Upload 47 files

MeshAnything/miche/LICENSE

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

MeshAnything/miche/encode.py

@@ -0,0 +1,73 @@
+# -*- coding: utf-8 -*-
+import argparse
+from omegaconf import OmegaConf
+import numpy as np
+import torch
+from .michelangelo.utils.misc import instantiate_from_config
+
+def load_surface(fp):
+    
+    with np.load(fp) as input_pc:
+        surface = input_pc['points']
+        normal = input_pc['normals']
+    
+    rng = np.random.default_rng()
+    ind = rng.choice(surface.shape[0], 4096, replace=False)
+    surface = torch.FloatTensor(surface[ind])
+    normal = torch.FloatTensor(normal[ind])
+    
+    surface = torch.cat([surface, normal], dim=-1).unsqueeze(0).cuda()
+    
+    return surface
+
+def reconstruction(args, model, bounds=(-1.25, -1.25, -1.25, 1.25, 1.25, 1.25), octree_depth=7, num_chunks=10000):
+
+    surface = load_surface(args.pointcloud_path)
+    # old_surface = surface.clone()
+
+    # surface[0,:,0]*=-1
+    # surface[0,:,1]*=-1
+    surface[0,:,2]*=-1
+
+    # encoding
+    shape_embed, shape_latents = model.model.encode_shape_embed(surface, return_latents=True)    
+    shape_zq, posterior = model.model.shape_model.encode_kl_embed(shape_latents)
+
+    # decoding
+    latents = model.model.shape_model.decode(shape_zq)
+    # geometric_func = partial(model.model.shape_model.query_geometry, latents=latents)
+    
+    return 0
+
+def load_model(ckpt_path="MeshAnything/miche/shapevae-256.ckpt"):
+    model_config = OmegaConf.load("MeshAnything/miche/shapevae-256.yaml")
+    # print(model_config)
+    if hasattr(model_config, "model"):
+        model_config = model_config.model
+
+    model = instantiate_from_config(model_config, ckpt_path=ckpt_path)
+    model = model.cuda()
+    model = model.eval()
+
+    return model
+if __name__ == "__main__":
+    '''
+    1. Reconstruct point cloud
+    2. Image-conditioned generation
+    3. Text-conditioned generation
+    '''
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--config_path", type=str, required=True)
+    parser.add_argument("--ckpt_path", type=str, required=True)
+    parser.add_argument("--pointcloud_path", type=str, default='./example_data/surface.npz', help='Path to the input point cloud')
+    parser.add_argument("--image_path", type=str, help='Path to the input image')
+    parser.add_argument("--text", type=str, help='Input text within a format: A 3D model of motorcar; Porsche 911.')
+    parser.add_argument("--output_dir", type=str, default='./output')
+    parser.add_argument("-s", "--seed", type=int, default=0)
+    args = parser.parse_args()
+    
+    print(f'-----------------------------------------------------------------------------')
+    print(f'>>> Output directory: {args.output_dir}')
+    print(f'-----------------------------------------------------------------------------')
+    
+    reconstruction(args, load_model(args))

MeshAnything/miche/michelangelo/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

MeshAnything/miche/michelangelo/data/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

MeshAnything/miche/michelangelo/data/templates.json

@@ -0,0 +1,69 @@
+{
+    "shape": [
+        "a point cloud model of {}.",
+        "There is a {} in the scene.",
+        "There is the {} in the scene.",
+        "a photo of a {} in the scene.",
+        "a photo of the {} in the scene.",
+        "a photo of one {} in the scene.",
+        "itap of a {}.",
+        "itap of my {}.",
+        "itap of the {}.",
+        "a photo of a {}.",
+        "a photo of my {}.",
+        "a photo of the {}.",
+        "a photo of one {}.",
+        "a photo of many {}.",
+        "a good photo of a {}.",
+        "a good photo of the {}.",
+        "a bad photo of a {}.",
+        "a bad photo of the {}.",
+        "a photo of a nice {}.",
+        "a photo of the nice {}.",
+        "a photo of a cool {}.",
+        "a photo of the cool {}.",
+        "a photo of a weird {}.",
+        "a photo of the weird {}.",
+        "a photo of a small {}.",
+        "a photo of the small {}.",
+        "a photo of a large {}.",
+        "a photo of the large {}.",
+        "a photo of a clean {}.",
+        "a photo of the clean {}.",
+        "a photo of a dirty {}.",
+        "a photo of the dirty {}.",
+        "a bright photo of a {}.",
+        "a bright photo of the {}.",
+        "a dark photo of a {}.",
+        "a dark photo of the {}.",
+        "a photo of a hard to see {}.",
+        "a photo of the hard to see {}.",
+        "a low resolution photo of a {}.",
+        "a low resolution photo of the {}.",
+        "a cropped photo of a {}.",
+        "a cropped photo of the {}.",
+        "a close-up photo of a {}.",
+        "a close-up photo of the {}.",
+        "a jpeg corrupted photo of a {}.",
+        "a jpeg corrupted photo of the {}.",
+        "a blurry photo of a {}.",
+        "a blurry photo of the {}.",
+        "a pixelated photo of a {}.",
+        "a pixelated photo of the {}.",
+        "a black and white photo of the {}.",
+        "a black and white photo of a {}",
+        "a plastic {}.",
+        "the plastic {}.",
+        "a toy {}.",
+        "the toy {}.",
+        "a plushie {}.",
+        "the plushie {}.",
+        "a cartoon {}.",
+        "the cartoon {}.",
+        "an embroidered {}.",
+        "the embroidered {}.",
+        "a painting of the {}.",
+        "a painting of a {}."
+    ]
+
+}

