for m in model.modules():
+        if not pretrained and hasattr(m, 'reset_parameters'):
+            m.reset_parameters()
+        if isinstance(m, torch.nn.Dropout) and opt.dropout is not None:
+            m.p = opt.dropout  # set dropout
+    for p in model.parameters():
+        p.requires_grad = True  # for training
+    model = model.to(device)
+
+    # Info
+    if RANK in {-1, 0}:
+        model.names = trainloader.dataset.classes  # attach class names
+        model.transforms = testloader.dataset.torch_transforms  # attach inference transforms
+        model_info(model)
+        if opt.verbose:
+            LOGGER.info(model)
+        images, labels = next(iter(trainloader))
+        file = imshow_cls(images[:25], labels[:25], names=model.names, f=save_dir / 'train_images.jpg')
+        logger.log_images(file, name='Train Examples')
+        logger.log_graph(model, imgsz)  # log model
+
+    # Optimizer
+    optimizer = smart_optimizer(model, opt.optimizer, opt.lr0, momentum=0.9, decay=opt.decay)
+
+    # Scheduler
+    lrf = 0.01  # final lr (fraction of lr0)
+    # lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - lrf) + lrf  # cosine
+    lf = lambda x: (1 - x / epochs) * (1 - lrf) + lrf  # linear
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
+    # scheduler = lr_scheduler.OneCycleLR(optimizer, max_lr=lr0, total_steps=epochs, pct_start=0.1,
+    #                                    final_div_factor=1 / 25 / lrf)
+
+    # EMA
+    ema = ModelEMA(model) if RANK in {-1, 0} else None
+
+    # DDP mode
+    if cuda and RANK != -1:
+        model = smart_DDP(model)
+
+    # Train
+    t0 = time.time()
+    criterion = smartCrossEntropyLoss(label_smoothing=opt.label_smoothing)  # loss function
+    best_fitness = 0.0
+    scaler = amp.GradScaler(enabled=cuda)
+    val = test_dir.stem  # 'val' or 'test'
+    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} test\n'
+                f'Using {nw * WORLD_SIZE} dataloader workers\n'
+                f"Logging results to {colorstr('bold', save_dir)}\n"
+                f'Starting {opt.model} training on {data} dataset with {nc} classes for {epochs} epochs...\n\n'
+                f"{'Epoch':>10}{'GPU_mem':>10}{'train_loss':>12}{f'{val}_loss':>12}{'top1_acc':>12}{'top5_acc':>12}")
+    for epoch in range(epochs):  # loop over the dataset multiple times
+        tloss, vloss, fitness = 0.0, 0.0, 0.0  # train loss, val loss, fitness
+        model.train()
+        if RANK != -1:
+            trainloader.sampler.set_epoch(epoch)
+        pbar = enumerate(trainloader)
+        if RANK in {-1, 0}:
+            pbar = tqdm(enumerate(trainloader), total=len(trainloader), bar_format=TQDM_BAR_FORMAT)
+        for i, (images, labels) in pbar:  # progress bar
+            images, labels = images.to(device, non_blocking=True), labels.to(device)
+
+            # Forward
+            with amp.autocast(enabled=cuda):  # stability issues when enabled
+                loss = criterion(model(images), labels)
+
+            # Backward
+            scaler.scale(loss).backward()
+
+            # Optimize
+            scaler.unscale_(optimizer)  # unscale gradients
+            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+            if ema:
+                ema.update(model)
+
+            if RANK in {-1, 0}:
+                # Print
+                tloss = (tloss * i + loss.item()) / (i + 1)  # update mean losses
+                mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0)  # (GB)
+                pbar.desc = f"{f'{epoch + 1}/{epochs}':>10}{mem:>10}{tloss:>12.3g}" + ' ' * 36
+
+                # Test
+                if i == len(pbar) - 1:  # last batch
+                    top1, top5, vloss = validate.run(model=ema.ema,
+                                                     dataloader=testloader,
+                                                     criterion=criterion,
+                                                     pbar=pbar)  # test accuracy, loss
+                    fitness = top1  # define fitness as top1 accuracy
+
+        # Scheduler
+        scheduler.step()
+
+        # Log metrics
+        if RANK in {-1, 0}:
+            # Best fitness
+            if fitness > best_fitness:
+                best_fitness = fitness
+
+            # Log
+            metrics = {
+                "train/loss": tloss,
+                f"{val}/loss": vloss,
+                "metrics/accuracy_top1": top1,
+                "metrics/accuracy_top5": top5,
+                "lr/0": optimizer.param_groups[0]['lr']}  # learning rate
+            logger.log_metrics(metrics, epoch)
+
+            # Save model
+            final_epoch = epoch + 1 == epochs
+            if (not opt.nosave) or final_epoch:
+                ckpt = {
+                    'epoch': epoch,
+                    'best_fitness': best_fitness,
+                    'model': deepcopy(ema.ema).half(),  # deepcopy(de_parallel(model)).half(),
+                    'ema': None,  # deepcopy(ema.ema).half(),
+                    'updates': ema.updates,
+                    'optimizer': None,  # optimizer.state_dict(),
+                    'opt': vars(opt),
+                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
+                    'date': datetime.now().isoformat()}
+
+                # Save last, best and delete
+                torch.save(ckpt, last)
+                if best_fitness == fitness:
+                    torch.save(ckpt, best)
+                del ckpt
+
+    # Train complete
+    if RANK in {-1, 0} and final_epoch:
+        LOGGER.info(f'\nTraining complete ({(time.time() - t0) / 3600:.3f} hours)'
+                    f"\nResults saved to {colorstr('bold', save_dir)}"
+                    f"\nPredict:         python classify/predict.py --weights {best} --source im.jpg"
+                    f"\nValidate:        python classify/val.py --weights {best} --data {data_dir}"
+                    f"\nExport:          python export.py --weights {best} --include onnx"
+                    f"\nPyTorch Hub:     model = torch.hub.load('ultralytics/yolov5', 'custom', '{best}')"
+                    f"\nVisualize:       https://netron.app\n")
+
+        # Plot examples
+        images, labels = (x[:25] for x in next(iter(testloader)))  # first 25 images and labels
+        pred = torch.max(ema.ema(images.to(device)), 1)[1]
+        file = imshow_cls(images, labels, pred, model.names, verbose=False, f=save_dir / 'test_images.jpg')
+
+        # Log results
+        meta = {"epochs": epochs, "top1_acc": best_fitness, "date": datetime.now().isoformat()}
+        logger.log_images(file, name='Test Examples (true-predicted)', epoch=epoch)
+        logger.log_model(best, epochs, metadata=meta)
+
+
+def parse_opt(known=False):
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--model', type=str, default='yolov5s-cls.pt', help='initial weights path')
+    parser.add_argument('--data', type=str, default='imagenette160', help='cifar10, cifar100, mnist, imagenet, ...')
+    parser.add_argument('--epochs', type=int, default=10, help='total training epochs')
+    parser.add_argument('--batch-size', type=int, default=64, help='total batch size for all GPUs')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=224, help='train, val image size (pixels)')
+    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
+    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='--cache images in "ram" (default) or "disk"')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--project', default=ROOT / 'runs/train-cls', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--pretrained', nargs='?', const=True, default=True, help='start from i.e. --pretrained False')
+    parser.add_argument('--optimizer', choices=['SGD', 'Adam', 'AdamW', 'RMSProp'], default='Adam', help='optimizer')
+    parser.add_argument('--lr0', type=float, default=0.001, help='initial learning rate')
+    parser.add_argument('--decay', type=float, default=5e-5, help='weight decay')
+    parser.add_argument('--label-smoothing', type=float, default=0.1, help='Label smoothing epsilon')
+    parser.add_argument('--cutoff', type=int, default=None, help='Model layer cutoff index for Classify() head')
+    parser.add_argument('--dropout', type=float, default=None, help='Dropout (fraction)')
+    parser.add_argument('--verbose', action='store_true', help='Verbose mode')
+    parser.add_argument('--seed', type=int, default=0, help='Global training seed')
+    parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
+    return parser.parse_known_args()[0] if known else parser.parse_args()
+
+
+def main(opt):
+    # Checks
+    if RANK in {-1, 0}:
+        print_args(vars(opt))
+        check_git_status()
+        check_requirements()
+
+    # DDP mode
+    device = select_device(opt.device, batch_size=opt.batch_size)
+    if LOCAL_RANK != -1:
+        assert opt.batch_size != -1, 'AutoBatch is coming soon for classification, please pass a valid --batch-size'
+        assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
+        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
+        torch.cuda.set_device(LOCAL_RANK)
+        device = torch.device('cuda', LOCAL_RANK)
+        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
+
+    # Parameters
+    opt.save_dir = increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok)  # increment run
+
+    # Train
+    train(opt, device)
+
+
+def run(**kwargs):
+    # Usage: from yolov5 import classify; classify.train.run(data=mnist, imgsz=320, model='yolov5m')
+    opt = parse_opt(True)
+    for k, v in kwargs.items():
+        setattr(opt, k, v)
+    main(opt)
+    return opt
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

yolov9/classify/val.py

@@ -0,0 +1,170 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Validate a trained YOLOv5 classification model on a classification dataset
+
+Usage:
+    $ bash data/scripts/get_imagenet.sh --val  # download ImageNet val split (6.3G, 50000 images)
+    $ python classify/val.py --weights yolov5m-cls.pt --data ../datasets/imagenet --img 224  # validate ImageNet
+
+Usage - formats:
+    $ python classify/val.py --weights yolov5s-cls.pt                 # PyTorch
+                                       yolov5s-cls.torchscript        # TorchScript
+                                       yolov5s-cls.onnx               # ONNX Runtime or OpenCV DNN with --dnn
+                                       yolov5s-cls_openvino_model     # OpenVINO
+                                       yolov5s-cls.engine             # TensorRT
+                                       yolov5s-cls.mlmodel            # CoreML (macOS-only)
+                                       yolov5s-cls_saved_model        # TensorFlow SavedModel
+                                       yolov5s-cls.pb                 # TensorFlow GraphDef
+                                       yolov5s-cls.tflite             # TensorFlow Lite
+                                       yolov5s-cls_edgetpu.tflite     # TensorFlow Edge TPU
+                                       yolov5s-cls_paddle_model       # PaddlePaddle
+"""
+
+import argparse
+import os
+import sys
+from pathlib import Path
+
+import torch
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLOv5 root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.common import DetectMultiBackend
+from utils.dataloaders import create_classification_dataloader
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, Profile, check_img_size, check_requirements, colorstr,
+                           increment_path, print_args)
+from utils.torch_utils import select_device, smart_inference_mode
+
+
+@smart_inference_mode()
+def run(
+    data=ROOT / '../datasets/mnist',  # dataset dir
+    weights=ROOT / 'yolov5s-cls.pt',  # model.pt path(s)
+    batch_size=128,  # batch size
+    imgsz=224,  # inference size (pixels)
+    device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+    workers=8,  # max dataloader workers (per RANK in DDP mode)
+    verbose=False,  # verbose output
+    project=ROOT / 'runs/val-cls',  # save to project/name
+    name='exp',  # save to project/name
+    exist_ok=False,  # existing project/name ok, do not increment
+    half=False,  # use FP16 half-precision inference
+    dnn=False,  # use OpenCV DNN for ONNX inference
+    model=None,
+    dataloader=None,
+    criterion=None,
+    pbar=None,
+):
+    # Initialize/load model and set device
+    training = model is not None
+    if training:  # called by train.py
+        device, pt, jit, engine = next(model.parameters()).device, True, False, False  # get model device, PyTorch model
+        half &= device.type != 'cpu'  # half precision only supported on CUDA
+        model.half() if half else model.float()
+    else:  # called directly
+        device = select_device(device, batch_size=batch_size)
+
+        # Directories
+        save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+        save_dir.mkdir(parents=True, exist_ok=True)  # make dir
+
+        # Load model
+        model = DetectMultiBackend(weights, device=device, dnn=dnn, fp16=half)
+        stride, pt, jit, engine = model.stride, model.pt, model.jit, model.engine
+        imgsz = check_img_size(imgsz, s=stride)  # check image size
+        half = model.fp16  # FP16 supported on limited backends with CUDA
+        if engine:
+            batch_size = model.batch_size
+        else:
+            device = model.device
+            if not (pt or jit):
+                batch_size = 1  # export.py models default to batch-size 1
+                LOGGER.info(f'Forcing --batch-size 1 square inference (1,3,{imgsz},{imgsz}) for non-PyTorch models')
+
+        # Dataloader
+        data = Path(data)
+        test_dir = data / 'test' if (data / 'test').exists() else data / 'val'  # data/test or data/val
+        dataloader = create_classification_dataloader(path=test_dir,
+                                                      imgsz=imgsz,
+                                                      batch_size=batch_size,
+                                                      augment=False,
+                                                      rank=-1,
+                                                      workers=workers)
+
+    model.eval()
+    pred, targets, loss, dt = [], [], 0, (Profile(), Profile(), Profile())
+    n = len(dataloader)  # number of batches
+    action = 'validating' if dataloader.dataset.root.stem == 'val' else 'testing'
+    desc = f"{pbar.desc[:-36]}{action:>36}" if pbar else f"{action}"
+    bar = tqdm(dataloader, desc, n, not training, bar_format=TQDM_BAR_FORMAT, position=0)
+    with torch.cuda.amp.autocast(enabled=device.type != 'cpu'):
+        for images, labels in bar:
+            with dt[0]:
+                images, labels = images.to(device, non_blocking=True), labels.to(device)
+
+            with dt[1]:
+                y = model(images)
+
+            with dt[2]:
+                pred.append(y.argsort(1, descending=True)[:, :5])
+                targets.append(labels)
+                if criterion:
+                    loss += criterion(y, labels)
+
+    loss /= n
+    pred, targets = torch.cat(pred), torch.cat(targets)
+    correct = (targets[:, None] == pred).float()
+    acc = torch.stack((correct[:, 0], correct.max(1).values), dim=1)  # (top1, top5) accuracy
+    top1, top5 = acc.mean(0).tolist()
+
+    if pbar:
+        pbar.desc = f"{pbar.desc[:-36]}{loss:>12.3g}{top1:>12.3g}{top5:>12.3g}"
+    if verbose:  # all classes
+        LOGGER.info(f"{'Class':>24}{'Images':>12}{'top1_acc':>12}{'top5_acc':>12}")
+        LOGGER.info(f"{'all':>24}{targets.shape[0]:>12}{top1:>12.3g}{top5:>12.3g}")
+        for i, c in model.names.items():
+            aci = acc[targets == i]
+            top1i, top5i = aci.mean(0).tolist()
+            LOGGER.info(f"{c:>24}{aci.shape[0]:>12}{top1i:>12.3g}{top5i:>12.3g}")
+
+        # Print results
+        t = tuple(x.t / len(dataloader.dataset.samples) * 1E3 for x in dt)  # speeds per image
+        shape = (1, 3, imgsz, imgsz)
+        LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms post-process per image at shape {shape}' % t)
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}")
+
+    return top1, top5, loss
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default=ROOT / '../datasets/mnist', help='dataset path')
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolov5s-cls.pt', help='model.pt path(s)')
+    parser.add_argument('--batch-size', type=int, default=128, help='batch size')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=224, help='inference size (pixels)')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--verbose', nargs='?', const=True, default=True, help='verbose output')
+    parser.add_argument('--project', default=ROOT / 'runs/val-cls', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    opt = parser.parse_args()
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    check_requirements(exclude=('tensorboard', 'thop'))
+    run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

yolov9/detect.py

@@ -0,0 +1,231 @@
+import argparse
+import os
+import platform
+import sys
+from pathlib import Path
+
+import torch
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.common import DetectMultiBackend
+from utils.dataloaders import IMG_FORMATS, VID_FORMATS, LoadImages, LoadScreenshots, LoadStreams
+from utils.general import (LOGGER, Profile, check_file, check_img_size, check_imshow, check_requirements, colorstr, cv2,
+                           increment_path, non_max_suppression, print_args, scale_boxes, strip_optimizer, xyxy2xywh)
+from utils.plots import Annotator, colors, save_one_box
+from utils.torch_utils import select_device, smart_inference_mode
+
+
+@smart_inference_mode()
+def run(
+        weights=ROOT / 'yolo.pt',  # model path or triton URL
+        source=ROOT / 'data/images',  # file/dir/URL/glob/screen/0(webcam)
+        data=ROOT / 'data/coco.yaml',  # dataset.yaml path
+        imgsz=(640, 640),  # inference size (height, width)
+        conf_thres=0.25,  # confidence threshold
+        iou_thres=0.45,  # NMS IOU threshold
+        max_det=1000,  # maximum detections per image
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        view_img=False,  # show results
+        save_txt=False,  # save results to *.txt
+        save_conf=False,  # save confidences in --save-txt labels
+        save_crop=False,  # save cropped prediction boxes
+        nosave=False,  # do not save images/videos
+        classes=None,  # filter by class: --class 0, or --class 0 2 3
+        agnostic_nms=False,  # class-agnostic NMS
+        augment=False,  # augmented inference
+        visualize=False,  # visualize features
+        update=False,  # update all models
+        project=ROOT / 'runs/detect',  # save results to project/name
+        name='exp',  # save results to project/name
+        exist_ok=False,  # existing project/name ok, do not increment
+        line_thickness=3,  # bounding box thickness (pixels)
+        hide_labels=False,  # hide labels
+        hide_conf=False,  # hide confidences
+        half=False,  # use FP16 half-precision inference
+        dnn=False,  # use OpenCV DNN for ONNX inference
+        vid_stride=1,  # video frame-rate stride
+):
+    source = str(source)
+    save_img = not nosave and not source.endswith('.txt')  # save inference images
+    is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS)
+    is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://'))
+    webcam = source.isnumeric() or source.endswith('.txt') or (is_url and not is_file)
+    screenshot = source.lower().startswith('screen')
+    if is_url and is_file:
+        source = check_file(source)  # download
+
+    # Directories
+    save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+    (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+    # Load model
+    device = select_device(device)
+    model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+    stride, names, pt = model.stride, model.names, model.pt
+    imgsz = check_img_size(imgsz, s=stride)  # check image size
+
+    # Dataloader
+    bs = 1  # batch_size
+    if webcam:
+        view_img = check_imshow(warn=True)
+        dataset = LoadStreams(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)
+        bs = len(dataset)
+    elif screenshot:
+        dataset = LoadScreenshots(source, img_size=imgsz, stride=stride, auto=pt)
+    else:
+        dataset = LoadImages(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)
+    vid_path, vid_writer = [None] * bs, [None] * bs
+
+    # Run inference
+    model.warmup(imgsz=(1 if pt or model.triton else bs, 3, *imgsz))  # warmup
+    seen, windows, dt = 0, [], (Profile(), Profile(), Profile())
+    for path, im, im0s, vid_cap, s in dataset:
+        with dt[0]:
+            im = torch.from_numpy(im).to(model.device)
+            im = im.half() if model.fp16 else im.float()  # uint8 to fp16/32
+            im /= 255  # 0 - 255 to 0.0 - 1.0
+            if len(im.shape) == 3:
+                im = im[None]  # expand for batch dim
+
+        # Inference
+        with dt[1]:
+            visualize = increment_path(save_dir / Path(path).stem, mkdir=True) if visualize else False
+            pred = model(im, augment=augment, visualize=visualize)
+
+        # NMS
+        with dt[2]:
+            pred = non_max_suppression(pred, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det)
+
+        # Second-stage classifier (optional)
+        # pred = utils.general.apply_classifier(pred, classifier_model, im, im0s)
+
+        # Process predictions
+        for i, det in enumerate(pred):  # per image
+            seen += 1
+            if webcam:  # batch_size >= 1
+                p, im0, frame = path[i], im0s[i].copy(), dataset.count
+                s += f'{i}: '
+            else:
+                p, im0, frame = path, im0s.copy(), getattr(dataset, 'frame', 0)
+
+            p = Path(p)  # to Path
+            save_path = str(save_dir / p.name)  # im.jpg
+            txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}')  # im.txt
+            s += '%gx%g ' % im.shape[2:]  # print string
+            gn = torch.tensor(im0.shape)[[1, 0, 1, 0]]  # normalization gain whwh
+            imc = im0.copy() if save_crop else im0  # for save_crop
+            annotator = Annotator(im0, line_width=line_thickness, example=str(names))
+            if len(det):
+                # Rescale boxes from img_size to im0 size
+                det[:, :4] = scale_boxes(im.shape[2:], det[:, :4], im0.shape).round()
+
+                # Print results
+                for c in det[:, 5].unique():
+                    n = (det[:, 5] == c).sum()  # detections per class
+                    s += f"{n} {names[int(c)]}{'s' * (n > 1)}, "  # add to string
+
+                # Write results
+                for *xyxy, conf, cls in reversed(det):
+                    if save_txt:  # Write to file
+                        xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
+                        line = (cls, *xywh, conf) if save_conf else (cls, *xywh)  # label format
+                        with open(f'{txt_path}.txt', 'a') as f:
+                            f.write(('%g ' * len(line)).rstrip() % line + '\n')
+
+                    if save_img or save_crop or view_img:  # Add bbox to image
+                        c = int(cls)  # integer class
+                        label = None if hide_labels else (names[c] if hide_conf else f'{names[c]} {conf:.2f}')
+                        annotator.box_label(xyxy, label, color=colors(c, True))
+                    if save_crop:
+                        save_one_box(xyxy, imc, file=save_dir / 'crops' / names[c] / f'{p.stem}.jpg', BGR=True)
+
+            # Stream results
+            im0 = annotator.result()
+            if view_img:
+                if platform.system() == 'Linux' and p not in windows:
+                    windows.append(p)
+                    cv2.namedWindow(str(p), cv2.WINDOW_NORMAL | cv2.WINDOW_KEEPRATIO)  # allow window resize (Linux)
+                    cv2.resizeWindow(str(p), im0.shape[1], im0.shape[0])
+                cv2.imshow(str(p), im0)
+                cv2.waitKey(1)  # 1 millisecond
+
+            # Save results (image with detections)
+            if save_img:
+                if dataset.mode == 'image':
+                    cv2.imwrite(save_path, im0)
+                else:  # 'video' or 'stream'
+                    if vid_path[i] != save_path:  # new video
+                        vid_path[i] = save_path
+                        if isinstance(vid_writer[i], cv2.VideoWriter):
+                            vid_writer[i].release()  # release previous video writer
+                        if vid_cap:  # video
+                            fps = vid_cap.get(cv2.CAP_PROP_FPS)
+                            w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+                            h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                        else:  # stream
+                            fps, w, h = 30, im0.shape[1], im0.shape[0]
+                        save_path = str(Path(save_path).with_suffix('.mp4'))  # force *.mp4 suffix on results videos
+                        vid_writer[i] = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
+                    vid_writer[i].write(im0)
+
+        # Print time (inference-only)
+        LOGGER.info(f"{s}{'' if len(det) else '(no detections), '}{dt[1].dt * 1E3:.1f}ms")
+
+    # Print results
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {(1, 3, *imgsz)}' % t)
+    if save_txt or save_img:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    if update:
+        strip_optimizer(weights[0])  # update model (to fix SourceChangeWarning)
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo.pt', help='model path or triton URL')
+    parser.add_argument('--source', type=str, default=ROOT / 'data/images', help='file/dir/URL/glob/screen/0(webcam)')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='(optional) dataset.yaml path')
+    parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w')
+    parser.add_argument('--conf-thres', type=float, default=0.25, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold')
+    parser.add_argument('--max-det', type=int, default=1000, help='maximum detections per image')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--view-img', action='store_true', help='show results')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
+    parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes')
+    parser.add_argument('--nosave', action='store_true', help='do not save images/videos')
+    parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --classes 0, or --classes 0 2 3')
+    parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--visualize', action='store_true', help='visualize features')
+    parser.add_argument('--update', action='store_true', help='update all models')
+    parser.add_argument('--project', default=ROOT / 'runs/detect', help='save results to project/name')
+    parser.add_argument('--name', default='exp', help='save results to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)')
+    parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels')
+    parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    parser.add_argument('--vid-stride', type=int, default=1, help='video frame-rate stride')
+    opt = parser.parse_args()
+    opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1  # expand
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    # check_requirements(exclude=('tensorboard', 'thop'))
+    run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

yolov9/detect_dual.py

@@ -0,0 +1,232 @@
+import argparse
+import os
+import platform
+import sys
+from pathlib import Path
+
+import torch
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.common import DetectMultiBackend
+from utils.dataloaders import IMG_FORMATS, VID_FORMATS, LoadImages, LoadScreenshots, LoadStreams
+from utils.general import (LOGGER, Profile, check_file, check_img_size, check_imshow, check_requirements, colorstr, cv2,
+                           increment_path, non_max_suppression, print_args, scale_boxes, strip_optimizer, xyxy2xywh)
+from utils.plots import Annotator, colors, save_one_box
+from utils.torch_utils import select_device, smart_inference_mode
+
+
+@smart_inference_mode()
+def run(
+        weights=ROOT / 'yolo.pt',  # model path or triton URL
+        source=ROOT / 'data/images',  # file/dir/URL/glob/screen/0(webcam)
+        data=ROOT / 'data/coco.yaml',  # dataset.yaml path
+        imgsz=(640, 640),  # inference size (height, width)
+        conf_thres=0.25,  # confidence threshold
+        iou_thres=0.45,  # NMS IOU threshold
+        max_det=1000,  # maximum detections per image
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        view_img=False,  # show results
+        save_txt=False,  # save results to *.txt
+        save_conf=False,  # save confidences in --save-txt labels
+        save_crop=False,  # save cropped prediction boxes
+        nosave=False,  # do not save images/videos
+        classes=None,  # filter by class: --class 0, or --class 0 2 3
+        agnostic_nms=False,  # class-agnostic NMS
+        augment=False,  # augmented inference
+        visualize=False,  # visualize features
+        update=False,  # update all models
+        project=ROOT / 'runs/detect',  # save results to project/name
+        name='exp',  # save results to project/name
+        exist_ok=False,  # existing project/name ok, do not increment
+        line_thickness=3,  # bounding box thickness (pixels)
+        hide_labels=False,  # hide labels
+        hide_conf=False,  # hide confidences
+        half=False,  # use FP16 half-precision inference
+        dnn=False,  # use OpenCV DNN for ONNX inference
+        vid_stride=1,  # video frame-rate stride
+):
+    source = str(source)
+    save_img = not nosave and not source.endswith('.txt')  # save inference images
+    is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS)
+    is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://'))
+    webcam = source.isnumeric() or source.endswith('.txt') or (is_url and not is_file)
+    screenshot = source.lower().startswith('screen')
+    if is_url and is_file:
+        source = check_file(source)  # download
+
+    # Directories
+    save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+    (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+    # Load model
+    device = select_device(device)
+    model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+    stride, names, pt = model.stride, model.names, model.pt
+    imgsz = check_img_size(imgsz, s=stride)  # check image size
+
+    # Dataloader
+    bs = 1  # batch_size
+    if webcam:
+        view_img = check_imshow(warn=True)
+        dataset = LoadStreams(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)
+        bs = len(dataset)
+    elif screenshot:
+        dataset = LoadScreenshots(source, img_size=imgsz, stride=stride, auto=pt)
+    else:
+        dataset = LoadImages(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)
+    vid_path, vid_writer = [None] * bs, [None] * bs
+
+    # Run inference
+    model.warmup(imgsz=(1 if pt or model.triton else bs, 3, *imgsz))  # warmup
+    seen, windows, dt = 0, [], (Profile(), Profile(), Profile())
+    for path, im, im0s, vid_cap, s in dataset:
+        with dt[0]:
+            im = torch.from_numpy(im).to(model.device)
+            im = im.half() if model.fp16 else im.float()  # uint8 to fp16/32
+            im /= 255  # 0 - 255 to 0.0 - 1.0
+            if len(im.shape) == 3:
+                im = im[None]  # expand for batch dim
+
+        # Inference
+        with dt[1]:
+            visualize = increment_path(save_dir / Path(path).stem, mkdir=True) if visualize else False
+            pred = model(im, augment=augment, visualize=visualize)
+            pred = pred[0][1]
+
+        # NMS
+        with dt[2]:
+            pred = non_max_suppression(pred, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det)
+
+        # Second-stage classifier (optional)
+        # pred = utils.general.apply_classifier(pred, classifier_model, im, im0s)
+
+        # Process predictions
+        for i, det in enumerate(pred):  # per image
+            seen += 1
+            if webcam:  # batch_size >= 1
+                p, im0, frame = path[i], im0s[i].copy(), dataset.count
+                s += f'{i}: '
+            else:
+                p, im0, frame = path, im0s.copy(), getattr(dataset, 'frame', 0)
+
+            p = Path(p)  # to Path
+            save_path = str(save_dir / p.name)  # im.jpg
+            txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}')  # im.txt
+            s += '%gx%g ' % im.shape[2:]  # print string
+            gn = torch.tensor(im0.shape)[[1, 0, 1, 0]]  # normalization gain whwh
+            imc = im0.copy() if save_crop else im0  # for save_crop
+            annotator = Annotator(im0, line_width=line_thickness, example=str(names))
+            if len(det):
+                # Rescale boxes from img_size to im0 size
+                det[:, :4] = scale_boxes(im.shape[2:], det[:, :4], im0.shape).round()
+
+                # Print results
+                for c in det[:, 5].unique():
+                    n = (det[:, 5] == c).sum()  # detections per class
+                    s += f"{n} {names[int(c)]}{'s' * (n > 1)}, "  # add to string
+
+                # Write results
+                for *xyxy, conf, cls in reversed(det):
+                    if save_txt:  # Write to file
+                        xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
+                        line = (cls, *xywh, conf) if save_conf else (cls, *xywh)  # label format
+                        with open(f'{txt_path}.txt', 'a') as f:
+                            f.write(('%g ' * len(line)).rstrip() % line + '\n')
+
+                    if save_img or save_crop or view_img:  # Add bbox to image
+                        c = int(cls)  # integer class
+                        label = None if hide_labels else (names[c] if hide_conf else f'{names[c]} {conf:.2f}')
+                        annotator.box_label(xyxy, label, color=colors(c, True))
+                    if save_crop:
+                        save_one_box(xyxy, imc, file=save_dir / 'crops' / names[c] / f'{p.stem}.jpg', BGR=True)
+
+            # Stream results
+            im0 = annotator.result()
+            if view_img:
+                if platform.system() == 'Linux' and p not in windows:
+                    windows.append(p)
+                    cv2.namedWindow(str(p), cv2.WINDOW_NORMAL | cv2.WINDOW_KEEPRATIO)  # allow window resize (Linux)
+                    cv2.resizeWindow(str(p), im0.shape[1], im0.shape[0])
+                cv2.imshow(str(p), im0)
+                cv2.waitKey(1)  # 1 millisecond
+
+            # Save results (image with detections)
+            if save_img:
+                if dataset.mode == 'image':
+                    cv2.imwrite(save_path, im0)
+                else:  # 'video' or 'stream'
+                    if vid_path[i] != save_path:  # new video
+                        vid_path[i] = save_path
+                        if isinstance(vid_writer[i], cv2.VideoWriter):
+                            vid_writer[i].release()  # release previous video writer
+                        if vid_cap:  # video
+                            fps = vid_cap.get(cv2.CAP_PROP_FPS)
+                            w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+                            h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                        else:  # stream
+                            fps, w, h = 30, im0.shape[1], im0.shape[0]
+                        save_path = str(Path(save_path).with_suffix('.mp4'))  # force *.mp4 suffix on results videos
+                        vid_writer[i] = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
+                    vid_writer[i].write(im0)
+
+        # Print time (inference-only)
+        LOGGER.info(f"{s}{'' if len(det) else '(no detections), '}{dt[1].dt * 1E3:.1f}ms")
+
+    # Print results
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {(1, 3, *imgsz)}' % t)
+    if save_txt or save_img:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    if update:
+        strip_optimizer(weights[0])  # update model (to fix SourceChangeWarning)
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo.pt', help='model path or triton URL')
+    parser.add_argument('--source', type=str, default=ROOT / 'data/images', help='file/dir/URL/glob/screen/0(webcam)')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='(optional) dataset.yaml path')
+    parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w')
+    parser.add_argument('--conf-thres', type=float, default=0.25, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold')
+    parser.add_argument('--max-det', type=int, default=1000, help='maximum detections per image')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--view-img', action='store_true', help='show results')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
+    parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes')
+    parser.add_argument('--nosave', action='store_true', help='do not save images/videos')
+    parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --classes 0, or --classes 0 2 3')
+    parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--visualize', action='store_true', help='visualize features')
+    parser.add_argument('--update', action='store_true', help='update all models')
+    parser.add_argument('--project', default=ROOT / 'runs/detect', help='save results to project/name')
+    parser.add_argument('--name', default='exp', help='save results to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)')
+    parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels')
+    parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    parser.add_argument('--vid-stride', type=int, default=1, help='video frame-rate stride')
+    opt = parser.parse_args()
+    opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1  # expand
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    # check_requirements(exclude=('tensorboard', 'thop'))
+    run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