MeshAnything/miche/michelangelo/data/transforms.py

@@ -0,0 +1,407 @@
+# -*- coding: utf-8 -*-
+import os
+import time
+import numpy as np
+import warnings
+import random
+from omegaconf.listconfig import ListConfig
+from webdataset import pipelinefilter
+import torch
+import torchvision.transforms.functional as TVF
+from torchvision.transforms import InterpolationMode
+from torchvision.transforms.transforms import _interpolation_modes_from_int
+from typing import Sequence
+
+from MeshAnything.miche.michelangelo.utils import instantiate_from_config
+
+
+def _uid_buffer_pick(buf_dict, rng):
+    uid_keys = list(buf_dict.keys())
+    selected_uid = rng.choice(uid_keys)
+    buf = buf_dict[selected_uid]
+
+    k = rng.randint(0, len(buf) - 1)
+    sample = buf[k]
+    buf[k] = buf[-1]
+    buf.pop()
+
+    if len(buf) == 0:
+        del buf_dict[selected_uid]
+
+    return sample
+
+
+def _add_to_buf_dict(buf_dict, sample):
+    key = sample["__key__"]
+    uid, uid_sample_id = key.split("_")
+    if uid not in buf_dict:
+        buf_dict[uid] = []
+    buf_dict[uid].append(sample)
+
+    return buf_dict
+
+
+def _uid_shuffle(data, bufsize=1000, initial=100, rng=None, handler=None):
+    """Shuffle the data in the stream.
+
+    This uses a buffer of size `bufsize`. Shuffling at
+    startup is less random; this is traded off against
+    yielding samples quickly.
+
+    data: iterator
+    bufsize: buffer size for shuffling
+    returns: iterator
+    rng: either random module or random.Random instance
+
+    """
+    if rng is None:
+        rng = random.Random(int((os.getpid() + time.time()) * 1e9))
+    initial = min(initial, bufsize)
+    buf_dict = dict()
+    current_samples = 0
+    for sample in data:
+        _add_to_buf_dict(buf_dict, sample)
+        current_samples += 1
+
+        if current_samples < bufsize:
+            try:
+                _add_to_buf_dict(buf_dict, next(data))  # skipcq: PYL-R1708
+                current_samples += 1
+            except StopIteration:
+                pass
+
+        if current_samples >= initial:
+            current_samples -= 1
+            yield _uid_buffer_pick(buf_dict, rng)
+
+    while current_samples > 0:
+        current_samples -= 1
+        yield _uid_buffer_pick(buf_dict, rng)
+
+
+uid_shuffle = pipelinefilter(_uid_shuffle)
+
+
+class RandomSample(object):
+    def __init__(self,
+                 num_volume_samples: int = 1024,
+                 num_near_samples: int = 1024):
+
+        super().__init__()
+
+        self.num_volume_samples = num_volume_samples
+        self.num_near_samples = num_near_samples
+
+    def __call__(self, sample):
+        rng = np.random.default_rng()
+
+        # 1. sample surface input
+        total_surface = sample["surface"]
+        ind = rng.choice(total_surface.shape[0], replace=False)
+        surface = total_surface[ind]
+
+        # 2. sample volume/near geometric points
+        vol_points = sample["vol_points"]
+        vol_label = sample["vol_label"]
+        near_points = sample["near_points"]
+        near_label = sample["near_label"]
+
+        ind = rng.choice(vol_points.shape[0], self.num_volume_samples, replace=False)
+        vol_points = vol_points[ind]
+        vol_label = vol_label[ind]
+        vol_points_labels = np.concatenate([vol_points, vol_label[:, np.newaxis]], axis=1)
+
+        ind = rng.choice(near_points.shape[0], self.num_near_samples, replace=False)
+        near_points = near_points[ind]
+        near_label = near_label[ind]
+        near_points_labels = np.concatenate([near_points, near_label[:, np.newaxis]], axis=1)
+
+        # concat sampled volume and near points
+        geo_points = np.concatenate([vol_points_labels, near_points_labels], axis=0)
+
+        sample = {
+            "surface": surface,
+            "geo_points": geo_points
+        }
+
+        return sample
+
+
+class SplitRandomSample(object):
+    def __init__(self,
+                 use_surface_sample: bool = False,
+                 num_surface_samples: int = 4096,
+                 num_volume_samples: int = 1024,
+                 num_near_samples: int = 1024):
+
+        super().__init__()
+
+        self.use_surface_sample = use_surface_sample
+        self.num_surface_samples = num_surface_samples
+        self.num_volume_samples = num_volume_samples
+        self.num_near_samples = num_near_samples
+
+    def __call__(self, sample):
+
+        rng = np.random.default_rng()
+
+        # 1. sample surface input
+        surface = sample["surface"]
+
+        if self.use_surface_sample:
+            replace = surface.shape[0] < self.num_surface_samples
+            ind = rng.choice(surface.shape[0], self.num_surface_samples, replace=replace)
+            surface = surface[ind]
+
+        # 2. sample volume/near geometric points
+        vol_points = sample["vol_points"]
+        vol_label = sample["vol_label"]
+        near_points = sample["near_points"]
+        near_label = sample["near_label"]
+
+        ind = rng.choice(vol_points.shape[0], self.num_volume_samples, replace=False)
+        vol_points = vol_points[ind]
+        vol_label = vol_label[ind]
+        vol_points_labels = np.concatenate([vol_points, vol_label[:, np.newaxis]], axis=1)
+
+        ind = rng.choice(near_points.shape[0], self.num_near_samples, replace=False)
+        near_points = near_points[ind]
+        near_label = near_label[ind]
+        near_points_labels = np.concatenate([near_points, near_label[:, np.newaxis]], axis=1)
+
+        # concat sampled volume and near points
+        geo_points = np.concatenate([vol_points_labels, near_points_labels], axis=0)
+
+        sample = {
+            "surface": surface,
+            "geo_points": geo_points
+        }
+
+        return sample
+
+
+class FeatureSelection(object):
+
+    VALID_SURFACE_FEATURE_DIMS = {
+        "none": [0, 1, 2],                              # xyz
+        "watertight_normal": [0, 1, 2, 3, 4, 5],        # xyz, normal
+        "normal": [0, 1, 2, 6, 7, 8]
+    }
+
+    def __init__(self, surface_feature_type: str):
+
+        self.surface_feature_type = surface_feature_type
+        self.surface_dims = self.VALID_SURFACE_FEATURE_DIMS[surface_feature_type]
+
+    def __call__(self, sample):
+        sample["surface"] = sample["surface"][:, self.surface_dims]
+        return sample
+
+
+class AxisScaleTransform(object):
+    def __init__(self, interval=(0.75, 1.25), jitter=True, jitter_scale=0.005):
+        assert isinstance(interval, (tuple, list, ListConfig))
+        self.interval = interval
+        self.min_val = interval[0]
+        self.max_val = interval[1]
+        self.inter_size = interval[1] - interval[0]
+        self.jitter = jitter
+        self.jitter_scale = jitter_scale
+
+    def __call__(self, sample):
+
+        surface = sample["surface"][..., 0:3]
+        geo_points = sample["geo_points"][..., 0:3]
+
+        scaling = torch.rand(1, 3) * self.inter_size + self.min_val
+        # print(scaling)
+        surface = surface * scaling
+        geo_points = geo_points * scaling
+
+        scale = (1 / torch.abs(surface).max().item()) * 0.999999
+        surface *= scale
+        geo_points *= scale
+
+        if self.jitter:
+            surface += self.jitter_scale * torch.randn_like(surface)
+            surface.clamp_(min=-1.015, max=1.015)
+
+        sample["surface"][..., 0:3] = surface
+        sample["geo_points"][..., 0:3] = geo_points
+
+        return sample
+
+
+class ToTensor(object):
+
+    def __init__(self, tensor_keys=("surface", "geo_points", "tex_points")):
+        self.tensor_keys = tensor_keys
+
+    def __call__(self, sample):
+        for key in self.tensor_keys:
+            if key not in sample:
+                continue
+
+            sample[key] = torch.tensor(sample[key], dtype=torch.float32)
+
+        return sample
+
+
+class AxisScale(object):
+    def __init__(self, interval=(0.75, 1.25), jitter=True, jitter_scale=0.005):
+        assert isinstance(interval, (tuple, list, ListConfig))
+        self.interval = interval
+        self.jitter = jitter
+        self.jitter_scale = jitter_scale
+
+    def __call__(self, surface, *args):
+        scaling = torch.rand(1, 3) * 0.5 + 0.75
+        # print(scaling)
+        surface = surface * scaling
+        scale = (1 / torch.abs(surface).max().item()) * 0.999999
+        surface *= scale
+
+        args_outputs = []
+        for _arg in args:
+            _arg = _arg * scaling * scale
+            args_outputs.append(_arg)
+
+        if self.jitter:
+            surface += self.jitter_scale * torch.randn_like(surface)
+            surface.clamp_(min=-1, max=1)
+
+        if len(args) == 0:
+            return surface
+        else:
+            return surface, *args_outputs
+
+
+class RandomResize(torch.nn.Module):
+    """Apply randomly Resize with a given probability."""
+
+    def __init__(
+        self,
+        size,
+        resize_radio=(0.5, 1),
+        allow_resize_interpolations=(InterpolationMode.BICUBIC, InterpolationMode.BILINEAR, InterpolationMode.BILINEAR),
+        interpolation=InterpolationMode.BICUBIC,
+        max_size=None,
+        antialias=None,
+    ):
+        super().__init__()
+        if not isinstance(size, (int, Sequence)):
+            raise TypeError(f"Size should be int or sequence. Got {type(size)}")
+        if isinstance(size, Sequence) and len(size) not in (1, 2):
+            raise ValueError("If size is a sequence, it should have 1 or 2 values")
+
+        self.size = size
+        self.max_size = max_size
+        # Backward compatibility with integer value
+        if isinstance(interpolation, int):
+            warnings.warn(
+                "Argument 'interpolation' of type int is deprecated since 0.13 and will be removed in 0.15. "
+                "Please use InterpolationMode enum."
+            )
+            interpolation = _interpolation_modes_from_int(interpolation)
+
+        self.interpolation = interpolation
+        self.antialias = antialias
+
+        self.resize_radio = resize_radio
+        self.allow_resize_interpolations = allow_resize_interpolations
+
+    def random_resize_params(self):
+        radio = torch.rand(1) * (self.resize_radio[1] - self.resize_radio[0]) + self.resize_radio[0]
+
+        if isinstance(self.size, int):
+            size = int(self.size * radio)
+        elif isinstance(self.size, Sequence):
+            size = list(self.size)
+            size = (int(size[0] * radio), int(size[1] * radio))
+        else:
+            raise RuntimeError()
+
+        interpolation = self.allow_resize_interpolations[
+            torch.randint(low=0, high=len(self.allow_resize_interpolations), size=(1,))
+        ]
+        return size, interpolation
+
+    def forward(self, img):
+        size, interpolation = self.random_resize_params()
+        img = TVF.resize(img, size, interpolation, self.max_size, self.antialias)
+        img = TVF.resize(img, self.size, self.interpolation, self.max_size, self.antialias)
+        return img
+
+    def __repr__(self) -> str:
+        detail = f"(size={self.size}, interpolation={self.interpolation.value},"
+        detail += f"max_size={self.max_size}, antialias={self.antialias}), resize_radio={self.resize_radio}"
+        return f"{self.__class__.__name__}{detail}"
+
+
+class Compose(object):
+    """Composes several transforms together. This transform does not support torchscript.
+    Please, see the note below.
+
+    Args:
+        transforms (list of ``Transform`` objects): list of transforms to compose.
+
+    Example:
+        >>> transforms.Compose([
+        >>>     transforms.CenterCrop(10),
+        >>>     transforms.ToTensor(),
+        >>> ])
+
+    .. note::
+        In order to script the transformations, please use ``torch.nn.Sequential`` as below.
+
+        >>> transforms = torch.nn.Sequential(
+        >>>     transforms.CenterCrop(10),
+        >>>     transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
+        >>> )
+        >>> scripted_transforms = torch.jit.script(transforms)
+
+        Make sure to use only scriptable transformations, i.e. that work with ``torch.Tensor``, does not require
+        `lambda` functions or ``PIL.Image``.
+
+    """
+
+    def __init__(self, transforms):
+        self.transforms = transforms
+
+    def __call__(self, *args):
+        for t in self.transforms:
+            args = t(*args)
+        return args
+
+    def __repr__(self):
+        format_string = self.__class__.__name__ + '('
+        for t in self.transforms:
+            format_string += '\n'
+            format_string += '    {0}'.format(t)
+        format_string += '\n)'
+        return format_string
+
+
+def identity(*args, **kwargs):
+    if len(args) == 1:
+        return args[0]
+    else:
+        return args
+
+
+def build_transforms(cfg):
+
+    if cfg is None:
+        return identity
+
+    transforms = []
+
+    for transform_name, cfg_instance in cfg.items():
+        transform_instance = instantiate_from_config(cfg_instance)
+        transforms.append(transform_instance)
+        print(f"Build transform: {transform_instance}")
+
+    transforms = Compose(transforms)
+
+    return transforms
+