yolov9/export.py

@@ -0,0 +1,686 @@
+import argparse
+import contextlib
+import json
+import os
+import platform
+import re
+import subprocess
+import sys
+import time
+import warnings
+from pathlib import Path
+
+import pandas as pd
+import torch
+from torch.utils.mobile_optimizer import optimize_for_mobile
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+if platform.system() != 'Windows':
+    ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.experimental import attempt_load, End2End
+from models.yolo import ClassificationModel, Detect, DDetect, DualDetect, DualDDetect, DetectionModel, SegmentationModel
+from utils.dataloaders import LoadImages
+from utils.general import (LOGGER, Profile, check_dataset, check_img_size, check_requirements, check_version,
+                           check_yaml, colorstr, file_size, get_default_args, print_args, url2file, yaml_save)
+from utils.torch_utils import select_device, smart_inference_mode
+
+MACOS = platform.system() == 'Darwin'  # macOS environment
+
+
+def export_formats():
+    # YOLO export formats
+    x = [
+        ['PyTorch', '-', '.pt', True, True],
+        ['TorchScript', 'torchscript', '.torchscript', True, True],
+        ['ONNX', 'onnx', '.onnx', True, True],
+        ['ONNX END2END', 'onnx_end2end', '_end2end.onnx', True, True],
+        ['OpenVINO', 'openvino', '_openvino_model', True, False],
+        ['TensorRT', 'engine', '.engine', False, True],
+        ['CoreML', 'coreml', '.mlmodel', True, False],
+        ['TensorFlow SavedModel', 'saved_model', '_saved_model', True, True],
+        ['TensorFlow GraphDef', 'pb', '.pb', True, True],
+        ['TensorFlow Lite', 'tflite', '.tflite', True, False],
+        ['TensorFlow Edge TPU', 'edgetpu', '_edgetpu.tflite', False, False],
+        ['TensorFlow.js', 'tfjs', '_web_model', False, False],
+        ['PaddlePaddle', 'paddle', '_paddle_model', True, True],]
+    return pd.DataFrame(x, columns=['Format', 'Argument', 'Suffix', 'CPU', 'GPU'])
+
+
+def try_export(inner_func):
+    # YOLO export decorator, i..e @try_export
+    inner_args = get_default_args(inner_func)
+
+    def outer_func(*args, **kwargs):
+        prefix = inner_args['prefix']
+        try:
+            with Profile() as dt:
+                f, model = inner_func(*args, **kwargs)
+            LOGGER.info(f'{prefix} export success ✅ {dt.t:.1f}s, saved as {f} ({file_size(f):.1f} MB)')
+            return f, model
+        except Exception as e:
+            LOGGER.info(f'{prefix} export failure ❌ {dt.t:.1f}s: {e}')
+            return None, None
+
+    return outer_func
+
+
+@try_export
+def export_torchscript(model, im, file, optimize, prefix=colorstr('TorchScript:')):
+    # YOLO TorchScript model export
+    LOGGER.info(f'\n{prefix} starting export with torch {torch.__version__}...')
+    f = file.with_suffix('.torchscript')
+
+    ts = torch.jit.trace(model, im, strict=False)
+    d = {"shape": im.shape, "stride": int(max(model.stride)), "names": model.names}
+    extra_files = {'config.txt': json.dumps(d)}  # torch._C.ExtraFilesMap()
+    if optimize:  # https://pytorch.org/tutorials/recipes/mobile_interpreter.html
+        optimize_for_mobile(ts)._save_for_lite_interpreter(str(f), _extra_files=extra_files)
+    else:
+        ts.save(str(f), _extra_files=extra_files)
+    return f, None
+
+
+@try_export
+def export_onnx(model, im, file, opset, dynamic, simplify, prefix=colorstr('ONNX:')):
+    # YOLO ONNX export
+    check_requirements('onnx')
+    import onnx
+
+    LOGGER.info(f'\n{prefix} starting export with onnx {onnx.__version__}...')
+    f = file.with_suffix('.onnx')
+
+    output_names = ['output0', 'output1'] if isinstance(model, SegmentationModel) else ['output0']
+    if dynamic:
+        dynamic = {'images': {0: 'batch', 2: 'height', 3: 'width'}}  # shape(1,3,640,640)
+        if isinstance(model, SegmentationModel):
+            dynamic['output0'] = {0: 'batch', 1: 'anchors'}  # shape(1,25200,85)
+            dynamic['output1'] = {0: 'batch', 2: 'mask_height', 3: 'mask_width'}  # shape(1,32,160,160)
+        elif isinstance(model, DetectionModel):
+            dynamic['output0'] = {0: 'batch', 1: 'anchors'}  # shape(1,25200,85)
+
+    torch.onnx.export(
+        model.cpu() if dynamic else model,  # --dynamic only compatible with cpu
+        im.cpu() if dynamic else im,
+        f,
+        verbose=False,
+        opset_version=opset,
+        do_constant_folding=True,
+        input_names=['images'],
+        output_names=output_names,
+        dynamic_axes=dynamic or None)
+
+    # Checks
+    model_onnx = onnx.load(f)  # load onnx model
+    onnx.checker.check_model(model_onnx)  # check onnx model
+
+    # Metadata
+    d = {'stride': int(max(model.stride)), 'names': model.names}
+    for k, v in d.items():
+        meta = model_onnx.metadata_props.add()
+        meta.key, meta.value = k, str(v)
+    onnx.save(model_onnx, f)
+
+    # Simplify
+    if simplify:
+        try:
+            cuda = torch.cuda.is_available()
+            check_requirements(('onnxruntime-gpu' if cuda else 'onnxruntime', 'onnx-simplifier>=0.4.1'))
+            import onnxsim
+
+            LOGGER.info(f'{prefix} simplifying with onnx-simplifier {onnxsim.__version__}...')
+            model_onnx, check = onnxsim.simplify(model_onnx)
+            assert check, 'assert check failed'
+            onnx.save(model_onnx, f)
+        except Exception as e:
+            LOGGER.info(f'{prefix} simplifier failure: {e}')
+    return f, model_onnx
+    
+
+@try_export
+def export_onnx_end2end(model, im, file, simplify, topk_all, iou_thres, conf_thres, device, labels, prefix=colorstr('ONNX END2END:')):
+    # YOLO ONNX export
+    check_requirements('onnx')
+    import onnx
+    LOGGER.info(f'\n{prefix} starting export with onnx {onnx.__version__}...')
+    f = os.path.splitext(file)[0] + "-end2end.onnx"
+    batch_size = 'batch'
+
+    dynamic_axes = {'images': {0 : 'batch', 2: 'height', 3:'width'}, } # variable length axes
+
+    output_axes = {
+                    'num_dets': {0: 'batch'},
+                    'det_boxes': {0: 'batch'},
+                    'det_scores': {0: 'batch'},
+                    'det_classes': {0: 'batch'},
+                }
+    dynamic_axes.update(output_axes)
+    model = End2End(model, topk_all, iou_thres, conf_thres, None ,device, labels)
+
+    output_names = ['num_dets', 'det_boxes', 'det_scores', 'det_classes']
+    shapes = [ batch_size, 1,  batch_size,  topk_all, 4,
+               batch_size,  topk_all,  batch_size,  topk_all]
+
+    torch.onnx.export(model, 
+                          im, 
+                          f, 
+                          verbose=False, 
+                          export_params=True,       # store the trained parameter weights inside the model file
+                          opset_version=12, 
+                          do_constant_folding=True, # whether to execute constant folding for optimization
+                          input_names=['images'],
+                          output_names=output_names,
+                          dynamic_axes=dynamic_axes)
+
+    # Checks
+    model_onnx = onnx.load(f)  # load onnx model
+    onnx.checker.check_model(model_onnx)  # check onnx model
+    for i in model_onnx.graph.output:
+        for j in i.type.tensor_type.shape.dim:
+            j.dim_param = str(shapes.pop(0))
+
+    if simplify:
+        try:
+            import onnxsim
+
+            print('\nStarting to simplify ONNX...')
+            model_onnx, check = onnxsim.simplify(model_onnx)
+            assert check, 'assert check failed'
+        except Exception as e:
+            print(f'Simplifier failure: {e}')
+
+        # print(onnx.helper.printable_graph(onnx_model.graph))  # print a human readable model
+        onnx.save(model_onnx,f)
+        print('ONNX export success, saved as %s' % f)
+    return f, model_onnx
+
+
+@try_export
+def export_openvino(file, metadata, half, prefix=colorstr('OpenVINO:')):
+    # YOLO OpenVINO export
+    check_requirements('openvino-dev')  # requires openvino-dev: https://pypi.org/project/openvino-dev/
+    import openvino.inference_engine as ie
+
+    LOGGER.info(f'\n{prefix} starting export with openvino {ie.__version__}...')
+    f = str(file).replace('.pt', f'_openvino_model{os.sep}')
+
+    #cmd = f"mo --input_model {file.with_suffix('.onnx')} --output_dir {f} --data_type {'FP16' if half else 'FP32'}"
+    #cmd = f"mo --input_model {file.with_suffix('.onnx')} --output_dir {f} {"--compress_to_fp16" if half else ""}"
+    half_arg = "--compress_to_fp16" if half else ""
+    cmd = f"mo --input_model {file.with_suffix('.onnx')} --output_dir {f} {half_arg}"
+    subprocess.run(cmd.split(), check=True, env=os.environ)  # export
+    yaml_save(Path(f) / file.with_suffix('.yaml').name, metadata)  # add metadata.yaml
+    return f, None
+
+
+@try_export
+def export_paddle(model, im, file, metadata, prefix=colorstr('PaddlePaddle:')):
+    # YOLO Paddle export
+    check_requirements(('paddlepaddle', 'x2paddle'))
+    import x2paddle
+    from x2paddle.convert import pytorch2paddle
+
+    LOGGER.info(f'\n{prefix} starting export with X2Paddle {x2paddle.__version__}...')
+    f = str(file).replace('.pt', f'_paddle_model{os.sep}')
+
+    pytorch2paddle(module=model, save_dir=f, jit_type='trace', input_examples=[im])  # export
+    yaml_save(Path(f) / file.with_suffix('.yaml').name, metadata)  # add metadata.yaml
+    return f, None
+
+
+@try_export
+def export_coreml(model, im, file, int8, half, prefix=colorstr('CoreML:')):
+    # YOLO CoreML export
+    check_requirements('coremltools')
+    import coremltools as ct
+
+    LOGGER.info(f'\n{prefix} starting export with coremltools {ct.__version__}...')
+    f = file.with_suffix('.mlmodel')
+
+    ts = torch.jit.trace(model, im, strict=False)  # TorchScript model
+    ct_model = ct.convert(ts, inputs=[ct.ImageType('image', shape=im.shape, scale=1 / 255, bias=[0, 0, 0])])
+    bits, mode = (8, 'kmeans_lut') if int8 else (16, 'linear') if half else (32, None)
+    if bits < 32:
+        if MACOS:  # quantization only supported on macOS
+            with warnings.catch_warnings():
+                warnings.filterwarnings("ignore", category=DeprecationWarning)  # suppress numpy==1.20 float warning
+                ct_model = ct.models.neural_network.quantization_utils.quantize_weights(ct_model, bits, mode)
+        else:
+            print(f'{prefix} quantization only supported on macOS, skipping...')
+    ct_model.save(f)
+    return f, ct_model
+
+
+@try_export
+def export_engine(model, im, file, half, dynamic, simplify, workspace=4, verbose=False, prefix=colorstr('TensorRT:')):
+    # YOLO TensorRT export https://developer.nvidia.com/tensorrt
+    assert im.device.type != 'cpu', 'export running on CPU but must be on GPU, i.e. `python export.py --device 0`'
+    try:
+        import tensorrt as trt
+    except Exception:
+        if platform.system() == 'Linux':
+            check_requirements('nvidia-tensorrt', cmds='-U --index-url https://pypi.ngc.nvidia.com')
+        import tensorrt as trt
+
+    if trt.__version__[0] == '7':  # TensorRT 7 handling https://github.com/ultralytics/yolov5/issues/6012
+        grid = model.model[-1].anchor_grid
+        model.model[-1].anchor_grid = [a[..., :1, :1, :] for a in grid]
+        export_onnx(model, im, file, 12, dynamic, simplify)  # opset 12
+        model.model[-1].anchor_grid = grid
+    else:  # TensorRT >= 8
+        check_version(trt.__version__, '8.0.0', hard=True)  # require tensorrt>=8.0.0
+        export_onnx(model, im, file, 12, dynamic, simplify)  # opset 12
+    onnx = file.with_suffix('.onnx')
+
+    LOGGER.info(f'\n{prefix} starting export with TensorRT {trt.__version__}...')
+    assert onnx.exists(), f'failed to export ONNX file: {onnx}'
+    f = file.with_suffix('.engine')  # TensorRT engine file
+    logger = trt.Logger(trt.Logger.INFO)
+    if verbose:
+        logger.min_severity = trt.Logger.Severity.VERBOSE
+
+    builder = trt.Builder(logger)
+    config = builder.create_builder_config()
+    config.max_workspace_size = workspace * 1 << 30
+    # config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, workspace << 30)  # fix TRT 8.4 deprecation notice
+
+    flag = (1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
+    network = builder.create_network(flag)
+    parser = trt.OnnxParser(network, logger)
+    if not parser.parse_from_file(str(onnx)):
+        raise RuntimeError(f'failed to load ONNX file: {onnx}')
+
+    inputs = [network.get_input(i) for i in range(network.num_inputs)]
+    outputs = [network.get_output(i) for i in range(network.num_outputs)]
+    for inp in inputs:
+        LOGGER.info(f'{prefix} input "{inp.name}" with shape{inp.shape} {inp.dtype}')
+    for out in outputs:
+        LOGGER.info(f'{prefix} output "{out.name}" with shape{out.shape} {out.dtype}')
+
+    if dynamic:
+        if im.shape[0] <= 1:
+            LOGGER.warning(f"{prefix} WARNING ⚠️ --dynamic model requires maximum --batch-size argument")
+        profile = builder.create_optimization_profile()
+        for inp in inputs:
+            profile.set_shape(inp.name, (1, *im.shape[1:]), (max(1, im.shape[0] // 2), *im.shape[1:]), im.shape)
+        config.add_optimization_profile(profile)
+
+    LOGGER.info(f'{prefix} building FP{16 if builder.platform_has_fast_fp16 and half else 32} engine as {f}')
+    if builder.platform_has_fast_fp16 and half:
+        config.set_flag(trt.BuilderFlag.FP16)
+    with builder.build_engine(network, config) as engine, open(f, 'wb') as t:
+        t.write(engine.serialize())
+    return f, None
+
+
+@try_export
+def export_saved_model(model,
+                       im,
+                       file,
+                       dynamic,
+                       tf_nms=False,
+                       agnostic_nms=False,
+                       topk_per_class=100,
+                       topk_all=100,
+                       iou_thres=0.45,
+                       conf_thres=0.25,
+                       keras=False,
+                       prefix=colorstr('TensorFlow SavedModel:')):
+    # YOLO TensorFlow SavedModel export
+    try:
+        import tensorflow as tf
+    except Exception:
+        check_requirements(f"tensorflow{'' if torch.cuda.is_available() else '-macos' if MACOS else '-cpu'}")
+        import tensorflow as tf
+    from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
+
+    from models.tf import TFModel
+
+    LOGGER.info(f'\n{prefix} starting export with tensorflow {tf.__version__}...')
+    f = str(file).replace('.pt', '_saved_model')
+    batch_size, ch, *imgsz = list(im.shape)  # BCHW
+
+    tf_model = TFModel(cfg=model.yaml, model=model, nc=model.nc, imgsz=imgsz)
+    im = tf.zeros((batch_size, *imgsz, ch))  # BHWC order for TensorFlow
+    _ = tf_model.predict(im, tf_nms, agnostic_nms, topk_per_class, topk_all, iou_thres, conf_thres)
+    inputs = tf.keras.Input(shape=(*imgsz, ch), batch_size=None if dynamic else batch_size)
+    outputs = tf_model.predict(inputs, tf_nms, agnostic_nms, topk_per_class, topk_all, iou_thres, conf_thres)
+    keras_model = tf.keras.Model(inputs=inputs, outputs=outputs)
+    keras_model.trainable = False
+    keras_model.summary()
+    if keras:
+        keras_model.save(f, save_format='tf')
+    else:
+        spec = tf.TensorSpec(keras_model.inputs[0].shape, keras_model.inputs[0].dtype)
+        m = tf.function(lambda x: keras_model(x))  # full model
+        m = m.get_concrete_function(spec)
+        frozen_func = convert_variables_to_constants_v2(m)
+        tfm = tf.Module()
+        tfm.__call__ = tf.function(lambda x: frozen_func(x)[:4] if tf_nms else frozen_func(x), [spec])
+        tfm.__call__(im)
+        tf.saved_model.save(tfm,
+                            f,
+                            options=tf.saved_model.SaveOptions(experimental_custom_gradients=False) if check_version(
+                                tf.__version__, '2.6') else tf.saved_model.SaveOptions())
+    return f, keras_model
+
+
+@try_export
+def export_pb(keras_model, file, prefix=colorstr('TensorFlow GraphDef:')):
+    # YOLO TensorFlow GraphDef *.pb export https://github.com/leimao/Frozen_Graph_TensorFlow
+    import tensorflow as tf
+    from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
+
+    LOGGER.info(f'\n{prefix} starting export with tensorflow {tf.__version__}...')
+    f = file.with_suffix('.pb')
+
+    m = tf.function(lambda x: keras_model(x))  # full model
+    m = m.get_concrete_function(tf.TensorSpec(keras_model.inputs[0].shape, keras_model.inputs[0].dtype))
+    frozen_func = convert_variables_to_constants_v2(m)
+    frozen_func.graph.as_graph_def()
+    tf.io.write_graph(graph_or_graph_def=frozen_func.graph, logdir=str(f.parent), name=f.name, as_text=False)
+    return f, None
+
+
+@try_export
+def export_tflite(keras_model, im, file, int8, data, nms, agnostic_nms, prefix=colorstr('TensorFlow Lite:')):
+    # YOLOv5 TensorFlow Lite export
+    import tensorflow as tf
+
+    LOGGER.info(f'\n{prefix} starting export with tensorflow {tf.__version__}...')
+    batch_size, ch, *imgsz = list(im.shape)  # BCHW
+    f = str(file).replace('.pt', '-fp16.tflite')
+
+    converter = tf.lite.TFLiteConverter.from_keras_model(keras_model)
+    converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS]
+    converter.target_spec.supported_types = [tf.float16]
+    converter.optimizations = [tf.lite.Optimize.DEFAULT]
+    if int8:
+        from models.tf import representative_dataset_gen
+        dataset = LoadImages(check_dataset(check_yaml(data))['train'], img_size=imgsz, auto=False)
+        converter.representative_dataset = lambda: representative_dataset_gen(dataset, ncalib=100)
+        converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+        converter.target_spec.supported_types = []
+        converter.inference_input_type = tf.uint8  # or tf.int8
+        converter.inference_output_type = tf.uint8  # or tf.int8
+        converter.experimental_new_quantizer = True
+        f = str(file).replace('.pt', '-int8.tflite')
+    if nms or agnostic_nms:
+        converter.target_spec.supported_ops.append(tf.lite.OpsSet.SELECT_TF_OPS)
+
+    tflite_model = converter.convert()
+    open(f, "wb").write(tflite_model)
+    return f, None
+
+
+@try_export
+def export_edgetpu(file, prefix=colorstr('Edge TPU:')):
+    # YOLO Edge TPU export https://coral.ai/docs/edgetpu/models-intro/
+    cmd = 'edgetpu_compiler --version'
+    help_url = 'https://coral.ai/docs/edgetpu/compiler/'
+    assert platform.system() == 'Linux', f'export only supported on Linux. See {help_url}'
+    if subprocess.run(f'{cmd} >/dev/null', shell=True).returncode != 0:
+        LOGGER.info(f'\n{prefix} export requires Edge TPU compiler. Attempting install from {help_url}')
+        sudo = subprocess.run('sudo --version >/dev/null', shell=True).returncode == 0  # sudo installed on system
+        for c in (
+                'curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add -',
+                'echo "deb https://packages.cloud.google.com/apt coral-edgetpu-stable main" | sudo tee /etc/apt/sources.list.d/coral-edgetpu.list',
+                'sudo apt-get update', 'sudo apt-get install edgetpu-compiler'):
+            subprocess.run(c if sudo else c.replace('sudo ', ''), shell=True, check=True)
+    ver = subprocess.run(cmd, shell=True, capture_output=True, check=True).stdout.decode().split()[-1]
+
+    LOGGER.info(f'\n{prefix} starting export with Edge TPU compiler {ver}...')
+    f = str(file).replace('.pt', '-int8_edgetpu.tflite')  # Edge TPU model
+    f_tfl = str(file).replace('.pt', '-int8.tflite')  # TFLite model
+
+    cmd = f"edgetpu_compiler -s -d -k 10 --out_dir {file.parent} {f_tfl}"
+    subprocess.run(cmd.split(), check=True)
+    return f, None
+
+
+@try_export
+def export_tfjs(file, prefix=colorstr('TensorFlow.js:')):
+    # YOLO TensorFlow.js export
+    check_requirements('tensorflowjs')
+    import tensorflowjs as tfjs
+
+    LOGGER.info(f'\n{prefix} starting export with tensorflowjs {tfjs.__version__}...')
+    f = str(file).replace('.pt', '_web_model')  # js dir
+    f_pb = file.with_suffix('.pb')  # *.pb path
+    f_json = f'{f}/model.json'  # *.json path
+
+    cmd = f'tensorflowjs_converter --input_format=tf_frozen_model ' \
+          f'--output_node_names=Identity,Identity_1,Identity_2,Identity_3 {f_pb} {f}'
+    subprocess.run(cmd.split())
+
+    json = Path(f_json).read_text()
+    with open(f_json, 'w') as j:  # sort JSON Identity_* in ascending order
+        subst = re.sub(
+            r'{"outputs": {"Identity.?.?": {"name": "Identity.?.?"}, '
+            r'"Identity.?.?": {"name": "Identity.?.?"}, '
+            r'"Identity.?.?": {"name": "Identity.?.?"}, '
+            r'"Identity.?.?": {"name": "Identity.?.?"}}}', r'{"outputs": {"Identity": {"name": "Identity"}, '
+            r'"Identity_1": {"name": "Identity_1"}, '
+            r'"Identity_2": {"name": "Identity_2"}, '
+            r'"Identity_3": {"name": "Identity_3"}}}', json)
+        j.write(subst)
+    return f, None
+
+
+def add_tflite_metadata(file, metadata, num_outputs):
+    # Add metadata to *.tflite models per https://www.tensorflow.org/lite/models/convert/metadata
+    with contextlib.suppress(ImportError):
+        # check_requirements('tflite_support')
+        from tflite_support import flatbuffers
+        from tflite_support import metadata as _metadata
+        from tflite_support import metadata_schema_py_generated as _metadata_fb
+
+        tmp_file = Path('/tmp/meta.txt')
+        with open(tmp_file, 'w') as meta_f:
+            meta_f.write(str(metadata))
+
+        model_meta = _metadata_fb.ModelMetadataT()
+        label_file = _metadata_fb.AssociatedFileT()
+        label_file.name = tmp_file.name
+        model_meta.associatedFiles = [label_file]
+
+        subgraph = _metadata_fb.SubGraphMetadataT()
+        subgraph.inputTensorMetadata = [_metadata_fb.TensorMetadataT()]
+        subgraph.outputTensorMetadata = [_metadata_fb.TensorMetadataT()] * num_outputs
+        model_meta.subgraphMetadata = [subgraph]
+
+        b = flatbuffers.Builder(0)
+        b.Finish(model_meta.Pack(b), _metadata.MetadataPopulator.METADATA_FILE_IDENTIFIER)
+        metadata_buf = b.Output()
+
+        populator = _metadata.MetadataPopulator.with_model_file(file)
+        populator.load_metadata_buffer(metadata_buf)
+        populator.load_associated_files([str(tmp_file)])
+        populator.populate()
+        tmp_file.unlink()
+
+
+@smart_inference_mode()
+def run(
+        data=ROOT / 'data/coco.yaml',  # 'dataset.yaml path'
+        weights=ROOT / 'yolo.pt',  # weights path
+        imgsz=(640, 640),  # image (height, width)
+        batch_size=1,  # batch size
+        device='cpu',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        include=('torchscript', 'onnx'),  # include formats
+        half=False,  # FP16 half-precision export
+        inplace=False,  # set YOLO Detect() inplace=True
+        keras=False,  # use Keras
+        optimize=False,  # TorchScript: optimize for mobile
+        int8=False,  # CoreML/TF INT8 quantization
+        dynamic=False,  # ONNX/TF/TensorRT: dynamic axes
+        simplify=False,  # ONNX: simplify model
+        opset=12,  # ONNX: opset version
+        verbose=False,  # TensorRT: verbose log
+        workspace=4,  # TensorRT: workspace size (GB)
+        nms=False,  # TF: add NMS to model
+        agnostic_nms=False,  # TF: add agnostic NMS to model
+        topk_per_class=100,  # TF.js NMS: topk per class to keep
+        topk_all=100,  # TF.js NMS: topk for all classes to keep
+        iou_thres=0.45,  # TF.js NMS: IoU threshold
+        conf_thres=0.25,  # TF.js NMS: confidence threshold
+):
+    t = time.time()
+    include = [x.lower() for x in include]  # to lowercase
+    fmts = tuple(export_formats()['Argument'][1:])  # --include arguments
+    flags = [x in include for x in fmts]
+    assert sum(flags) == len(include), f'ERROR: Invalid --include {include}, valid --include arguments are {fmts}'
+    jit, onnx, onnx_end2end, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, paddle = flags  # export booleans
+    file = Path(url2file(weights) if str(weights).startswith(('http:/', 'https:/')) else weights)  # PyTorch weights
+
+    # Load PyTorch model
+    device = select_device(device)
+    if half:
+        assert device.type != 'cpu' or coreml, '--half only compatible with GPU export, i.e. use --device 0'
+        assert not dynamic, '--half not compatible with --dynamic, i.e. use either --half or --dynamic but not both'
+    model = attempt_load(weights, device=device, inplace=True, fuse=True)  # load FP32 model
+
+    # Checks
+    imgsz *= 2 if len(imgsz) == 1 else 1  # expand
+    if optimize:
+        assert device.type == 'cpu', '--optimize not compatible with cuda devices, i.e. use --device cpu'
+
+    # Input
+    gs = int(max(model.stride))  # grid size (max stride)
+    imgsz = [check_img_size(x, gs) for x in imgsz]  # verify img_size are gs-multiples
+    im = torch.zeros(batch_size, 3, *imgsz).to(device)  # image size(1,3,320,192) BCHW iDetection
+
+    # Update model
+    model.eval()
+    for k, m in model.named_modules():
+        if isinstance(m, (Detect, DDetect, DualDetect, DualDDetect)):
+            m.inplace = inplace
+            m.dynamic = dynamic
+            m.export = True
+
+    for _ in range(2):
+        y = model(im)  # dry runs
+    if half and not coreml:
+        im, model = im.half(), model.half()  # to FP16
+    shape = tuple((y[0] if isinstance(y, (tuple, list)) else y).shape)  # model output shape
+    metadata = {'stride': int(max(model.stride)), 'names': model.names}  # model metadata
+    LOGGER.info(f"\n{colorstr('PyTorch:')} starting from {file} with output shape {shape} ({file_size(file):.1f} MB)")
+
+    # Exports
+    f = [''] * len(fmts)  # exported filenames
+    warnings.filterwarnings(action='ignore', category=torch.jit.TracerWarning)  # suppress TracerWarning
+    if jit:  # TorchScript
+        f[0], _ = export_torchscript(model, im, file, optimize)
+    if engine:  # TensorRT required before ONNX
+        f[1], _ = export_engine(model, im, file, half, dynamic, simplify, workspace, verbose)
+    if onnx or xml:  # OpenVINO requires ONNX
+        f[2], _ = export_onnx(model, im, file, opset, dynamic, simplify)
+    if onnx_end2end:
+        if isinstance(model, DetectionModel):
+            labels = model.names
+            f[2], _ = export_onnx_end2end(model, im, file, simplify, topk_all, iou_thres, conf_thres, device, len(labels))
+        else:
+            raise RuntimeError("The model is not a DetectionModel.")
+    if xml:  # OpenVINO
+        f[3], _ = export_openvino(file, metadata, half)
+    if coreml:  # CoreML
+        f[4], _ = export_coreml(model, im, file, int8, half)
+    if any((saved_model, pb, tflite, edgetpu, tfjs)):  # TensorFlow formats
+        assert not tflite or not tfjs, 'TFLite and TF.js models must be exported separately, please pass only one type.'
+        assert not isinstance(model, ClassificationModel), 'ClassificationModel export to TF formats not yet supported.'
+        f[5], s_model = export_saved_model(model.cpu(),
+                                           im,
+                                           file,
+                                           dynamic,
+                                           tf_nms=nms or agnostic_nms or tfjs,
+                                           agnostic_nms=agnostic_nms or tfjs,
+                                           topk_per_class=topk_per_class,
+                                           topk_all=topk_all,
+                                           iou_thres=iou_thres,
+                                           conf_thres=conf_thres,
+                                           keras=keras)
+        if pb or tfjs:  # pb prerequisite to tfjs
+            f[6], _ = export_pb(s_model, file)
+        if tflite or edgetpu:
+            f[7], _ = export_tflite(s_model, im, file, int8 or edgetpu, data=data, nms=nms, agnostic_nms=agnostic_nms)
+            if edgetpu:
+                f[8], _ = export_edgetpu(file)
+            add_tflite_metadata(f[8] or f[7], metadata, num_outputs=len(s_model.outputs))
+        if tfjs:
+            f[9], _ = export_tfjs(file)
+    if paddle:  # PaddlePaddle
+        f[10], _ = export_paddle(model, im, file, metadata)
+
+    # Finish
+    f = [str(x) for x in f if x]  # filter out '' and None
+    if any(f):
+        cls, det, seg = (isinstance(model, x) for x in (ClassificationModel, DetectionModel, SegmentationModel))  # type
+        dir = Path('segment' if seg else 'classify' if cls else '')
+        h = '--half' if half else ''  # --half FP16 inference arg
+        s = "# WARNING ⚠️ ClassificationModel not yet supported for PyTorch Hub AutoShape inference" if cls else \
+            "# WARNING ⚠️ SegmentationModel not yet supported for PyTorch Hub AutoShape inference" if seg else ''
+        if onnx_end2end:
+            LOGGER.info(f'\nExport complete ({time.time() - t:.1f}s)'
+                        f"\nResults saved to {colorstr('bold', file.parent.resolve())}"
+                        f"\nVisualize:       https://netron.app")
+        else:
+            LOGGER.info(f'\nExport complete ({time.time() - t:.1f}s)'
+                        f"\nResults saved to {colorstr('bold', file.parent.resolve())}"
+                        f"\nDetect:          python {dir / ('detect.py' if det else 'predict.py')} --weights {f[-1]} {h}"
+                        f"\nValidate:        python {dir / 'val.py'} --weights {f[-1]} {h}"
+                        f"\nPyTorch Hub:     model = torch.hub.load('ultralytics/yolov5', 'custom', '{f[-1]}')  {s}"
+                        f"\nVisualize:       https://netron.app")
+    return f  # return list of exported files/dirs
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco.yaml', help='dataset.yaml path')
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo.pt', help='model.pt path(s)')
+    parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640, 640], help='image (h, w)')
+    parser.add_argument('--batch-size', type=int, default=1, help='batch size')
+    parser.add_argument('--device', default='cpu', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--half', action='store_true', help='FP16 half-precision export')
+    parser.add_argument('--inplace', action='store_true', help='set YOLO Detect() inplace=True')
+    parser.add_argument('--keras', action='store_true', help='TF: use Keras')
+    parser.add_argument('--optimize', action='store_true', help='TorchScript: optimize for mobile')
+    parser.add_argument('--int8', action='store_true', help='CoreML/TF INT8 quantization')
+    parser.add_argument('--dynamic', action='store_true', help='ONNX/TF/TensorRT: dynamic axes')
+    parser.add_argument('--simplify', action='store_true', help='ONNX: simplify model')
+    parser.add_argument('--opset', type=int, default=12, help='ONNX: opset version')
+    parser.add_argument('--verbose', action='store_true', help='TensorRT: verbose log')
+    parser.add_argument('--workspace', type=int, default=4, help='TensorRT: workspace size (GB)')
+    parser.add_argument('--nms', action='store_true', help='TF: add NMS to model')
+    parser.add_argument('--agnostic-nms', action='store_true', help='TF: add agnostic NMS to model')
+    parser.add_argument('--topk-per-class', type=int, default=100, help='TF.js NMS: topk per class to keep')
+    parser.add_argument('--topk-all', type=int, default=100, help='ONNX END2END/TF.js NMS: topk for all classes to keep')
+    parser.add_argument('--iou-thres', type=float, default=0.45, help='ONNX END2END/TF.js NMS: IoU threshold')
+    parser.add_argument('--conf-thres', type=float, default=0.25, help='ONNX END2END/TF.js NMS: confidence threshold')
+    parser.add_argument(
+        '--include',
+        nargs='+',
+        default=['torchscript'],
+        help='torchscript, onnx, onnx_end2end, openvino, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, paddle')
+    opt = parser.parse_args()
+
+    if 'onnx_end2end' in opt.include:  
+        opt.simplify = True
+        opt.dynamic = True
+        opt.inplace = True
+        opt.half = False
+
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    for opt.weights in (opt.weights if isinstance(opt.weights, list) else [opt.weights]):
+        run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

yolov9/hubconf.py

@@ -0,0 +1,107 @@
+import torch
+
+
+def _create(name, pretrained=True, channels=3, classes=80, autoshape=True, verbose=True, device=None):
+    """Creates or loads a YOLO model
+
+    Arguments:
+        name (str): model name 'yolov3' or path 'path/to/best.pt'
+        pretrained (bool): load pretrained weights into the model
+        channels (int): number of input channels
+        classes (int): number of model classes
+        autoshape (bool): apply YOLO .autoshape() wrapper to model
+        verbose (bool): print all information to screen
+        device (str, torch.device, None): device to use for model parameters
+
+    Returns:
+        YOLO model
+    """
+    from pathlib import Path
+
+    from models.common import AutoShape, DetectMultiBackend
+    from models.experimental import attempt_load
+    from models.yolo import ClassificationModel, DetectionModel, SegmentationModel
+    from utils.downloads import attempt_download
+    from utils.general import LOGGER, check_requirements, intersect_dicts, logging
+    from utils.torch_utils import select_device
+
+    if not verbose:
+        LOGGER.setLevel(logging.WARNING)
+    check_requirements(exclude=('opencv-python', 'tensorboard', 'thop'))
+    name = Path(name)
+    path = name.with_suffix('.pt') if name.suffix == '' and not name.is_dir() else name  # checkpoint path
+    try:
+        device = select_device(device)
+        if pretrained and channels == 3 and classes == 80:
+            try:
+                model = DetectMultiBackend(path, device=device, fuse=autoshape)  # detection model
+                if autoshape:
+                    if model.pt and isinstance(model.model, ClassificationModel):
+                        LOGGER.warning('WARNING ⚠️ YOLO ClassificationModel is not yet AutoShape compatible. '
+                                       'You must pass torch tensors in BCHW to this model, i.e. shape(1,3,224,224).')
+                    elif model.pt and isinstance(model.model, SegmentationModel):
+                        LOGGER.warning('WARNING ⚠️ YOLO SegmentationModel is not yet AutoShape compatible. '
+                                       'You will not be able to run inference with this model.')
+                    else:
+                        model = AutoShape(model)  # for file/URI/PIL/cv2/np inputs and NMS
+            except Exception:
+                model = attempt_load(path, device=device, fuse=False)  # arbitrary model
+        else:
+            cfg = list((Path(__file__).parent / 'models').rglob(f'{path.stem}.yaml'))[0]  # model.yaml path
+            model = DetectionModel(cfg, channels, classes)  # create model
+            if pretrained:
+                ckpt = torch.load(attempt_download(path), map_location=device)  # load
+                csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+                csd = intersect_dicts(csd, model.state_dict(), exclude=['anchors'])  # intersect
+                model.load_state_dict(csd, strict=False)  # load
+                if len(ckpt['model'].names) == classes:
+                    model.names = ckpt['model'].names  # set class names attribute
+        if not verbose:
+            LOGGER.setLevel(logging.INFO)  # reset to default
+        return model.to(device)
+
+    except Exception as e:
+        help_url = 'https://github.com/ultralytics/yolov5/issues/36'
+        s = f'{e}. Cache may be out of date, try `force_reload=True` or see {help_url} for help.'
+        raise Exception(s) from e
+
+
+def custom(path='path/to/model.pt', autoshape=True, _verbose=True, device=None):
+    # YOLO custom or local model
+    return _create(path, autoshape=autoshape, verbose=_verbose, device=device)
+
+
+if __name__ == '__main__':
+    import argparse
+    from pathlib import Path
+
+    import numpy as np
+    from PIL import Image
+
+    from utils.general import cv2, print_args
+
+    # Argparser
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--model', type=str, default='yolo', help='model name')
+    opt = parser.parse_args()
+    print_args(vars(opt))
+
+    # Model
+    model = _create(name=opt.model, pretrained=True, channels=3, classes=80, autoshape=True, verbose=True)
+    # model = custom(path='path/to/model.pt')  # custom
+
+    # Images
+    imgs = [
+        'data/images/zidane.jpg',  # filename
+        Path('data/images/zidane.jpg'),  # Path
+        'https://ultralytics.com/images/zidane.jpg',  # URI
+        cv2.imread('data/images/bus.jpg')[:, :, ::-1],  # OpenCV
+        Image.open('data/images/bus.jpg'),  # PIL
+        np.zeros((320, 640, 3))]  # numpy
+
+    # Inference
+    results = model(imgs, size=320)  # batched inference
+
+    # Results
+    results.print()
+    results.save()