MeshAnything/miche/michelangelo/data/utils.py

@@ -0,0 +1,59 @@
+# -*- coding: utf-8 -*-
+
+import torch
+import numpy as np
+
+
+def worker_init_fn(_):
+    worker_info = torch.utils.data.get_worker_info()
+    worker_id = worker_info.id
+
+    # dataset = worker_info.dataset
+    # split_size = dataset.num_records // worker_info.num_workers
+    # # reset num_records to the true number to retain reliable length information
+    # dataset.sample_ids = dataset.valid_ids[worker_id * split_size:(worker_id + 1) * split_size]
+    # current_id = np.random.choice(len(np.random.get_state()[1]), 1)
+    # return np.random.seed(np.random.get_state()[1][current_id] + worker_id)
+
+    return np.random.seed(np.random.get_state()[1][0] + worker_id)
+
+
+def collation_fn(samples, combine_tensors=True, combine_scalars=True):
+    """
+
+    Args:
+        samples (list[dict]):
+        combine_tensors:
+        combine_scalars:
+
+    Returns:
+
+    """
+
+    result = {}
+
+    keys = samples[0].keys()
+
+    for key in keys:
+        result[key] = []
+
+    for sample in samples:
+        for key in keys:
+            val = sample[key]
+            result[key].append(val)
+
+    for key in keys:
+        val_list = result[key]
+        if isinstance(val_list[0], (int, float)):
+            if combine_scalars:
+                result[key] = np.array(result[key])
+
+        elif isinstance(val_list[0], torch.Tensor):
+            if combine_tensors:
+                result[key] = torch.stack(val_list)
+
+        elif isinstance(val_list[0], np.ndarray):
+            if combine_tensors:
+                result[key] = np.stack(val_list)
+
+    return result

MeshAnything/miche/michelangelo/graphics/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

MeshAnything/miche/michelangelo/graphics/primitives/__init__.py

@@ -0,0 +1,9 @@
+# -*- coding: utf-8 -*-
+
+from .volume import generate_dense_grid_points
+
+from .mesh import (
+    MeshOutput,
+    save_obj,
+    savemeshtes2
+)

MeshAnything/miche/michelangelo/graphics/primitives/mesh.py

@@ -0,0 +1,114 @@
+# -*- coding: utf-8 -*-
+
+import os
+import cv2
+import numpy as np
+import PIL.Image
+from typing import Optional
+
+import trimesh
+
+
+def save_obj(pointnp_px3, facenp_fx3, fname):
+    fid = open(fname, "w")
+    write_str = ""
+    for pidx, p in enumerate(pointnp_px3):
+        pp = p
+        write_str += "v %f %f %f\n" % (pp[0], pp[1], pp[2])
+
+    for i, f in enumerate(facenp_fx3):
+        f1 = f + 1
+        write_str += "f %d %d %d\n" % (f1[0], f1[1], f1[2])
+    fid.write(write_str)
+    fid.close()
+    return
+
+
+def savemeshtes2(pointnp_px3, tcoords_px2, facenp_fx3, facetex_fx3, tex_map, fname):
+    fol, na = os.path.split(fname)
+    na, _ = os.path.splitext(na)
+
+    matname = "%s/%s.mtl" % (fol, na)
+    fid = open(matname, "w")
+    fid.write("newmtl material_0\n")
+    fid.write("Kd 1 1 1\n")
+    fid.write("Ka 0 0 0\n")
+    fid.write("Ks 0.4 0.4 0.4\n")
+    fid.write("Ns 10\n")
+    fid.write("illum 2\n")
+    fid.write("map_Kd %s.png\n" % na)
+    fid.close()
+    ####
+
+    fid = open(fname, "w")
+    fid.write("mtllib %s.mtl\n" % na)
+
+    for pidx, p in enumerate(pointnp_px3):
+        pp = p
+        fid.write("v %f %f %f\n" % (pp[0], pp[1], pp[2]))
+
+    for pidx, p in enumerate(tcoords_px2):
+        pp = p
+        fid.write("vt %f %f\n" % (pp[0], pp[1]))
+
+    fid.write("usemtl material_0\n")
+    for i, f in enumerate(facenp_fx3):
+        f1 = f + 1
+        f2 = facetex_fx3[i] + 1
+        fid.write("f %d/%d %d/%d %d/%d\n" % (f1[0], f2[0], f1[1], f2[1], f1[2], f2[2]))
+    fid.close()
+
+    PIL.Image.fromarray(np.ascontiguousarray(tex_map), "RGB").save(
+        os.path.join(fol, "%s.png" % na))
+
+    return
+
+
+class MeshOutput(object):
+
+    def __init__(self,
+                 mesh_v: np.ndarray,
+                 mesh_f: np.ndarray,
+                 vertex_colors: Optional[np.ndarray] = None,
+                 uvs: Optional[np.ndarray] = None,
+                 mesh_tex_idx: Optional[np.ndarray] = None,
+                 tex_map: Optional[np.ndarray] = None):
+
+        self.mesh_v = mesh_v
+        self.mesh_f = mesh_f
+        self.vertex_colors = vertex_colors
+        self.uvs = uvs
+        self.mesh_tex_idx = mesh_tex_idx
+        self.tex_map = tex_map
+
+    def contain_uv_texture(self):
+        return (self.uvs is not None) and (self.mesh_tex_idx is not None) and (self.tex_map is not None)
+
+    def contain_vertex_colors(self):
+        return self.vertex_colors is not None
+
+    def export(self, fname):
+
+        if self.contain_uv_texture():
+            savemeshtes2(
+                self.mesh_v,
+                self.uvs,
+                self.mesh_f,
+                self.mesh_tex_idx,
+                self.tex_map,
+                fname
+            )
+
+        elif self.contain_vertex_colors():
+            mesh_obj = trimesh.Trimesh(vertices=self.mesh_v, faces=self.mesh_f, vertex_colors=self.vertex_colors)
+            mesh_obj.export(fname)
+
+        else:
+            save_obj(
+                self.mesh_v,
+                self.mesh_f,
+                fname
+            )
+
+
+

MeshAnything/miche/michelangelo/graphics/primitives/volume.py

@@ -0,0 +1,21 @@
+# -*- coding: utf-8 -*-
+
+import numpy as np
+
+
+def generate_dense_grid_points(bbox_min: np.ndarray,
+                               bbox_max: np.ndarray,
+                               octree_depth: int,
+                               indexing: str = "ij"):
+    length = bbox_max - bbox_min
+    num_cells = np.exp2(octree_depth)
+    x = np.linspace(bbox_min[0], bbox_max[0], int(num_cells) + 1, dtype=np.float32)
+    y = np.linspace(bbox_min[1], bbox_max[1], int(num_cells) + 1, dtype=np.float32)
+    z = np.linspace(bbox_min[2], bbox_max[2], int(num_cells) + 1, dtype=np.float32)
+    [xs, ys, zs] = np.meshgrid(x, y, z, indexing=indexing)
+    xyz = np.stack((xs, ys, zs), axis=-1)
+    xyz = xyz.reshape(-1, 3)
+    grid_size = [int(num_cells) + 1, int(num_cells) + 1, int(num_cells) + 1]
+
+    return xyz, grid_size, length
+