yolov9/models/__init__.py

@@ -0,0 +1 @@
+# init

yolov9/models/common.py

@@ -0,0 +1,1233 @@
+import ast
+import contextlib
+import json
+import math
+import platform
+import warnings
+import zipfile
+from collections import OrderedDict, namedtuple
+from copy import copy
+from pathlib import Path
+from urllib.parse import urlparse
+
+from typing import Optional
+
+import cv2
+import numpy as np
+import pandas as pd
+import requests
+import torch
+import torch.nn as nn
+from IPython.display import display
+from PIL import Image
+from torch.cuda import amp
+
+from utils import TryExcept
+from utils.dataloaders import exif_transpose, letterbox
+from utils.general import (LOGGER, ROOT, Profile, check_requirements, check_suffix, check_version, colorstr,
+                           increment_path, is_notebook, make_divisible, non_max_suppression, scale_boxes,
+                           xywh2xyxy, xyxy2xywh, yaml_load)
+from utils.plots import Annotator, colors, save_one_box
+from utils.torch_utils import copy_attr, smart_inference_mode
+
+
+def autopad(k, p=None, d=1):  # kernel, padding, dilation
+    # Pad to 'same' shape outputs
+    if d > 1:
+        k = d * (k - 1) + 1 if isinstance(k, int) else [d * (x - 1) + 1 for x in k]  # actual kernel-size
+    if p is None:
+        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # auto-pad
+    return p
+
+
+class Conv(nn.Module):
+    # Standard convolution with args(ch_in, ch_out, kernel, stride, padding, groups, dilation, activation)
+    default_act = nn.SiLU()  # default activation
+
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, d=1, act=True):
+        super().__init__()
+        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p, d), groups=g, dilation=d, bias=False)
+        self.bn = nn.BatchNorm2d(c2)
+        self.act = self.default_act if act is True else act if isinstance(act, nn.Module) else nn.Identity()
+
+    def forward(self, x):
+        return self.act(self.bn(self.conv(x)))
+
+    def forward_fuse(self, x):
+        return self.act(self.conv(x))
+
+
+class AConv(nn.Module):
+    def __init__(self, c1, c2):  # ch_in, ch_out, shortcut, kernels, groups, expand
+        super().__init__()
+        self.cv1 = Conv(c1, c2, 3, 2, 1)
+
+    def forward(self, x):
+        x = torch.nn.functional.avg_pool2d(x, 2, 1, 0, False, True)
+        return self.cv1(x)
+
+
+class ADown(nn.Module):
+    def __init__(self, c1, c2):  # ch_in, ch_out, shortcut, kernels, groups, expand
+        super().__init__()
+        self.c = c2 // 2
+        self.cv1 = Conv(c1 // 2, self.c, 3, 2, 1)
+        self.cv2 = Conv(c1 // 2, self.c, 1, 1, 0)
+
+    def forward(self, x):
+        x = torch.nn.functional.avg_pool2d(x, 2, 1, 0, False, True)
+        x1,x2 = x.chunk(2, 1)
+        x1 = self.cv1(x1)
+        x2 = torch.nn.functional.max_pool2d(x2, 3, 2, 1)
+        x2 = self.cv2(x2)
+        return torch.cat((x1, x2), 1)
+
+
+class RepConvN(nn.Module):
+    """RepConv is a basic rep-style block, including training and deploy status
+    This code is based on https://github.com/DingXiaoH/RepVGG/blob/main/repvgg.py
+    """
+    default_act = nn.SiLU()  # default activation
+
+    def __init__(self, c1, c2, k=3, s=1, p=1, g=1, d=1, act=True, bn=False, deploy=False):
+        super().__init__()
+        assert k == 3 and p == 1
+        self.g = g
+        self.c1 = c1
+        self.c2 = c2
+        self.act = self.default_act if act is True else act if isinstance(act, nn.Module) else nn.Identity()
+
+        self.bn = None
+        self.conv1 = Conv(c1, c2, k, s, p=p, g=g, act=False)
+        self.conv2 = Conv(c1, c2, 1, s, p=(p - k // 2), g=g, act=False)
+
+    def forward_fuse(self, x):
+        """Forward process"""
+        return self.act(self.conv(x))
+
+    def forward(self, x):
+        """Forward process"""
+        id_out = 0 if self.bn is None else self.bn(x)
+        return self.act(self.conv1(x) + self.conv2(x) + id_out)
+
+    def get_equivalent_kernel_bias(self):
+        kernel3x3, bias3x3 = self._fuse_bn_tensor(self.conv1)
+        kernel1x1, bias1x1 = self._fuse_bn_tensor(self.conv2)
+        kernelid, biasid = self._fuse_bn_tensor(self.bn)
+        return kernel3x3 + self._pad_1x1_to_3x3_tensor(kernel1x1) + kernelid, bias3x3 + bias1x1 + biasid
+
+    def _avg_to_3x3_tensor(self, avgp):
+        channels = self.c1
+        groups = self.g
+        kernel_size = avgp.kernel_size
+        input_dim = channels // groups
+        k = torch.zeros((channels, input_dim, kernel_size, kernel_size))
+        k[np.arange(channels), np.tile(np.arange(input_dim), groups), :, :] = 1.0 / kernel_size ** 2
+        return k
+
+    def _pad_1x1_to_3x3_tensor(self, kernel1x1):
+        if kernel1x1 is None:
+            return 0
+        else:
+            return torch.nn.functional.pad(kernel1x1, [1, 1, 1, 1])
+
+    def _fuse_bn_tensor(self, branch):
+        if branch is None:
+            return 0, 0
+        if isinstance(branch, Conv):
+            kernel = branch.conv.weight
+            running_mean = branch.bn.running_mean
+            running_var = branch.bn.running_var
+            gamma = branch.bn.weight
+            beta = branch.bn.bias
+            eps = branch.bn.eps
+        elif isinstance(branch, nn.BatchNorm2d):
+            if not hasattr(self, 'id_tensor'):
+                input_dim = self.c1 // self.g
+                kernel_value = np.zeros((self.c1, input_dim, 3, 3), dtype=np.float32)
+                for i in range(self.c1):
+                    kernel_value[i, i % input_dim, 1, 1] = 1
+                self.id_tensor = torch.from_numpy(kernel_value).to(branch.weight.device)
+            kernel = self.id_tensor
+            running_mean = branch.running_mean
+            running_var = branch.running_var
+            gamma = branch.weight
+            beta = branch.bias
+            eps = branch.eps
+        std = (running_var + eps).sqrt()
+        t = (gamma / std).reshape(-1, 1, 1, 1)
+        return kernel * t, beta - running_mean * gamma / std
+
+    def fuse_convs(self):
+        if hasattr(self, 'conv'):
+            return
+        kernel, bias = self.get_equivalent_kernel_bias()
+        self.conv = nn.Conv2d(in_channels=self.conv1.conv.in_channels,
+                              out_channels=self.conv1.conv.out_channels,
+                              kernel_size=self.conv1.conv.kernel_size,
+                              stride=self.conv1.conv.stride,
+                              padding=self.conv1.conv.padding,
+                              dilation=self.conv1.conv.dilation,
+                              groups=self.conv1.conv.groups,
+                              bias=True).requires_grad_(False)
+        self.conv.weight.data = kernel
+        self.conv.bias.data = bias
+        for para in self.parameters():
+            para.detach_()
+        self.__delattr__('conv1')
+        self.__delattr__('conv2')
+        if hasattr(self, 'nm'):
+            self.__delattr__('nm')
+        if hasattr(self, 'bn'):
+            self.__delattr__('bn')
+        if hasattr(self, 'id_tensor'):
+            self.__delattr__('id_tensor')
+
+
+class SP(nn.Module):
+    def __init__(self, k=3, s=1):
+        super(SP, self).__init__()
+        self.m = nn.MaxPool2d(kernel_size=k, stride=s, padding=k // 2)
+
+    def forward(self, x):
+        return self.m(x)
+
+
+class MP(nn.Module):
+    # Max pooling
+    def __init__(self, k=2):
+        super(MP, self).__init__()
+        self.m = nn.MaxPool2d(kernel_size=k, stride=k)
+
+    def forward(self, x):
+        return self.m(x)
+
+
+class ConvTranspose(nn.Module):
+    # Convolution transpose 2d layer
+    default_act = nn.SiLU()  # default activation
+
+    def __init__(self, c1, c2, k=2, s=2, p=0, bn=True, act=True):
+        super().__init__()
+        self.conv_transpose = nn.ConvTranspose2d(c1, c2, k, s, p, bias=not bn)
+        self.bn = nn.BatchNorm2d(c2) if bn else nn.Identity()
+        self.act = self.default_act if act is True else act if isinstance(act, nn.Module) else nn.Identity()
+
+    def forward(self, x):
+        return self.act(self.bn(self.conv_transpose(x)))
+
+
+class DWConv(Conv):
+    # Depth-wise convolution
+    def __init__(self, c1, c2, k=1, s=1, d=1, act=True):  # ch_in, ch_out, kernel, stride, dilation, activation
+        super().__init__(c1, c2, k, s, g=math.gcd(c1, c2), d=d, act=act)
+
+
+class DWConvTranspose2d(nn.ConvTranspose2d):
+    # Depth-wise transpose convolution
+    def __init__(self, c1, c2, k=1, s=1, p1=0, p2=0):  # ch_in, ch_out, kernel, stride, padding, padding_out
+        super().__init__(c1, c2, k, s, p1, p2, groups=math.gcd(c1, c2))
+
+
+class DFL(nn.Module):
+    # DFL module
+    def __init__(self, c1=17):
+        super().__init__()
+        self.conv = nn.Conv2d(c1, 1, 1, bias=False).requires_grad_(False)
+        self.conv.weight.data[:] = nn.Parameter(torch.arange(c1, dtype=torch.float).view(1, c1, 1, 1)) # / 120.0
+        self.c1 = c1
+        # self.bn = nn.BatchNorm2d(4)
+
+    def forward(self, x):
+        b, c, a = x.shape  # batch, channels, anchors
+        return self.conv(x.view(b, 4, self.c1, a).transpose(2, 1).softmax(1)).view(b, 4, a)
+        # return self.conv(x.view(b, self.c1, 4, a).softmax(1)).view(b, 4, a)
+
+
+class BottleneckBase(nn.Module):
+    # Standard bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, k=(1, 3), e=0.5):  # ch_in, ch_out, shortcut, kernels, groups, expand
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, k[0], 1)
+        self.cv2 = Conv(c_, c2, k[1], 1, g=g)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
+
+
+class RBottleneckBase(nn.Module):
+    # Standard bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 1), e=0.5):  # ch_in, ch_out, shortcut, kernels, groups, expand
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, k[0], 1)
+        self.cv2 = Conv(c_, c2, k[1], 1, g=g)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
+
+
+class RepNRBottleneckBase(nn.Module):
+    # Standard bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 1), e=0.5):  # ch_in, ch_out, shortcut, kernels, groups, expand
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = RepConvN(c1, c_, k[0], 1)
+        self.cv2 = Conv(c_, c2, k[1], 1, g=g)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
+
+
+class Bottleneck(nn.Module):
+    # Standard bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5):  # ch_in, ch_out, shortcut, kernels, groups, expand
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, k[0], 1)
+        self.cv2 = Conv(c_, c2, k[1], 1, g=g)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
+
+
+class RepNBottleneck(nn.Module):
+    # Standard bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5):  # ch_in, ch_out, shortcut, kernels, groups, expand
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = RepConvN(c1, c_, k[0], 1)
+        self.cv2 = Conv(c_, c2, k[1], 1, g=g)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
+
+
+class Res(nn.Module):
+    # ResNet bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
+        super(Res, self).__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_, c_, 3, 1, g=g)
+        self.cv3 = Conv(c_, c2, 1, 1)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv3(self.cv2(self.cv1(x))) if self.add else self.cv3(self.cv2(self.cv1(x)))
+
+
+class RepNRes(nn.Module):
+    # ResNet bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
+        super(RepNRes, self).__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = RepConvN(c_, c_, 3, 1, g=g)
+        self.cv3 = Conv(c_, c2, 1, 1)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv3(self.cv2(self.cv1(x))) if self.add else self.cv3(self.cv2(self.cv1(x)))
+
+
+class BottleneckCSP(nn.Module):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False)
+        self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
+        self.cv4 = Conv(2 * c_, c2, 1, 1)
+        self.bn = nn.BatchNorm2d(2 * c_)  # applied to cat(cv2, cv3)
+        self.act = nn.SiLU()
+        self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
+
+    def forward(self, x):
+        y1 = self.cv3(self.m(self.cv1(x)))
+        y2 = self.cv2(x)
+        return self.cv4(self.act(self.bn(torch.cat((y1, y2), 1))))
+
+
+class CSP(nn.Module):
+    # CSP Bottleneck with 3 convolutions
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c1, c_, 1, 1)
+        self.cv3 = Conv(2 * c_, c2, 1)  # optional act=FReLU(c2)
+        self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
+
+    def forward(self, x):
+        return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1))
+
+
+class RepNCSP(nn.Module):
+    # CSP Bottleneck with 3 convolutions
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c1, c_, 1, 1)
+        self.cv3 = Conv(2 * c_, c2, 1)  # optional act=FReLU(c2)
+        self.m = nn.Sequential(*(RepNBottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
+
+    def forward(self, x):
+        return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1))
+
+
+class CSPBase(nn.Module):
+    # CSP Bottleneck with 3 convolutions
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c1, c_, 1, 1)
+        self.cv3 = Conv(2 * c_, c2, 1)  # optional act=FReLU(c2)
+        self.m = nn.Sequential(*(BottleneckBase(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
+
+    def forward(self, x):
+        return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1))
+
+
+class SPP(nn.Module):
+    # Spatial Pyramid Pooling (SPP) layer https://arxiv.org/abs/1406.4729
+    def __init__(self, c1, c2, k=(5, 9, 13)):
+        super().__init__()
+        c_ = c1 // 2  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1)
+        self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])
+
+    def forward(self, x):
+        x = self.cv1(x)
+        with warnings.catch_warnings():
+            warnings.simplefilter('ignore')  # suppress torch 1.9.0 max_pool2d() warning
+            return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1))
+
+        
+class ASPP(torch.nn.Module):
+
+    def __init__(self, in_channels, out_channels):
+        super().__init__()
+        kernel_sizes = [1, 3, 3, 1]
+        dilations = [1, 3, 6, 1]
+        paddings = [0, 3, 6, 0]
+        self.aspp = torch.nn.ModuleList()
+        for aspp_idx in range(len(kernel_sizes)):
+            conv = torch.nn.Conv2d(
+                in_channels,
+                out_channels,
+                kernel_size=kernel_sizes[aspp_idx],
+                stride=1,
+                dilation=dilations[aspp_idx],
+                padding=paddings[aspp_idx],
+                bias=True)
+            self.aspp.append(conv)
+        self.gap = torch.nn.AdaptiveAvgPool2d(1)
+        self.aspp_num = len(kernel_sizes)
+        for m in self.modules():
+            if isinstance(m, torch.nn.Conv2d):
+                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+                m.weight.data.normal_(0, math.sqrt(2. / n))
+                m.bias.data.fill_(0)
+
+    def forward(self, x):
+        avg_x = self.gap(x)
+        out = []
+        for aspp_idx in range(self.aspp_num):
+            inp = avg_x if (aspp_idx == self.aspp_num - 1) else x
+            out.append(F.relu_(self.aspp[aspp_idx](inp)))
+        out[-1] = out[-1].expand_as(out[-2])
+        out = torch.cat(out, dim=1)
+        return out
+
+
+class SPPCSPC(nn.Module):
+    # CSP SPP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5, k=(5, 9, 13)):
+        super(SPPCSPC, self).__init__()
+        c_ = int(2 * c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c1, c_, 1, 1)
+        self.cv3 = Conv(c_, c_, 3, 1)
+        self.cv4 = Conv(c_, c_, 1, 1)
+        self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])
+        self.cv5 = Conv(4 * c_, c_, 1, 1)
+        self.cv6 = Conv(c_, c_, 3, 1)
+        self.cv7 = Conv(2 * c_, c2, 1, 1)
+
+    def forward(self, x):
+        x1 = self.cv4(self.cv3(self.cv1(x)))
+        y1 = self.cv6(self.cv5(torch.cat([x1] + [m(x1) for m in self.m], 1)))
+        y2 = self.cv2(x)
+        return self.cv7(torch.cat((y1, y2), dim=1))
+
+
+class SPPF(nn.Module):
+    # Spatial Pyramid Pooling - Fast (SPPF) layer by Glenn Jocher
+    def __init__(self, c1, c2, k=5):  # equivalent to SPP(k=(5, 9, 13))
+        super().__init__()
+        c_ = c1 // 2  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_ * 4, c2, 1, 1)
+        self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2)
+        # self.m = SoftPool2d(kernel_size=k, stride=1, padding=k // 2)
+
+    def forward(self, x):
+        x = self.cv1(x)
+        with warnings.catch_warnings():
+            warnings.simplefilter('ignore')  # suppress torch 1.9.0 max_pool2d() warning
+            y1 = self.m(x)
+            y2 = self.m(y1)
+            return self.cv2(torch.cat((x, y1, y2, self.m(y2)), 1))
+
+
+import torch.nn.functional as F
+from torch.nn.modules.utils import _pair
+    
+    
+class ReOrg(nn.Module):
+    # yolo
+    def __init__(self):
+        super(ReOrg, self).__init__()
+
+    def forward(self, x):  # x(b,c,w,h) -> y(b,4c,w/2,h/2)
+        return torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1)
+
+
+class Contract(nn.Module):
+    # Contract width-height into channels, i.e. x(1,64,80,80) to x(1,256,40,40)
+    def __init__(self, gain=2):
+        super().__init__()
+        self.gain = gain
+
+    def forward(self, x):
+        b, c, h, w = x.size()  # assert (h / s == 0) and (W / s == 0), 'Indivisible gain'
+        s = self.gain
+        x = x.view(b, c, h // s, s, w // s, s)  # x(1,64,40,2,40,2)
+        x = x.permute(0, 3, 5, 1, 2, 4).contiguous()  # x(1,2,2,64,40,40)
+        return x.view(b, c * s * s, h // s, w // s)  # x(1,256,40,40)
+
+
+class Expand(nn.Module):
+    # Expand channels into width-height, i.e. x(1,64,80,80) to x(1,16,160,160)
+    def __init__(self, gain=2):
+        super().__init__()
+        self.gain = gain
+
+    def forward(self, x):
+        b, c, h, w = x.size()  # assert C / s ** 2 == 0, 'Indivisible gain'
+        s = self.gain
+        x = x.view(b, s, s, c // s ** 2, h, w)  # x(1,2,2,16,80,80)
+        x = x.permute(0, 3, 4, 1, 5, 2).contiguous()  # x(1,16,80,2,80,2)
+        return x.view(b, c // s ** 2, h * s, w * s)  # x(1,16,160,160)
+
+
+class Concat(nn.Module):
+    # Concatenate a list of tensors along dimension
+    def __init__(self, dimension=1):
+        super().__init__()
+        self.d = dimension
+
+    def forward(self, x):
+        return torch.cat(x, self.d)
+
+
+class Shortcut(nn.Module):
+    def __init__(self, dimension=0):
+        super(Shortcut, self).__init__()
+        self.d = dimension
+
+    def forward(self, x):
+        return x[0]+x[1]
+    
+    
+class Silence(nn.Module):
+    def __init__(self):
+        super(Silence, self).__init__()
+    def forward(self, x):    
+        return x
+
+
+##### GELAN #####        
+        
+class SPPELAN(nn.Module):
+    # spp-elan
+    def __init__(self, c1, c2, c3):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        self.c = c3
+        self.cv1 = Conv(c1, c3, 1, 1)
+        self.cv2 = SP(5)
+        self.cv3 = SP(5)
+        self.cv4 = SP(5)
+        self.cv5 = Conv(4*c3, c2, 1, 1)
+
+    def forward(self, x):
+        y = [self.cv1(x)]
+        y.extend(m(y[-1]) for m in [self.cv2, self.cv3, self.cv4])
+        return self.cv5(torch.cat(y, 1))
+        
+        
+class ELAN1(nn.Module):
+
+    def __init__(self, c1, c2, c3, c4):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        self.c = c3//2
+        self.cv1 = Conv(c1, c3, 1, 1)
+        self.cv2 = Conv(c3//2, c4, 3, 1)
+        self.cv3 = Conv(c4, c4, 3, 1)
+        self.cv4 = Conv(c3+(2*c4), c2, 1, 1)
+
+    def forward(self, x):
+        y = list(self.cv1(x).chunk(2, 1))
+        y.extend(m(y[-1]) for m in [self.cv2, self.cv3])
+        return self.cv4(torch.cat(y, 1))
+
+    def forward_split(self, x):
+        y = list(self.cv1(x).split((self.c, self.c), 1))
+        y.extend(m(y[-1]) for m in [self.cv2, self.cv3])
+        return self.cv4(torch.cat(y, 1))
+        
+        
+class RepNCSPELAN4(nn.Module):
+    # csp-elan
+    def __init__(self, c1, c2, c3, c4, c5=1):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        self.c = c3//2
+        self.cv1 = Conv(c1, c3, 1, 1)
+        self.cv2 = nn.Sequential(RepNCSP(c3//2, c4, c5), Conv(c4, c4, 3, 1))
+        self.cv3 = nn.Sequential(RepNCSP(c4, c4, c5), Conv(c4, c4, 3, 1))
+        self.cv4 = Conv(c3+(2*c4), c2, 1, 1)
+
+    def forward(self, x):
+        y = list(self.cv1(x).chunk(2, 1))
+        y.extend((m(y[-1])) for m in [self.cv2, self.cv3])
+        return self.cv4(torch.cat(y, 1))
+
+    def forward_split(self, x):
+        y = list(self.cv1(x).split((self.c, self.c), 1))
+        y.extend(m(y[-1]) for m in [self.cv2, self.cv3])
+        return self.cv4(torch.cat(y, 1))
+
+#################
+
+
+##### YOLOR #####
+
+class ImplicitA(nn.Module):
+    def __init__(self, channel):
+        super(ImplicitA, self).__init__()
+        self.channel = channel
+        self.implicit = nn.Parameter(torch.zeros(1, channel, 1, 1))
+        nn.init.normal_(self.implicit, std=.02)        
+
+    def forward(self, x):
+        return self.implicit + x
+
+
+class ImplicitM(nn.Module):
+    def __init__(self, channel):
+        super(ImplicitM, self).__init__()
+        self.channel = channel
+        self.implicit = nn.Parameter(torch.ones(1, channel, 1, 1))
+        nn.init.normal_(self.implicit, mean=1., std=.02)        
+
+    def forward(self, x):
+        return self.implicit * x
+
+#################
+
+
+##### CBNet #####
+
+class CBLinear(nn.Module):
+    def __init__(self, c1, c2s, k=1, s=1, p=None, g=1):  # ch_in, ch_outs, kernel, stride, padding, groups
+        super(CBLinear, self).__init__()
+        self.c2s = c2s
+        self.conv = nn.Conv2d(c1, sum(c2s), k, s, autopad(k, p), groups=g, bias=True)
+
+    def forward(self, x):
+        outs = self.conv(x).split(self.c2s, dim=1)
+        return outs
+
+class CBFuse(nn.Module):
+    def __init__(self, idx):
+        super(CBFuse, self).__init__()
+        self.idx = idx
+
+    def forward(self, xs):
+        target_size = xs[-1].shape[2:]
+        res = [F.interpolate(x[self.idx[i]], size=target_size, mode='nearest') for i, x in enumerate(xs[:-1])]
+        out = torch.sum(torch.stack(res + xs[-1:]), dim=0)
+        return out
+
+#################
+
+
+class DetectMultiBackend(nn.Module):
+    # YOLO MultiBackend class for python inference on various backends
+    def __init__(self, weights='yolo.pt', device=torch.device('cpu'), dnn=False, data=None, fp16=False, fuse=True):
+        # Usage:
+        #   PyTorch:              weights = *.pt
+        #   TorchScript:                    *.torchscript
+        #   ONNX Runtime:                   *.onnx
+        #   ONNX OpenCV DNN:                *.onnx --dnn
+        #   OpenVINO:                       *_openvino_model
+        #   CoreML:                         *.mlmodel
+        #   TensorRT:                       *.engine
+        #   TensorFlow SavedModel:          *_saved_model
+        #   TensorFlow GraphDef:            *.pb
+        #   TensorFlow Lite:                *.tflite
+        #   TensorFlow Edge TPU:            *_edgetpu.tflite
+        #   PaddlePaddle:                   *_paddle_model
+        from models.experimental import attempt_download, attempt_load  # scoped to avoid circular import
+
+        super().__init__()
+        w = str(weights[0] if isinstance(weights, list) else weights)
+        pt, jit, onnx, onnx_end2end, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, paddle, triton = self._model_type(w)
+        fp16 &= pt or jit or onnx or engine  # FP16
+        nhwc = coreml or saved_model or pb or tflite or edgetpu  # BHWC formats (vs torch BCWH)
+        stride = 32  # default stride
+        cuda = torch.cuda.is_available() and device.type != 'cpu'  # use CUDA
+        if not (pt or triton):
+            w = attempt_download(w)  # download if not local
+
+        if pt:  # PyTorch
+            model = attempt_load(weights if isinstance(weights, list) else w, device=device, inplace=True, fuse=fuse)
+            stride = max(int(model.stride.max()), 32)  # model stride
+            names = model.module.names if hasattr(model, 'module') else model.names  # get class names
+            model.half() if fp16 else model.float()
+            self.model = model  # explicitly assign for to(), cpu(), cuda(), half()
+        elif jit:  # TorchScript
+            LOGGER.info(f'Loading {w} for TorchScript inference...')
+            extra_files = {'config.txt': ''}  # model metadata
+            model = torch.jit.load(w, _extra_files=extra_files, map_location=device)
+            model.half() if fp16 else model.float()
+            if extra_files['config.txt']:  # load metadata dict
+                d = json.loads(extra_files['config.txt'],
+                               object_hook=lambda d: {int(k) if k.isdigit() else k: v
+                                                      for k, v in d.items()})
+                stride, names = int(d['stride']), d['names']
+        elif dnn:  # ONNX OpenCV DNN
+            LOGGER.info(f'Loading {w} for ONNX OpenCV DNN inference...')
+            check_requirements('opencv-python>=4.5.4')
+            net = cv2.dnn.readNetFromONNX(w)
+        elif onnx:  # ONNX Runtime
+            LOGGER.info(f'Loading {w} for ONNX Runtime inference...')
+            check_requirements(('onnx', 'onnxruntime-gpu' if cuda else 'onnxruntime'))
+            import onnxruntime
+            providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] if cuda else ['CPUExecutionProvider']
+            session = onnxruntime.InferenceSession(w, providers=providers)
+            output_names = [x.name for x in session.get_outputs()]
+            meta = session.get_modelmeta().custom_metadata_map  # metadata
+            if 'stride' in meta:
+                stride, names = int(meta['stride']), eval(meta['names'])
+        elif xml:  # OpenVINO
+            LOGGER.info(f'Loading {w} for OpenVINO inference...')
+            check_requirements('openvino')  # requires openvino-dev: https://pypi.org/project/openvino-dev/
+            from openvino.runtime import Core, Layout, get_batch
+            ie = Core()
+            if not Path(w).is_file():  # if not *.xml
+                w = next(Path(w).glob('*.xml'))  # get *.xml file from *_openvino_model dir
+            network = ie.read_model(model=w, weights=Path(w).with_suffix('.bin'))
+            if network.get_parameters()[0].get_layout().empty:
+                network.get_parameters()[0].set_layout(Layout("NCHW"))
+            batch_dim = get_batch(network)
+            if batch_dim.is_static:
+                batch_size = batch_dim.get_length()
+            executable_network = ie.compile_model(network, device_name="CPU")  # device_name="MYRIAD" for Intel NCS2
+            stride, names = self._load_metadata(Path(w).with_suffix('.yaml'))  # load metadata
+        elif engine:  # TensorRT
+            LOGGER.info(f'Loading {w} for TensorRT inference...')
+            import tensorrt as trt  # https://developer.nvidia.com/nvidia-tensorrt-download
+            check_version(trt.__version__, '7.0.0', hard=True)  # require tensorrt>=7.0.0
+            if device.type == 'cpu':
+                device = torch.device('cuda:0')
+            Binding = namedtuple('Binding', ('name', 'dtype', 'shape', 'data', 'ptr'))
+            logger = trt.Logger(trt.Logger.INFO)
+            with open(w, 'rb') as f, trt.Runtime(logger) as runtime:
+                model = runtime.deserialize_cuda_engine(f.read())
+            context = model.create_execution_context()
+            bindings = OrderedDict()
+            output_names = []
+            fp16 = False  # default updated below
+            dynamic = False
+            for i in range(model.num_bindings):
+                name = model.get_binding_name(i)
+                dtype = trt.nptype(model.get_binding_dtype(i))
+                if model.binding_is_input(i):
+                    if -1 in tuple(model.get_binding_shape(i)):  # dynamic
+                        dynamic = True
+                        context.set_binding_shape(i, tuple(model.get_profile_shape(0, i)[2]))
+                    if dtype == np.float16:
+                        fp16 = True
+                else:  # output
+                    output_names.append(name)
+                shape = tuple(context.get_binding_shape(i))
+                im = torch.from_numpy(np.empty(shape, dtype=dtype)).to(device)
+                bindings[name] = Binding(name, dtype, shape, im, int(im.data_ptr()))
+            binding_addrs = OrderedDict((n, d.ptr) for n, d in bindings.items())
+            batch_size = bindings['images'].shape[0]  # if dynamic, this is instead max batch size
+        elif coreml:  # CoreML
+            LOGGER.info(f'Loading {w} for CoreML inference...')
+            import coremltools as ct
+            model = ct.models.MLModel(w)
+        elif saved_model:  # TF SavedModel
+            LOGGER.info(f'Loading {w} for TensorFlow SavedModel inference...')
+            import tensorflow as tf
+            keras = False  # assume TF1 saved_model
+            model = tf.keras.models.load_model(w) if keras else tf.saved_model.load(w)
+        elif pb:  # GraphDef https://www.tensorflow.org/guide/migrate#a_graphpb_or_graphpbtxt
+            LOGGER.info(f'Loading {w} for TensorFlow GraphDef inference...')
+            import tensorflow as tf
+
+            def wrap_frozen_graph(gd, inputs, outputs):
+                x = tf.compat.v1.wrap_function(lambda: tf.compat.v1.import_graph_def(gd, name=""), [])  # wrapped
+                ge = x.graph.as_graph_element
+                return x.prune(tf.nest.map_structure(ge, inputs), tf.nest.map_structure(ge, outputs))
+
+            def gd_outputs(gd):
+                name_list, input_list = [], []
+                for node in gd.node:  # tensorflow.core.framework.node_def_pb2.NodeDef
+                    name_list.append(node.name)
+                    input_list.extend(node.input)
+                return sorted(f'{x}:0' for x in list(set(name_list) - set(input_list)) if not x.startswith('NoOp'))
+
+            gd = tf.Graph().as_graph_def()  # TF GraphDef
+            with open(w, 'rb') as f:
+                gd.ParseFromString(f.read())
+            frozen_func = wrap_frozen_graph(gd, inputs="x:0", outputs=gd_outputs(gd))
+        elif tflite or edgetpu:  # https://www.tensorflow.org/lite/guide/python#install_tensorflow_lite_for_python
+            try:  # https://coral.ai/docs/edgetpu/tflite-python/#update-existing-tf-lite-code-for-the-edge-tpu
+                from tflite_runtime.interpreter import Interpreter, load_delegate
+            except ImportError:
+                import tensorflow as tf
+                Interpreter, load_delegate = tf.lite.Interpreter, tf.lite.experimental.load_delegate,
+            if edgetpu:  # TF Edge TPU https://coral.ai/software/#edgetpu-runtime
+                LOGGER.info(f'Loading {w} for TensorFlow Lite Edge TPU inference...')
+                delegate = {
+                    'Linux': 'libedgetpu.so.1',
+                    'Darwin': 'libedgetpu.1.dylib',
+                    'Windows': 'edgetpu.dll'}[platform.system()]
+                interpreter = Interpreter(model_path=w, experimental_delegates=[load_delegate(delegate)])
+            else:  # TFLite
+                LOGGER.info(f'Loading {w} for TensorFlow Lite inference...')
+                interpreter = Interpreter(model_path=w)  # load TFLite model
+            interpreter.allocate_tensors()  # allocate
+            input_details = interpreter.get_input_details()  # inputs
+            output_details = interpreter.get_output_details()  # outputs
+            # load metadata
+            with contextlib.suppress(zipfile.BadZipFile):
+                with zipfile.ZipFile(w, "r") as model:
+                    meta_file = model.namelist()[0]
+                    meta = ast.literal_eval(model.read(meta_file).decode("utf-8"))
+                    stride, names = int(meta['stride']), meta['names']
+        elif tfjs:  # TF.js
+            raise NotImplementedError('ERROR: YOLO TF.js inference is not supported')
+        elif paddle:  # PaddlePaddle
+            LOGGER.info(f'Loading {w} for PaddlePaddle inference...')
+            check_requirements('paddlepaddle-gpu' if cuda else 'paddlepaddle')
+            import paddle.inference as pdi
+            if not Path(w).is_file():  # if not *.pdmodel
+                w = next(Path(w).rglob('*.pdmodel'))  # get *.pdmodel file from *_paddle_model dir
+            weights = Path(w).with_suffix('.pdiparams')
+            config = pdi.Config(str(w), str(weights))
+            if cuda:
+                config.enable_use_gpu(memory_pool_init_size_mb=2048, device_id=0)
+            predictor = pdi.create_predictor(config)
+            input_handle = predictor.get_input_handle(predictor.get_input_names()[0])
+            output_names = predictor.get_output_names()
+        elif triton:  # NVIDIA Triton Inference Server
+            LOGGER.info(f'Using {w} as Triton Inference Server...')
+            check_requirements('tritonclient[all]')
+            from utils.triton import TritonRemoteModel
+            model = TritonRemoteModel(url=w)
+            nhwc = model.runtime.startswith("tensorflow")
+        else:
+            raise NotImplementedError(f'ERROR: {w} is not a supported format')
+
+        # class names
+        if 'names' not in locals():
+            names = yaml_load(data)['names'] if data else {i: f'class{i}' for i in range(999)}
+        if names[0] == 'n01440764' and len(names) == 1000:  # ImageNet
+            names = yaml_load(ROOT / 'data/ImageNet.yaml')['names']  # human-readable names
+
+        self.__dict__.update(locals())  # assign all variables to self
+
+    def forward(self, im, augment=False, visualize=False):
+        # YOLO MultiBackend inference
+        b, ch, h, w = im.shape  # batch, channel, height, width
+        if self.fp16 and im.dtype != torch.float16:
+            im = im.half()  # to FP16
+        if self.nhwc:
+            im = im.permute(0, 2, 3, 1)  # torch BCHW to numpy BHWC shape(1,320,192,3)
+
+        if self.pt:  # PyTorch
+            y = self.model(im, augment=augment, visualize=visualize) if augment or visualize else self.model(im)
+        elif self.jit:  # TorchScript
+            y = self.model(im)
+        elif self.dnn:  # ONNX OpenCV DNN
+            im = im.cpu().numpy()  # torch to numpy
+            self.net.setInput(im)
+            y = self.net.forward()
+        elif self.onnx:  # ONNX Runtime
+            im = im.cpu().numpy()  # torch to numpy
+            y = self.session.run(self.output_names, {self.session.get_inputs()[0].name: im})
+        elif self.xml:  # OpenVINO
+            im = im.cpu().numpy()  # FP32
+            y = list(self.executable_network([im]).values())
+        elif self.engine:  # TensorRT
+            if self.dynamic and im.shape != self.bindings['images'].shape:
+                i = self.model.get_binding_index('images')
+                self.context.set_binding_shape(i, im.shape)  # reshape if dynamic
+                self.bindings['images'] = self.bindings['images']._replace(shape=im.shape)
+                for name in self.output_names:
+                    i = self.model.get_binding_index(name)
+                    self.bindings[name].data.resize_(tuple(self.context.get_binding_shape(i)))
+            s = self.bindings['images'].shape
+            assert im.shape == s, f"input size {im.shape} {'>' if self.dynamic else 'not equal to'} max model size {s}"
+            self.binding_addrs['images'] = int(im.data_ptr())
+            self.context.execute_v2(list(self.binding_addrs.values()))
+            y = [self.bindings[x].data for x in sorted(self.output_names)]
+        elif self.coreml:  # CoreML
+            im = im.cpu().numpy()
+            im = Image.fromarray((im[0] * 255).astype('uint8'))
+            # im = im.resize((192, 320), Image.ANTIALIAS)
+            y = self.model.predict({'image': im})  # coordinates are xywh normalized
+            if 'confidence' in y:
+                box = xywh2xyxy(y['coordinates'] * [[w, h, w, h]])  # xyxy pixels
+                conf, cls = y['confidence'].max(1), y['confidence'].argmax(1).astype(np.float)
+                y = np.concatenate((box, conf.reshape(-1, 1), cls.reshape(-1, 1)), 1)
+            else:
+                y = list(reversed(y.values()))  # reversed for segmentation models (pred, proto)
+        elif self.paddle:  # PaddlePaddle
+            im = im.cpu().numpy().astype(np.float32)
+            self.input_handle.copy_from_cpu(im)
+            self.predictor.run()
+            y = [self.predictor.get_output_handle(x).copy_to_cpu() for x in self.output_names]
+        elif self.triton:  # NVIDIA Triton Inference Server
+            y = self.model(im)
+        else:  # TensorFlow (SavedModel, GraphDef, Lite, Edge TPU)
+            im = im.cpu().numpy()
+            if self.saved_model:  # SavedModel
+                y = self.model(im, training=False) if self.keras else self.model(im)
+            elif self.pb:  # GraphDef
+                y = self.frozen_func(x=self.tf.constant(im))
+            else:  # Lite or Edge TPU
+                input = self.input_details[0]
+                int8 = input['dtype'] == np.uint8  # is TFLite quantized uint8 model
+                if int8:
+                    scale, zero_point = input['quantization']
+                    im = (im / scale + zero_point).astype(np.uint8)  # de-scale
+                self.interpreter.set_tensor(input['index'], im)
+                self.interpreter.invoke()
+                y = []
+                for output in self.output_details:
+                    x = self.interpreter.get_tensor(output['index'])
+                    if int8:
+                        scale, zero_point = output['quantization']
+                        x = (x.astype(np.float32) - zero_point) * scale  # re-scale
+                    y.append(x)
+            y = [x if isinstance(x, np.ndarray) else x.numpy() for x in y]
+            y[0][..., :4] *= [w, h, w, h]  # xywh normalized to pixels
+
+        if isinstance(y, (list, tuple)):
+            return self.from_numpy(y[0]) if len(y) == 1 else [self.from_numpy(x) for x in y]
+        else:
+            return self.from_numpy(y)
+
+    def from_numpy(self, x):
+        return torch.from_numpy(x).to(self.device) if isinstance(x, np.ndarray) else x
+
+    def warmup(self, imgsz=(1, 3, 640, 640)):
+        # Warmup model by running inference once
+        warmup_types = self.pt, self.jit, self.onnx, self.engine, self.saved_model, self.pb, self.triton
+        if any(warmup_types) and (self.device.type != 'cpu' or self.triton):
+            im = torch.empty(*imgsz, dtype=torch.half if self.fp16 else torch.float, device=self.device)  # input
+            for _ in range(2 if self.jit else 1):  #
+                self.forward(im)  # warmup
+
+    @staticmethod
+    def _model_type(p='path/to/model.pt'):
+        # Return model type from model path, i.e. path='path/to/model.onnx' -> type=onnx
+        # types = [pt, jit, onnx, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, paddle]
+        from export import export_formats
+        from utils.downloads import is_url
+        sf = list(export_formats().Suffix)  # export suffixes
+        if not is_url(p, check=False):
+            check_suffix(p, sf)  # checks
+        url = urlparse(p)  # if url may be Triton inference server
+        types = [s in Path(p).name for s in sf]
+        types[8] &= not types[9]  # tflite &= not edgetpu
+        triton = not any(types) and all([any(s in url.scheme for s in ["http", "grpc"]), url.netloc])
+        return types + [triton]
+
+    @staticmethod
+    def _load_metadata(f=Path('path/to/meta.yaml')):
+        # Load metadata from meta.yaml if it exists
+        if f.exists():
+            d = yaml_load(f)
+            return d['stride'], d['names']  # assign stride, names
+        return None, None
+
+
+class AutoShape(nn.Module):
+    # YOLO input-robust model wrapper for passing cv2/np/PIL/torch inputs. Includes preprocessing, inference and NMS
+    conf = 0.25  # NMS confidence threshold
+    iou = 0.45  # NMS IoU threshold
+    agnostic = False  # NMS class-agnostic
+    multi_label = False  # NMS multiple labels per box
+    classes = None  # (optional list) filter by class, i.e. = [0, 15, 16] for COCO persons, cats and dogs
+    max_det = 1000  # maximum number of detections per image
+    amp = False  # Automatic Mixed Precision (AMP) inference
+
+    def __init__(self, model, verbose=True):
+        super().__init__()
+        if verbose:
+            LOGGER.info('Adding AutoShape... ')
+        copy_attr(self, model, include=('yaml', 'nc', 'hyp', 'names', 'stride', 'abc'), exclude=())  # copy attributes
+        self.dmb = isinstance(model, DetectMultiBackend)  # DetectMultiBackend() instance
+        self.pt = not self.dmb or model.pt  # PyTorch model
+        self.model = model.eval()
+        if self.pt:
+            m = self.model.model.model[-1] if self.dmb else self.model.model[-1]  # Detect()
+            m.inplace = False  # Detect.inplace=False for safe multithread inference
+            m.export = True  # do not output loss values
+
+    def _apply(self, fn):
+        # Apply to(), cpu(), cuda(), half() to model tensors that are not parameters or registered buffers
+        self = super()._apply(fn)
+        from models.yolo import Detect, Segment
+        if self.pt:
+            m = self.model.model.model[-1] if self.dmb else self.model.model[-1]  # Detect()
+            if isinstance(m, (Detect, Segment)):
+                for k in 'stride', 'anchor_grid', 'stride_grid', 'grid':
+                    x = getattr(m, k)
+                    setattr(m, k, list(map(fn, x))) if isinstance(x, (list, tuple)) else setattr(m, k, fn(x))
+        return self
+
+    @smart_inference_mode()
+    def forward(self, ims, size=640, augment=False, profile=False):
+        # Inference from various sources. For size(height=640, width=1280), RGB images example inputs are:
+        #   file:        ims = 'data/images/zidane.jpg'  # str or PosixPath
+        #   URI:             = 'https://ultralytics.com/images/zidane.jpg'
+        #   OpenCV:          = cv2.imread('image.jpg')[:,:,::-1]  # HWC BGR to RGB x(640,1280,3)
+        #   PIL:             = Image.open('image.jpg') or ImageGrab.grab()  # HWC x(640,1280,3)
+        #   numpy:           = np.zeros((640,1280,3))  # HWC
+        #   torch:           = torch.zeros(16,3,320,640)  # BCHW (scaled to size=640, 0-1 values)
+        #   multiple:        = [Image.open('image1.jpg'), Image.open('image2.jpg'), ...]  # list of images
+
+        dt = (Profile(), Profile(), Profile())
+        with dt[0]:
+            if isinstance(size, int):  # expand
+                size = (size, size)
+            p = next(self.model.parameters()) if self.pt else torch.empty(1, device=self.model.device)  # param
+            autocast = self.amp and (p.device.type != 'cpu')  # Automatic Mixed Precision (AMP) inference
+            if isinstance(ims, torch.Tensor):  # torch
+                with amp.autocast(autocast):
+                    return self.model(ims.to(p.device).type_as(p), augment=augment)  # inference
+
+            # Pre-process
+            n, ims = (len(ims), list(ims)) if isinstance(ims, (list, tuple)) else (1, [ims])  # number, list of images
+            shape0, shape1, files = [], [], []  # image and inference shapes, filenames
+            for i, im in enumerate(ims):
+                f = f'image{i}'  # filename
+                if isinstance(im, (str, Path)):  # filename or uri
+                    im, f = Image.open(requests.get(im, stream=True).raw if str(im).startswith('http') else im), im
+                    im = np.asarray(exif_transpose(im))
+                elif isinstance(im, Image.Image):  # PIL Image
+                    im, f = np.asarray(exif_transpose(im)), getattr(im, 'filename', f) or f
+                files.append(Path(f).with_suffix('.jpg').name)
+                if im.shape[0] < 5:  # image in CHW
+                    im = im.transpose((1, 2, 0))  # reverse dataloader .transpose(2, 0, 1)
+                im = im[..., :3] if im.ndim == 3 else cv2.cvtColor(im, cv2.COLOR_GRAY2BGR)  # enforce 3ch input
+                s = im.shape[:2]  # HWC
+                shape0.append(s)  # image shape
+                g = max(size) / max(s)  # gain
+                shape1.append([int(y * g) for y in s])
+                ims[i] = im if im.data.contiguous else np.ascontiguousarray(im)  # update
+            shape1 = [make_divisible(x, self.stride) for x in np.array(shape1).max(0)]  # inf shape
+            x = [letterbox(im, shape1, auto=False)[0] for im in ims]  # pad
+            x = np.ascontiguousarray(np.array(x).transpose((0, 3, 1, 2)))  # stack and BHWC to BCHW
+            x = torch.from_numpy(x).to(p.device).type_as(p) / 255  # uint8 to fp16/32
+
+        with amp.autocast(autocast):
+            # Inference
+            with dt[1]:
+                y = self.model(x, augment=augment)  # forward
+
+            # Post-process
+            with dt[2]:
+                y = non_max_suppression(y if self.dmb else y[0],
+                                        self.conf,
+                                        self.iou,
+                                        self.classes,
+                                        self.agnostic,
+                                        self.multi_label,
+                                        max_det=self.max_det)  # NMS
+                for i in range(n):
+                    scale_boxes(shape1, y[i][:, :4], shape0[i])
+
+            return Detections(ims, y, files, dt, self.names, x.shape)
+
+
+class Detections:
+    # YOLO detections class for inference results
+    def __init__(self, ims, pred, files, times=(0, 0, 0), names=None, shape=None):
+        super().__init__()
+        d = pred[0].device  # device
+        gn = [torch.tensor([*(im.shape[i] for i in [1, 0, 1, 0]), 1, 1], device=d) for im in ims]  # normalizations
+        self.ims = ims  # list of images as numpy arrays
+        self.pred = pred  # list of tensors pred[0] = (xyxy, conf, cls)
+        self.names = names  # class names
+        self.files = files  # image filenames
+        self.times = times  # profiling times
+        self.xyxy = pred  # xyxy pixels
+        self.xywh = [xyxy2xywh(x) for x in pred]  # xywh pixels
+        self.xyxyn = [x / g for x, g in zip(self.xyxy, gn)]  # xyxy normalized
+        self.xywhn = [x / g for x, g in zip(self.xywh, gn)]  # xywh normalized
+        self.n = len(self.pred)  # number of images (batch size)
+        self.t = tuple(x.t / self.n * 1E3 for x in times)  # timestamps (ms)
+        self.s = tuple(shape)  # inference BCHW shape
+
+    def _run(self, pprint=False, show=False, save=False, crop=False, render=False, labels=True, save_dir=Path('')):
+        s, crops = '', []
+        for i, (im, pred) in enumerate(zip(self.ims, self.pred)):
+            s += f'\nimage {i + 1}/{len(self.pred)}: {im.shape[0]}x{im.shape[1]} '  # string
+            if pred.shape[0]:
+                for c in pred[:, -1].unique():
+                    n = (pred[:, -1] == c).sum()  # detections per class
+                    s += f"{n} {self.names[int(c)]}{'s' * (n > 1)}, "  # add to string
+                s = s.rstrip(', ')
+                if show or save or render or crop:
+                    annotator = Annotator(im, example=str(self.names))
+                    for *box, conf, cls in reversed(pred):  # xyxy, confidence, class
+                        label = f'{self.names[int(cls)]} {conf:.2f}'
+                        if crop:
+                            file = save_dir / 'crops' / self.names[int(cls)] / self.files[i] if save else None
+                            crops.append({
+                                'box': box,
+                                'conf': conf,
+                                'cls': cls,
+                                'label': label,
+                                'im': save_one_box(box, im, file=file, save=save)})
+                        else:  # all others
+                            annotator.box_label(box, label if labels else '', color=colors(cls))
+                    im = annotator.im
+            else:
+                s += '(no detections)'
+
+            im = Image.fromarray(im.astype(np.uint8)) if isinstance(im, np.ndarray) else im  # from np
+            if show:
+                display(im) if is_notebook() else im.show(self.files[i])
+            if save:
+                f = self.files[i]
+                im.save(save_dir / f)  # save
+                if i == self.n - 1:
+                    LOGGER.info(f"Saved {self.n} image{'s' * (self.n > 1)} to {colorstr('bold', save_dir)}")
+            if render:
+                self.ims[i] = np.asarray(im)
+        if pprint:
+            s = s.lstrip('\n')
+            return f'{s}\nSpeed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {self.s}' % self.t
+        if crop:
+            if save:
+                LOGGER.info(f'Saved results to {save_dir}\n')
+            return crops
+
+    @TryExcept('Showing images is not supported in this environment')
+    def show(self, labels=True):
+        self._run(show=True, labels=labels)  # show results
+
+    def save(self, labels=True, save_dir='runs/detect/exp', exist_ok=False):
+        save_dir = increment_path(save_dir, exist_ok, mkdir=True)  # increment save_dir
+        self._run(save=True, labels=labels, save_dir=save_dir)  # save results
+
+    def crop(self, save=True, save_dir='runs/detect/exp', exist_ok=False):
+        save_dir = increment_path(save_dir, exist_ok, mkdir=True) if save else None
+        return self._run(crop=True, save=save, save_dir=save_dir)  # crop results
+
+    def render(self, labels=True):
+        self._run(render=True, labels=labels)  # render results
+        return self.ims
+
+    def pandas(self):
+        # return detections as pandas DataFrames, i.e. print(results.pandas().xyxy[0])
+        new = copy(self)  # return copy
+        ca = 'xmin', 'ymin', 'xmax', 'ymax', 'confidence', 'class', 'name'  # xyxy columns
+        cb = 'xcenter', 'ycenter', 'width', 'height', 'confidence', 'class', 'name'  # xywh columns
+        for k, c in zip(['xyxy', 'xyxyn', 'xywh', 'xywhn'], [ca, ca, cb, cb]):
+            a = [[x[:5] + [int(x[5]), self.names[int(x[5])]] for x in x.tolist()] for x in getattr(self, k)]  # update
+            setattr(new, k, [pd.DataFrame(x, columns=c) for x in a])
+        return new
+
+    def tolist(self):
+        # return a list of Detections objects, i.e. 'for result in results.tolist():'
+        r = range(self.n)  # iterable
+        x = [Detections([self.ims[i]], [self.pred[i]], [self.files[i]], self.times, self.names, self.s) for i in r]
+        # for d in x:
+        #    for k in ['ims', 'pred', 'xyxy', 'xyxyn', 'xywh', 'xywhn']:
+        #        setattr(d, k, getattr(d, k)[0])  # pop out of list
+        return x
+
+    def print(self):
+        LOGGER.info(self.__str__())
+
+    def __len__(self):  # override len(results)
+        return self.n
+
+    def __str__(self):  # override print(results)
+        return self._run(pprint=True)  # print results
+
+    def __repr__(self):
+        return f'YOLO {self.__class__} instance\n' + self.__str__()
+
+
+class Proto(nn.Module):
+    # YOLO mask Proto module for segmentation models
+    def __init__(self, c1, c_=256, c2=32):  # ch_in, number of protos, number of masks
+        super().__init__()
+        self.cv1 = Conv(c1, c_, k=3)
+        self.upsample = nn.Upsample(scale_factor=2, mode='nearest')
+        self.cv2 = Conv(c_, c_, k=3)
+        self.cv3 = Conv(c_, c2)
+
+    def forward(self, x):
+        return self.cv3(self.cv2(self.upsample(self.cv1(x))))
+
+
+class UConv(nn.Module):
+    def __init__(self, c1, c_=256, c2=256):  # ch_in, number of protos, number of masks
+        super().__init__()
+        
+        self.cv1 = Conv(c1, c_, k=3)
+        self.cv2 = nn.Conv2d(c_, c2, 1, 1)
+        self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
+
+    def forward(self, x):
+        return self.up(self.cv2(self.cv1(x)))
+
+
+class Classify(nn.Module):
+    # YOLO classification head, i.e. x(b,c1,20,20) to x(b,c2)
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1):  # ch_in, ch_out, kernel, stride, padding, groups
+        super().__init__()
+        c_ = 1280  # efficientnet_b0 size
+        self.conv = Conv(c1, c_, k, s, autopad(k, p), g)
+        self.pool = nn.AdaptiveAvgPool2d(1)  # to x(b,c_,1,1)
+        self.drop = nn.Dropout(p=0.0, inplace=True)
+        self.linear = nn.Linear(c_, c2)  # to x(b,c2)
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, 1)
+        return self.linear(self.drop(self.pool(self.conv(x)).flatten(1)))