MeshAnything/miche/michelangelo/models/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

MeshAnything/miche/michelangelo/models/asl_diffusion/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

MeshAnything/miche/michelangelo/models/asl_diffusion/asl_diffuser_pl_module.py

@@ -0,0 +1,483 @@
+# -*- coding: utf-8 -*-
+
+from omegaconf import DictConfig
+from typing import List, Tuple, Dict, Optional, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.optim import lr_scheduler
+import pytorch_lightning as pl
+from pytorch_lightning.utilities import rank_zero_only
+
+from einops import rearrange
+
+from diffusers.schedulers import (
+    DDPMScheduler,
+    DDIMScheduler,
+    KarrasVeScheduler,
+    DPMSolverMultistepScheduler
+)
+
+from MeshAnything.miche.michelangelo.utils import instantiate_from_config
+# from MeshAnything.miche.michelangelo.models.tsal.tsal_base import ShapeAsLatentPLModule
+from MeshAnything.miche.michelangelo.models.tsal.tsal_base import AlignedShapeAsLatentPLModule
+from MeshAnything.miche.michelangelo.models.asl_diffusion.inference_utils import ddim_sample
+
+SchedulerType = Union[DDIMScheduler, KarrasVeScheduler, DPMSolverMultistepScheduler]
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class ASLDiffuser(pl.LightningModule):
+    first_stage_model: Optional[AlignedShapeAsLatentPLModule]
+    # cond_stage_model: Optional[Union[nn.Module, pl.LightningModule]]
+    model: nn.Module
+
+    def __init__(self, *,
+                 first_stage_config,
+                 denoiser_cfg,
+                 scheduler_cfg,
+                 optimizer_cfg,
+                 loss_cfg,
+                 first_stage_key: str = "surface",
+                 cond_stage_key: str = "image",
+                 cond_stage_trainable: bool = True,
+                 scale_by_std: bool = False,
+                 z_scale_factor: float = 1.0,
+                 ckpt_path: Optional[str] = None,
+                 ignore_keys: Union[Tuple[str], List[str]] = ()):
+
+        super().__init__()
+
+        self.first_stage_key = first_stage_key
+        self.cond_stage_key = cond_stage_key
+        self.cond_stage_trainable = cond_stage_trainable
+
+        # 1. initialize first stage. 
+        # Note: the condition model contained in the first stage model.
+        self.first_stage_config = first_stage_config
+        self.first_stage_model = None
+        # self.instantiate_first_stage(first_stage_config)
+
+        # 2. initialize conditional stage
+        # self.instantiate_cond_stage(cond_stage_config)
+        self.cond_stage_model = {
+            "image": self.encode_image,
+            "image_unconditional_embedding": self.empty_img_cond,
+            "text": self.encode_text,
+            "text_unconditional_embedding": self.empty_text_cond,
+            "surface": self.encode_surface,
+            "surface_unconditional_embedding": self.empty_surface_cond,
+        }
+
+        # 3. diffusion model
+        self.model = instantiate_from_config(
+            denoiser_cfg, device=None, dtype=None
+        )
+
+        self.optimizer_cfg = optimizer_cfg
+
+        # 4. scheduling strategy
+        self.scheduler_cfg = scheduler_cfg
+
+        self.noise_scheduler: DDPMScheduler = instantiate_from_config(scheduler_cfg.noise)
+        self.denoise_scheduler: SchedulerType = instantiate_from_config(scheduler_cfg.denoise)
+
+        # 5. loss configures
+        self.loss_cfg = loss_cfg
+
+        self.scale_by_std = scale_by_std
+        if scale_by_std:
+            self.register_buffer("z_scale_factor", torch.tensor(z_scale_factor))
+        else:
+            self.z_scale_factor = z_scale_factor
+
+        self.ckpt_path = ckpt_path
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+
+    def instantiate_first_stage(self, config):
+        model = instantiate_from_config(config)
+        self.first_stage_model = model.eval()
+        self.first_stage_model.train = disabled_train
+        for param in self.first_stage_model.parameters():
+            param.requires_grad = False
+
+        self.first_stage_model = self.first_stage_model.to(self.device)
+
+    # def instantiate_cond_stage(self, config):
+    #     if not self.cond_stage_trainable:
+    #         if config == "__is_first_stage__":
+    #             print("Using first stage also as cond stage.")
+    #             self.cond_stage_model = self.first_stage_model
+    #         elif config == "__is_unconditional__":
+    #             print(f"Training {self.__class__.__name__} as an unconditional model.")
+    #             self.cond_stage_model = None
+    #             # self.be_unconditional = True
+    #         else:
+    #             model = instantiate_from_config(config)
+    #             self.cond_stage_model = model.eval()
+    #             self.cond_stage_model.train = disabled_train
+    #             for param in self.cond_stage_model.parameters():
+    #                 param.requires_grad = False
+    #     else:
+    #         assert config != "__is_first_stage__"
+    #         assert config != "__is_unconditional__"
+    #         model = instantiate_from_config(config)
+    #         self.cond_stage_model = model
+
+    def init_from_ckpt(self, path, ignore_keys=()):
+        state_dict = torch.load(path, map_location="cpu")["state_dict"]
+
+        keys = list(state_dict.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del state_dict[k]
+
+        missing, unexpected = self.load_state_dict(state_dict, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+            print(f"Unexpected Keys: {unexpected}")
+
+    @property
+    def zero_rank(self):
+        if self._trainer:
+            zero_rank = self.trainer.local_rank == 0
+        else:
+            zero_rank = True
+
+        return zero_rank
+
+    def configure_optimizers(self) -> Tuple[List, List]:
+
+        lr = self.learning_rate
+
+        trainable_parameters = list(self.model.parameters())
+        # if the conditional encoder is trainable
+
+        # if self.cond_stage_trainable:
+        #     conditioner_params = [p for p in self.cond_stage_model.parameters() if p.requires_grad]
+        #     trainable_parameters += conditioner_params
+        #     print(f"number of trainable conditional parameters: {len(conditioner_params)}.")
+
+        if self.optimizer_cfg is None:
+            optimizers = [torch.optim.AdamW(trainable_parameters, lr=lr, betas=(0.9, 0.99), weight_decay=1e-3)]
+            schedulers = []
+        else:
+            optimizer = instantiate_from_config(self.optimizer_cfg.optimizer, params=trainable_parameters)
+            scheduler_func = instantiate_from_config(
+                self.optimizer_cfg.scheduler,
+                max_decay_steps=self.trainer.max_steps,
+                lr_max=lr
+            )
+            scheduler = {
+                "scheduler": lr_scheduler.LambdaLR(optimizer, lr_lambda=scheduler_func.schedule),
+                "interval": "step",
+                "frequency": 1
+            }
+            optimizers = [optimizer]
+            schedulers = [scheduler]
+
+        return optimizers, schedulers
+
+    @torch.no_grad()
+    def encode_text(self, text):
+
+        b = text.shape[0]
+        text_tokens = rearrange(text, "b t l -> (b t) l")
+        text_embed = self.first_stage_model.model.encode_text_embed(text_tokens)
+        text_embed = rearrange(text_embed, "(b t) d -> b t d", b=b)
+        text_embed = text_embed.mean(dim=1)
+        text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)
+
+        return text_embed
+
+    @torch.no_grad()
+    def encode_image(self, img):
+
+        return self.first_stage_model.model.encode_image_embed(img)
+
+    @torch.no_grad()
+    def encode_surface(self, surface):
+
+        return self.first_stage_model.model.encode_shape_embed(surface, return_latents=False)
+
+    @torch.no_grad()
+    def empty_text_cond(self, cond):
+
+        return torch.zeros_like(cond, device=cond.device)
+
+    @torch.no_grad()
+    def empty_img_cond(self, cond):
+
+        return torch.zeros_like(cond, device=cond.device)
+
+    @torch.no_grad()
+    def empty_surface_cond(self, cond):
+
+        return torch.zeros_like(cond, device=cond.device)
+
+    @torch.no_grad()
+    def encode_first_stage(self, surface: torch.FloatTensor, sample_posterior=True):
+
+        z_q = self.first_stage_model.encode(surface, sample_posterior)
+        z_q = self.z_scale_factor * z_q
+
+        return z_q
+
+    @torch.no_grad()
+    def decode_first_stage(self, z_q: torch.FloatTensor, **kwargs):
+