yolov9/models/detect/gelan-c.yaml

@@ -0,0 +1,80 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 0-P1/2
+
+   # conv down
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 2
+
+   # avg-conv down
+   [-1, 1, ADown, [256]],  # 3-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 4
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 5-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 6
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 7-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 8
+  ]
+
+# gelan head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [512, 256]],  # 9
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 12
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 15 (P3/8-small)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [256]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 18 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [512]],
+   [[-1, 9], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 21 (P5/32-large)
+
+   # detect
+   [[15, 18, 21], 1, DDetect, [nc]],  # DDetect(P3, P4, P5)
+  ]

yolov9/models/detect/gelan-e.yaml

@@ -0,0 +1,121 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   [-1, 1, Silence, []],
+  
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
+
+   # conv down
+   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 2]],  # 3
+
+   # avg-conv down
+   [-1, 1, ADown, [256]],  # 4-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 2]],  # 5
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 6-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [1024, 512, 256, 2]],  # 7
+
+   # avg-conv down
+   [-1, 1, ADown, [1024]],  # 8-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [1024, 512, 256, 2]],  # 9
+   
+   # routing
+   [1, 1, CBLinear, [[64]]], # 10
+   [3, 1, CBLinear, [[64, 128]]], # 11
+   [5, 1, CBLinear, [[64, 128, 256]]], # 12
+   [7, 1, CBLinear, [[64, 128, 256, 512]]], # 13
+   [9, 1, CBLinear, [[64, 128, 256, 512, 1024]]], # 14
+  
+   # conv down fuse
+   [0, 1, Conv, [64, 3, 2]],  # 15-P1/2
+   [[10, 11, 12, 13, 14, -1], 1, CBFuse, [[0, 0, 0, 0, 0]]], # 16
+
+   # conv down fuse
+   [-1, 1, Conv, [128, 3, 2]],  # 17-P2/4
+   [[11, 12, 13, 14, -1], 1, CBFuse, [[1, 1, 1, 1]]], # 18  
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 2]],  # 19
+
+   # avg-conv down fuse
+   [-1, 1, ADown, [256]],  # 20-P3/8
+   [[12, 13, 14, -1], 1, CBFuse, [[2, 2, 2]]], # 21  
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 2]],  # 22
+
+   # avg-conv down fuse
+   [-1, 1, ADown, [512]],  # 23-P4/16
+   [[13, 14, -1], 1, CBFuse, [[3, 3]]], # 24 
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [1024, 512, 256, 2]],  # 25
+
+   # avg-conv down fuse
+   [-1, 1, ADown, [1024]],  # 26-P5/32
+   [[14, -1], 1, CBFuse, [[4]]], # 27
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [1024, 512, 256, 2]],  # 28
+  ]
+
+# gelan head
+head:
+  [
+   # elan-spp block
+   [28, 1, SPPELAN, [512, 256]],  # 29
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 25], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 2]],  # 32
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 22], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 2]],  # 35 (P3/8-small)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [256]],
+   [[-1, 32], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 2]],  # 38 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [512]],
+   [[-1, 29], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 1024, 512, 2]],  # 41 (P5/32-large)
+
+   # detect
+   [[35, 38, 41], 1, DDetect, [nc]],  # Detect(P3, P4, P5)
+  ]

yolov9/models/detect/gelan-m.yaml

@@ -0,0 +1,80 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   # conv down
+   [-1, 1, Conv, [32, 3, 2]],  # 0-P1/2
+
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 1-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [128, 128, 64, 1]],  # 2
+
+   # avg-conv down
+   [-1, 1, AConv, [240]],  # 3-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [240, 240, 120, 1]],  # 4
+
+   # avg-conv down
+   [-1, 1, AConv, [360]],  # 5-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [360, 360, 180, 1]],  # 6
+
+   # avg-conv down
+   [-1, 1, AConv, [480]],  # 7-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [480, 480, 240, 1]],  # 8
+  ]
+
+# elan head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [480, 240]],  # 9
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [360, 360, 180, 1]],  # 12
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [240, 240, 120, 1]],  # 15
+
+   # avg-conv-down merge
+   [-1, 1, AConv, [180]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [360, 360, 180, 1]],  # 18 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, AConv, [240]],
+   [[-1, 9], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [480, 480, 240, 1]],  # 21 (P5/32-large)
+
+   # detect
+   [[15, 18, 21], 1, DDetect, [nc]],  # DDetect(P3, P4, P5)
+  ]

yolov9/models/detect/gelan-s.yaml

@@ -0,0 +1,80 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   # conv down
+   [-1, 1, Conv, [32, 3, 2]],  # 0-P1/2
+
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 1-P2/4
+
+   # elan-1 block
+   [-1, 1, ELAN1, [64, 64, 32]],  # 2
+
+   # avg-conv down
+   [-1, 1, AConv, [128]],  # 3-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [128, 128, 64, 3]],  # 4
+
+   # avg-conv down
+   [-1, 1, AConv, [192]],  # 5-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [192, 192, 96, 3]],  # 6
+
+   # avg-conv down
+   [-1, 1, AConv, [256]],  # 7-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 3]],  # 8
+  ]
+
+# elan head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [256, 128]],  # 9
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [192, 192, 96, 3]],  # 12
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [128, 128, 64, 3]],  # 15
+
+   # avg-conv-down merge
+   [-1, 1, AConv, [96]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [192, 192, 96, 3]],  # 18 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, AConv, [128]],
+   [[-1, 9], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 3]],  # 21 (P5/32-large)
+
+   # detect
+   [[15, 18, 21], 1, DDetect, [nc]],  # DDetect(P3, P4, P5)
+  ]

yolov9/models/detect/gelan-t.yaml

@@ -0,0 +1,80 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   # conv down
+   [-1, 1, Conv, [16, 3, 2]],  # 0-P1/2
+
+   # conv down
+   [-1, 1, Conv, [32, 3, 2]],  # 1-P2/4
+
+   # elan-1 block
+   [-1, 1, ELAN1, [32, 32, 16]],  # 2
+
+   # avg-conv down
+   [-1, 1, AConv, [64]],  # 3-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [64, 64, 32, 3]],  # 4
+
+   # avg-conv down
+   [-1, 1, AConv, [96]],  # 5-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [96, 96, 48, 3]],  # 6
+
+   # avg-conv down
+   [-1, 1, AConv, [128]],  # 7-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [128, 128, 64, 3]],  # 8
+  ]
+
+# elan head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [128, 64]],  # 9
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [96, 96, 48, 3]],  # 12
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [64, 64, 32, 3]],  # 15
+
+   # avg-conv-down merge
+   [-1, 1, AConv, [48]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [96, 96, 48, 3]],  # 18 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, AConv, [64]],
+   [[-1, 9], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [128, 128, 64, 3]],  # 21 (P5/32-large)
+
+   # detect
+   [[15, 18, 21], 1, DDetect, [nc]],  # DDetect(P3, P4, P5)
+  ]

yolov9/models/detect/gelan.yaml

@@ -0,0 +1,80 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 0-P1/2
+
+   # conv down
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 2
+
+   # avg-conv down
+   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 4
+
+   # avg-conv down
+   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 6
+
+   # avg-conv down
+   [-1, 1, Conv, [512, 3, 2]],  # 7-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 8
+  ]
+
+# gelan head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [512, 256]],  # 9
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 12
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 15 (P3/8-small)
+
+   # avg-conv-down merge
+   [-1, 1, Conv, [256, 3, 2]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 18 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, Conv, [512, 3, 2]],
+   [[-1, 9], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 21 (P5/32-large)
+
+   # detect
+   [[15, 18, 21], 1, DDetect, [nc]],  # Detect(P3, P4, P5)
+  ]

yolov9/models/detect/yolov7-af.yaml

@@ -0,0 +1,137 @@
+# YOLOv7
+
+# Parameters
+nc: 80  # number of classes
+depth_multiple: 1.  # model depth multiple
+width_multiple: 1.  # layer channel multiple
+anchors: 3
+
+# YOLOv7 backbone
+backbone:
+  # [from, number, module, args]
+  [[-1, 1, Conv, [32, 3, 1]],  # 0
+  
+   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2      
+   [-1, 1, Conv, [64, 3, 1]],
+   
+   [-1, 1, Conv, [128, 3, 2]],  # 3-P2/4  
+   [-1, 1, Conv, [64, 1, 1]],
+   [-2, 1, Conv, [64, 1, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [[-1, -3, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [256, 1, 1]],  # 11
+         
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [128, 1, 1]],
+   [-3, 1, Conv, [128, 1, 1]],
+   [-1, 1, Conv, [128, 3, 2]],
+   [[-1, -3], 1, Concat, [1]],  # 16-P3/8  
+   [-1, 1, Conv, [128, 1, 1]],
+   [-2, 1, Conv, [128, 1, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [[-1, -3, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [512, 1, 1]],  # 24
+         
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [256, 1, 1]],
+   [-3, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [256, 3, 2]],
+   [[-1, -3], 1, Concat, [1]],  # 29-P4/16  
+   [-1, 1, Conv, [256, 1, 1]],
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [[-1, -3, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [1024, 1, 1]],  # 37
+         
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [512, 1, 1]],
+   [-3, 1, Conv, [512, 1, 1]],
+   [-1, 1, Conv, [512, 3, 2]],
+   [[-1, -3], 1, Concat, [1]],  # 42-P5/32  
+   [-1, 1, Conv, [256, 1, 1]],
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [[-1, -3, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [1024, 1, 1]],  # 50
+  ]
+
+# yolov7 head
+head:
+  [[-1, 1, SPPCSPC, [512]], # 51
+  
+   [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [37, 1, Conv, [256, 1, 1]], # route backbone P4
+   [[-1, -2], 1, Concat, [1]],
+   
+   [-1, 1, Conv, [256, 1, 1]],
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [256, 1, 1]], # 63
+   
+   [-1, 1, Conv, [128, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [24, 1, Conv, [128, 1, 1]], # route backbone P3
+   [[-1, -2], 1, Concat, [1]],
+   
+   [-1, 1, Conv, [128, 1, 1]],
+   [-2, 1, Conv, [128, 1, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [128, 1, 1]], # 75
+      
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [128, 1, 1]],
+   [-3, 1, Conv, [128, 1, 1]],
+   [-1, 1, Conv, [128, 3, 2]],
+   [[-1, -3, 63], 1, Concat, [1]],
+   
+   [-1, 1, Conv, [256, 1, 1]],
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [256, 1, 1]], # 88
+      
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [256, 1, 1]],
+   [-3, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [256, 3, 2]],
+   [[-1, -3, 51], 1, Concat, [1]],
+   
+   [-1, 1, Conv, [512, 1, 1]],
+   [-2, 1, Conv, [512, 1, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [512, 1, 1]], # 101
+   
+   [75, 1, Conv, [256, 3, 1]],
+   [88, 1, Conv, [512, 3, 1]],
+   [101, 1, Conv, [1024, 3, 1]],
+
+   [[102, 103, 104], 1, Detect, [nc]],  # Detect(P3, P4, P5)
+  ]

yolov9/models/detect/yolov9-cf.yaml

@@ -0,0 +1,124 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# YOLOv9 backbone
+backbone:
+  [
+   [-1, 1, Silence, []],  
+   
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
+
+   # conv down
+   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 3
+
+   # avg-conv down
+   [-1, 1, ADown, [256]],  # 4-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 5
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 6-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 7
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 8-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 9
+  ]
+
+# YOLOv9 head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [512, 256]],  # 10
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 7], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 13
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 5], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 16 (P3/8-small)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [256]],
+   [[-1, 13], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 19 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [512]],
+   [[-1, 10], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 22 (P5/32-large)
+   
+   
+   # multi-level reversible auxiliary branch
+   
+   # routing
+   [5, 1, CBLinear, [[256]]], # 23
+   [7, 1, CBLinear, [[256, 512]]], # 24
+   [9, 1, CBLinear, [[256, 512, 512]]], # 25
+   
+   # conv down
+   [0, 1, Conv, [64, 3, 2]],  # 26-P1/2
+
+   # conv down
+   [-1, 1, Conv, [128, 3, 2]],  # 27-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 28
+
+   # avg-conv down fuse
+   [-1, 1, ADown, [256]],  # 29-P3/8
+   [[23, 24, 25, -1], 1, CBFuse, [[0, 0, 0]]], # 30  
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 31
+
+   # avg-conv down fuse
+   [-1, 1, ADown, [512]],  # 32-P4/16
+   [[24, 25, -1], 1, CBFuse, [[1, 1]]], # 33 
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 34
+
+   # avg-conv down fuse
+   [-1, 1, ADown, [512]],  # 35-P5/32
+   [[25, -1], 1, CBFuse, [[2]]], # 36
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 37
+   
+   
+   
+   # detection head
+
+   # detect
+   [[31, 34, 37, 16, 19, 22, 16, 19, 22], 1, TripleDDetect, [nc]],  # TripleDDetect(A3, A4, A5, P3, P4, P5, P3, P4, P5) Auxiliary/Coarse(NMS-based)/Fine(NMS-free)
+  ]

yolov9/models/detect/yolov9-m.yaml

@@ -0,0 +1,117 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   [-1, 1, Silence, []],  
+   
+   # conv down
+   [-1, 1, Conv, [32, 3, 2]],  # 1-P1/2
+
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 2-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [128, 128, 64, 1]],  # 3
+
+   # avg-conv down
+   [-1, 1, AConv, [240]],  # 4-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [240, 240, 120, 1]],  # 5
+
+   # avg-conv down
+   [-1, 1, AConv, [360]],  # 6-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [360, 360, 180, 1]],  # 7
+
+   # avg-conv down
+   [-1, 1, AConv, [480]],  # 8-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [480, 480, 240, 1]],  # 9
+  ]
+
+# elan head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [480, 240]],  # 10
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 7], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [360, 360, 180, 1]],  # 13
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 5], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [240, 240, 120, 1]],  # 16
+
+   # avg-conv-down merge
+   [-1, 1, AConv, [180]],
+   [[-1, 13], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [360, 360, 180, 1]],  # 19 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, AConv, [240]],
+   [[-1, 10], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [480, 480, 240, 1]],  # 22 (P5/32-large)
+   
+   # routing
+   [5, 1, CBLinear, [[240]]], # 23
+   [7, 1, CBLinear, [[240, 360]]], # 24
+   [9, 1, CBLinear, [[240, 360, 480]]], # 25
+   
+   # conv down
+   [0, 1, Conv, [32, 3, 2]],  # 26-P1/2
+
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 27-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [128, 128, 64, 1]],  # 28
+
+   # avg-conv down
+   [-1, 1, AConv, [240]],   # 29-P3/8
+   [[23, 24, 25, -1], 1, CBFuse, [[0, 0, 0]]], # 30  
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [240, 240, 120, 1]],  # 31
+
+   # avg-conv down
+   [-1, 1, AConv, [360]],   # 32-P4/16
+   [[24, 25, -1], 1, CBFuse, [[1, 1]]], # 33 
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [360, 360, 180, 1]],  # 34
+
+   # avg-conv down
+   [-1, 1, AConv, [480]],   # 35-P5/32
+   [[25, -1], 1, CBFuse, [[2]]], # 36
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [480, 480, 240, 1]],  # 37
+
+   # detect
+   [[31, 34, 37, 16, 19, 22], 1, DualDDetect, [nc]],  # Detect(P3, P4, P5)
+  ]

yolov9/models/detect/yolov9-s.yaml

@@ -0,0 +1,97 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   # conv down
+   [-1, 1, Conv, [32, 3, 2]],  # 0-P1/2
+
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 1-P2/4
+
+   # elan-1 block
+   [-1, 1, ELAN1, [64, 64, 32]],  # 2
+
+   # avg-conv down
+   [-1, 1, AConv, [128]],  # 3-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [128, 128, 64, 3]],  # 4
+
+   # avg-conv down
+   [-1, 1, AConv, [192]],  # 5-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [192, 192, 96, 3]],  # 6
+
+   # avg-conv down
+   [-1, 1, AConv, [256]],  # 7-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 3]],  # 8
+  ]
+
+# elan head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [256, 128]],  # 9
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [192, 192, 96, 3]],  # 12
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [128, 128, 64, 3]],  # 15
+
+   # avg-conv-down merge
+   [-1, 1, AConv, [96]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [192, 192, 96, 3]],  # 18 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, AConv, [128]],
+   [[-1, 9], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 3]],  # 21 (P5/32-large)
+   
+   # elan-spp block
+   [8, 1, SPPELAN, [256, 128]],  # 22
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [192, 192, 96, 3]],  # 25
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [128, 128, 64, 3]],  # 28
+
+   # detect
+   [[28, 25, 22, 15, 18, 21], 1, DualDDetect, [nc]],  # Detect(P3, P4, P5)
+  ]

yolov9/models/detect/yolov9-t.yaml

@@ -0,0 +1,97 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   # conv down
+   [-1, 1, Conv, [16, 3, 2]],  # 0-P1/2
+
+   # conv down
+   [-1, 1, Conv, [32, 3, 2]],  # 1-P2/4
+
+   # elan-1 block
+   [-1, 1, ELAN1, [32, 32, 16]],  # 2
+
+   # avg-conv down
+   [-1, 1, AConv, [64]],  # 3-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [64, 64, 32, 3]],  # 4
+
+   # avg-conv down
+   [-1, 1, AConv, [96]],  # 5-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [96, 96, 48, 3]],  # 6
+
+   # avg-conv down
+   [-1, 1, AConv, [128]],  # 7-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [128, 128, 64, 3]],  # 8
+  ]
+
+# elan head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [128, 64]],  # 9
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [96, 96, 48, 3]],  # 12
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [64, 64, 32, 3]],  # 15
+
+   # avg-conv-down merge
+   [-1, 1, AConv, [48]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [96, 96, 48, 3]],  # 18 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, AConv, [64]],
+   [[-1, 9], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [128, 128, 64, 3]],  # 21 (P5/32-large)
+   
+   # elan-spp block
+   [8, 1, SPPELAN, [128, 64]],  # 22
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [96, 96, 48, 3]],  # 25
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [64, 64, 32, 3]],  # 28
+
+   # detect
+   [[28, 25, 22, 15, 18, 21], 1, DualDDetect, [nc]],  # Detect(P3, P4, P5)
+  ]

yolov9/models/detect/yolov9.yaml

@@ -0,0 +1,117 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# YOLOv9 backbone
+backbone:
+  [
+   [-1, 1, Silence, []],  
+   
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
+
+   # conv down
+   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 3
+
+   # conv down
+   [-1, 1, Conv, [256, 3, 2]],  # 4-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 5
+
+   # conv down
+   [-1, 1, Conv, [512, 3, 2]],  # 6-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 7
+
+   # conv down
+   [-1, 1, Conv, [512, 3, 2]],  # 8-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 9
+  ]
+
+# YOLOv9 head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [512, 256]],  # 10
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 7], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 13
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 5], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 16 (P3/8-small)
+
+   # conv-down merge
+   [-1, 1, Conv, [256, 3, 2]],
+   [[-1, 13], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 19 (P4/16-medium)
+
+   # conv-down merge
+   [-1, 1, Conv, [512, 3, 2]],
+   [[-1, 10], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 22 (P5/32-large)
+   
+   # routing
+   [5, 1, CBLinear, [[256]]], # 23
+   [7, 1, CBLinear, [[256, 512]]], # 24
+   [9, 1, CBLinear, [[256, 512, 512]]], # 25
+   
+   # conv down
+   [0, 1, Conv, [64, 3, 2]],  # 26-P1/2
+
+   # conv down
+   [-1, 1, Conv, [128, 3, 2]],  # 27-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 28
+
+   # conv down fuse
+   [-1, 1, Conv, [256, 3, 2]],  # 29-P3/8
+   [[23, 24, 25, -1], 1, CBFuse, [[0, 0, 0]]], # 30  
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 31
+
+   # conv down fuse
+   [-1, 1, Conv, [512, 3, 2]],  # 32-P4/16
+   [[24, 25, -1], 1, CBFuse, [[1, 1]]], # 33 
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 34
+
+   # conv down fuse
+   [-1, 1, Conv, [512, 3, 2]],  # 35-P5/32
+   [[25, -1], 1, CBFuse, [[2]]], # 36
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 37
+
+   # detect
+   [[31, 34, 37, 16, 19, 22], 1, DualDDetect, [nc]],  # DualDDetect(A3, A4, A5, P3, P4, P5)
+  ]

yolov9/models/experimental.py

@@ -0,0 +1,275 @@
+import math
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+from utils.downloads import attempt_download
+
+
+class Sum(nn.Module):
+    # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070
+    def __init__(self, n, weight=False):  # n: number of inputs
+        super().__init__()
+        self.weight = weight  # apply weights boolean
+        self.iter = range(n - 1)  # iter object
+        if weight:
+            self.w = nn.Parameter(-torch.arange(1.0, n) / 2, requires_grad=True)  # layer weights
+
+    def forward(self, x):
+        y = x[0]  # no weight
+        if self.weight:
+            w = torch.sigmoid(self.w) * 2
+            for i in self.iter:
+                y = y + x[i + 1] * w[i]
+        else:
+            for i in self.iter:
+                y = y + x[i + 1]
+        return y
+
+
+class MixConv2d(nn.Module):
+    # Mixed Depth-wise Conv https://arxiv.org/abs/1907.09595
+    def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True):  # ch_in, ch_out, kernel, stride, ch_strategy
+        super().__init__()
+        n = len(k)  # number of convolutions
+        if equal_ch:  # equal c_ per group
+            i = torch.linspace(0, n - 1E-6, c2).floor()  # c2 indices
+            c_ = [(i == g).sum() for g in range(n)]  # intermediate channels
+        else:  # equal weight.numel() per group
+            b = [c2] + [0] * n
+            a = np.eye(n + 1, n, k=-1)
+            a -= np.roll(a, 1, axis=1)
+            a *= np.array(k) ** 2
+            a[0] = 1
+            c_ = np.linalg.lstsq(a, b, rcond=None)[0].round()  # solve for equal weight indices, ax = b
+
+        self.m = nn.ModuleList([
+            nn.Conv2d(c1, int(c_), k, s, k // 2, groups=math.gcd(c1, int(c_)), bias=False) for k, c_ in zip(k, c_)])
+        self.bn = nn.BatchNorm2d(c2)
+        self.act = nn.SiLU()
+
+    def forward(self, x):
+        return self.act(self.bn(torch.cat([m(x) for m in self.m], 1)))
+
+
+class Ensemble(nn.ModuleList):
+    # Ensemble of models
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x, augment=False, profile=False, visualize=False):
+        y = [module(x, augment, profile, visualize)[0] for module in self]
+        # y = torch.stack(y).max(0)[0]  # max ensemble
+        # y = torch.stack(y).mean(0)  # mean ensemble
+        y = torch.cat(y, 1)  # nms ensemble
+        return y, None  # inference, train output
+
+
+class ORT_NMS(torch.autograd.Function):
+    '''ONNX-Runtime NMS operation'''
+    @staticmethod
+    def forward(ctx,
+                boxes,
+                scores,
+                max_output_boxes_per_class=torch.tensor([100]),
+                iou_threshold=torch.tensor([0.45]),
+                score_threshold=torch.tensor([0.25])):
+        device = boxes.device
+        batch = scores.shape[0]
+        num_det = random.randint(0, 100)
+        batches = torch.randint(0, batch, (num_det,)).sort()[0].to(device)
+        idxs = torch.arange(100, 100 + num_det).to(device)
+        zeros = torch.zeros((num_det,), dtype=torch.int64).to(device)
+        selected_indices = torch.cat([batches[None], zeros[None], idxs[None]], 0).T.contiguous()
+        selected_indices = selected_indices.to(torch.int64)
+        return selected_indices
+
+    @staticmethod
+    def symbolic(g, boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold):
+        return g.op("NonMaxSuppression", boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold)
+
+
+class TRT_NMS(torch.autograd.Function):
+    '''TensorRT NMS operation'''
+    @staticmethod
+    def forward(
+        ctx,
+        boxes,
+        scores,
+        background_class=-1,
+        box_coding=1,
+        iou_threshold=0.45,
+        max_output_boxes=100,
+        plugin_version="1",
+        score_activation=0,
+        score_threshold=0.25,
+    ):
+
+        batch_size, num_boxes, num_classes = scores.shape
+        num_det = torch.randint(0, max_output_boxes, (batch_size, 1), dtype=torch.int32)
+        det_boxes = torch.randn(batch_size, max_output_boxes, 4)
+        det_scores = torch.randn(batch_size, max_output_boxes)
+        det_classes = torch.randint(0, num_classes, (batch_size, max_output_boxes), dtype=torch.int32)
+        return num_det, det_boxes, det_scores, det_classes
+
+    @staticmethod
+    def symbolic(g,
+                 boxes,
+                 scores,
+                 background_class=-1,
+                 box_coding=1,
+                 iou_threshold=0.45,
+                 max_output_boxes=100,
+                 plugin_version="1",
+                 score_activation=0,
+                 score_threshold=0.25):
+        out = g.op("TRT::EfficientNMS_TRT",
+                   boxes,
+                   scores,
+                   background_class_i=background_class,
+                   box_coding_i=box_coding,
+                   iou_threshold_f=iou_threshold,
+                   max_output_boxes_i=max_output_boxes,
+                   plugin_version_s=plugin_version,
+                   score_activation_i=score_activation,
+                   score_threshold_f=score_threshold,
+                   outputs=4)
+        nums, boxes, scores, classes = out
+        return nums, boxes, scores, classes
+
+
+class ONNX_ORT(nn.Module):
+    '''onnx module with ONNX-Runtime NMS operation.'''
+    def __init__(self, max_obj=100, iou_thres=0.45, score_thres=0.25, max_wh=640, device=None, n_classes=80):
+        super().__init__()
+        self.device = device if device else torch.device("cpu")
+        self.max_obj = torch.tensor([max_obj]).to(device)
+        self.iou_threshold = torch.tensor([iou_thres]).to(device)
+        self.score_threshold = torch.tensor([score_thres]).to(device)
+        self.max_wh = max_wh # if max_wh != 0 : non-agnostic else : agnostic
+        self.convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]],
+                                           dtype=torch.float32,
+                                           device=self.device)
+        self.n_classes=n_classes
+
+    def forward(self, x):
+        ## https://github.com/thaitc-hust/yolov9-tensorrt/blob/main/torch2onnx.py
+        ## thanks https://github.com/thaitc-hust
+        if isinstance(x, list):  ## yolov9-c.pt and yolov9-e.pt return list
+            x = x[1]
+        x = x.permute(0, 2, 1)
+        bboxes_x = x[..., 0:1]
+        bboxes_y = x[..., 1:2]
+        bboxes_w = x[..., 2:3]
+        bboxes_h = x[..., 3:4]
+        bboxes = torch.cat([bboxes_x, bboxes_y, bboxes_w, bboxes_h], dim = -1)
+        bboxes = bboxes.unsqueeze(2) # [n_batch, n_bboxes, 4] -> [n_batch, n_bboxes, 1, 4]
+        obj_conf = x[..., 4:]
+        scores = obj_conf
+        bboxes @= self.convert_matrix
+        max_score, category_id = scores.max(2, keepdim=True)
+        dis = category_id.float() * self.max_wh
+        nmsbox = bboxes + dis
+        max_score_tp = max_score.transpose(1, 2).contiguous()
+        selected_indices = ORT_NMS.apply(nmsbox, max_score_tp, self.max_obj, self.iou_threshold, self.score_threshold)
+        X, Y = selected_indices[:, 0], selected_indices[:, 2]
+        selected_boxes = bboxes[X, Y, :]
+        selected_categories = category_id[X, Y, :].float()
+        selected_scores = max_score[X, Y, :]
+        X = X.unsqueeze(1).float()
+        return torch.cat([X, selected_boxes, selected_categories, selected_scores], 1)
+
+
+class ONNX_TRT(nn.Module):
+    '''onnx module with TensorRT NMS operation.'''
+    def __init__(self, max_obj=100, iou_thres=0.45, score_thres=0.25, max_wh=None ,device=None, n_classes=80):
+        super().__init__()
+        assert max_wh is None
+        self.device = device if device else torch.device('cpu')
+        self.background_class = -1,
+        self.box_coding = 1,
+        self.iou_threshold = iou_thres
+        self.max_obj = max_obj
+        self.plugin_version = '1'
+        self.score_activation = 0
+        self.score_threshold = score_thres
+        self.n_classes=n_classes
+
+    def forward(self, x):
+        ## https://github.com/thaitc-hust/yolov9-tensorrt/blob/main/torch2onnx.py
+        ## thanks https://github.com/thaitc-hust
+        if isinstance(x, list):  ## yolov9-c.pt and yolov9-e.pt return list
+            x = x[1]
+        x = x.permute(0, 2, 1)
+        bboxes_x = x[..., 0:1]
+        bboxes_y = x[..., 1:2]
+        bboxes_w = x[..., 2:3]
+        bboxes_h = x[..., 3:4]
+        bboxes = torch.cat([bboxes_x, bboxes_y, bboxes_w, bboxes_h], dim = -1)
+        bboxes = bboxes.unsqueeze(2) # [n_batch, n_bboxes, 4] -> [n_batch, n_bboxes, 1, 4]
+        obj_conf = x[..., 4:]
+        scores = obj_conf
+        num_det, det_boxes, det_scores, det_classes = TRT_NMS.apply(bboxes, scores, self.background_class, self.box_coding,
+                                                                    self.iou_threshold, self.max_obj,
+                                                                    self.plugin_version, self.score_activation,
+                                                                    self.score_threshold)
+        return num_det, det_boxes, det_scores, det_classes
+
+class End2End(nn.Module):
+    '''export onnx or tensorrt model with NMS operation.'''
+    def __init__(self, model, max_obj=100, iou_thres=0.45, score_thres=0.25, max_wh=None, device=None, n_classes=80):
+        super().__init__()
+        device = device if device else torch.device('cpu')
+        assert isinstance(max_wh,(int)) or max_wh is None
+        self.model = model.to(device)
+        self.model.model[-1].end2end = True
+        self.patch_model = ONNX_TRT if max_wh is None else ONNX_ORT
+        self.end2end = self.patch_model(max_obj, iou_thres, score_thres, max_wh, device, n_classes)
+        self.end2end.eval()
+
+    def forward(self, x):
+        x = self.model(x)
+        x = self.end2end(x)
+        return x
+
+
+def attempt_load(weights, device=None, inplace=True, fuse=True):
+    # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a
+    from models.yolo import Detect, Model
+
+    model = Ensemble()
+    for w in weights if isinstance(weights, list) else [weights]:
+        ckpt = torch.load(attempt_download(w), map_location='cpu')  # load
+        ckpt = (ckpt.get('ema') or ckpt['model']).to(device).float()  # FP32 model
+
+        # Model compatibility updates
+        if not hasattr(ckpt, 'stride'):
+            ckpt.stride = torch.tensor([32.])
+        if hasattr(ckpt, 'names') and isinstance(ckpt.names, (list, tuple)):
+            ckpt.names = dict(enumerate(ckpt.names))  # convert to dict
+
+        model.append(ckpt.fuse().eval() if fuse and hasattr(ckpt, 'fuse') else ckpt.eval())  # model in eval mode
+
+    # Module compatibility updates
+    for m in model.modules():
+        t = type(m)
+        if t in (nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU, Detect, Model):
+            m.inplace = inplace  # torch 1.7.0 compatibility
+            # if t is Detect and not isinstance(m.anchor_grid, list):
+            #    delattr(m, 'anchor_grid')
+            #    setattr(m, 'anchor_grid', [torch.zeros(1)] * m.nl)
+        elif t is nn.Upsample and not hasattr(m, 'recompute_scale_factor'):
+            m.recompute_scale_factor = None  # torch 1.11.0 compatibility
+
+    # Return model
+    if len(model) == 1:
+        return model[-1]
+
+    # Return detection ensemble
+    print(f'Ensemble created with {weights}\n')
+    for k in 'names', 'nc', 'yaml':
+        setattr(model, k, getattr(model[0], k))
+    model.stride = model[torch.argmax(torch.tensor([m.stride.max() for m in model])).int()].stride  # max stride
+    assert all(model[0].nc == m.nc for m in model), f'Models have different class counts: {[m.nc for m in model]}'
+    return model

yolov9/models/hub/anchors.yaml

@@ -0,0 +1,59 @@
+# YOLOv3 & YOLOv5
+# Default anchors for COCO data
+
+
+# P5 -------------------------------------------------------------------------------------------------------------------
+# P5-640:
+anchors_p5_640:
+  - [10,13, 16,30, 33,23]  # P3/8
+  - [30,61, 62,45, 59,119]  # P4/16
+  - [116,90, 156,198, 373,326]  # P5/32
+
+
+# P6 -------------------------------------------------------------------------------------------------------------------
+# P6-640:  thr=0.25: 0.9964 BPR, 5.54 anchors past thr, n=12, img_size=640, metric_all=0.281/0.716-mean/best, past_thr=0.469-mean: 9,11,  21,19,  17,41,  43,32,  39,70,  86,64,  65,131,  134,130,  120,265,  282,180,  247,354,  512,387
+anchors_p6_640:
+  - [9,11,  21,19,  17,41]  # P3/8
+  - [43,32,  39,70,  86,64]  # P4/16
+  - [65,131,  134,130,  120,265]  # P5/32
+  - [282,180,  247,354,  512,387]  # P6/64
+
+# P6-1280:  thr=0.25: 0.9950 BPR, 5.55 anchors past thr, n=12, img_size=1280, metric_all=0.281/0.714-mean/best, past_thr=0.468-mean: 19,27,  44,40,  38,94,  96,68,  86,152,  180,137,  140,301,  303,264,  238,542,  436,615,  739,380,  925,792
+anchors_p6_1280:
+  - [19,27,  44,40,  38,94]  # P3/8
+  - [96,68,  86,152,  180,137]  # P4/16
+  - [140,301,  303,264,  238,542]  # P5/32
+  - [436,615,  739,380,  925,792]  # P6/64
+
+# P6-1920:  thr=0.25: 0.9950 BPR, 5.55 anchors past thr, n=12, img_size=1920, metric_all=0.281/0.714-mean/best, past_thr=0.468-mean: 28,41,  67,59,  57,141,  144,103,  129,227,  270,205,  209,452,  455,396,  358,812,  653,922,  1109,570,  1387,1187
+anchors_p6_1920:
+  - [28,41,  67,59,  57,141]  # P3/8
+  - [144,103,  129,227,  270,205]  # P4/16
+  - [209,452,  455,396,  358,812]  # P5/32
+  - [653,922,  1109,570,  1387,1187]  # P6/64
+
+
+# P7 -------------------------------------------------------------------------------------------------------------------
+# P7-640:  thr=0.25: 0.9962 BPR, 6.76 anchors past thr, n=15, img_size=640, metric_all=0.275/0.733-mean/best, past_thr=0.466-mean: 11,11,  13,30,  29,20,  30,46,  61,38,  39,92,  78,80,  146,66,  79,163,  149,150,  321,143,  157,303,  257,402,  359,290,  524,372
+anchors_p7_640:
+  - [11,11,  13,30,  29,20]  # P3/8
+  - [30,46,  61,38,  39,92]  # P4/16
+  - [78,80,  146,66,  79,163]  # P5/32
+  - [149,150,  321,143,  157,303]  # P6/64
+  - [257,402,  359,290,  524,372]  # P7/128
+
+# P7-1280:  thr=0.25: 0.9968 BPR, 6.71 anchors past thr, n=15, img_size=1280, metric_all=0.273/0.732-mean/best, past_thr=0.463-mean: 19,22,  54,36,  32,77,  70,83,  138,71,  75,173,  165,159,  148,334,  375,151,  334,317,  251,626,  499,474,  750,326,  534,814,  1079,818
+anchors_p7_1280:
+  - [19,22,  54,36,  32,77]  # P3/8
+  - [70,83,  138,71,  75,173]  # P4/16
+  - [165,159,  148,334,  375,151]  # P5/32
+  - [334,317,  251,626,  499,474]  # P6/64
+  - [750,326,  534,814,  1079,818]  # P7/128
+
+# P7-1920:  thr=0.25: 0.9968 BPR, 6.71 anchors past thr, n=15, img_size=1920, metric_all=0.273/0.732-mean/best, past_thr=0.463-mean: 29,34,  81,55,  47,115,  105,124,  207,107,  113,259,  247,238,  222,500,  563,227,  501,476,  376,939,  749,711,  1126,489,  801,1222,  1618,1227
+anchors_p7_1920:
+  - [29,34,  81,55,  47,115]  # P3/8
+  - [105,124,  207,107,  113,259]  # P4/16
+  - [247,238,  222,500,  563,227]  # P5/32
+  - [501,476,  376,939,  749,711]  # P6/64
+  - [1126,489,  801,1222,  1618,1227]  # P7/128

yolov9/models/hub/yolov3-spp.yaml

@@ -0,0 +1,51 @@
+# YOLOv3
+
+# Parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+anchors:
+  - [10,13, 16,30, 33,23]  # P3/8
+  - [30,61, 62,45, 59,119]  # P4/16
+  - [116,90, 156,198, 373,326]  # P5/32
+
+# darknet53 backbone
+backbone:
+  # [from, number, module, args]
+  [[-1, 1, Conv, [32, 3, 1]],  # 0
+   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
+   [-1, 1, Bottleneck, [64]],
+   [-1, 1, Conv, [128, 3, 2]],  # 3-P2/4
+   [-1, 2, Bottleneck, [128]],
+   [-1, 1, Conv, [256, 3, 2]],  # 5-P3/8
+   [-1, 8, Bottleneck, [256]],
+   [-1, 1, Conv, [512, 3, 2]],  # 7-P4/16
+   [-1, 8, Bottleneck, [512]],
+   [-1, 1, Conv, [1024, 3, 2]],  # 9-P5/32
+   [-1, 4, Bottleneck, [1024]],  # 10
+  ]
+
+# YOLOv3-SPP head
+head:
+  [[-1, 1, Bottleneck, [1024, False]],
+   [-1, 1, SPP, [512, [5, 9, 13]]],
+   [-1, 1, Conv, [1024, 3, 1]],
+   [-1, 1, Conv, [512, 1, 1]],
+   [-1, 1, Conv, [1024, 3, 1]],  # 15 (P5/32-large)
+
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 8], 1, Concat, [1]],  # cat backbone P4
+   [-1, 1, Bottleneck, [512, False]],
+   [-1, 1, Bottleneck, [512, False]],
+   [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [512, 3, 1]],  # 22 (P4/16-medium)
+
+   [-2, 1, Conv, [128, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P3
+   [-1, 1, Bottleneck, [256, False]],
+   [-1, 2, Bottleneck, [256, False]],  # 27 (P3/8-small)
+
+   [[27, 22, 15], 1, Detect, [nc, anchors]],   # Detect(P3, P4, P5)
+  ]

yolov9/models/hub/yolov3-tiny.yaml

@@ -0,0 +1,41 @@
+# YOLOv3
+
+# Parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+anchors:
+  - [10,14, 23,27, 37,58]  # P4/16
+  - [81,82, 135,169, 344,319]  # P5/32
+
+# YOLOv3-tiny backbone
+backbone:
+  # [from, number, module, args]
+  [[-1, 1, Conv, [16, 3, 1]],  # 0
+   [-1, 1, nn.MaxPool2d, [2, 2, 0]],  # 1-P1/2
+   [-1, 1, Conv, [32, 3, 1]],
+   [-1, 1, nn.MaxPool2d, [2, 2, 0]],  # 3-P2/4
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, nn.MaxPool2d, [2, 2, 0]],  # 5-P3/8
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, nn.MaxPool2d, [2, 2, 0]],  # 7-P4/16
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, nn.MaxPool2d, [2, 2, 0]],  # 9-P5/32
+   [-1, 1, Conv, [512, 3, 1]],
+   [-1, 1, nn.ZeroPad2d, [[0, 1, 0, 1]]],  # 11
+   [-1, 1, nn.MaxPool2d, [2, 1, 0]],  # 12
+  ]
+
+# YOLOv3-tiny head
+head:
+  [[-1, 1, Conv, [1024, 3, 1]],
+   [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [512, 3, 1]],  # 15 (P5/32-large)
+
+   [-2, 1, Conv, [128, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 8], 1, Concat, [1]],  # cat backbone P4
+   [-1, 1, Conv, [256, 3, 1]],  # 19 (P4/16-medium)
+
+   [[19, 15], 1, Detect, [nc, anchors]],  # Detect(P4, P5)
+  ]

yolov9/models/hub/yolov3.yaml

@@ -0,0 +1,51 @@
+# YOLOv3
+
+# Parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+anchors:
+  - [10,13, 16,30, 33,23]  # P3/8
+  - [30,61, 62,45, 59,119]  # P4/16
+  - [116,90, 156,198, 373,326]  # P5/32
+
+# darknet53 backbone
+backbone:
+  # [from, number, module, args]
+  [[-1, 1, Conv, [32, 3, 1]],  # 0
+   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
+   [-1, 1, Bottleneck, [64]],
+   [-1, 1, Conv, [128, 3, 2]],  # 3-P2/4
+   [-1, 2, Bottleneck, [128]],
+   [-1, 1, Conv, [256, 3, 2]],  # 5-P3/8
+   [-1, 8, Bottleneck, [256]],
+   [-1, 1, Conv, [512, 3, 2]],  # 7-P4/16
+   [-1, 8, Bottleneck, [512]],
+   [-1, 1, Conv, [1024, 3, 2]],  # 9-P5/32
+   [-1, 4, Bottleneck, [1024]],  # 10
+  ]
+
+# YOLOv3 head
+head:
+  [[-1, 1, Bottleneck, [1024, False]],
+   [-1, 1, Conv, [512, 1, 1]],
+   [-1, 1, Conv, [1024, 3, 1]],
+   [-1, 1, Conv, [512, 1, 1]],
+   [-1, 1, Conv, [1024, 3, 1]],  # 15 (P5/32-large)
+
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 8], 1, Concat, [1]],  # cat backbone P4
+   [-1, 1, Bottleneck, [512, False]],
+   [-1, 1, Bottleneck, [512, False]],
+   [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [512, 3, 1]],  # 22 (P4/16-medium)
+
+   [-2, 1, Conv, [128, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P3
+   [-1, 1, Bottleneck, [256, False]],
+   [-1, 2, Bottleneck, [256, False]],  # 27 (P3/8-small)
+
+   [[27, 22, 15], 1, Detect, [nc, anchors]],   # Detect(P3, P4, P5)
+  ]

yolov9/models/panoptic/gelan-c-pan.yaml

@@ -0,0 +1,80 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 0-P1/2
+
+   # conv down
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 2
+
+   # avg-conv down
+   [-1, 1, ADown, [256]],  # 3-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 4
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 5-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 6
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 7-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 8
+  ]
+
+# gelan head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [512, 256]],  # 9
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 12
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 15 (P3/8-small)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [256]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 18 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [512]],
+   [[-1, 9], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 21 (P5/32-large)
+
+   # panoptic
+   [[15, 18, 21], 1, Panoptic, [nc, 93, 32, 256]],  # Panoptic(P3, P4, P5)
+  ]

yolov9/models/panoptic/yolov7-af-pan.yaml

@@ -0,0 +1,137 @@
+# YOLOv7
+
+# Parameters
+nc: 80  # number of classes
+sem_nc: 93  # number of stuff classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+anchors: 3
+
+# YOLOv7 backbone
+backbone:
+  [[-1, 1, Conv, [32, 3, 1]],  # 0
+  
+   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2      
+   [-1, 1, Conv, [64, 3, 1]],
+   
+   [-1, 1, Conv, [128, 3, 2]],  # 3-P2/4  
+   [-1, 1, Conv, [64, 1, 1]],
+   [-2, 1, Conv, [64, 1, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [[-1, -3, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [256, 1, 1]],  # 11
+         
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [128, 1, 1]],
+   [-3, 1, Conv, [128, 1, 1]],
+   [-1, 1, Conv, [128, 3, 2]],
+   [[-1, -3], 1, Concat, [1]],  # 16-P3/8  
+   [-1, 1, Conv, [128, 1, 1]],
+   [-2, 1, Conv, [128, 1, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [[-1, -3, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [512, 1, 1]],  # 24
+         
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [256, 1, 1]],
+   [-3, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [256, 3, 2]],
+   [[-1, -3], 1, Concat, [1]],  # 29-P4/16  
+   [-1, 1, Conv, [256, 1, 1]],
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [[-1, -3, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [1024, 1, 1]],  # 37
+         
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [512, 1, 1]],
+   [-3, 1, Conv, [512, 1, 1]],
+   [-1, 1, Conv, [512, 3, 2]],
+   [[-1, -3], 1, Concat, [1]],  # 42-P5/32  
+   [-1, 1, Conv, [256, 1, 1]],
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [[-1, -3, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [1024, 1, 1]],  # 50
+  ]
+
+# yolov7 head
+head:
+  [[-1, 1, SPPCSPC, [512]], # 51
+  
+   [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [37, 1, Conv, [256, 1, 1]], # route backbone P4
+   [[-1, -2], 1, Concat, [1]],
+   
+   [-1, 1, Conv, [256, 1, 1]],
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [256, 1, 1]], # 63
+   
+   [-1, 1, Conv, [128, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [24, 1, Conv, [128, 1, 1]], # route backbone P3
+   [[-1, -2], 1, Concat, [1]],
+   
+   [-1, 1, Conv, [128, 1, 1]],
+   [-2, 1, Conv, [128, 1, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [128, 1, 1]], # 75
+      
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [128, 1, 1]],
+   [-3, 1, Conv, [128, 1, 1]],
+   [-1, 1, Conv, [128, 3, 2]],
+   [[-1, -3, 63], 1, Concat, [1]],
+   
+   [-1, 1, Conv, [256, 1, 1]],
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [256, 1, 1]], # 88
+      
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [256, 1, 1]],
+   [-3, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [256, 3, 2]],
+   [[-1, -3, 51], 1, Concat, [1]],
+   
+   [-1, 1, Conv, [512, 1, 1]],
+   [-2, 1, Conv, [512, 1, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [512, 1, 1]], # 101
+   
+   [75, 1, Conv, [256, 3, 1]],
+   [88, 1, Conv, [512, 3, 1]],
+   [101, 1, Conv, [1024, 3, 1]],
+
+   [[102, 103, 104], 1, Panoptic, [nc, 93, 32, 256]],  # Panoptic(P3, P4, P5)
+  ]

yolov9/models/segment/gelan-c-dseg.yaml

@@ -0,0 +1,84 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 0-P1/2
+
+   # conv down
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 2
+
+   # avg-conv down
+   [-1, 1, ADown, [256]],  # 3-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 4
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 5-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 6
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 7-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 8
+  ]
+
+# gelan head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [512, 256]],  # 9
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 12
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 15 (P3/8-small)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [256]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 18 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [512]],
+   [[-1, 9], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 21 (P5/32-large)
+
+   [15, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 22   
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [-1, 1, Conv, [256, 3, 1]], # 24
+
+   # segment
+   [[15, 18, 21, 24], 1, DSegment, [nc, 32, 256]],  # Segment(P3, P4, P5)
+  ]

yolov9/models/segment/gelan-c-seg.yaml

@@ -0,0 +1,80 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 0-P1/2
+
+   # conv down
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 2
+
+   # avg-conv down
+   [-1, 1, ADown, [256]],  # 3-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 4
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 5-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 6
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 7-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 8
+  ]
+
+# gelan head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [512, 256]],  # 9
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 12
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 15 (P3/8-small)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [256]],
+   [[-1, 12], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 18 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [512]],
+   [[-1, 9], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 21 (P5/32-large)
+
+   # segment
+   [[15, 18, 21], 1, Segment, [nc, 32, 256]],  # Segment(P3, P4, P5)
+  ]

yolov9/models/segment/yolov7-af-seg.yaml

@@ -0,0 +1,136 @@
+# YOLOv7
+
+# Parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+anchors: 3
+
+# YOLOv7 backbone
+backbone:
+  [[-1, 1, Conv, [32, 3, 1]],  # 0
+  
+   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2      
+   [-1, 1, Conv, [64, 3, 1]],
+   
+   [-1, 1, Conv, [128, 3, 2]],  # 3-P2/4  
+   [-1, 1, Conv, [64, 1, 1]],
+   [-2, 1, Conv, [64, 1, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [[-1, -3, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [256, 1, 1]],  # 11
+         
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [128, 1, 1]],
+   [-3, 1, Conv, [128, 1, 1]],
+   [-1, 1, Conv, [128, 3, 2]],
+   [[-1, -3], 1, Concat, [1]],  # 16-P3/8  
+   [-1, 1, Conv, [128, 1, 1]],
+   [-2, 1, Conv, [128, 1, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [[-1, -3, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [512, 1, 1]],  # 24
+         
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [256, 1, 1]],
+   [-3, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [256, 3, 2]],
+   [[-1, -3], 1, Concat, [1]],  # 29-P4/16  
+   [-1, 1, Conv, [256, 1, 1]],
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [[-1, -3, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [1024, 1, 1]],  # 37
+         
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [512, 1, 1]],
+   [-3, 1, Conv, [512, 1, 1]],
+   [-1, 1, Conv, [512, 3, 2]],
+   [[-1, -3], 1, Concat, [1]],  # 42-P5/32  
+   [-1, 1, Conv, [256, 1, 1]],
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [[-1, -3, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [1024, 1, 1]],  # 50
+  ]
+
+# yolov7 head
+head:
+  [[-1, 1, SPPCSPC, [512]], # 51
+  
+   [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [37, 1, Conv, [256, 1, 1]], # route backbone P4
+   [[-1, -2], 1, Concat, [1]],
+   
+   [-1, 1, Conv, [256, 1, 1]],
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [256, 1, 1]], # 63
+   
+   [-1, 1, Conv, [128, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [24, 1, Conv, [128, 1, 1]], # route backbone P3
+   [[-1, -2], 1, Concat, [1]],
+   
+   [-1, 1, Conv, [128, 1, 1]],
+   [-2, 1, Conv, [128, 1, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [-1, 1, Conv, [64, 3, 1]],
+   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [128, 1, 1]], # 75
+      
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [128, 1, 1]],
+   [-3, 1, Conv, [128, 1, 1]],
+   [-1, 1, Conv, [128, 3, 2]],
+   [[-1, -3, 63], 1, Concat, [1]],
+   
+   [-1, 1, Conv, [256, 1, 1]],
+   [-2, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [-1, 1, Conv, [128, 3, 1]],
+   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [256, 1, 1]], # 88
+      
+   [-1, 1, MP, []],
+   [-1, 1, Conv, [256, 1, 1]],
+   [-3, 1, Conv, [256, 1, 1]],
+   [-1, 1, Conv, [256, 3, 2]],
+   [[-1, -3, 51], 1, Concat, [1]],
+   
+   [-1, 1, Conv, [512, 1, 1]],
+   [-2, 1, Conv, [512, 1, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [-1, 1, Conv, [256, 3, 1]],
+   [[-1, -2, -3, -4, -5, -6], 1, Concat, [1]],
+   [-1, 1, Conv, [512, 1, 1]], # 101
+   
+   [75, 1, Conv, [256, 3, 1]],
+   [88, 1, Conv, [512, 3, 1]],
+   [101, 1, Conv, [1024, 3, 1]],
+
+   [[102, 103, 104], 1, Segment, [nc, 32, 256]],  # Segment(P3, P4, P5)
+  ]

yolov9/models/segment/yolov9-c-dseg.yaml

@@ -0,0 +1,130 @@
+# YOLOv9
+
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+#activation: nn.LeakyReLU(0.1)
+#activation: nn.ReLU()
+
+# anchors
+anchors: 3
+
+# gelan backbone
+backbone:
+  [
+   [-1, 1, Silence, []],  
+   
+   # conv down
+   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
+
+   # conv down
+   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 3
+
+   # avg-conv down
+   [-1, 1, ADown, [256]],  # 4-P3/8
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 5
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 6-P4/16
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 7
+
+   # avg-conv down
+   [-1, 1, ADown, [512]],  # 8-P5/32
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 9
+  ]
+
+# YOLOv9 head
+head:
+  [
+   # elan-spp block
+   [-1, 1, SPPELAN, [512, 256]],  # 10
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 7], 1, Concat, [1]],  # cat backbone P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 13
+
+   # up-concat merge
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 5], 1, Concat, [1]],  # cat backbone P3
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 16 (P3/8-small)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [256]],
+   [[-1, 13], 1, Concat, [1]],  # cat head P4
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 19 (P4/16-medium)
+
+   # avg-conv-down merge
+   [-1, 1, ADown, [512]],
+   [[-1, 10], 1, Concat, [1]],  # cat head P5
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 22 (P5/32-large)
+   
+   
+   # multi-level reversible auxiliary branch
+   
+   # routing
+   [5, 1, CBLinear, [[256]]], # 23
+   [7, 1, CBLinear, [[256, 512]]], # 24
+   [9, 1, CBLinear, [[256, 512, 512]]], # 25
+   
+   # conv down
+   [0, 1, Conv, [64, 3, 2]],  # 26-P1/2
+
+   # conv down
+   [-1, 1, Conv, [128, 3, 2]],  # 27-P2/4
+
+   # elan-1 block
+   [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 28
+
+   # avg-conv down fuse
+   [-1, 1, ADown, [256]],  # 29-P3/8
+   [[23, 24, 25, -1], 1, CBFuse, [[0, 0, 0]]], # 30  
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 31
+
+   # avg-conv down fuse
+   [-1, 1, ADown, [512]],  # 32-P4/16
+   [[24, 25, -1], 1, CBFuse, [[1, 1]]], # 33 
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 34
+
+   # avg-conv down fuse
+   [-1, 1, ADown, [512]],  # 35-P5/32
+   [[25, -1], 1, CBFuse, [[2]]], # 36
+
+   # elan-2 block
+   [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 37
+   
+   [31, 1, RepNCSPELAN4, [512, 256, 128, 2]],  # 38 
+   
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [-1, 1, Conv, [256, 3, 1]], # 40
+   
+   [16, 1, RepNCSPELAN4, [256, 256, 128, 2]],  # 41  
+   
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [-1, 1, Conv, [256, 3, 1]], # 43
+
+   # segment
+   [[31, 34, 37, 16, 19, 22, 40, 43], 1, DualDSegment, [nc, 32, 256]],  # Segment(P3, P4, P5)
+  ]