+        z_q = 1. / self.z_scale_factor * z_q
+        latents = self.first_stage_model.decode(z_q, **kwargs)
+        return latents
+
+    @rank_zero_only
+    @torch.no_grad()
+    def on_train_batch_start(self, batch, batch_idx):
+        # only for very first batch
+        if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 \
+                and batch_idx == 0 and self.ckpt_path is None:
+            # set rescale weight to 1./std of encodings
+            print("### USING STD-RESCALING ###")
+
+            z_q = self.encode_first_stage(batch[self.first_stage_key])
+            z = z_q.detach()
+
+            del self.z_scale_factor
+            self.register_buffer("z_scale_factor", 1. / z.flatten().std())
+            print(f"setting self.z_scale_factor to {self.z_scale_factor}")
+
+            print("### USING STD-RESCALING ###")
+
+    def compute_loss(self, model_outputs, split):
+        """
+
+        Args:
+            model_outputs (dict):
+                - x_0:
+                - noise:
+                - noise_prior:
+                - noise_pred:
+                - noise_pred_prior:
+
+            split (str):
+
+        Returns:
+
+        """
+
+        pred = model_outputs["pred"]
+
+        if self.noise_scheduler.prediction_type == "epsilon":
+            target = model_outputs["noise"]
+        elif self.noise_scheduler.prediction_type == "sample":
+            target = model_outputs["x_0"]
+        else:
+            raise NotImplementedError(f"Prediction Type: {self.noise_scheduler.prediction_type} not yet supported.")
+
+        if self.loss_cfg.loss_type == "l1":
+            simple = F.l1_loss(pred, target, reduction="mean")
+        elif self.loss_cfg.loss_type in ["mse", "l2"]:
+            simple = F.mse_loss(pred, target, reduction="mean")
+        else:
+            raise NotImplementedError(f"Loss Type: {self.loss_cfg.loss_type} not yet supported.")
+
+        total_loss = simple
+
+        loss_dict = {
+            f"{split}/total_loss": total_loss.clone().detach(),
+            f"{split}/simple": simple.detach(),
+        }
+
+        return total_loss, loss_dict
+
+    def forward(self, batch):
+        """
+
+        Args:
+            batch:
+
+        Returns:
+
+        """
+
+        if self.first_stage_model is None:
+            self.instantiate_first_stage(self.first_stage_config)
+
+        latents = self.encode_first_stage(batch[self.first_stage_key])
+
+        # conditions = self.cond_stage_model.encode(batch[self.cond_stage_key])
+
+        conditions = self.cond_stage_model[self.cond_stage_key](batch[self.cond_stage_key]).unsqueeze(1)
+
+        mask = torch.rand((len(conditions), 1, 1), device=conditions.device, dtype=conditions.dtype) >= 0.1
+        conditions = conditions * mask.to(conditions)
+
+        # Sample noise that we"ll add to the latents
+        # [batch_size, n_token, latent_dim]
+        noise = torch.randn_like(latents)
+        bs = latents.shape[0]
+        # Sample a random timestep for each motion
+        timesteps = torch.randint(
+            0,
+            self.noise_scheduler.config.num_train_timesteps,
+            (bs,),
+            device=latents.device,
+        )
+        timesteps = timesteps.long()
+        # Add noise to the latents according to the noise magnitude at each timestep
+        noisy_z = self.noise_scheduler.add_noise(latents, noise, timesteps)
+
+        # diffusion model forward
+        noise_pred = self.model(noisy_z, timesteps, conditions)
+
+        diffusion_outputs = {
+            "x_0": noisy_z,
+            "noise": noise,
+            "pred": noise_pred
+        }
+
+        return diffusion_outputs
+
+    def training_step(self, batch: Dict[str, Union[torch.FloatTensor, List[str]]],
+                      batch_idx: int, optimizer_idx: int = 0) -> torch.FloatTensor:
+        """
+
+        Args:
+            batch (dict): the batch sample, and it contains:
+                - surface (torch.FloatTensor):
+                - image (torch.FloatTensor): if provide, [bs, 3, h, w], item range [0, 1]
+                - depth (torch.FloatTensor): if provide, [bs, 1, h, w], item range [-1, 1]
+                - normal (torch.FloatTensor): if provide, [bs, 3, h, w], item range [-1, 1]
+                - text (list of str):
+
+            batch_idx (int):
+
+            optimizer_idx (int):
+
+        Returns:
+            loss (torch.FloatTensor):
+
+        """
+
+        diffusion_outputs = self(batch)
+
+        loss, loss_dict = self.compute_loss(diffusion_outputs, "train")
+        self.log_dict(loss_dict, prog_bar=True, logger=True, sync_dist=False, rank_zero_only=True)
+
+        return loss
+
+    def validation_step(self, batch: Dict[str, torch.FloatTensor],
+                        batch_idx: int, optimizer_idx: int = 0) -> torch.FloatTensor:
+        """
+
+        Args:
+            batch (dict): the batch sample, and it contains:
+                - surface_pc (torch.FloatTensor): [n_pts, 4]
+                - surface_feats (torch.FloatTensor): [n_pts, c]
+                - text (list of str):
+
+            batch_idx (int):
+
+            optimizer_idx (int):
+
+        Returns:
+            loss (torch.FloatTensor):
+
+        """
+
+        diffusion_outputs = self(batch)
+
+        loss, loss_dict = self.compute_loss(diffusion_outputs, "val")
+        self.log_dict(loss_dict, prog_bar=True, logger=True, sync_dist=False, rank_zero_only=True)
+
+        return loss
+
+    @torch.no_grad()
+    def sample(self,
+               batch: Dict[str, Union[torch.FloatTensor, List[str]]],
+               sample_times: int = 1,
+               steps: Optional[int] = None,
+               guidance_scale: Optional[float] = None,
+               eta: float = 0.0,
+               return_intermediates: bool = False, **kwargs):
+
+        if self.first_stage_model is None:
+            self.instantiate_first_stage(self.first_stage_config)
+
+        if steps is None:
+            steps = self.scheduler_cfg.num_inference_steps
+
+        if guidance_scale is None:
+            guidance_scale = self.scheduler_cfg.guidance_scale
+        do_classifier_free_guidance = guidance_scale > 0
+
+        # conditional encode
+        xc = batch[self.cond_stage_key]
+        # cond = self.cond_stage_model[self.cond_stage_key](xc)
+        cond = self.cond_stage_model[self.cond_stage_key](xc).unsqueeze(1)
+
+        if do_classifier_free_guidance:
+            """
+            Note: There are two kinds of uncond for text. 
+            1: using "" as uncond text; (in SAL diffusion)
+            2: zeros_like(cond) as uncond text; (in MDM)
+            """
+            # un_cond = self.cond_stage_model.unconditional_embedding(batch_size=len(xc))
+            un_cond = self.cond_stage_model[f"{self.cond_stage_key}_unconditional_embedding"](cond)
+            # un_cond = torch.zeros_like(cond, device=cond.device)
+            cond = torch.cat([un_cond, cond], dim=0)
+
+        outputs = []
+        latents = None
+
+        if not return_intermediates:
+            for _ in range(sample_times):
+                sample_loop = ddim_sample(
+                    self.denoise_scheduler,
+                    self.model,
+                    shape=self.first_stage_model.latent_shape,
+                    cond=cond,
+                    steps=steps,
+                    guidance_scale=guidance_scale,
+                    do_classifier_free_guidance=do_classifier_free_guidance,
+                    device=self.device,
+                    eta=eta,
+                    disable_prog=not self.zero_rank
+                )
+                for sample, t in sample_loop:
+                    latents = sample
+                outputs.append(self.decode_first_stage(latents, **kwargs))
+        else:
+
+            sample_loop = ddim_sample(
+                self.denoise_scheduler,
+                self.model,
+                shape=self.first_stage_model.latent_shape,
+                cond=cond,
+                steps=steps,
+                guidance_scale=guidance_scale,
+                do_classifier_free_guidance=do_classifier_free_guidance,
+                device=self.device,
+                eta=eta,
+                disable_prog=not self.zero_rank
+            )
+
+            iter_size = steps // sample_times
+            i = 0
+            for sample, t in sample_loop:
+                latents = sample
+                if i % iter_size == 0 or i == steps - 1:
+                    outputs.append(self.decode_first_stage(latents, **kwargs))
+                i += 1
+
+        return outputs