yolov9/models/tf.py

@@ -0,0 +1,596 @@
+import argparse
+import sys
+from copy import deepcopy
+from pathlib import Path
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+# ROOT = ROOT.relative_to(Path.cwd())  # relative
+
+import numpy as np
+import tensorflow as tf
+import torch
+import torch.nn as nn
+from tensorflow import keras
+
+from models.common import (C3, SPP, SPPF, Bottleneck, BottleneckCSP, C3x, Concat, Conv, CrossConv, DWConv,
+                           DWConvTranspose2d, Focus, autopad)
+from models.experimental import MixConv2d, attempt_load
+from models.yolo import Detect, Segment
+from utils.activations import SiLU
+from utils.general import LOGGER, make_divisible, print_args
+
+
+class TFBN(keras.layers.Layer):
+    # TensorFlow BatchNormalization wrapper
+    def __init__(self, w=None):
+        super().__init__()
+        self.bn = keras.layers.BatchNormalization(
+            beta_initializer=keras.initializers.Constant(w.bias.numpy()),
+            gamma_initializer=keras.initializers.Constant(w.weight.numpy()),
+            moving_mean_initializer=keras.initializers.Constant(w.running_mean.numpy()),
+            moving_variance_initializer=keras.initializers.Constant(w.running_var.numpy()),
+            epsilon=w.eps)
+
+    def call(self, inputs):
+        return self.bn(inputs)
+
+
+class TFPad(keras.layers.Layer):
+    # Pad inputs in spatial dimensions 1 and 2
+    def __init__(self, pad):
+        super().__init__()
+        if isinstance(pad, int):
+            self.pad = tf.constant([[0, 0], [pad, pad], [pad, pad], [0, 0]])
+        else:  # tuple/list
+            self.pad = tf.constant([[0, 0], [pad[0], pad[0]], [pad[1], pad[1]], [0, 0]])
+
+    def call(self, inputs):
+        return tf.pad(inputs, self.pad, mode='constant', constant_values=0)
+
+
+class TFConv(keras.layers.Layer):
+    # Standard convolution
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True, w=None):
+        # ch_in, ch_out, weights, kernel, stride, padding, groups
+        super().__init__()
+        assert g == 1, "TF v2.2 Conv2D does not support 'groups' argument"
+        # TensorFlow convolution padding is inconsistent with PyTorch (e.g. k=3 s=2 'SAME' padding)
+        # see https://stackoverflow.com/questions/52975843/comparing-conv2d-with-padding-between-tensorflow-and-pytorch
+        conv = keras.layers.Conv2D(
+            filters=c2,
+            kernel_size=k,
+            strides=s,
+            padding='SAME' if s == 1 else 'VALID',
+            use_bias=not hasattr(w, 'bn'),
+            kernel_initializer=keras.initializers.Constant(w.conv.weight.permute(2, 3, 1, 0).numpy()),
+            bias_initializer='zeros' if hasattr(w, 'bn') else keras.initializers.Constant(w.conv.bias.numpy()))
+        self.conv = conv if s == 1 else keras.Sequential([TFPad(autopad(k, p)), conv])
+        self.bn = TFBN(w.bn) if hasattr(w, 'bn') else tf.identity
+        self.act = activations(w.act) if act else tf.identity
+
+    def call(self, inputs):
+        return self.act(self.bn(self.conv(inputs)))
+
+
+class TFDWConv(keras.layers.Layer):
+    # Depthwise convolution
+    def __init__(self, c1, c2, k=1, s=1, p=None, act=True, w=None):
+        # ch_in, ch_out, weights, kernel, stride, padding, groups
+        super().__init__()
+        assert c2 % c1 == 0, f'TFDWConv() output={c2} must be a multiple of input={c1} channels'
+        conv = keras.layers.DepthwiseConv2D(
+            kernel_size=k,
+            depth_multiplier=c2 // c1,
+            strides=s,
+            padding='SAME' if s == 1 else 'VALID',
+            use_bias=not hasattr(w, 'bn'),
+            depthwise_initializer=keras.initializers.Constant(w.conv.weight.permute(2, 3, 1, 0).numpy()),
+            bias_initializer='zeros' if hasattr(w, 'bn') else keras.initializers.Constant(w.conv.bias.numpy()))
+        self.conv = conv if s == 1 else keras.Sequential([TFPad(autopad(k, p)), conv])
+        self.bn = TFBN(w.bn) if hasattr(w, 'bn') else tf.identity
+        self.act = activations(w.act) if act else tf.identity
+
+    def call(self, inputs):
+        return self.act(self.bn(self.conv(inputs)))
+
+
+class TFDWConvTranspose2d(keras.layers.Layer):
+    # Depthwise ConvTranspose2d
+    def __init__(self, c1, c2, k=1, s=1, p1=0, p2=0, w=None):
+        # ch_in, ch_out, weights, kernel, stride, padding, groups
+        super().__init__()
+        assert c1 == c2, f'TFDWConv() output={c2} must be equal to input={c1} channels'
+        assert k == 4 and p1 == 1, 'TFDWConv() only valid for k=4 and p1=1'
+        weight, bias = w.weight.permute(2, 3, 1, 0).numpy(), w.bias.numpy()
+        self.c1 = c1
+        self.conv = [
+            keras.layers.Conv2DTranspose(filters=1,
+                                         kernel_size=k,
+                                         strides=s,
+                                         padding='VALID',
+                                         output_padding=p2,
+                                         use_bias=True,
+                                         kernel_initializer=keras.initializers.Constant(weight[..., i:i + 1]),
+                                         bias_initializer=keras.initializers.Constant(bias[i])) for i in range(c1)]
+
+    def call(self, inputs):
+        return tf.concat([m(x) for m, x in zip(self.conv, tf.split(inputs, self.c1, 3))], 3)[:, 1:-1, 1:-1]
+
+
+class TFFocus(keras.layers.Layer):
+    # Focus wh information into c-space
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True, w=None):
+        # ch_in, ch_out, kernel, stride, padding, groups
+        super().__init__()
+        self.conv = TFConv(c1 * 4, c2, k, s, p, g, act, w.conv)
+
+    def call(self, inputs):  # x(b,w,h,c) -> y(b,w/2,h/2,4c)
+        # inputs = inputs / 255  # normalize 0-255 to 0-1
+        inputs = [inputs[:, ::2, ::2, :], inputs[:, 1::2, ::2, :], inputs[:, ::2, 1::2, :], inputs[:, 1::2, 1::2, :]]
+        return self.conv(tf.concat(inputs, 3))
+
+
+class TFBottleneck(keras.layers.Layer):
+    # Standard bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5, w=None):  # ch_in, ch_out, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1)
+        self.cv2 = TFConv(c_, c2, 3, 1, g=g, w=w.cv2)
+        self.add = shortcut and c1 == c2
+
+    def call(self, inputs):
+        return inputs + self.cv2(self.cv1(inputs)) if self.add else self.cv2(self.cv1(inputs))
+
+
+class TFCrossConv(keras.layers.Layer):
+    # Cross Convolution
+    def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False, w=None):
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = TFConv(c1, c_, (1, k), (1, s), w=w.cv1)
+        self.cv2 = TFConv(c_, c2, (k, 1), (s, 1), g=g, w=w.cv2)
+        self.add = shortcut and c1 == c2
+
+    def call(self, inputs):
+        return inputs + self.cv2(self.cv1(inputs)) if self.add else self.cv2(self.cv1(inputs))
+
+
+class TFConv2d(keras.layers.Layer):
+    # Substitution for PyTorch nn.Conv2D
+    def __init__(self, c1, c2, k, s=1, g=1, bias=True, w=None):
+        super().__init__()
+        assert g == 1, "TF v2.2 Conv2D does not support 'groups' argument"
+        self.conv = keras.layers.Conv2D(filters=c2,
+                                        kernel_size=k,
+                                        strides=s,
+                                        padding='VALID',
+                                        use_bias=bias,
+                                        kernel_initializer=keras.initializers.Constant(
+                                            w.weight.permute(2, 3, 1, 0).numpy()),
+                                        bias_initializer=keras.initializers.Constant(w.bias.numpy()) if bias else None)
+
+    def call(self, inputs):
+        return self.conv(inputs)
+
+
+class TFBottleneckCSP(keras.layers.Layer):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, w=None):
+        # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1)
+        self.cv2 = TFConv2d(c1, c_, 1, 1, bias=False, w=w.cv2)
+        self.cv3 = TFConv2d(c_, c_, 1, 1, bias=False, w=w.cv3)
+        self.cv4 = TFConv(2 * c_, c2, 1, 1, w=w.cv4)
+        self.bn = TFBN(w.bn)
+        self.act = lambda x: keras.activations.swish(x)
+        self.m = keras.Sequential([TFBottleneck(c_, c_, shortcut, g, e=1.0, w=w.m[j]) for j in range(n)])
+
+    def call(self, inputs):
+        y1 = self.cv3(self.m(self.cv1(inputs)))
+        y2 = self.cv2(inputs)
+        return self.cv4(self.act(self.bn(tf.concat((y1, y2), axis=3))))
+
+
+class TFC3(keras.layers.Layer):
+    # CSP Bottleneck with 3 convolutions
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, w=None):
+        # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1)
+        self.cv2 = TFConv(c1, c_, 1, 1, w=w.cv2)
+        self.cv3 = TFConv(2 * c_, c2, 1, 1, w=w.cv3)
+        self.m = keras.Sequential([TFBottleneck(c_, c_, shortcut, g, e=1.0, w=w.m[j]) for j in range(n)])
+
+    def call(self, inputs):
+        return self.cv3(tf.concat((self.m(self.cv1(inputs)), self.cv2(inputs)), axis=3))
+
+
+class TFC3x(keras.layers.Layer):
+    # 3 module with cross-convolutions
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, w=None):
+        # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1)
+        self.cv2 = TFConv(c1, c_, 1, 1, w=w.cv2)
+        self.cv3 = TFConv(2 * c_, c2, 1, 1, w=w.cv3)
+        self.m = keras.Sequential([
+            TFCrossConv(c_, c_, k=3, s=1, g=g, e=1.0, shortcut=shortcut, w=w.m[j]) for j in range(n)])
+
+    def call(self, inputs):
+        return self.cv3(tf.concat((self.m(self.cv1(inputs)), self.cv2(inputs)), axis=3))
+
+
+class TFSPP(keras.layers.Layer):
+    # Spatial pyramid pooling layer used in YOLOv3-SPP
+    def __init__(self, c1, c2, k=(5, 9, 13), w=None):
+        super().__init__()
+        c_ = c1 // 2  # hidden channels
+        self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1)
+        self.cv2 = TFConv(c_ * (len(k) + 1), c2, 1, 1, w=w.cv2)
+        self.m = [keras.layers.MaxPool2D(pool_size=x, strides=1, padding='SAME') for x in k]
+
+    def call(self, inputs):
+        x = self.cv1(inputs)
+        return self.cv2(tf.concat([x] + [m(x) for m in self.m], 3))
+
+
+class TFSPPF(keras.layers.Layer):
+    # Spatial pyramid pooling-Fast layer
+    def __init__(self, c1, c2, k=5, w=None):
+        super().__init__()
+        c_ = c1 // 2  # hidden channels
+        self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1)
+        self.cv2 = TFConv(c_ * 4, c2, 1, 1, w=w.cv2)
+        self.m = keras.layers.MaxPool2D(pool_size=k, strides=1, padding='SAME')
+
+    def call(self, inputs):
+        x = self.cv1(inputs)
+        y1 = self.m(x)
+        y2 = self.m(y1)
+        return self.cv2(tf.concat([x, y1, y2, self.m(y2)], 3))
+
+
+class TFDetect(keras.layers.Layer):
+    # TF YOLO Detect layer
+    def __init__(self, nc=80, anchors=(), ch=(), imgsz=(640, 640), w=None):  # detection layer
+        super().__init__()
+        self.stride = tf.convert_to_tensor(w.stride.numpy(), dtype=tf.float32)
+        self.nc = nc  # number of classes
+        self.no = nc + 5  # number of outputs per anchor
+        self.nl = len(anchors)  # number of detection layers
+        self.na = len(anchors[0]) // 2  # number of anchors
+        self.grid = [tf.zeros(1)] * self.nl  # init grid
+        self.anchors = tf.convert_to_tensor(w.anchors.numpy(), dtype=tf.float32)
+        self.anchor_grid = tf.reshape(self.anchors * tf.reshape(self.stride, [self.nl, 1, 1]), [self.nl, 1, -1, 1, 2])
+        self.m = [TFConv2d(x, self.no * self.na, 1, w=w.m[i]) for i, x in enumerate(ch)]
+        self.training = False  # set to False after building model
+        self.imgsz = imgsz
+        for i in range(self.nl):
+            ny, nx = self.imgsz[0] // self.stride[i], self.imgsz[1] // self.stride[i]
+            self.grid[i] = self._make_grid(nx, ny)
+
+    def call(self, inputs):
+        z = []  # inference output
+        x = []
+        for i in range(self.nl):
+            x.append(self.m[i](inputs[i]))
+            # x(bs,20,20,255) to x(bs,3,20,20,85)
+            ny, nx = self.imgsz[0] // self.stride[i], self.imgsz[1] // self.stride[i]
+            x[i] = tf.reshape(x[i], [-1, ny * nx, self.na, self.no])
+
+            if not self.training:  # inference
+                y = x[i]
+                grid = tf.transpose(self.grid[i], [0, 2, 1, 3]) - 0.5
+                anchor_grid = tf.transpose(self.anchor_grid[i], [0, 2, 1, 3]) * 4
+                xy = (tf.sigmoid(y[..., 0:2]) * 2 + grid) * self.stride[i]  # xy
+                wh = tf.sigmoid(y[..., 2:4]) ** 2 * anchor_grid
+                # Normalize xywh to 0-1 to reduce calibration error
+                xy /= tf.constant([[self.imgsz[1], self.imgsz[0]]], dtype=tf.float32)
+                wh /= tf.constant([[self.imgsz[1], self.imgsz[0]]], dtype=tf.float32)
+                y = tf.concat([xy, wh, tf.sigmoid(y[..., 4:5 + self.nc]), y[..., 5 + self.nc:]], -1)
+                z.append(tf.reshape(y, [-1, self.na * ny * nx, self.no]))
+
+        return tf.transpose(x, [0, 2, 1, 3]) if self.training else (tf.concat(z, 1),)
+
+    @staticmethod
+    def _make_grid(nx=20, ny=20):
+        # yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
+        # return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
+        xv, yv = tf.meshgrid(tf.range(nx), tf.range(ny))
+        return tf.cast(tf.reshape(tf.stack([xv, yv], 2), [1, 1, ny * nx, 2]), dtype=tf.float32)
+
+
+class TFSegment(TFDetect):
+    # YOLO Segment head for segmentation models
+    def __init__(self, nc=80, anchors=(), nm=32, npr=256, ch=(), imgsz=(640, 640), w=None):
+        super().__init__(nc, anchors, ch, imgsz, w)
+        self.nm = nm  # number of masks
+        self.npr = npr  # number of protos
+        self.no = 5 + nc + self.nm  # number of outputs per anchor
+        self.m = [TFConv2d(x, self.no * self.na, 1, w=w.m[i]) for i, x in enumerate(ch)]  # output conv
+        self.proto = TFProto(ch[0], self.npr, self.nm, w=w.proto)  # protos
+        self.detect = TFDetect.call
+
+    def call(self, x):
+        p = self.proto(x[0])
+        # p = TFUpsample(None, scale_factor=4, mode='nearest')(self.proto(x[0]))  # (optional) full-size protos
+        p = tf.transpose(p, [0, 3, 1, 2])  # from shape(1,160,160,32) to shape(1,32,160,160)
+        x = self.detect(self, x)
+        return (x, p) if self.training else (x[0], p)
+
+
+class TFProto(keras.layers.Layer):
+
+    def __init__(self, c1, c_=256, c2=32, w=None):
+        super().__init__()
+        self.cv1 = TFConv(c1, c_, k=3, w=w.cv1)
+        self.upsample = TFUpsample(None, scale_factor=2, mode='nearest')
+        self.cv2 = TFConv(c_, c_, k=3, w=w.cv2)
+        self.cv3 = TFConv(c_, c2, w=w.cv3)
+
+    def call(self, inputs):
+        return self.cv3(self.cv2(self.upsample(self.cv1(inputs))))
+
+
+class TFUpsample(keras.layers.Layer):
+    # TF version of torch.nn.Upsample()
+    def __init__(self, size, scale_factor, mode, w=None):  # warning: all arguments needed including 'w'
+        super().__init__()
+        assert scale_factor % 2 == 0, "scale_factor must be multiple of 2"
+        self.upsample = lambda x: tf.image.resize(x, (x.shape[1] * scale_factor, x.shape[2] * scale_factor), mode)
+        # self.upsample = keras.layers.UpSampling2D(size=scale_factor, interpolation=mode)
+        # with default arguments: align_corners=False, half_pixel_centers=False
+        # self.upsample = lambda x: tf.raw_ops.ResizeNearestNeighbor(images=x,
+        #                                                            size=(x.shape[1] * 2, x.shape[2] * 2))
+
+    def call(self, inputs):
+        return self.upsample(inputs)
+
+
+class TFConcat(keras.layers.Layer):
+    # TF version of torch.concat()
+    def __init__(self, dimension=1, w=None):
+        super().__init__()
+        assert dimension == 1, "convert only NCHW to NHWC concat"
+        self.d = 3
+
+    def call(self, inputs):
+        return tf.concat(inputs, self.d)
+
+
+def parse_model(d, ch, model, imgsz):  # model_dict, input_channels(3)
+    LOGGER.info(f"\n{'':>3}{'from':>18}{'n':>3}{'params':>10}  {'module':<40}{'arguments':<30}")
+    anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple']
+    na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors  # number of anchors
+    no = na * (nc + 5)  # number of outputs = anchors * (classes + 5)
+
+    layers, save, c2 = [], [], ch[-1]  # layers, savelist, ch out
+    for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']):  # from, number, module, args
+        m_str = m
+        m = eval(m) if isinstance(m, str) else m  # eval strings
+        for j, a in enumerate(args):
+            try:
+                args[j] = eval(a) if isinstance(a, str) else a  # eval strings
+            except NameError:
+                pass
+
+        n = max(round(n * gd), 1) if n > 1 else n  # depth gain
+        if m in [
+                nn.Conv2d, Conv, DWConv, DWConvTranspose2d, Bottleneck, SPP, SPPF, MixConv2d, Focus, CrossConv,
+                BottleneckCSP, C3, C3x]:
+            c1, c2 = ch[f], args[0]
+            c2 = make_divisible(c2 * gw, 8) if c2 != no else c2
+
+            args = [c1, c2, *args[1:]]
+            if m in [BottleneckCSP, C3, C3x]:
+                args.insert(2, n)
+                n = 1
+        elif m is nn.BatchNorm2d:
+            args = [ch[f]]
+        elif m is Concat:
+            c2 = sum(ch[-1 if x == -1 else x + 1] for x in f)
+        elif m in [Detect, Segment]:
+            args.append([ch[x + 1] for x in f])
+            if isinstance(args[1], int):  # number of anchors
+                args[1] = [list(range(args[1] * 2))] * len(f)
+            if m is Segment:
+                args[3] = make_divisible(args[3] * gw, 8)
+            args.append(imgsz)
+        else:
+            c2 = ch[f]
+
+        tf_m = eval('TF' + m_str.replace('nn.', ''))
+        m_ = keras.Sequential([tf_m(*args, w=model.model[i][j]) for j in range(n)]) if n > 1 \
+            else tf_m(*args, w=model.model[i])  # module
+
+        torch_m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args)  # module
+        t = str(m)[8:-2].replace('__main__.', '')  # module type
+        np = sum(x.numel() for x in torch_m_.parameters())  # number params
+        m_.i, m_.f, m_.type, m_.np = i, f, t, np  # attach index, 'from' index, type, number params
+        LOGGER.info(f'{i:>3}{str(f):>18}{str(n):>3}{np:>10}  {t:<40}{str(args):<30}')  # print
+        save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1)  # append to savelist
+        layers.append(m_)
+        ch.append(c2)
+    return keras.Sequential(layers), sorted(save)
+
+
+class TFModel:
+    # TF YOLO model
+    def __init__(self, cfg='yolo.yaml', ch=3, nc=None, model=None, imgsz=(640, 640)):  # model, channels, classes
+        super().__init__()
+        if isinstance(cfg, dict):
+            self.yaml = cfg  # model dict
+        else:  # is *.yaml
+            import yaml  # for torch hub
+            self.yaml_file = Path(cfg).name
+            with open(cfg) as f:
+                self.yaml = yaml.load(f, Loader=yaml.FullLoader)  # model dict
+
+        # Define model
+        if nc and nc != self.yaml['nc']:
+            LOGGER.info(f"Overriding {cfg} nc={self.yaml['nc']} with nc={nc}")
+            self.yaml['nc'] = nc  # override yaml value
+        self.model, self.savelist = parse_model(deepcopy(self.yaml), ch=[ch], model=model, imgsz=imgsz)
+
+    def predict(self,
+                inputs,
+                tf_nms=False,
+                agnostic_nms=False,
+                topk_per_class=100,
+                topk_all=100,
+                iou_thres=0.45,
+                conf_thres=0.25):
+        y = []  # outputs
+        x = inputs
+        for m in self.model.layers:
+            if m.f != -1:  # if not from previous layer
+                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
+
+            x = m(x)  # run
+            y.append(x if m.i in self.savelist else None)  # save output
+
+        # Add TensorFlow NMS
+        if tf_nms:
+            boxes = self._xywh2xyxy(x[0][..., :4])
+            probs = x[0][:, :, 4:5]
+            classes = x[0][:, :, 5:]
+            scores = probs * classes
+            if agnostic_nms:
+                nms = AgnosticNMS()((boxes, classes, scores), topk_all, iou_thres, conf_thres)
+            else:
+                boxes = tf.expand_dims(boxes, 2)
+                nms = tf.image.combined_non_max_suppression(boxes,
+                                                            scores,
+                                                            topk_per_class,
+                                                            topk_all,
+                                                            iou_thres,
+                                                            conf_thres,
+                                                            clip_boxes=False)
+            return (nms,)
+        return x  # output [1,6300,85] = [xywh, conf, class0, class1, ...]
+        # x = x[0]  # [x(1,6300,85), ...] to x(6300,85)
+        # xywh = x[..., :4]  # x(6300,4) boxes
+        # conf = x[..., 4:5]  # x(6300,1) confidences
+        # cls = tf.reshape(tf.cast(tf.argmax(x[..., 5:], axis=1), tf.float32), (-1, 1))  # x(6300,1)  classes
+        # return tf.concat([conf, cls, xywh], 1)
+
+    @staticmethod
+    def _xywh2xyxy(xywh):
+        # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right
+        x, y, w, h = tf.split(xywh, num_or_size_splits=4, axis=-1)
+        return tf.concat([x - w / 2, y - h / 2, x + w / 2, y + h / 2], axis=-1)
+
+
+class AgnosticNMS(keras.layers.Layer):
+    # TF Agnostic NMS
+    def call(self, input, topk_all, iou_thres, conf_thres):
+        # wrap map_fn to avoid TypeSpec related error https://stackoverflow.com/a/65809989/3036450
+        return tf.map_fn(lambda x: self._nms(x, topk_all, iou_thres, conf_thres),
+                         input,
+                         fn_output_signature=(tf.float32, tf.float32, tf.float32, tf.int32),
+                         name='agnostic_nms')
+
+    @staticmethod
+    def _nms(x, topk_all=100, iou_thres=0.45, conf_thres=0.25):  # agnostic NMS
+        boxes, classes, scores = x
+        class_inds = tf.cast(tf.argmax(classes, axis=-1), tf.float32)
+        scores_inp = tf.reduce_max(scores, -1)
+        selected_inds = tf.image.non_max_suppression(boxes,
+                                                     scores_inp,
+                                                     max_output_size=topk_all,
+                                                     iou_threshold=iou_thres,
+                                                     score_threshold=conf_thres)
+        selected_boxes = tf.gather(boxes, selected_inds)
+        padded_boxes = tf.pad(selected_boxes,
+                              paddings=[[0, topk_all - tf.shape(selected_boxes)[0]], [0, 0]],
+                              mode="CONSTANT",
+                              constant_values=0.0)
+        selected_scores = tf.gather(scores_inp, selected_inds)
+        padded_scores = tf.pad(selected_scores,
+                               paddings=[[0, topk_all - tf.shape(selected_boxes)[0]]],
+                               mode="CONSTANT",
+                               constant_values=-1.0)
+        selected_classes = tf.gather(class_inds, selected_inds)
+        padded_classes = tf.pad(selected_classes,
+                                paddings=[[0, topk_all - tf.shape(selected_boxes)[0]]],
+                                mode="CONSTANT",
+                                constant_values=-1.0)
+        valid_detections = tf.shape(selected_inds)[0]
+        return padded_boxes, padded_scores, padded_classes, valid_detections
+
+
+def activations(act=nn.SiLU):
+    # Returns TF activation from input PyTorch activation
+    if isinstance(act, nn.LeakyReLU):
+        return lambda x: keras.activations.relu(x, alpha=0.1)
+    elif isinstance(act, nn.Hardswish):
+        return lambda x: x * tf.nn.relu6(x + 3) * 0.166666667
+    elif isinstance(act, (nn.SiLU, SiLU)):
+        return lambda x: keras.activations.swish(x)
+    else:
+        raise Exception(f'no matching TensorFlow activation found for PyTorch activation {act}')
+
+
+def representative_dataset_gen(dataset, ncalib=100):
+    # Representative dataset generator for use with converter.representative_dataset, returns a generator of np arrays
+    for n, (path, img, im0s, vid_cap, string) in enumerate(dataset):
+        im = np.transpose(img, [1, 2, 0])
+        im = np.expand_dims(im, axis=0).astype(np.float32)
+        im /= 255
+        yield [im]
+        if n >= ncalib:
+            break
+
+
+def run(
+        weights=ROOT / 'yolo.pt',  # weights path
+        imgsz=(640, 640),  # inference size h,w
+        batch_size=1,  # batch size
+        dynamic=False,  # dynamic batch size
+):
+    # PyTorch model
+    im = torch.zeros((batch_size, 3, *imgsz))  # BCHW image
+    model = attempt_load(weights, device=torch.device('cpu'), inplace=True, fuse=False)
+    _ = model(im)  # inference
+    model.info()
+
+    # TensorFlow model
+    im = tf.zeros((batch_size, *imgsz, 3))  # BHWC image
+    tf_model = TFModel(cfg=model.yaml, model=model, nc=model.nc, imgsz=imgsz)
+    _ = tf_model.predict(im)  # inference
+
+    # Keras model
+    im = keras.Input(shape=(*imgsz, 3), batch_size=None if dynamic else batch_size)
+    keras_model = keras.Model(inputs=im, outputs=tf_model.predict(im))
+    keras_model.summary()
+
+    LOGGER.info('PyTorch, TensorFlow and Keras models successfully verified.\nUse export.py for TF model export.')
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', type=str, default=ROOT / 'yolo.pt', help='weights path')
+    parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w')
+    parser.add_argument('--batch-size', type=int, default=1, help='batch size')
+    parser.add_argument('--dynamic', action='store_true', help='dynamic batch size')
+    opt = parser.parse_args()
+    opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1  # expand
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

yolov9/models/yolo.py

@@ -0,0 +1,818 @@
+import argparse
+import os
+import platform
+import sys
+from copy import deepcopy
+from pathlib import Path
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+if platform.system() != 'Windows':
+    ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.common import *
+from models.experimental import *
+from utils.general import LOGGER, check_version, check_yaml, make_divisible, print_args
+from utils.plots import feature_visualization
+from utils.torch_utils import (fuse_conv_and_bn, initialize_weights, model_info, profile, scale_img, select_device,
+                               time_sync)
+from utils.tal.anchor_generator import make_anchors, dist2bbox
+
+try:
+    import thop  # for FLOPs computation
+except ImportError:
+    thop = None
+
+
+class Detect(nn.Module):
+    # YOLO Detect head for detection models
+    dynamic = False  # force grid reconstruction
+    export = False  # export mode
+    shape = None
+    anchors = torch.empty(0)  # init
+    strides = torch.empty(0)  # init
+
+    def __init__(self, nc=80, ch=(), inplace=True):  # detection layer
+        super().__init__()
+        self.nc = nc  # number of classes
+        self.nl = len(ch)  # number of detection layers
+        self.reg_max = 16
+        self.no = nc + self.reg_max * 4  # number of outputs per anchor
+        self.inplace = inplace  # use inplace ops (e.g. slice assignment)
+        self.stride = torch.zeros(self.nl)  # strides computed during build
+
+        c2, c3 = max((ch[0] // 4, self.reg_max * 4, 16)), max((ch[0], min((self.nc * 2, 128))))  # channels
+        self.cv2 = nn.ModuleList(
+            nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3), nn.Conv2d(c2, 4 * self.reg_max, 1)) for x in ch)
+        self.cv3 = nn.ModuleList(
+            nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch)
+        self.dfl = DFL(self.reg_max) if self.reg_max > 1 else nn.Identity()
+
+    def forward(self, x):
+        shape = x[0].shape  # BCHW
+        for i in range(self.nl):
+            x[i] = torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1)
+        if self.training:
+            return x
+        elif self.dynamic or self.shape != shape:
+            self.anchors, self.strides = (x.transpose(0, 1) for x in make_anchors(x, self.stride, 0.5))
+            self.shape = shape
+
+        box, cls = torch.cat([xi.view(shape[0], self.no, -1) for xi in x], 2).split((self.reg_max * 4, self.nc), 1)
+        dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+        y = torch.cat((dbox, cls.sigmoid()), 1)
+        return y if self.export else (y, x)
+
+    def bias_init(self):
+        # Initialize Detect() biases, WARNING: requires stride availability
+        m = self  # self.model[-1]  # Detect() module
+        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
+        # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
+        for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
+            a[-1].bias.data[:] = 1.0  # box
+            b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
+
+
+class DDetect(nn.Module):
+    # YOLO Detect head for detection models
+    dynamic = False  # force grid reconstruction
+    export = False  # export mode
+    shape = None
+    anchors = torch.empty(0)  # init
+    strides = torch.empty(0)  # init
+
+    def __init__(self, nc=80, ch=(), inplace=True):  # detection layer
+        super().__init__()
+        self.nc = nc  # number of classes
+        self.nl = len(ch)  # number of detection layers
+        self.reg_max = 16
+        self.no = nc + self.reg_max * 4  # number of outputs per anchor
+        self.inplace = inplace  # use inplace ops (e.g. slice assignment)
+        self.stride = torch.zeros(self.nl)  # strides computed during build
+
+        c2, c3 = make_divisible(max((ch[0] // 4, self.reg_max * 4, 16)), 4), max((ch[0], min((self.nc * 2, 128))))  # channels
+        self.cv2 = nn.ModuleList(
+            nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3, g=4), nn.Conv2d(c2, 4 * self.reg_max, 1, groups=4)) for x in ch)
+        self.cv3 = nn.ModuleList(
+            nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch)
+        self.dfl = DFL(self.reg_max) if self.reg_max > 1 else nn.Identity()
+
+    def forward(self, x):
+        shape = x[0].shape  # BCHW
+        for i in range(self.nl):
+            x[i] = torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1)
+        if self.training:
+            return x
+        elif self.dynamic or self.shape != shape:
+            self.anchors, self.strides = (x.transpose(0, 1) for x in make_anchors(x, self.stride, 0.5))
+            self.shape = shape
+
+        box, cls = torch.cat([xi.view(shape[0], self.no, -1) for xi in x], 2).split((self.reg_max * 4, self.nc), 1)
+        dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+        y = torch.cat((dbox, cls.sigmoid()), 1)
+        return y if self.export else (y, x)
+
+    def bias_init(self):
+        # Initialize Detect() biases, WARNING: requires stride availability
+        m = self  # self.model[-1]  # Detect() module
+        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
+        # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
+        for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
+            a[-1].bias.data[:] = 1.0  # box
+            b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
+
+
+class DualDetect(nn.Module):
+    # YOLO Detect head for detection models
+    dynamic = False  # force grid reconstruction
+    export = False  # export mode
+    shape = None
+    anchors = torch.empty(0)  # init
+    strides = torch.empty(0)  # init
+
+    def __init__(self, nc=80, ch=(), inplace=True):  # detection layer
+        super().__init__()
+        self.nc = nc  # number of classes
+        self.nl = len(ch) // 2  # number of detection layers
+        self.reg_max = 16
+        self.no = nc + self.reg_max * 4  # number of outputs per anchor
+        self.inplace = inplace  # use inplace ops (e.g. slice assignment)
+        self.stride = torch.zeros(self.nl)  # strides computed during build
+
+        c2, c3 = max((ch[0] // 4, self.reg_max * 4, 16)), max((ch[0], min((self.nc * 2, 128))))  # channels
+        c4, c5 = max((ch[self.nl] // 4, self.reg_max * 4, 16)), max((ch[self.nl], min((self.nc * 2, 128))))  # channels
+        self.cv2 = nn.ModuleList(
+            nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3), nn.Conv2d(c2, 4 * self.reg_max, 1)) for x in ch[:self.nl])
+        self.cv3 = nn.ModuleList(
+            nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch[:self.nl])
+        self.cv4 = nn.ModuleList(
+            nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, 4 * self.reg_max, 1)) for x in ch[self.nl:])
+        self.cv5 = nn.ModuleList(
+            nn.Sequential(Conv(x, c5, 3), Conv(c5, c5, 3), nn.Conv2d(c5, self.nc, 1)) for x in ch[self.nl:])
+        self.dfl = DFL(self.reg_max)
+        self.dfl2 = DFL(self.reg_max)
+
+    def forward(self, x):
+        shape = x[0].shape  # BCHW
+        d1 = []
+        d2 = []
+        for i in range(self.nl):
+            d1.append(torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1))
+            d2.append(torch.cat((self.cv4[i](x[self.nl+i]), self.cv5[i](x[self.nl+i])), 1))
+        if self.training:
+            return [d1, d2]
+        elif self.dynamic or self.shape != shape:
+            self.anchors, self.strides = (d1.transpose(0, 1) for d1 in make_anchors(d1, self.stride, 0.5))
+            self.shape = shape
+
+        box, cls = torch.cat([di.view(shape[0], self.no, -1) for di in d1], 2).split((self.reg_max * 4, self.nc), 1)
+        dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+        box2, cls2 = torch.cat([di.view(shape[0], self.no, -1) for di in d2], 2).split((self.reg_max * 4, self.nc), 1)
+        dbox2 = dist2bbox(self.dfl2(box2), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+        y = [torch.cat((dbox, cls.sigmoid()), 1), torch.cat((dbox2, cls2.sigmoid()), 1)]
+        return y if self.export else (y, [d1, d2])
+
+    def bias_init(self):
+        # Initialize Detect() biases, WARNING: requires stride availability
+        m = self  # self.model[-1]  # Detect() module
+        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
+        # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
+        for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
+            a[-1].bias.data[:] = 1.0  # box
+            b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
+        for a, b, s in zip(m.cv4, m.cv5, m.stride):  # from
+            a[-1].bias.data[:] = 1.0  # box
+            b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
+
+
+class DualDDetect(nn.Module):
+    # YOLO Detect head for detection models
+    dynamic = False  # force grid reconstruction
+    export = False  # export mode
+    shape = None
+    anchors = torch.empty(0)  # init
+    strides = torch.empty(0)  # init
+
+    def __init__(self, nc=80, ch=(), inplace=True):  # detection layer
+        super().__init__()
+        self.nc = nc  # number of classes
+        self.nl = len(ch) // 2  # number of detection layers
+        self.reg_max = 16
+        self.no = nc + self.reg_max * 4  # number of outputs per anchor
+        self.inplace = inplace  # use inplace ops (e.g. slice assignment)
+        self.stride = torch.zeros(self.nl)  # strides computed during build
+
+        c2, c3 = make_divisible(max((ch[0] // 4, self.reg_max * 4, 16)), 4), max((ch[0], min((self.nc * 2, 128))))  # channels
+        c4, c5 = make_divisible(max((ch[self.nl] // 4, self.reg_max * 4, 16)), 4), max((ch[self.nl], min((self.nc * 2, 128))))  # channels
+        self.cv2 = nn.ModuleList(
+            nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3, g=4), nn.Conv2d(c2, 4 * self.reg_max, 1, groups=4)) for x in ch[:self.nl])
+        self.cv3 = nn.ModuleList(
+            nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch[:self.nl])
+        self.cv4 = nn.ModuleList(
+            nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3, g=4), nn.Conv2d(c4, 4 * self.reg_max, 1, groups=4)) for x in ch[self.nl:])
+        self.cv5 = nn.ModuleList(
+            nn.Sequential(Conv(x, c5, 3), Conv(c5, c5, 3), nn.Conv2d(c5, self.nc, 1)) for x in ch[self.nl:])
+        self.dfl = DFL(self.reg_max)
+        self.dfl2 = DFL(self.reg_max)
+
+    def forward(self, x):
+        shape = x[0].shape  # BCHW
+        d1 = []
+        d2 = []
+        for i in range(self.nl):
+            d1.append(torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1))
+            d2.append(torch.cat((self.cv4[i](x[self.nl+i]), self.cv5[i](x[self.nl+i])), 1))
+        if self.training:
+            return [d1, d2]
+        elif self.dynamic or self.shape != shape:
+            self.anchors, self.strides = (d1.transpose(0, 1) for d1 in make_anchors(d1, self.stride, 0.5))
+            self.shape = shape
+
+        box, cls = torch.cat([di.view(shape[0], self.no, -1) for di in d1], 2).split((self.reg_max * 4, self.nc), 1)
+        dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+        box2, cls2 = torch.cat([di.view(shape[0], self.no, -1) for di in d2], 2).split((self.reg_max * 4, self.nc), 1)
+        dbox2 = dist2bbox(self.dfl2(box2), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+        y = [torch.cat((dbox, cls.sigmoid()), 1), torch.cat((dbox2, cls2.sigmoid()), 1)]
+        return y if self.export else (y, [d1, d2])
+        #y = torch.cat((dbox2, cls2.sigmoid()), 1)
+        #return y if self.export else (y, d2)
+        #y1 = torch.cat((dbox, cls.sigmoid()), 1)
+        #y2 = torch.cat((dbox2, cls2.sigmoid()), 1)
+        #return [y1, y2] if self.export else [(y1, d1), (y2, d2)]
+        #return [y1, y2] if self.export else [(y1, y2), (d1, d2)]
+
+    def bias_init(self):
+        # Initialize Detect() biases, WARNING: requires stride availability
+        m = self  # self.model[-1]  # Detect() module
+        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
+        # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
+        for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
+            a[-1].bias.data[:] = 1.0  # box
+            b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
+        for a, b, s in zip(m.cv4, m.cv5, m.stride):  # from
+            a[-1].bias.data[:] = 1.0  # box
+            b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
+
+
+class TripleDetect(nn.Module):
+    # YOLO Detect head for detection models
+    dynamic = False  # force grid reconstruction
+    export = False  # export mode
+    shape = None
+    anchors = torch.empty(0)  # init
+    strides = torch.empty(0)  # init
+
+    def __init__(self, nc=80, ch=(), inplace=True):  # detection layer
+        super().__init__()
+        self.nc = nc  # number of classes
+        self.nl = len(ch) // 3  # number of detection layers
+        self.reg_max = 16
+        self.no = nc + self.reg_max * 4  # number of outputs per anchor
+        self.inplace = inplace  # use inplace ops (e.g. slice assignment)
+        self.stride = torch.zeros(self.nl)  # strides computed during build
+
+        c2, c3 = max((ch[0] // 4, self.reg_max * 4, 16)), max((ch[0], min((self.nc * 2, 128))))  # channels
+        c4, c5 = max((ch[self.nl] // 4, self.reg_max * 4, 16)), max((ch[self.nl], min((self.nc * 2, 128))))  # channels
+        c6, c7 = max((ch[self.nl * 2] // 4, self.reg_max * 4, 16)), max((ch[self.nl * 2], min((self.nc * 2, 128))))  # channels
+        self.cv2 = nn.ModuleList(
+            nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3), nn.Conv2d(c2, 4 * self.reg_max, 1)) for x in ch[:self.nl])
+        self.cv3 = nn.ModuleList(
+            nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch[:self.nl])
+        self.cv4 = nn.ModuleList(
+            nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, 4 * self.reg_max, 1)) for x in ch[self.nl:self.nl*2])
+        self.cv5 = nn.ModuleList(
+            nn.Sequential(Conv(x, c5, 3), Conv(c5, c5, 3), nn.Conv2d(c5, self.nc, 1)) for x in ch[self.nl:self.nl*2])
+        self.cv6 = nn.ModuleList(
+            nn.Sequential(Conv(x, c6, 3), Conv(c6, c6, 3), nn.Conv2d(c6, 4 * self.reg_max, 1)) for x in ch[self.nl*2:self.nl*3])
+        self.cv7 = nn.ModuleList(
+            nn.Sequential(Conv(x, c7, 3), Conv(c7, c7, 3), nn.Conv2d(c7, self.nc, 1)) for x in ch[self.nl*2:self.nl*3])
+        self.dfl = DFL(self.reg_max)
+        self.dfl2 = DFL(self.reg_max)
+        self.dfl3 = DFL(self.reg_max)
+
+    def forward(self, x):
+        shape = x[0].shape  # BCHW
+        d1 = []
+        d2 = []
+        d3 = []
+        for i in range(self.nl):
+            d1.append(torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1))
+            d2.append(torch.cat((self.cv4[i](x[self.nl+i]), self.cv5[i](x[self.nl+i])), 1))
+            d3.append(torch.cat((self.cv6[i](x[self.nl*2+i]), self.cv7[i](x[self.nl*2+i])), 1))
+        if self.training:
+            return [d1, d2, d3]
+        elif self.dynamic or self.shape != shape:
+            self.anchors, self.strides = (d1.transpose(0, 1) for d1 in make_anchors(d1, self.stride, 0.5))
+            self.shape = shape
+
+        box, cls = torch.cat([di.view(shape[0], self.no, -1) for di in d1], 2).split((self.reg_max * 4, self.nc), 1)
+        dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+        box2, cls2 = torch.cat([di.view(shape[0], self.no, -1) for di in d2], 2).split((self.reg_max * 4, self.nc), 1)
+        dbox2 = dist2bbox(self.dfl2(box2), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+        box3, cls3 = torch.cat([di.view(shape[0], self.no, -1) for di in d3], 2).split((self.reg_max * 4, self.nc), 1)
+        dbox3 = dist2bbox(self.dfl3(box3), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+        y = [torch.cat((dbox, cls.sigmoid()), 1), torch.cat((dbox2, cls2.sigmoid()), 1), torch.cat((dbox3, cls3.sigmoid()), 1)]
+        return y if self.export else (y, [d1, d2, d3])
+
+    def bias_init(self):
+        # Initialize Detect() biases, WARNING: requires stride availability
+        m = self  # self.model[-1]  # Detect() module
+        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
+        # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
+        for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
+            a[-1].bias.data[:] = 1.0  # box
+            b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
+        for a, b, s in zip(m.cv4, m.cv5, m.stride):  # from
+            a[-1].bias.data[:] = 1.0  # box
+            b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
+        for a, b, s in zip(m.cv6, m.cv7, m.stride):  # from
+            a[-1].bias.data[:] = 1.0  # box
+            b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
+
+
+class TripleDDetect(nn.Module):
+    # YOLO Detect head for detection models
+    dynamic = False  # force grid reconstruction
+    export = False  # export mode
+    shape = None
+    anchors = torch.empty(0)  # init
+    strides = torch.empty(0)  # init
+
+    def __init__(self, nc=80, ch=(), inplace=True):  # detection layer
+        super().__init__()
+        self.nc = nc  # number of classes
+        self.nl = len(ch) // 3  # number of detection layers
+        self.reg_max = 16
+        self.no = nc + self.reg_max * 4  # number of outputs per anchor
+        self.inplace = inplace  # use inplace ops (e.g. slice assignment)
+        self.stride = torch.zeros(self.nl)  # strides computed during build
+
+        c2, c3 = make_divisible(max((ch[0] // 4, self.reg_max * 4, 16)), 4), \
+                                max((ch[0], min((self.nc * 2, 128))))  # channels
+        c4, c5 = make_divisible(max((ch[self.nl] // 4, self.reg_max * 4, 16)), 4), \
+                                max((ch[self.nl], min((self.nc * 2, 128))))  # channels
+        c6, c7 = make_divisible(max((ch[self.nl * 2] // 4, self.reg_max * 4, 16)), 4), \
+                                max((ch[self.nl * 2], min((self.nc * 2, 128))))  # channels
+        self.cv2 = nn.ModuleList(
+            nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3, g=4), 
+                          nn.Conv2d(c2, 4 * self.reg_max, 1, groups=4)) for x in ch[:self.nl])
+        self.cv3 = nn.ModuleList(
+            nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch[:self.nl])
+        self.cv4 = nn.ModuleList(
+            nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3, g=4), 
+                          nn.Conv2d(c4, 4 * self.reg_max, 1, groups=4)) for x in ch[self.nl:self.nl*2])
+        self.cv5 = nn.ModuleList(
+            nn.Sequential(Conv(x, c5, 3), Conv(c5, c5, 3), nn.Conv2d(c5, self.nc, 1)) for x in ch[self.nl:self.nl*2])
+        self.cv6 = nn.ModuleList(
+            nn.Sequential(Conv(x, c6, 3), Conv(c6, c6, 3, g=4), 
+                          nn.Conv2d(c6, 4 * self.reg_max, 1, groups=4)) for x in ch[self.nl*2:self.nl*3])
+        self.cv7 = nn.ModuleList(
+            nn.Sequential(Conv(x, c7, 3), Conv(c7, c7, 3), nn.Conv2d(c7, self.nc, 1)) for x in ch[self.nl*2:self.nl*3])
+        self.dfl = DFL(self.reg_max)
+        self.dfl2 = DFL(self.reg_max)
+        self.dfl3 = DFL(self.reg_max)
+
+    def forward(self, x):
+        shape = x[0].shape  # BCHW
+        d1 = []
+        d2 = []
+        d3 = []
+        for i in range(self.nl):
+            d1.append(torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1))
+            d2.append(torch.cat((self.cv4[i](x[self.nl+i]), self.cv5[i](x[self.nl+i])), 1))
+            d3.append(torch.cat((self.cv6[i](x[self.nl*2+i]), self.cv7[i](x[self.nl*2+i])), 1))
+        if self.training:
+            return [d1, d2, d3]
+        elif self.dynamic or self.shape != shape:
+            self.anchors, self.strides = (d1.transpose(0, 1) for d1 in make_anchors(d1, self.stride, 0.5))
+            self.shape = shape
+
+        box, cls = torch.cat([di.view(shape[0], self.no, -1) for di in d1], 2).split((self.reg_max * 4, self.nc), 1)
+        dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+        box2, cls2 = torch.cat([di.view(shape[0], self.no, -1) for di in d2], 2).split((self.reg_max * 4, self.nc), 1)
+        dbox2 = dist2bbox(self.dfl2(box2), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+        box3, cls3 = torch.cat([di.view(shape[0], self.no, -1) for di in d3], 2).split((self.reg_max * 4, self.nc), 1)
+        dbox3 = dist2bbox(self.dfl3(box3), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
+        #y = [torch.cat((dbox, cls.sigmoid()), 1), torch.cat((dbox2, cls2.sigmoid()), 1), torch.cat((dbox3, cls3.sigmoid()), 1)]
+        #return y if self.export else (y, [d1, d2, d3])
+        y = torch.cat((dbox3, cls3.sigmoid()), 1)
+        return y if self.export else (y, d3)
+
+    def bias_init(self):
+        # Initialize Detect() biases, WARNING: requires stride availability
+        m = self  # self.model[-1]  # Detect() module
+        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
+        # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
+        for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
+            a[-1].bias.data[:] = 1.0  # box
+            b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
+        for a, b, s in zip(m.cv4, m.cv5, m.stride):  # from
+            a[-1].bias.data[:] = 1.0  # box
+            b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
+        for a, b, s in zip(m.cv6, m.cv7, m.stride):  # from
+            a[-1].bias.data[:] = 1.0  # box
+            b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
+
+
+class Segment(Detect):
+    # YOLO Segment head for segmentation models
+    def __init__(self, nc=80, nm=32, npr=256, ch=(), inplace=True):
+        super().__init__(nc, ch, inplace)
+        self.nm = nm  # number of masks
+        self.npr = npr  # number of protos
+        self.proto = Proto(ch[0], self.npr, self.nm)  # protos
+        self.detect = Detect.forward
+
+        c4 = max(ch[0] // 4, self.nm)
+        self.cv4 = nn.ModuleList(nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, self.nm, 1)) for x in ch)
+
+    def forward(self, x):
+        p = self.proto(x[0])
+        bs = p.shape[0]
+
+        mc = torch.cat([self.cv4[i](x[i]).view(bs, self.nm, -1) for i in range(self.nl)], 2)  # mask coefficients
+        x = self.detect(self, x)
+        if self.training:
+            return x, mc, p
+        return (torch.cat([x, mc], 1), p) if self.export else (torch.cat([x[0], mc], 1), (x[1], mc, p))
+
+
+class DSegment(DDetect):
+    # YOLO Segment head for segmentation models
+    def __init__(self, nc=80, nm=32, npr=256, ch=(), inplace=True):
+        super().__init__(nc, ch[:-1], inplace)
+        self.nl = len(ch)-1
+        self.nm = nm  # number of masks
+        self.npr = npr  # number of protos
+        self.proto = Conv(ch[-1], self.nm, 1)  # protos
+        self.detect = DDetect.forward
+
+        c4 = max(ch[0] // 4, self.nm)
+        self.cv4 = nn.ModuleList(nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, self.nm, 1)) for x in ch[:-1])
+
+    def forward(self, x):
+        p = self.proto(x[-1])
+        bs = p.shape[0]
+
+        mc = torch.cat([self.cv4[i](x[i]).view(bs, self.nm, -1) for i in range(self.nl)], 2)  # mask coefficients
+        x = self.detect(self, x[:-1])
+        if self.training:
+            return x, mc, p
+        return (torch.cat([x, mc], 1), p) if self.export else (torch.cat([x[0], mc], 1), (x[1], mc, p))
+
+
+class DualDSegment(DualDDetect):
+    # YOLO Segment head for segmentation models
+    def __init__(self, nc=80, nm=32, npr=256, ch=(), inplace=True):
+        super().__init__(nc, ch[:-2], inplace)
+        self.nl = (len(ch)-2) // 2
+        self.nm = nm  # number of masks
+        self.npr = npr  # number of protos
+        self.proto = Conv(ch[-2], self.nm, 1)  # protos
+        self.proto2 = Conv(ch[-1], self.nm, 1)  # protos
+        self.detect = DualDDetect.forward
+
+        c6 = max(ch[0] // 4, self.nm)
+        c7 = max(ch[self.nl] // 4, self.nm)
+        self.cv6 = nn.ModuleList(nn.Sequential(Conv(x, c6, 3), Conv(c6, c6, 3), nn.Conv2d(c6, self.nm, 1)) for x in ch[:self.nl])
+        self.cv7 = nn.ModuleList(nn.Sequential(Conv(x, c7, 3), Conv(c7, c7, 3), nn.Conv2d(c7, self.nm, 1)) for x in ch[self.nl:self.nl*2])
+
+    def forward(self, x):
+        p = [self.proto(x[-2]), self.proto2(x[-1])]
+        bs = p[0].shape[0]
+
+        mc = [torch.cat([self.cv6[i](x[i]).view(bs, self.nm, -1) for i in range(self.nl)], 2),
+              torch.cat([self.cv7[i](x[self.nl+i]).view(bs, self.nm, -1) for i in range(self.nl)], 2)]  # mask coefficients
+        d = self.detect(self, x[:-2])
+        if self.training:
+            return d, mc, p
+        return (torch.cat([d[0][1], mc[1]], 1), (d[1][1], mc[1], p[1]))
+
+
+class Panoptic(Detect):
+    # YOLO Panoptic head for panoptic segmentation models
+    def __init__(self, nc=80, sem_nc=93, nm=32, npr=256, ch=(), inplace=True):
+        super().__init__(nc, ch, inplace)
+        self.sem_nc = sem_nc
+        self.nm = nm  # number of masks
+        self.npr = npr  # number of protos
+        self.proto = Proto(ch[0], self.npr, self.nm)  # protos
+        self.uconv = UConv(ch[0], ch[0]//4, self.sem_nc+self.nc)
+        self.detect = Detect.forward
+
+        c4 = max(ch[0] // 4, self.nm)
+        self.cv4 = nn.ModuleList(nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, self.nm, 1)) for x in ch)
+
+
+    def forward(self, x):
+        p = self.proto(x[0])
+        s = self.uconv(x[0])
+        bs = p.shape[0]
+
+        mc = torch.cat([self.cv4[i](x[i]).view(bs, self.nm, -1) for i in range(self.nl)], 2)  # mask coefficients
+        x = self.detect(self, x)
+        if self.training:
+            return x, mc, p, s
+        return (torch.cat([x, mc], 1), p, s) if self.export else (torch.cat([x[0], mc], 1), (x[1], mc, p, s))
+    
+
+class BaseModel(nn.Module):
+    # YOLO base model
+    def forward(self, x, profile=False, visualize=False):
+        return self._forward_once(x, profile, visualize)  # single-scale inference, train
+
+    def _forward_once(self, x, profile=False, visualize=False):
+        y, dt = [], []  # outputs
+        for m in self.model:
+            if m.f != -1:  # if not from previous layer
+                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
+            if profile:
+                self._profile_one_layer(m, x, dt)
+            x = m(x)  # run
+            y.append(x if m.i in self.save else None)  # save output
+            if visualize:
+                feature_visualization(x, m.type, m.i, save_dir=visualize)
+        return x
+
+    def _profile_one_layer(self, m, x, dt):
+        c = m == self.model[-1]  # is final layer, copy input as inplace fix
+        o = thop.profile(m, inputs=(x.copy() if c else x,), verbose=False)[0] / 1E9 * 2 if thop else 0  # FLOPs
+        t = time_sync()
+        for _ in range(10):
+            m(x.copy() if c else x)
+        dt.append((time_sync() - t) * 100)
+        if m == self.model[0]:
+            LOGGER.info(f"{'time (ms)':>10s} {'GFLOPs':>10s} {'params':>10s}  module")
+        LOGGER.info(f'{dt[-1]:10.2f} {o:10.2f} {m.np:10.0f}  {m.type}')
+        if c:
+            LOGGER.info(f"{sum(dt):10.2f} {'-':>10s} {'-':>10s}  Total")
+
+    def fuse(self):  # fuse model Conv2d() + BatchNorm2d() layers
+        LOGGER.info('Fusing layers... ')
+        for m in self.model.modules():
+            if isinstance(m, (RepConvN)) and hasattr(m, 'fuse_convs'):
+                m.fuse_convs()
+                m.forward = m.forward_fuse  # update forward
+            if isinstance(m, (Conv, DWConv)) and hasattr(m, 'bn'):
+                m.conv = fuse_conv_and_bn(m.conv, m.bn)  # update conv
+                delattr(m, 'bn')  # remove batchnorm
+                m.forward = m.forward_fuse  # update forward
+        self.info()
+        return self
+
+    def info(self, verbose=False, img_size=640):  # print model information
+        model_info(self, verbose, img_size)
+
+    def _apply(self, fn):
+        # Apply to(), cpu(), cuda(), half() to model tensors that are not parameters or registered buffers
+        self = super()._apply(fn)
+        m = self.model[-1]  # Detect()
+        if isinstance(m, (Detect, DualDetect, TripleDetect, DDetect, DualDDetect, TripleDDetect, Segment, DSegment, DualDSegment, Panoptic)):
+            m.stride = fn(m.stride)
+            m.anchors = fn(m.anchors)
+            m.strides = fn(m.strides)
+            # m.grid = list(map(fn, m.grid))
+        return self
+
+
+class DetectionModel(BaseModel):
+    # YOLO detection model
+    def __init__(self, cfg='yolo.yaml', ch=3, nc=None, anchors=None):  # model, input channels, number of classes
+        super().__init__()
+        if isinstance(cfg, dict):
+            self.yaml = cfg  # model dict
+        else:  # is *.yaml
+            import yaml  # for torch hub
+            self.yaml_file = Path(cfg).name
+            with open(cfg, encoding='ascii', errors='ignore') as f:
+                self.yaml = yaml.safe_load(f)  # model dict
+
+        # Define model
+        ch = self.yaml['ch'] = self.yaml.get('ch', ch)  # input channels
+        if nc and nc != self.yaml['nc']:
+            LOGGER.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}")
+            self.yaml['nc'] = nc  # override yaml value
+        if anchors:
+            LOGGER.info(f'Overriding model.yaml anchors with anchors={anchors}')
+            self.yaml['anchors'] = round(anchors)  # override yaml value
+        self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch])  # model, savelist
+        self.names = [str(i) for i in range(self.yaml['nc'])]  # default names
+        self.inplace = self.yaml.get('inplace', True)
+
+        # Build strides, anchors
+        m = self.model[-1]  # Detect()
+        if isinstance(m, (Detect, DDetect, Segment, DSegment, Panoptic)):
+            s = 256  # 2x min stride
+            m.inplace = self.inplace
+            forward = lambda x: self.forward(x)[0] if isinstance(m, (Segment, DSegment, Panoptic)) else self.forward(x)
+            m.stride = torch.tensor([s / x.shape[-2] for x in forward(torch.zeros(1, ch, s, s))])  # forward
+            # check_anchor_order(m)
+            # m.anchors /= m.stride.view(-1, 1, 1)
+            self.stride = m.stride
+            m.bias_init()  # only run once
+        if isinstance(m, (DualDetect, TripleDetect, DualDDetect, TripleDDetect, DualDSegment)):
+            s = 256  # 2x min stride
+            m.inplace = self.inplace
+            forward = lambda x: self.forward(x)[0][0] if isinstance(m, (DualDSegment)) else self.forward(x)[0]
+            m.stride = torch.tensor([s / x.shape[-2] for x in forward(torch.zeros(1, ch, s, s))])  # forward
+            # check_anchor_order(m)
+            # m.anchors /= m.stride.view(-1, 1, 1)
+            self.stride = m.stride
+            m.bias_init()  # only run once
+
+        # Init weights, biases
+        initialize_weights(self)
+        self.info()
+        LOGGER.info('')
+
+    def forward(self, x, augment=False, profile=False, visualize=False):
+        if augment:
+            return self._forward_augment(x)  # augmented inference, None
+        return self._forward_once(x, profile, visualize)  # single-scale inference, train
+
+    def _forward_augment(self, x):
+        img_size = x.shape[-2:]  # height, width
+        s = [1, 0.83, 0.67]  # scales
+        f = [None, 3, None]  # flips (2-ud, 3-lr)
+        y = []  # outputs
+        for si, fi in zip(s, f):
+            xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max()))
+            yi = self._forward_once(xi)[0]  # forward
+            # cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1])  # save
+            yi = self._descale_pred(yi, fi, si, img_size)
+            y.append(yi)
+        y = self._clip_augmented(y)  # clip augmented tails
+        return torch.cat(y, 1), None  # augmented inference, train
+
+    def _descale_pred(self, p, flips, scale, img_size):
+        # de-scale predictions following augmented inference (inverse operation)
+        if self.inplace:
+            p[..., :4] /= scale  # de-scale
+            if flips == 2:
+                p[..., 1] = img_size[0] - p[..., 1]  # de-flip ud
+            elif flips == 3:
+                p[..., 0] = img_size[1] - p[..., 0]  # de-flip lr
+        else:
+            x, y, wh = p[..., 0:1] / scale, p[..., 1:2] / scale, p[..., 2:4] / scale  # de-scale
+            if flips == 2:
+                y = img_size[0] - y  # de-flip ud
+            elif flips == 3:
+                x = img_size[1] - x  # de-flip lr
+            p = torch.cat((x, y, wh, p[..., 4:]), -1)
+        return p
+
+    def _clip_augmented(self, y):
+        # Clip YOLO augmented inference tails
+        nl = self.model[-1].nl  # number of detection layers (P3-P5)
+        g = sum(4 ** x for x in range(nl))  # grid points
+        e = 1  # exclude layer count
+        i = (y[0].shape[1] // g) * sum(4 ** x for x in range(e))  # indices
+        y[0] = y[0][:, :-i]  # large
+        i = (y[-1].shape[1] // g) * sum(4 ** (nl - 1 - x) for x in range(e))  # indices
+        y[-1] = y[-1][:, i:]  # small
+        return y
+
+
+Model = DetectionModel  # retain YOLO 'Model' class for backwards compatibility
+
+
+class SegmentationModel(DetectionModel):
+    # YOLO segmentation model
+    def __init__(self, cfg='yolo-seg.yaml', ch=3, nc=None, anchors=None):
+        super().__init__(cfg, ch, nc, anchors)
+
+
+class ClassificationModel(BaseModel):
+    # YOLO classification model
+    def __init__(self, cfg=None, model=None, nc=1000, cutoff=10):  # yaml, model, number of classes, cutoff index
+        super().__init__()
+        self._from_detection_model(model, nc, cutoff) if model is not None else self._from_yaml(cfg)
+
+    def _from_detection_model(self, model, nc=1000, cutoff=10):
+        # Create a YOLO classification model from a YOLO detection model
+        if isinstance(model, DetectMultiBackend):
+            model = model.model  # unwrap DetectMultiBackend
+        model.model = model.model[:cutoff]  # backbone
+        m = model.model[-1]  # last layer
+        ch = m.conv.in_channels if hasattr(m, 'conv') else m.cv1.conv.in_channels  # ch into module
+        c = Classify(ch, nc)  # Classify()
+        c.i, c.f, c.type = m.i, m.f, 'models.common.Classify'  # index, from, type
+        model.model[-1] = c  # replace
+        self.model = model.model
+        self.stride = model.stride
+        self.save = []
+        self.nc = nc
+
+    def _from_yaml(self, cfg):
+        # Create a YOLO classification model from a *.yaml file
+        self.model = None
+
+
+def parse_model(d, ch):  # model_dict, input_channels(3)
+    # Parse a YOLO model.yaml dictionary
+    LOGGER.info(f"\n{'':>3}{'from':>18}{'n':>3}{'params':>10}  {'module':<40}{'arguments':<30}")
+    anchors, nc, gd, gw, act = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple'], d.get('activation')
+    if act:
+        Conv.default_act = eval(act)  # redefine default activation, i.e. Conv.default_act = nn.SiLU()
+        RepConvN.default_act = eval(act)  # redefine default activation, i.e. Conv.default_act = nn.SiLU()
+        LOGGER.info(f"{colorstr('activation:')} {act}")  # print
+    na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors  # number of anchors
+    no = na * (nc + 5)  # number of outputs = anchors * (classes + 5)
+
+    layers, save, c2 = [], [], ch[-1]  # layers, savelist, ch out
+    for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']):  # from, number, module, args
+        m = eval(m) if isinstance(m, str) else m  # eval strings
+        for j, a in enumerate(args):
+            with contextlib.suppress(NameError):
+                args[j] = eval(a) if isinstance(a, str) else a  # eval strings
+
+        n = n_ = max(round(n * gd), 1) if n > 1 else n  # depth gain
+        if m in {
+            Conv, AConv, ConvTranspose, 
+            Bottleneck, SPP, SPPF, DWConv, BottleneckCSP, nn.ConvTranspose2d, DWConvTranspose2d, SPPCSPC, ADown,
+            ELAN1, RepNCSPELAN4, SPPELAN}:
+            c1, c2 = ch[f], args[0]
+            if c2 != no:  # if not output
+                c2 = make_divisible(c2 * gw, 8)
+
+            args = [c1, c2, *args[1:]]
+            if m in {BottleneckCSP, SPPCSPC}:
+                args.insert(2, n)  # number of repeats
+                n = 1
+        elif m is nn.BatchNorm2d:
+            args = [ch[f]]
+        elif m is Concat:
+            c2 = sum(ch[x] for x in f)
+        elif m is Shortcut:
+            c2 = ch[f[0]]
+        elif m is ReOrg:
+            c2 = ch[f] * 4
+        elif m is CBLinear:
+            c2 = args[0]
+            c1 = ch[f]
+            args = [c1, c2, *args[1:]]
+        elif m is CBFuse:
+            c2 = ch[f[-1]]
+        # TODO: channel, gw, gd
+        elif m in {Detect, DualDetect, TripleDetect, DDetect, DualDDetect, TripleDDetect, Segment, DSegment, DualDSegment, Panoptic}:
+            args.append([ch[x] for x in f])
+            # if isinstance(args[1], int):  # number of anchors
+            #     args[1] = [list(range(args[1] * 2))] * len(f)
+            if m in {Segment, DSegment, DualDSegment, Panoptic}:
+                args[2] = make_divisible(args[2] * gw, 8)
+        elif m is Contract:
+            c2 = ch[f] * args[0] ** 2
+        elif m is Expand:
+            c2 = ch[f] // args[0] ** 2
+        else:
+            c2 = ch[f]
+
+        m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args)  # module
+        t = str(m)[8:-2].replace('__main__.', '')  # module type
+        np = sum(x.numel() for x in m_.parameters())  # number params
+        m_.i, m_.f, m_.type, m_.np = i, f, t, np  # attach index, 'from' index, type, number params
+        LOGGER.info(f'{i:>3}{str(f):>18}{n_:>3}{np:10.0f}  {t:<40}{str(args):<30}')  # print
+        save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1)  # append to savelist
+        layers.append(m_)
+        if i == 0:
+            ch = []
+        ch.append(c2)
+    return nn.Sequential(*layers), sorted(save)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--cfg', type=str, default='yolo.yaml', help='model.yaml')
+    parser.add_argument('--batch-size', type=int, default=1, help='total batch size for all GPUs')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--profile', action='store_true', help='profile model speed')
+    parser.add_argument('--line-profile', action='store_true', help='profile model speed layer by layer')
+    parser.add_argument('--test', action='store_true', help='test all yolo*.yaml')
+    opt = parser.parse_args()
+    opt.cfg = check_yaml(opt.cfg)  # check YAML
+    print_args(vars(opt))
+    device = select_device(opt.device)
+
+    # Create model
+    im = torch.rand(opt.batch_size, 3, 640, 640).to(device)
+    model = Model(opt.cfg).to(device)
+    model.eval()
+
+    # Options
+    if opt.line_profile:  # profile layer by layer
+        model(im, profile=True)
+
+    elif opt.profile:  # profile forward-backward
+        results = profile(input=im, ops=[model], n=3)
+
+    elif opt.test:  # test all models
+        for cfg in Path(ROOT / 'models').rglob('yolo*.yaml'):
+            try:
+                _ = Model(cfg)
+            except Exception as e:
+                print(f'Error in {cfg}: {e}')
+
+    else:  # report fused model summary
+        model.fuse()