MeshAnything/miche/michelangelo/models/asl_diffusion/asl_udt.py

@@ -0,0 +1,104 @@
+# -*- coding: utf-8 -*-
+
+import torch
+import torch.nn as nn
+from typing import Optional
+from diffusers.models.embeddings import Timesteps
+import math
+
+from MeshAnything.miche.michelangelo.models.modules.transformer_blocks import MLP
+from MeshAnything.miche.michelangelo.models.modules.diffusion_transformer import UNetDiffusionTransformer
+
+
+class ConditionalASLUDTDenoiser(nn.Module):
+
+    def __init__(self, *,
+                 device: Optional[torch.device],
+                 dtype: Optional[torch.dtype],
+                 input_channels: int,
+                 output_channels: int,
+                 n_ctx: int,
+                 width: int,
+                 layers: int,
+                 heads: int,
+                 context_dim: int,
+                 context_ln: bool = True,
+                 skip_ln: bool = False,
+                 init_scale: float = 0.25,
+                 flip_sin_to_cos: bool = False,
+                 use_checkpoint: bool = False):
+        super().__init__()
+
+        self.use_checkpoint = use_checkpoint
+
+        init_scale = init_scale * math.sqrt(1.0 / width)
+
+        self.backbone = UNetDiffusionTransformer(
+            device=device,
+            dtype=dtype,
+            n_ctx=n_ctx,
+            width=width,
+            layers=layers,
+            heads=heads,
+            skip_ln=skip_ln,
+            init_scale=init_scale,
+            use_checkpoint=use_checkpoint
+        )
+        self.ln_post = nn.LayerNorm(width, device=device, dtype=dtype)
+        self.input_proj = nn.Linear(input_channels, width, device=device, dtype=dtype)
+        self.output_proj = nn.Linear(width, output_channels, device=device, dtype=dtype)
+
+        # timestep embedding
+        self.time_embed = Timesteps(width, flip_sin_to_cos=flip_sin_to_cos, downscale_freq_shift=0)
+        self.time_proj = MLP(
+            device=device, dtype=dtype, width=width, init_scale=init_scale
+        )
+
+        self.context_embed = nn.Sequential(
+            nn.LayerNorm(context_dim, device=device, dtype=dtype),
+            nn.Linear(context_dim, width, device=device, dtype=dtype),
+        )
+
+        if context_ln:
+            self.context_embed = nn.Sequential(
+                nn.LayerNorm(context_dim, device=device, dtype=dtype),
+                nn.Linear(context_dim, width, device=device, dtype=dtype),
+            )
+        else:
+            self.context_embed = nn.Linear(context_dim, width, device=device, dtype=dtype)
+
+    def forward(self,
+                model_input: torch.FloatTensor,
+                timestep: torch.LongTensor,
+                context: torch.FloatTensor):
+
+        r"""
+        Args:
+            model_input (torch.FloatTensor): [bs, n_data, c]
+            timestep (torch.LongTensor): [bs,]
+            context (torch.FloatTensor): [bs, context_tokens, c]
+
+        Returns:
+            sample (torch.FloatTensor): [bs, n_data, c]
+
+        """
+
+        _, n_data, _ = model_input.shape
+
+        # 1. time
+        t_emb = self.time_proj(self.time_embed(timestep)).unsqueeze(dim=1)
+
+        # 2. conditions projector
+        context = self.context_embed(context)
+
+        # 3. denoiser
+        x = self.input_proj(model_input)
+        x = torch.cat([t_emb, context, x], dim=1)
+        x = self.backbone(x)
+        x = self.ln_post(x)
+        x = x[:, -n_data:]
+        sample = self.output_proj(x)
+
+        return sample
+
+