yolov9/panoptic/predict.py

@@ -0,0 +1,246 @@
+import argparse
+import os
+import platform
+import sys
+from pathlib import Path
+
+import torch
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.common import DetectMultiBackend
+from utils.dataloaders import IMG_FORMATS, VID_FORMATS, LoadImages, LoadScreenshots, LoadStreams
+from utils.general import (LOGGER, Profile, check_file, check_img_size, check_imshow, check_requirements, colorstr, cv2,
+                           increment_path, non_max_suppression, print_args, scale_boxes, scale_segments,
+                           strip_optimizer, xyxy2xywh)
+from utils.plots import Annotator, colors, save_one_box
+from utils.segment.general import masks2segments, process_mask
+from utils.torch_utils import select_device, smart_inference_mode
+
+
+@smart_inference_mode()
+def run(
+    weights=ROOT / 'yolo-pan.pt',  # model.pt path(s)
+    source=ROOT / 'data/images',  # file/dir/URL/glob/screen/0(webcam)
+    data=ROOT / 'data/coco128.yaml',  # dataset.yaml path
+    imgsz=(640, 640),  # inference size (height, width)
+    conf_thres=0.25,  # confidence threshold
+    iou_thres=0.45,  # NMS IOU threshold
+    max_det=1000,  # maximum detections per image
+    device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+    view_img=False,  # show results
+    save_txt=False,  # save results to *.txt
+    save_conf=False,  # save confidences in --save-txt labels
+    save_crop=False,  # save cropped prediction boxes
+    nosave=False,  # do not save images/videos
+    classes=None,  # filter by class: --class 0, or --class 0 2 3
+    agnostic_nms=False,  # class-agnostic NMS
+    augment=False,  # augmented inference
+    visualize=False,  # visualize features
+    update=False,  # update all models
+    project=ROOT / 'runs/predict-seg',  # save results to project/name
+    name='exp',  # save results to project/name
+    exist_ok=False,  # existing project/name ok, do not increment
+    line_thickness=3,  # bounding box thickness (pixels)
+    hide_labels=False,  # hide labels
+    hide_conf=False,  # hide confidences
+    half=False,  # use FP16 half-precision inference
+    dnn=False,  # use OpenCV DNN for ONNX inference
+    vid_stride=1,  # video frame-rate stride
+    retina_masks=False,
+):
+    source = str(source)
+    save_img = not nosave and not source.endswith('.txt')  # save inference images
+    is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS)
+    is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://'))
+    webcam = source.isnumeric() or source.endswith('.txt') or (is_url and not is_file)
+    screenshot = source.lower().startswith('screen')
+    if is_url and is_file:
+        source = check_file(source)  # download
+
+    # Directories
+    save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+    (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+    # Load model
+    device = select_device(device)
+    model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+    stride, names, pt = model.stride, model.names, model.pt
+    imgsz = check_img_size(imgsz, s=stride)  # check image size
+
+    # Dataloader
+    bs = 1  # batch_size
+    if webcam:
+        view_img = check_imshow(warn=True)
+        dataset = LoadStreams(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)
+        bs = len(dataset)
+    elif screenshot:
+        dataset = LoadScreenshots(source, img_size=imgsz, stride=stride, auto=pt)
+    else:
+        dataset = LoadImages(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)
+    vid_path, vid_writer = [None] * bs, [None] * bs
+
+    # Run inference
+    model.warmup(imgsz=(1 if pt else bs, 3, *imgsz))  # warmup
+    seen, windows, dt = 0, [], (Profile(), Profile(), Profile())
+    for path, im, im0s, vid_cap, s in dataset:
+        with dt[0]:
+            im = torch.from_numpy(im).to(model.device)
+            im = im.half() if model.fp16 else im.float()  # uint8 to fp16/32
+            im /= 255  # 0 - 255 to 0.0 - 1.0
+            if len(im.shape) == 3:
+                im = im[None]  # expand for batch dim
+
+        # Inference
+        with dt[1]:
+            visualize = increment_path(save_dir / Path(path).stem, mkdir=True) if visualize else False
+            pred, proto = model(im, augment=augment, visualize=visualize)[:2]
+
+        # NMS
+        with dt[2]:
+            pred = non_max_suppression(pred, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det, nm=32)
+
+        # Second-stage classifier (optional)
+        # pred = utils.general.apply_classifier(pred, classifier_model, im, im0s)
+
+        # Process predictions
+        for i, det in enumerate(pred):  # per image
+            seen += 1
+            if webcam:  # batch_size >= 1
+                p, im0, frame = path[i], im0s[i].copy(), dataset.count
+                s += f'{i}: '
+            else:
+                p, im0, frame = path, im0s.copy(), getattr(dataset, 'frame', 0)
+
+            p = Path(p)  # to Path
+            save_path = str(save_dir / p.name)  # im.jpg
+            txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}')  # im.txt
+            s += '%gx%g ' % im.shape[2:]  # print string
+            imc = im0.copy() if save_crop else im0  # for save_crop
+            annotator = Annotator(im0, line_width=line_thickness, example=str(names))
+            if len(det):
+                masks = process_mask(proto[i], det[:, 6:], det[:, :4], im.shape[2:], upsample=True)  # HWC
+                det[:, :4] = scale_boxes(im.shape[2:], det[:, :4], im0.shape).round()  # rescale boxes to im0 size
+
+                # Segments
+                if save_txt:
+                    segments = reversed(masks2segments(masks))
+                    segments = [scale_segments(im.shape[2:], x, im0.shape, normalize=True) for x in segments]
+
+                # Print results
+                for c in det[:, 5].unique():
+                    n = (det[:, 5] == c).sum()  # detections per class
+                    s += f"{n} {names[int(c)]}{'s' * (n > 1)}, "  # add to string
+
+                # Mask plotting
+                annotator.masks(masks,
+                                colors=[colors(x, True) for x in det[:, 5]],
+                                im_gpu=None if retina_masks else im[i])
+
+                # Write results
+                for j, (*xyxy, conf, cls) in enumerate(reversed(det[:, :6])):
+                    if save_txt:  # Write to file
+                        segj = segments[j].reshape(-1)  # (n,2) to (n*2)
+                        line = (cls, *segj, conf) if save_conf else (cls, *segj)  # label format
+                        with open(f'{txt_path}.txt', 'a') as f:
+                            f.write(('%g ' * len(line)).rstrip() % line + '\n')
+
+                    if save_img or save_crop or view_img:  # Add bbox to image
+                        c = int(cls)  # integer class
+                        label = None if hide_labels else (names[c] if hide_conf else f'{names[c]} {conf:.2f}')
+                        annotator.box_label(xyxy, label, color=colors(c, True))
+                        # annotator.draw.polygon(segments[j], outline=colors(c, True), width=3)
+                    if save_crop:
+                        save_one_box(xyxy, imc, file=save_dir / 'crops' / names[c] / f'{p.stem}.jpg', BGR=True)
+
+            # Stream results
+            im0 = annotator.result()
+            if view_img:
+                if platform.system() == 'Linux' and p not in windows:
+                    windows.append(p)
+                    cv2.namedWindow(str(p), cv2.WINDOW_NORMAL | cv2.WINDOW_KEEPRATIO)  # allow window resize (Linux)
+                    cv2.resizeWindow(str(p), im0.shape[1], im0.shape[0])
+                cv2.imshow(str(p), im0)
+                if cv2.waitKey(1) == ord('q'):  # 1 millisecond
+                    exit()
+
+            # Save results (image with detections)
+            if save_img:
+                if dataset.mode == 'image':
+                    cv2.imwrite(save_path, im0)
+                else:  # 'video' or 'stream'
+                    if vid_path[i] != save_path:  # new video
+                        vid_path[i] = save_path
+                        if isinstance(vid_writer[i], cv2.VideoWriter):
+                            vid_writer[i].release()  # release previous video writer
+                        if vid_cap:  # video
+                            fps = vid_cap.get(cv2.CAP_PROP_FPS)
+                            w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+                            h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                        else:  # stream
+                            fps, w, h = 30, im0.shape[1], im0.shape[0]
+                        save_path = str(Path(save_path).with_suffix('.mp4'))  # force *.mp4 suffix on results videos
+                        vid_writer[i] = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
+                    vid_writer[i].write(im0)
+
+        # Print time (inference-only)
+        LOGGER.info(f"{s}{'' if len(det) else '(no detections), '}{dt[1].dt * 1E3:.1f}ms")
+
+    # Print results
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {(1, 3, *imgsz)}' % t)
+    if save_txt or save_img:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    if update:
+        strip_optimizer(weights[0])  # update model (to fix SourceChangeWarning)
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo-pan.pt', help='model path(s)')
+    parser.add_argument('--source', type=str, default=ROOT / 'data/images', help='file/dir/URL/glob/screen/0(webcam)')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='(optional) dataset.yaml path')
+    parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w')
+    parser.add_argument('--conf-thres', type=float, default=0.25, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold')
+    parser.add_argument('--max-det', type=int, default=1000, help='maximum detections per image')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--view-img', action='store_true', help='show results')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
+    parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes')
+    parser.add_argument('--nosave', action='store_true', help='do not save images/videos')
+    parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --classes 0, or --classes 0 2 3')
+    parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--visualize', action='store_true', help='visualize features')
+    parser.add_argument('--update', action='store_true', help='update all models')
+    parser.add_argument('--project', default=ROOT / 'runs/predict-seg', help='save results to project/name')
+    parser.add_argument('--name', default='exp', help='save results to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)')
+    parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels')
+    parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    parser.add_argument('--vid-stride', type=int, default=1, help='video frame-rate stride')
+    parser.add_argument('--retina-masks', action='store_true', help='whether to plot masks in native resolution')
+    opt = parser.parse_args()
+    opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1  # expand
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    check_requirements(exclude=('tensorboard', 'thop'))
+    run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

yolov9/panoptic/train.py

@@ -0,0 +1,662 @@
+import argparse
+import math
+import os
+import random
+import sys
+import time
+from copy import deepcopy
+from datetime import datetime
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import yaml
+from torch.optim import lr_scheduler
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import panoptic.val as validate  # for end-of-epoch mAP
+from models.experimental import attempt_load
+from models.yolo import SegmentationModel
+from utils.autoanchor import check_anchors
+from utils.autobatch import check_train_batch_size
+from utils.callbacks import Callbacks
+from utils.downloads import attempt_download, is_url
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, check_amp, check_dataset, check_file, check_git_info,
+                           check_git_status, check_img_size, check_requirements, check_suffix, check_yaml, colorstr,
+                           get_latest_run, increment_path, init_seeds, intersect_dicts, labels_to_class_weights,
+                           labels_to_image_weights, one_cycle, one_flat_cycle, print_args, print_mutation, strip_optimizer, yaml_save)
+from utils.loggers import GenericLogger
+from utils.plots import plot_evolve, plot_labels
+from utils.panoptic.dataloaders import create_dataloader
+from utils.panoptic.loss_tal import ComputeLoss
+from utils.panoptic.metrics import KEYS, fitness
+from utils.panoptic.plots import plot_images_and_masks, plot_results_with_masks
+from utils.torch_utils import (EarlyStopping, ModelEMA, de_parallel, select_device, smart_DDP, smart_optimizer,
+                               smart_resume, torch_distributed_zero_first)
+
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+GIT_INFO = None#check_git_info()
+
+
+def train(hyp, opt, device, callbacks):  # hyp is path/to/hyp.yaml or hyp dictionary
+    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze, mask_ratio = \
+        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
+        opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze, opt.mask_ratio
+    # callbacks.run('on_pretrain_routine_start')
+
+    # Directories
+    w = save_dir / 'weights'  # weights dir
+    (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir
+    last, best = w / 'last.pt', w / 'best.pt'
+
+    # Hyperparameters
+    if isinstance(hyp, str):
+        with open(hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+    LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
+    opt.hyp = hyp.copy()  # for saving hyps to checkpoints
+
+    # Save run settings
+    if not evolve:
+        yaml_save(save_dir / 'hyp.yaml', hyp)
+        yaml_save(save_dir / 'opt.yaml', vars(opt))
+
+    # Loggers
+    data_dict = None
+    if RANK in {-1, 0}:
+        logger = GenericLogger(opt=opt, console_logger=LOGGER)
+
+    # Config
+    plots = not evolve and not opt.noplots  # create plots
+    overlap = not opt.no_overlap
+    cuda = device.type != 'cpu'
+    init_seeds(opt.seed + 1 + RANK, deterministic=True)
+    with torch_distributed_zero_first(LOCAL_RANK):
+        data_dict = data_dict or check_dataset(data)  # check if None
+    train_path, val_path = data_dict['train'], data_dict['val']
+    nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
+    names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
+    #is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt')  # COCO dataset
+    is_coco = isinstance(val_path, str) and val_path.endswith('val2017.txt')  # COCO dataset
+
+    # Model
+    check_suffix(weights, '.pt')  # check weights
+    pretrained = weights.endswith('.pt')
+    if pretrained:
+        with torch_distributed_zero_first(LOCAL_RANK):
+            weights = attempt_download(weights)  # download if not found locally
+        ckpt = torch.load(weights, map_location='cpu')  # load checkpoint to CPU to avoid CUDA memory leak
+        model = SegmentationModel(cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device)
+        exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
+        csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+        csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect
+        model.load_state_dict(csd, strict=False)  # load
+        LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}')  # report
+    else:
+        model = SegmentationModel(cfg, ch=3, nc=nc).to(device)  # create
+    amp = check_amp(model)  # check AMP
+
+    # Freeze
+    freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
+    for k, v in model.named_parameters():
+        #v.requires_grad = True  # train all layers
+        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
+        if any(x in k for x in freeze):
+            LOGGER.info(f'freezing {k}')
+            v.requires_grad = False
+
+    # Image size
+    gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple
+
+    # Batch size
+    if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size
+        batch_size = check_train_batch_size(model, imgsz, amp)
+        logger.update_params({"batch_size": batch_size})
+        # loggers.on_params_update({"batch_size": batch_size})
+
+    # Optimizer
+    nbs = 64  # nominal batch size
+    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
+    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
+    optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])
+
+    # Scheduler
+    if opt.cos_lr:
+        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    elif opt.flat_cos_lr:
+        lf = one_flat_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    elif opt.fixed_lr:
+        lf = lambda x: 1.0
+    elif opt.poly_lr:
+        power = 0.9
+        lf = lambda x: ((1 - (x / epochs)) ** power) * (1.0 - hyp['lrf']) + hyp['lrf']
+    else:
+        lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)  # plot_lr_scheduler(optimizer, scheduler, epochs)
+
+    # EMA
+    ema = ModelEMA(model) if RANK in {-1, 0} else None
+
+    # Resume
+    best_fitness, start_epoch = 0.0, 0
+    if pretrained:
+        if resume:
+            best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)
+        del ckpt, csd
+
+    # DP mode
+    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
+        LOGGER.warning('WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.')
+        model = torch.nn.DataParallel(model)
+
+    # SyncBatchNorm
+    if opt.sync_bn and cuda and RANK != -1:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
+        LOGGER.info('Using SyncBatchNorm()')
+
+    # Trainloader
+    train_loader, dataset = create_dataloader(
+        train_path,
+        imgsz,
+        batch_size // WORLD_SIZE,
+        gs,
+        single_cls,
+        hyp=hyp,
+        augment=True,
+        cache=None if opt.cache == 'val' else opt.cache,
+        rect=opt.rect,
+        rank=LOCAL_RANK,
+        workers=workers,
+        image_weights=opt.image_weights,
+        close_mosaic=opt.close_mosaic != 0,
+        quad=opt.quad,
+        prefix=colorstr('train: '),
+        shuffle=True,
+        mask_downsample_ratio=mask_ratio,
+        overlap_mask=overlap,
+    )
+    labels = np.concatenate(dataset.labels, 0)
+    mlc = int(labels[:, 0].max())  # max label class
+    assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
+
+    # Process 0
+    if RANK in {-1, 0}:
+        val_loader = create_dataloader(val_path,
+                                       imgsz,
+                                       batch_size // WORLD_SIZE * 2,
+                                       gs,
+                                       single_cls,
+                                       hyp=hyp,
+                                       cache=None if noval else opt.cache,
+                                       rect=True,
+                                       rank=-1,
+                                       workers=workers * 2,
+                                       pad=0.5,
+                                       mask_downsample_ratio=mask_ratio,
+                                       overlap_mask=overlap,
+                                       prefix=colorstr('val: '))[0]
+
+        if not resume:
+            #if not opt.noautoanchor:
+            #    check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor
+            model.half().float()  # pre-reduce anchor precision
+
+            if plots:
+                plot_labels(labels, names, save_dir)
+        # callbacks.run('on_pretrain_routine_end', labels, names)
+
+    # DDP mode
+    if cuda and RANK != -1:
+        model = smart_DDP(model)
+
+    # Model attributes
+    nl = de_parallel(model).model[-1].nl  # number of detection layers (to scale hyps)
+    #hyp['box'] *= 3 / nl  # scale to layers
+    #hyp['cls'] *= nc / 80 * 3 / nl  # scale to classes and layers
+    #hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers
+    hyp['label_smoothing'] = opt.label_smoothing
+    model.nc = nc  # attach number of classes to model
+    model.hyp = hyp  # attach hyperparameters to model
+    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
+    model.names = names
+
+    # Start training
+    t0 = time.time()
+    nb = len(train_loader)  # number of batches
+    nw = max(round(hyp['warmup_epochs'] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)
+    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
+    last_opt_step = -1
+    maps = np.zeros(nc)  # mAP per class
+    results = (0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0)  # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls)
+    scheduler.last_epoch = start_epoch - 1  # do not move
+    scaler = torch.cuda.amp.GradScaler(enabled=amp)
+    stopper, stop = EarlyStopping(patience=opt.patience), False
+    compute_loss = ComputeLoss(model, overlap=overlap)  # init loss class
+    # callbacks.run('on_train_start')
+    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
+                f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
+                f"Logging results to {colorstr('bold', save_dir)}\n"
+                f'Starting training for {epochs} epochs...')
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        # callbacks.run('on_train_epoch_start')
+        model.train()
+
+        # Update image weights (optional, single-GPU only)
+        if opt.image_weights:
+            cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
+            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
+            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx
+        if epoch == (epochs - opt.close_mosaic):
+            LOGGER.info("Closing dataloader mosaic")
+            dataset.mosaic = False
+
+        # Update mosaic border (optional)
+        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
+        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders
+
+        mloss = torch.zeros(6, device=device)  # mean losses
+        if RANK != -1:
+            train_loader.sampler.set_epoch(epoch)
+        pbar = enumerate(train_loader)
+        LOGGER.info(('\n' + '%11s' * 10) %
+                    ('Epoch', 'GPU_mem', 'box_loss', 'seg_loss', 'cls_loss', 'dfl_loss', 'fcl_loss', 'dic_loss', 'Instances', 'Size'))
+        if RANK in {-1, 0}:
+            pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar
+        optimizer.zero_grad()
+        for i, (imgs, targets, paths, _, masks, semasks) in pbar:  # batch ------------------------------------------------------
+            # callbacks.run('on_train_batch_start')
+            #print(imgs.shape)
+            #print(semasks.shape)
+            #print(masks.shape)
+            ni = i + nb * epoch  # number integrated batches (since train start)
+            imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0
+
+            # Warmup
+            if ni <= nw:
+                xi = [0, nw]  # x interp
+                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
+                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
+                for j, x in enumerate(optimizer.param_groups):
+                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
+                    x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
+                    if 'momentum' in x:
+                        x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
+
+            # Multi-scale
+            if opt.multi_scale:
+                sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # size
+                sf = sz / max(imgs.shape[2:])  # scale factor
+                if sf != 1:
+                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
+                    imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
+
+            # Forward
+            with torch.cuda.amp.autocast(amp):
+                pred = model(imgs)  # forward
+                loss, loss_items = compute_loss(pred, targets.to(device), masks=masks.to(device).float(),
+                                                semasks=semasks.to(device).float())
+                if RANK != -1:
+                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
+                if opt.quad:
+                    loss *= 4.
+
+            # Backward
+            torch.use_deterministic_algorithms(False)
+            scaler.scale(loss).backward()
+
+            # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
+            if ni - last_opt_step >= accumulate:
+                scaler.unscale_(optimizer)  # unscale gradients
+                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients
+                scaler.step(optimizer)  # optimizer.step
+                scaler.update()
+                optimizer.zero_grad()
+                if ema:
+                    ema.update(model)
+                last_opt_step = ni
+
+            # Log
+            if RANK in {-1, 0}:
+                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
+                mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
+                pbar.set_description(('%11s' * 2 + '%11.4g' * 8) %
+                                     (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
+                # callbacks.run('on_train_batch_end', model, ni, imgs, targets, paths)
+                # if callbacks.stop_training:
+                #    return
+
+                # Mosaic plots
+                if plots:
+                    if ni < 10:
+                        plot_images_and_masks(imgs, targets, masks, semasks, paths, save_dir / f"train_batch{ni}.jpg")
+                    if ni == 10:
+                        files = sorted(save_dir.glob('train*.jpg'))
+                        logger.log_images(files, "Mosaics", epoch)
+            # end batch ------------------------------------------------------------------------------------------------
+
+        # Scheduler
+        lr = [x['lr'] for x in optimizer.param_groups]  # for loggers
+        scheduler.step()
+
+        if RANK in {-1, 0}:
+            # mAP
+            # callbacks.run('on_train_epoch_end', epoch=epoch)
+            ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
+            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
+            if not noval or final_epoch:  # Calculate mAP
+                if (opt.save_period > 0 and epoch % opt.save_period == 0) or (epoch > (epochs - 2 * opt.close_mosaic)):
+                    results, maps, _ = validate.run(data_dict,
+                                                    batch_size=batch_size // WORLD_SIZE * 2,
+                                                    imgsz=imgsz,
+                                                    half=amp,
+                                                    model=ema.ema,
+                                                    single_cls=single_cls,
+                                                    dataloader=val_loader,
+                                                    save_dir=save_dir,
+                                                    plots=False,
+                                                    callbacks=callbacks,
+                                                    compute_loss=compute_loss,
+                                                    mask_downsample_ratio=mask_ratio,
+                                                    overlap=overlap)
+
+            # Update best mAP
+            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, mAP@.5, mAP@.5-.95]
+            stop = stopper(epoch=epoch, fitness=fi)  # early stop check
+            if fi > best_fitness:
+                best_fitness = fi
+            log_vals = list(mloss) + list(results) + lr
+            # callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
+            # Log val metrics and media
+            metrics_dict = dict(zip(KEYS, log_vals))
+            logger.log_metrics(metrics_dict, epoch)
+
+            # Save model
+            if (not nosave) or (final_epoch and not evolve):  # if save
+                ckpt = {
+                    'epoch': epoch,
+                    'best_fitness': best_fitness,
+                    'model': deepcopy(de_parallel(model)).half(),
+                    'ema': deepcopy(ema.ema).half(),
+                    'updates': ema.updates,
+                    'optimizer': optimizer.state_dict(),
+                    'opt': vars(opt),
+                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
+                    'date': datetime.now().isoformat()}
+
+                # Save last, best and delete
+                torch.save(ckpt, last)
+                if best_fitness == fi:
+                    torch.save(ckpt, best)
+                if opt.save_period > 0 and epoch % opt.save_period == 0:
+                    torch.save(ckpt, w / f'epoch{epoch}.pt')
+                    logger.log_model(w / f'epoch{epoch}.pt')
+                del ckpt
+                # callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
+
+        # EarlyStopping
+        if RANK != -1:  # if DDP training
+            broadcast_list = [stop if RANK == 0 else None]
+            dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks
+            if RANK != 0:
+                stop = broadcast_list[0]
+        if stop:
+            break  # must break all DDP ranks
+
+        # end epoch ----------------------------------------------------------------------------------------------------
+    # end training -----------------------------------------------------------------------------------------------------
+    if RANK in {-1, 0}:
+        LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
+        for f in last, best:
+            if f.exists():
+                strip_optimizer(f)  # strip optimizers
+                if f is best:
+                    LOGGER.info(f'\nValidating {f}...')
+                    results, _, _ = validate.run(
+                        data_dict,
+                        batch_size=batch_size // WORLD_SIZE * 2,
+                        imgsz=imgsz,
+                        model=attempt_load(f, device).half(),
+                        iou_thres=0.65 if is_coco else 0.60,  # best pycocotools at iou 0.65
+                        single_cls=single_cls,
+                        dataloader=val_loader,
+                        save_dir=save_dir,
+                        save_json=is_coco,
+                        verbose=True,
+                        plots=plots,
+                        callbacks=callbacks,
+                        compute_loss=compute_loss,
+                        mask_downsample_ratio=mask_ratio,
+                        overlap=overlap)  # val best model with plots
+                    if is_coco:
+                        # callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
+                        metrics_dict = dict(zip(KEYS, list(mloss) + list(results) + lr))
+                        logger.log_metrics(metrics_dict, epoch)
+
+        # callbacks.run('on_train_end', last, best, epoch, results)
+        # on train end callback using genericLogger
+        logger.log_metrics(dict(zip(KEYS[6:22], results)), epochs)
+        if not opt.evolve:
+            logger.log_model(best, epoch)
+        if plots:
+            plot_results_with_masks(file=save_dir / 'results.csv')  # save results.png
+            files = ['results.png', 'confusion_matrix.png', *(f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R'))]
+            files = [(save_dir / f) for f in files if (save_dir / f).exists()]  # filter
+            LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}")
+            logger.log_images(files, "Results", epoch + 1)
+            logger.log_images(sorted(save_dir.glob('val*.jpg')), "Validation", epoch + 1)
+    torch.cuda.empty_cache()
+    return results
+
+
+def parse_opt(known=False):
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', type=str, default=ROOT / 'yolo-pan.pt', help='initial weights path')
+    parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128-seg.yaml', help='dataset.yaml path')
+    parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path')
+    parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
+    parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
+    parser.add_argument('--rect', action='store_true', help='rectangular training')
+    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
+    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
+    parser.add_argument('--noval', action='store_true', help='only validate final epoch')
+    parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
+    parser.add_argument('--noplots', action='store_true', help='save no plot files')
+    parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
+    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
+    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
+    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
+    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
+    parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW', 'LION'], default='SGD', help='optimizer')
+    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--project', default=ROOT / 'runs/train-pan', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--quad', action='store_true', help='quad dataloader')
+    parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
+    parser.add_argument('--flat-cos-lr', action='store_true', help='cosine LR scheduler')
+    parser.add_argument('--fixed-lr', action='store_true', help='fixed LR scheduler')
+    parser.add_argument('--poly-lr', action='store_true', help='fixed LR scheduler')
+    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
+    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
+    parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
+    parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
+    parser.add_argument('--seed', type=int, default=0, help='Global training seed')
+    parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
+    parser.add_argument('--close-mosaic', type=int, default=0, help='Experimental')
+
+    # Instance Segmentation Args
+    parser.add_argument('--mask-ratio', type=int, default=4, help='Downsample the truth masks to saving memory')
+    parser.add_argument('--no-overlap', action='store_true', help='Overlap masks train faster at slightly less mAP')
+
+    return parser.parse_known_args()[0] if known else parser.parse_args()
+
+
+def main(opt, callbacks=Callbacks()):
+    # Checks
+    if RANK in {-1, 0}:
+        print_args(vars(opt))
+        #check_git_status()
+        #check_requirements()
+
+    # Resume
+    if opt.resume and not opt.evolve:  # resume from specified or most recent last.pt
+        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
+        opt_yaml = last.parent.parent / 'opt.yaml'  # train options yaml
+        opt_data = opt.data  # original dataset
+        if opt_yaml.is_file():
+            with open(opt_yaml, errors='ignore') as f:
+                d = yaml.safe_load(f)
+        else:
+            d = torch.load(last, map_location='cpu')['opt']
+        opt = argparse.Namespace(**d)  # replace
+        opt.cfg, opt.weights, opt.resume = '', str(last), True  # reinstate
+        if is_url(opt_data):
+            opt.data = check_file(opt_data)  # avoid HUB resume auth timeout
+    else:
+        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
+            check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project)  # checks
+        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
+        if opt.evolve:
+            if opt.project == str(ROOT / 'runs/train'):  # if default project name, rename to runs/evolve
+                opt.project = str(ROOT / 'runs/evolve')
+            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
+        if opt.name == 'cfg':
+            opt.name = Path(opt.cfg).stem  # use model.yaml as name
+        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))
+
+    # DDP mode
+    device = select_device(opt.device, batch_size=opt.batch_size)
+    if LOCAL_RANK != -1:
+        msg = 'is not compatible with YOLO Multi-GPU DDP training'
+        assert not opt.image_weights, f'--image-weights {msg}'
+        assert not opt.evolve, f'--evolve {msg}'
+        assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
+        assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
+        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
+        torch.cuda.set_device(LOCAL_RANK)
+        device = torch.device('cuda', LOCAL_RANK)
+        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
+
+    # Train
+    if not opt.evolve:
+        train(opt.hyp, opt, device, callbacks)
+
+    # Evolve hyperparameters (optional)
+    else:
+        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
+        meta = {
+            'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
+            'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
+            'momentum': (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1
+            'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
+            'warmup_epochs': (1, 0.0, 5.0),  # warmup epochs (fractions ok)
+            'warmup_momentum': (1, 0.0, 0.95),  # warmup initial momentum
+            'warmup_bias_lr': (1, 0.0, 0.2),  # warmup initial bias lr
+            'box': (1, 0.02, 0.2),  # box loss gain
+            'cls': (1, 0.2, 4.0),  # cls loss gain
+            'cls_pw': (1, 0.5, 2.0),  # cls BCELoss positive_weight
+            'obj': (1, 0.2, 4.0),  # obj loss gain (scale with pixels)
+            'obj_pw': (1, 0.5, 2.0),  # obj BCELoss positive_weight
+            'iou_t': (0, 0.1, 0.7),  # IoU training threshold
+            'anchor_t': (1, 2.0, 8.0),  # anchor-multiple threshold
+            'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
+            'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
+            'hsv_h': (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
+            'hsv_s': (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
+            'hsv_v': (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
+            'degrees': (1, 0.0, 45.0),  # image rotation (+/- deg)
+            'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
+            'scale': (1, 0.0, 0.9),  # image scale (+/- gain)
+            'shear': (1, 0.0, 10.0),  # image shear (+/- deg)
+            'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
+            'flipud': (1, 0.0, 1.0),  # image flip up-down (probability)
+            'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
+            'mosaic': (1, 0.0, 1.0),  # image mixup (probability)
+            'mixup': (1, 0.0, 1.0),  # image mixup (probability)
+            'copy_paste': (1, 0.0, 1.0)}  # segment copy-paste (probability)
+
+        with open(opt.hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+            if 'anchors' not in hyp:  # anchors commented in hyp.yaml
+                hyp['anchors'] = 3
+        if opt.noautoanchor:
+            del hyp['anchors'], meta['anchors']
+        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
+        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
+        evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
+        if opt.bucket:
+            os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {evolve_csv}')  # download evolve.csv if exists
+
+        for _ in range(opt.evolve):  # generations to evolve
+            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate
+                # Select parent(s)
+                parent = 'single'  # parent selection method: 'single' or 'weighted'
+                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
+                n = min(5, len(x))  # number of previous results to consider
+                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
+                w = fitness(x) - fitness(x).min() + 1E-6  # weights (sum > 0)
+                if parent == 'single' or len(x) == 1:
+                    # x = x[random.randint(0, n - 1)]  # random selection
+                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
+                elif parent == 'weighted':
+                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination
+
+                # Mutate
+                mp, s = 0.8, 0.2  # mutation probability, sigma
+                npr = np.random
+                npr.seed(int(time.time()))
+                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
+                ng = len(meta)
+                v = np.ones(ng)
+                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
+                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
+                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
+                    hyp[k] = float(x[i + 7] * v[i])  # mutate
+
+            # Constrain to limits
+            for k, v in meta.items():
+                hyp[k] = max(hyp[k], v[1])  # lower limit
+                hyp[k] = min(hyp[k], v[2])  # upper limit
+                hyp[k] = round(hyp[k], 5)  # significant digits
+
+            # Train mutation
+            results = train(hyp.copy(), opt, device, callbacks)
+            callbacks = Callbacks()
+            # Write mutation results
+            print_mutation(KEYS, results, hyp.copy(), save_dir, opt.bucket)
+
+        # Plot results
+        plot_evolve(evolve_csv)
+        LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n'
+                    f"Results saved to {colorstr('bold', save_dir)}\n"
+                    f'Usage example: $ python train.py --hyp {evolve_yaml}')
+
+
+def run(**kwargs):
+    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolo.pt')
+    opt = parse_opt(True)
+    for k, v in kwargs.items():
+        setattr(opt, k, v)
+    main(opt)
+    return opt
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

yolov9/panoptic/val.py

@@ -0,0 +1,597 @@
+import argparse
+import json
+import os
+import sys
+from multiprocessing.pool import ThreadPool
+from pathlib import Path
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import torch.nn.functional as F
+import torchvision.transforms as transforms
+from pycocotools import mask as maskUtils
+from models.common import DetectMultiBackend
+from models.yolo import SegmentationModel
+from utils.callbacks import Callbacks
+from utils.coco_utils import getCocoIds, getMappingId, getMappingIndex
+from utils.general import (LOGGER, NUM_THREADS, TQDM_BAR_FORMAT, Profile, check_dataset, check_img_size,
+                           check_requirements, check_yaml, coco80_to_coco91_class, colorstr, increment_path,
+                           non_max_suppression, print_args, scale_boxes, xywh2xyxy, xyxy2xywh)
+from utils.metrics import ConfusionMatrix, box_iou
+from utils.plots import output_to_target, plot_val_study
+from utils.panoptic.dataloaders import create_dataloader
+from utils.panoptic.general import mask_iou, process_mask, process_mask_upsample, scale_image
+from utils.panoptic.metrics import Metrics, ap_per_class_box_and_mask, Semantic_Metrics
+from utils.panoptic.plots import plot_images_and_masks
+from utils.torch_utils import de_parallel, select_device, smart_inference_mode
+
+
+def save_one_txt(predn, save_conf, shape, file):
+    # Save one txt result
+    gn = torch.tensor(shape)[[1, 0, 1, 0]]  # normalization gain whwh
+    for *xyxy, conf, cls in predn.tolist():
+        xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
+        line = (cls, *xywh, conf) if save_conf else (cls, *xywh)  # label format
+        with open(file, 'a') as f:
+            f.write(('%g ' * len(line)).rstrip() % line + '\n')
+
+
+def save_one_json(predn, jdict, path, class_map, pred_masks):
+    # Save one JSON result {"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}
+    from pycocotools.mask import encode
+
+    def single_encode(x):
+        rle = encode(np.asarray(x[:, :, None], order="F", dtype="uint8"))[0]
+        rle["counts"] = rle["counts"].decode("utf-8")
+        return rle
+
+    image_id = int(path.stem) if path.stem.isnumeric() else path.stem
+    box = xyxy2xywh(predn[:, :4])  # xywh
+    box[:, :2] -= box[:, 2:] / 2  # xy center to top-left corner
+    pred_masks = np.transpose(pred_masks, (2, 0, 1))
+    with ThreadPool(NUM_THREADS) as pool:
+        rles = pool.map(single_encode, pred_masks)
+    for i, (p, b) in enumerate(zip(predn.tolist(), box.tolist())):
+        jdict.append({
+            'image_id': image_id,
+            'category_id': class_map[int(p[5])],
+            'bbox': [round(x, 3) for x in b],
+            'score': round(p[4], 5),
+            'segmentation': rles[i]})
+
+
+def process_batch(detections, labels, iouv, pred_masks=None, gt_masks=None, overlap=False, masks=False):
+    """
+    Return correct prediction matrix
+    Arguments:
+        detections (array[N, 6]), x1, y1, x2, y2, conf, class
+        labels (array[M, 5]), class, x1, y1, x2, y2
+    Returns:
+        correct (array[N, 10]), for 10 IoU levels
+    """
+    if masks:
+        if overlap:
+            nl = len(labels)
+            index = torch.arange(nl, device=gt_masks.device).view(nl, 1, 1) + 1
+            gt_masks = gt_masks.repeat(nl, 1, 1)  # shape(1,640,640) -> (n,640,640)
+            gt_masks = torch.where(gt_masks == index, 1.0, 0.0)
+        if gt_masks.shape[1:] != pred_masks.shape[1:]:
+            gt_masks = F.interpolate(gt_masks[None], pred_masks.shape[1:], mode="bilinear", align_corners=False)[0]
+            gt_masks = gt_masks.gt_(0.5)
+        iou = mask_iou(gt_masks.view(gt_masks.shape[0], -1), pred_masks.view(pred_masks.shape[0], -1))
+    else:  # boxes
+        iou = box_iou(labels[:, 1:], detections[:, :4])
+
+    correct = np.zeros((detections.shape[0], iouv.shape[0])).astype(bool)
+    correct_class = labels[:, 0:1] == detections[:, 5]
+    for i in range(len(iouv)):
+        x = torch.where((iou >= iouv[i]) & correct_class)  # IoU > threshold and classes match
+        if x[0].shape[0]:
+            matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()  # [label, detect, iou]
+            if x[0].shape[0] > 1:
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
+                # matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
+            correct[matches[:, 1].astype(int), i] = True
+    return torch.tensor(correct, dtype=torch.bool, device=iouv.device)
+
+
+@smart_inference_mode()
+def run(
+        data,
+        weights=None,  # model.pt path(s)
+        batch_size=32,  # batch size
+        imgsz=640,  # inference size (pixels)
+        conf_thres=0.001,  # confidence threshold
+        iou_thres=0.6,  # NMS IoU threshold
+        max_det=300,  # maximum detections per image
+        task='val',  # train, val, test, speed or study
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        workers=8,  # max dataloader workers (per RANK in DDP mode)
+        single_cls=False,  # treat as single-class dataset
+        augment=False,  # augmented inference
+        verbose=False,  # verbose output
+        save_txt=False,  # save results to *.txt
+        save_hybrid=False,  # save label+prediction hybrid results to *.txt
+        save_conf=False,  # save confidences in --save-txt labels
+        save_json=False,  # save a COCO-JSON results file
+        project=ROOT / 'runs/val-pan',  # save to project/name
+        name='exp',  # save to project/name
+        exist_ok=False,  # existing project/name ok, do not increment
+        half=True,  # use FP16 half-precision inference
+        dnn=False,  # use OpenCV DNN for ONNX inference
+        model=None,
+        dataloader=None,
+        save_dir=Path(''),
+        plots=True,
+        overlap=False,
+        mask_downsample_ratio=1,
+        compute_loss=None,
+        callbacks=Callbacks(),
+):
+    if save_json:
+        check_requirements(['pycocotools'])
+        process = process_mask_upsample  # more accurate
+    else:
+        process = process_mask  # faster
+
+    # Initialize/load model and set device
+    training = model is not None
+    if training:  # called by train.py
+        device, pt, jit, engine = next(model.parameters()).device, True, False, False  # get model device, PyTorch model
+        half &= device.type != 'cpu'  # half precision only supported on CUDA
+        model.half() if half else model.float()
+        nm = de_parallel(model).model[-1].nm  # number of masks
+    else:  # called directly
+        device = select_device(device, batch_size=batch_size)
+
+        # Directories
+        save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+        (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+        # Load model
+        model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+        stride, pt, jit, engine = model.stride, model.pt, model.jit, model.engine
+        imgsz = check_img_size(imgsz, s=stride)  # check image size
+        half = model.fp16  # FP16 supported on limited backends with CUDA
+        nm = de_parallel(model).model.model[-1].nm if isinstance(model, SegmentationModel) else 32  # number of masks
+        if engine:
+            batch_size = model.batch_size
+        else:
+            device = model.device
+            if not (pt or jit):
+                batch_size = 1  # export.py models default to batch-size 1
+                LOGGER.info(f'Forcing --batch-size 1 square inference (1,3,{imgsz},{imgsz}) for non-PyTorch models')
+
+        # Data
+        data = check_dataset(data)  # check
+
+    # Configure
+    model.eval()
+    cuda = device.type != 'cpu'
+    #is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'coco{os.sep}val2017.txt')  # COCO dataset
+    is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'val2017.txt')  # COCO dataset
+    nc = 1 if single_cls else int(data['nc'])  # number of classes
+    stuff_names = data.get('stuff_names', [])  # names of stuff classes
+    stuff_nc = len(stuff_names)  # number of stuff classes
+    iouv = torch.linspace(0.5, 0.95, 10, device=device)  # iou vector for mAP@0.5:0.95
+    niou = iouv.numel()
+
+    # Semantic Segmentation
+    img_id_list = []
+
+    # Dataloader
+    if not training:
+        if pt and not single_cls:  # check --weights are trained on --data
+            ncm = model.model.nc
+            assert ncm == nc, f'{weights} ({ncm} classes) trained on different --data than what you passed ({nc} ' \
+                              f'classes). Pass correct combination of --weights and --data that are trained together.'
+        model.warmup(imgsz=(1 if pt else batch_size, 3, imgsz, imgsz))  # warmup
+        pad, rect = (0.0, False) if task == 'speed' else (0.5, pt)  # square inference for benchmarks
+        task = task if task in ('train', 'val', 'test') else 'val'  # path to train/val/test images
+        dataloader = create_dataloader(data[task],
+                                       imgsz,
+                                       batch_size,
+                                       stride,
+                                       single_cls,
+                                       pad=pad,
+                                       rect=rect,
+                                       workers=workers,
+                                       prefix=colorstr(f'{task}: '),
+                                       overlap_mask=overlap,
+                                       mask_downsample_ratio=mask_downsample_ratio)[0]
+
+    seen = 0
+    confusion_matrix = ConfusionMatrix(nc=nc)
+    names = model.names if hasattr(model, 'names') else model.module.names  # get class names
+    if isinstance(names, (list, tuple)):  # old format
+        names = dict(enumerate(names))
+    class_map = coco80_to_coco91_class() if is_coco else list(range(1000))
+    s = ('%22s' + '%11s' * 12) % ('Class', 'Images', 'Instances', 'Box(P', "R", "mAP50", "mAP50-95)", "Mask(P", "R",
+                                  "mAP50", "mAP50-95)", 'S(MIoU', 'FWIoU)')
+    dt = Profile(), Profile(), Profile()
+    metrics = Metrics()
+    semantic_metrics = Semantic_Metrics(nc = (nc + stuff_nc), device = device)
+    loss = torch.zeros(6, device=device)
+    jdict, stats = [], []
+    semantic_jdict = []
+    # callbacks.run('on_val_start')
+    pbar = tqdm(dataloader, desc=s, bar_format=TQDM_BAR_FORMAT)  # progress bar
+    for batch_i, (im, targets, paths, shapes, masks, semasks) in enumerate(pbar):
+        # callbacks.run('on_val_batch_start')
+        with dt[0]:
+            if cuda:
+                im = im.to(device, non_blocking=True)
+                targets = targets.to(device)
+                masks = masks.to(device)
+                semasks = semasks.to(device)
+            masks = masks.float()
+            semasks = semasks.float()
+            im = im.half() if half else im.float()  # uint8 to fp16/32
+            im /= 255  # 0 - 255 to 0.0 - 1.0
+            nb, _, height, width = im.shape  # batch size, channels, height, width
+
+        # Inference
+        with dt[1]:
+            preds, train_out = model(im)# if compute_loss else (*model(im, augment=augment)[:2], None)
+            #train_out, preds, protos = p if len(p) == 3 else p[1]
+            #preds = p
+            #train_out = p[1][0] if len(p[1]) == 3 else p[0]
+            # protos = train_out[-1]
+            #print(preds.shape)
+            #print(train_out[0].shape)
+            #print(train_out[1].shape)
+            #print(train_out[2].shape)
+            _, pred_masks, protos, psemasks = train_out
+
+        # Loss
+        if compute_loss:
+            loss += compute_loss(train_out, targets, masks, semasks = semasks)[1]  # box, obj, cls
+
+        # NMS
+        targets[:, 2:] *= torch.tensor((width, height, width, height), device=device)  # to pixels
+        lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else []  # for autolabelling
+        with dt[2]:
+            preds = non_max_suppression(preds,
+                                        conf_thres,
+                                        iou_thres,
+                                        labels=lb,
+                                        multi_label=True,
+                                        agnostic=single_cls,
+                                        max_det=max_det,
+                                        nm=nm)
+
+        # Metrics
+        plot_masks = []  # masks for plotting
+        plot_semasks = []  # masks for plotting
+
+        if training:
+            semantic_metrics.update(psemasks, semasks)
+        else:
+            _, _, smh, smw = semasks.shape
+            semantic_metrics.update(torch.nn.functional.interpolate(psemasks, size = (smh, smw), mode = 'bilinear', align_corners = False), semasks)
+
+        if plots and batch_i < 3:
+            plot_semasks.append(psemasks.clone().detach().cpu())
+
+        for si, (pred, proto, psemask) in enumerate(zip(preds, protos, psemasks)):
+            labels = targets[targets[:, 0] == si, 1:]
+            nl, npr = labels.shape[0], pred.shape[0]  # number of labels, predictions
+            path, shape = Path(paths[si]), shapes[si][0]
+            image_id = path.stem
+            img_id_list.append(image_id)
+            correct_masks = torch.zeros(npr, niou, dtype=torch.bool, device=device)  # init
+            correct_bboxes = torch.zeros(npr, niou, dtype=torch.bool, device=device)  # init
+            seen += 1
+
+            if npr == 0:
+                if nl:
+                    stats.append((correct_masks, correct_bboxes, *torch.zeros((2, 0), device=device), labels[:, 0]))
+                    if plots:
+                        confusion_matrix.process_batch(detections=None, labels=labels[:, 0])
+            else:
+                # Masks
+                midx = [si] if overlap else targets[:, 0] == si
+                gt_masks = masks[midx]
+                pred_masks = process(proto, pred[:, 6:], pred[:, :4], shape=im[si].shape[1:])
+
+                # Predictions
+                if single_cls:
+                    pred[:, 5] = 0
+                predn = pred.clone()
+                scale_boxes(im[si].shape[1:], predn[:, :4], shape, shapes[si][1])  # native-space pred
+
+                # Evaluate
+                if nl:
+                    tbox = xywh2xyxy(labels[:, 1:5])  # target boxes
+                    scale_boxes(im[si].shape[1:], tbox, shape, shapes[si][1])  # native-space labels
+                    labelsn = torch.cat((labels[:, 0:1], tbox), 1)  # native-space labels
+                    correct_bboxes = process_batch(predn, labelsn, iouv)
+                    correct_masks = process_batch(predn, labelsn, iouv, pred_masks, gt_masks, overlap=overlap, masks=True)
+                    if plots:
+                        confusion_matrix.process_batch(predn, labelsn)
+                stats.append((correct_masks, correct_bboxes, pred[:, 4], pred[:, 5], labels[:, 0]))  # (conf, pcls, tcls)
+
+                pred_masks = torch.as_tensor(pred_masks, dtype=torch.uint8)
+                if plots and batch_i < 3:
+                    plot_masks.append(pred_masks[:15].cpu())  # filter top 15 to plot
+
+                # Save/log
+                if save_txt:
+                    save_one_txt(predn, save_conf, shape, file=save_dir / 'labels' / f'{path.stem}.txt')
+                if save_json:
+                    pred_masks = scale_image(im[si].shape[1:],
+                                            pred_masks.permute(1, 2, 0).contiguous().cpu().numpy(), shape, shapes[si][1])
+                    save_one_json(predn, jdict, path, class_map, pred_masks)  # append to COCO-JSON dictionary
+                # callbacks.run('on_val_image_end', pred, predn, path, names, im[si])
+
+            # Semantic Segmentation
+            h0, w0 = shape
+
+            # resize
+            _, mask_h, mask_w = psemask.shape
+            h_ratio = mask_h / h0
+            w_ratio = mask_w / w0
+
+            if h_ratio == w_ratio:
+                psemask = torch.nn.functional.interpolate(psemask[None, :], size = (h0, w0), mode = 'bilinear', align_corners = False)
+            else:
+                transform = transforms.CenterCrop((h0, w0))
+
+                if (1 != h_ratio) and (1 != w_ratio):
+                    h_new = h0 if (h_ratio < w_ratio) else int(mask_h / w_ratio)
+                    w_new = w0 if (h_ratio > w_ratio) else int(mask_w / h_ratio)
+                    psemask = torch.nn.functional.interpolate(psemask[None, :], size = (h_new, w_new), mode = 'bilinear', align_corners = False)
+
+                psemask = transform(psemask)
+
+            psemask = torch.squeeze(psemask)
+
+            nc, h, w = psemask.shape
+
+            semantic_mask = torch.flatten(psemask, start_dim = 1).permute(1, 0) # class x h x w -> (h x w) x class
+
+            max_idx = semantic_mask.argmax(1)
+            output_masks = torch.zeros(semantic_mask.shape).scatter(1, max_idx.cpu().unsqueeze(1), 1.0) # one hot: (h x w) x class
+            output_masks = torch.reshape(output_masks.permute(1, 0), (nc, h, w)) # (h x w) x class -> class x h x w
+            psemask = output_masks.to(device = device)
+
+            # TODO: check is_coco
+            instances_ids = getCocoIds(name = 'instances')
+            stuff_mask = torch.zeros((h, w), device = device)
+            check_semantic_mask = False
+            for idx, pred_semantic_mask in enumerate(psemask):
+                category_id = int(getMappingId(idx))
+                if 183 == category_id:
+                    # set all non-stuff pixels to other
+                    pred_semantic_mask = (torch.logical_xor(stuff_mask, torch.ones((h, w), device = device))).int()
+
+                # ignore the classes which all zeros / unlabeled class
+                if (0 >= torch.max(pred_semantic_mask)) or (0 >= category_id):
+                    continue
+
+                if category_id not in instances_ids:
+                    # record all stuff mask
+                    stuff_mask = torch.logical_or(stuff_mask, pred_semantic_mask)
+
+                if (category_id not in instances_ids):
+                    rle = maskUtils.encode(np.asfortranarray(pred_semantic_mask.cpu(), dtype = np.uint8))
+                    rle['counts'] = rle['counts'].decode('utf-8')
+
+                    temp_d = {
+                        'image_id': int(image_id) if image_id.isnumeric() else image_id,
+                        'category_id': category_id,
+                        'segmentation': rle,
+                        'score': 1
+                    }
+
+                    semantic_jdict.append(temp_d)
+                    check_semantic_mask = True
+
+            if not check_semantic_mask:
+                # append a other mask for evaluation if the image without any mask
+                other_mask = (torch.ones((h, w), device = device)).int()
+
+                rle = maskUtils.encode(np.asfortranarray(other_mask.cpu(), dtype = np.uint8))
+                rle['counts'] = rle['counts'].decode('utf-8')
+
+                temp_d = {
+                    'image_id': int(image_id) if image_id.isnumeric() else image_id,
+                    'category_id': 183,
+                    'segmentation': rle,
+                    'score': 1
+                }
+
+                semantic_jdict.append(temp_d)
+
+        # Plot images
+        if plots and batch_i < 3:
+            if len(plot_masks):
+                plot_masks = torch.cat(plot_masks, dim=0)
+            if len(plot_semasks):
+                plot_semasks = torch.cat(plot_semasks, dim = 0)
+            plot_images_and_masks(im, targets, masks, semasks, paths, save_dir / f'val_batch{batch_i}_labels.jpg', names)
+            plot_images_and_masks(im, output_to_target(preds, max_det=15), plot_masks, plot_semasks, paths,
+                                  save_dir / f'val_batch{batch_i}_pred.jpg', names)  # pred
+
+        # callbacks.run('on_val_batch_end')
+
+    # Compute metrics
+    stats = [torch.cat(x, 0).cpu().numpy() for x in zip(*stats)]  # to numpy
+    if len(stats) and stats[0].any():
+        results = ap_per_class_box_and_mask(*stats, plot=plots, save_dir=save_dir, names=names)
+        metrics.update(results)
+    nt = np.bincount(stats[4].astype(int), minlength=nc)  # number of targets per class
+
+    # Print results
+    pf = '%22s' + '%11i' * 2 + '%11.3g' * 10  # print format
+    LOGGER.info(pf % ("all", seen, nt.sum(), *metrics.mean_results(), *semantic_metrics.results()))
+    if nt.sum() == 0:
+        LOGGER.warning(f'WARNING ⚠️ no labels found in {task} set, can not compute metrics without labels')
+
+    # Print results per class
+    if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats):
+        for i, c in enumerate(metrics.ap_class_index):
+            LOGGER.info(pf % (names[c], seen, nt[c], *metrics.class_result(i), *semantic_metrics.results()))
+
+    # Print speeds
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    if not training:
+        shape = (batch_size, 3, imgsz, imgsz)
+        LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {shape}' % t)
+
+    # Plots
+    if plots:
+        confusion_matrix.plot(save_dir=save_dir, names=list(names.values()))
+    # callbacks.run('on_val_end')
+
+    mp_bbox, mr_bbox, map50_bbox, map_bbox, mp_mask, mr_mask, map50_mask, map_mask = metrics.mean_results()
+    miou_sem, fwiou_sem = semantic_metrics.results()
+    semantic_metrics.reset()
+
+    # Save JSON
+    if save_json and len(jdict):
+        w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else ''  # weights
+        anno_path = Path(data.get('path', '../coco'))
+        anno_json = str(anno_path / 'annotations/instances_val2017.json')  # annotations json
+        pred_json = str(save_dir / f"{w}_predictions.json")  # predictions json
+        LOGGER.info(f'\nEvaluating pycocotools mAP... saving {pred_json}...')
+        with open(pred_json, 'w') as f:
+            json.dump(jdict, f)
+
+        semantic_anno_json = str(anno_path / 'annotations/stuff_val2017.json')  # annotations json
+        semantic_pred_json = str(save_dir / f"{w}_predictions_stuff.json")  # predictions json
+        LOGGER.info(f'\nsaving {semantic_pred_json}...')
+        with open(semantic_pred_json, 'w') as f:
+            json.dump(semantic_jdict, f)
+
+        try:  # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
+            from pycocotools.coco import COCO
+            from pycocotools.cocoeval import COCOeval
+
+            anno = COCO(anno_json)  # init annotations api
+            pred = anno.loadRes(pred_json)  # init predictions api
+            results = []
+            for eval in COCOeval(anno, pred, 'bbox'), COCOeval(anno, pred, 'segm'):
+                if is_coco:
+                    eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.im_files]  # img ID to evaluate
+                eval.evaluate()
+                eval.accumulate()
+                eval.summarize()
+                results.extend(eval.stats[:2])  # update results (mAP@0.5:0.95, mAP@0.5)
+            map_bbox, map50_bbox, map_mask, map50_mask = results
+
+            # Semantic Segmentation
+            from utils.stuff_seg.cocostuffeval import COCOStuffeval
+
+            LOGGER.info(f'\nEvaluating pycocotools stuff... ')
+            imgIds = [int(x) for x in img_id_list]
+
+            stuffGt = COCO(semantic_anno_json)  # initialize COCO ground truth api
+            stuffDt = stuffGt.loadRes(semantic_pred_json)  # initialize COCO pred api
+
+            cocoStuffEval = COCOStuffeval(stuffGt, stuffDt)
+            cocoStuffEval.params.imgIds = imgIds  # image IDs to evaluate
+            cocoStuffEval.evaluate()
+            stats, statsClass = cocoStuffEval.summarize()
+            stuffIds = getCocoIds(name = 'stuff')
+            title = ' {:<5} | {:^6} | {:^6} '.format('class', 'iou', 'macc') if (0 >= len(stuff_names)) else \
+                    ' {:<5} | {:<20} | {:^6} | {:^6} '.format('class', 'class name', 'iou', 'macc')
+            print(title)
+            for idx, (iou, macc) in enumerate(zip(statsClass['ious'], statsClass['maccs'])):
+                id = (idx + 1)
+                if id not in stuffIds:
+                    continue
+                content = ' {:<5} | {:0.4f} | {:0.4f} '.format(str(id), iou, macc) if (0 >= len(stuff_names)) else \
+                            ' {:<5} | {:<20} | {:0.4f} | {:0.4f} '.format(str(id), str(stuff_names[getMappingIndex(id, name = 'stuff')]), iou, macc)
+                print(content)
+
+        except Exception as e:
+            LOGGER.info(f'pycocotools unable to run: {e}')
+
+    # Return results
+    model.float()  # for training
+    if not training:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    final_metric = mp_bbox, mr_bbox, map50_bbox, map_bbox, mp_mask, mr_mask, map50_mask, map_mask, miou_sem, fwiou_sem
+    return (*final_metric, *(loss.cpu() / len(dataloader)).tolist()), metrics.get_maps(nc), t
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128-pan.yaml', help='dataset.yaml path')
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo-pan.pt', help='model path(s)')
+    parser.add_argument('--batch-size', type=int, default=32, help='batch size')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)')
+    parser.add_argument('--conf-thres', type=float, default=0.001, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.6, help='NMS IoU threshold')
+    parser.add_argument('--max-det', type=int, default=300, help='maximum detections per image')
+    parser.add_argument('--task', default='val', help='train, val, test, speed or study')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--verbose', action='store_true', help='report mAP by class')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--save-hybrid', action='store_true', help='save label+prediction hybrid results to *.txt')
+    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
+    parser.add_argument('--save-json', action='store_true', help='save a COCO-JSON results file')
+    parser.add_argument('--project', default=ROOT / 'runs/val-pan', help='save results to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    opt = parser.parse_args()
+    opt.data = check_yaml(opt.data)  # check YAML
+    # opt.save_json |= opt.data.endswith('coco.yaml')
+    opt.save_txt |= opt.save_hybrid
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    #check_requirements(requirements=ROOT / 'requirements.txt', exclude=('tensorboard', 'thop'))
+
+    if opt.task in ('train', 'val', 'test'):  # run normally
+        if opt.conf_thres > 0.001:  # https://github.com/
+            LOGGER.warning(f'WARNING ⚠️ confidence threshold {opt.conf_thres} > 0.001 produces invalid results')
+        if opt.save_hybrid:
+            LOGGER.warning('WARNING ⚠️ --save-hybrid returns high mAP from hybrid labels, not from predictions alone')
+        run(**vars(opt))
+
+    else:
+        weights = opt.weights if isinstance(opt.weights, list) else [opt.weights]
+        opt.half = torch.cuda.is_available() and opt.device != 'cpu'  # FP16 for fastest results
+        if opt.task == 'speed':  # speed benchmarks
+            # python val.py --task speed --data coco.yaml --batch 1 --weights yolo.pt...
+            opt.conf_thres, opt.iou_thres, opt.save_json = 0.25, 0.45, False
+            for opt.weights in weights:
+                run(**vars(opt), plots=False)
+
+        elif opt.task == 'study':  # speed vs mAP benchmarks
+            # python val.py --task study --data coco.yaml --iou 0.7 --weights yolo.pt...
+            for opt.weights in weights:
+                f = f'study_{Path(opt.data).stem}_{Path(opt.weights).stem}.txt'  # filename to save to
+                x, y = list(range(256, 1536 + 128, 128)), []  # x axis (image sizes), y axis
+                for opt.imgsz in x:  # img-size
+                    LOGGER.info(f'\nRunning {f} --imgsz {opt.imgsz}...')
+                    r, _, t = run(**vars(opt), plots=False)
+                    y.append(r + t)  # results and times
+                np.savetxt(f, y, fmt='%10.4g')  # save
+            os.system('zip -r study.zip study_*.txt')
+            plot_val_study(x=x)  # plot
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)