MeshAnything/miche/michelangelo/models/asl_diffusion/base.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+
+import torch
+import torch.nn as nn
+
+
+class BaseDenoiser(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x, t, context):
+        raise NotImplementedError

MeshAnything/miche/michelangelo/models/asl_diffusion/clip_asl_diffuser_pl_module.py

@@ -0,0 +1,393 @@
+# -*- coding: utf-8 -*-
+
+from omegaconf import DictConfig
+from typing import List, Tuple, Dict, Optional, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.optim import lr_scheduler
+import pytorch_lightning as pl
+from pytorch_lightning.utilities import rank_zero_only
+
+from diffusers.schedulers import (
+    DDPMScheduler,
+    DDIMScheduler,
+    KarrasVeScheduler,
+    DPMSolverMultistepScheduler
+)
+
+from MeshAnything.miche.michelangelo.utils import instantiate_from_config
+from MeshAnything.miche.michelangelo.models.tsal.tsal_base import AlignedShapeAsLatentPLModule
+from MeshAnything.miche.michelangelo.models.asl_diffusion.inference_utils import ddim_sample
+
+SchedulerType = Union[DDIMScheduler, KarrasVeScheduler, DPMSolverMultistepScheduler]
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class ClipASLDiffuser(pl.LightningModule):
+    first_stage_model: Optional[AlignedShapeAsLatentPLModule]
+    cond_stage_model: Optional[Union[nn.Module, pl.LightningModule]]
+    model: nn.Module
+
+    def __init__(self, *,
+                 first_stage_config,
+                 cond_stage_config,
+                 denoiser_cfg,
+                 scheduler_cfg,
+                 optimizer_cfg,
+                 loss_cfg,
+                 first_stage_key: str = "surface",
+                 cond_stage_key: str = "image",
+                 scale_by_std: bool = False,
+                 z_scale_factor: float = 1.0,
+                 ckpt_path: Optional[str] = None,
+                 ignore_keys: Union[Tuple[str], List[str]] = ()):
+
+        super().__init__()
+
+        self.first_stage_key = first_stage_key
+        self.cond_stage_key = cond_stage_key
+
+        # 1. lazy initialize first stage
+        self.instantiate_first_stage(first_stage_config)
+
+        # 2. initialize conditional stage
+        self.instantiate_cond_stage(cond_stage_config)
+
+        # 3. diffusion model
+        self.model = instantiate_from_config(
+            denoiser_cfg, device=None, dtype=None
+        )
+
+        self.optimizer_cfg = optimizer_cfg
+
+        # 4. scheduling strategy
+        self.scheduler_cfg = scheduler_cfg
+
+        self.noise_scheduler: DDPMScheduler = instantiate_from_config(scheduler_cfg.noise)
+        self.denoise_scheduler: SchedulerType = instantiate_from_config(scheduler_cfg.denoise)
+
+        # 5. loss configures
+        self.loss_cfg = loss_cfg
+
+        self.scale_by_std = scale_by_std
+        if scale_by_std:
+            self.register_buffer("z_scale_factor", torch.tensor(z_scale_factor))
+        else:
+            self.z_scale_factor = z_scale_factor
+
+        self.ckpt_path = ckpt_path
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+
+    def instantiate_non_trainable_model(self, config):
+        model = instantiate_from_config(config)
+        model = model.eval()
+        model.train = disabled_train
+        for param in model.parameters():
+            param.requires_grad = False
+
+        return model
+
+    def instantiate_first_stage(self, first_stage_config):
+        self.first_stage_model = self.instantiate_non_trainable_model(first_stage_config)
+        self.first_stage_model.set_shape_model_only()
+
+    def instantiate_cond_stage(self, cond_stage_config):
+        self.cond_stage_model = self.instantiate_non_trainable_model(cond_stage_config)
+
+    def init_from_ckpt(self, path, ignore_keys=()):
+        state_dict = torch.load(path, map_location="cpu")["state_dict"]
+
+        keys = list(state_dict.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del state_dict[k]
+
+        missing, unexpected = self.load_state_dict(state_dict, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+            print(f"Unexpected Keys: {unexpected}")
+
+    @property
+    def zero_rank(self):
+        if self._trainer:
+            zero_rank = self.trainer.local_rank == 0
+        else:
+            zero_rank = True
+
+        return zero_rank
+
+    def configure_optimizers(self) -> Tuple[List, List]:
+
+        lr = self.learning_rate
+
+        trainable_parameters = list(self.model.parameters())
+        if self.optimizer_cfg is None:
+            optimizers = [torch.optim.AdamW(trainable_parameters, lr=lr, betas=(0.9, 0.99), weight_decay=1e-3)]
+            schedulers = []
+        else:
+            optimizer = instantiate_from_config(self.optimizer_cfg.optimizer, params=trainable_parameters)
+            scheduler_func = instantiate_from_config(
+                self.optimizer_cfg.scheduler,
+                max_decay_steps=self.trainer.max_steps,
+                lr_max=lr
+            )
+            scheduler = {
+                "scheduler": lr_scheduler.LambdaLR(optimizer, lr_lambda=scheduler_func.schedule),
+                "interval": "step",
+                "frequency": 1
+            }
+            optimizers = [optimizer]
+            schedulers = [scheduler]
+
+        return optimizers, schedulers
+
+    @torch.no_grad()
+    def encode_first_stage(self, surface: torch.FloatTensor, sample_posterior=True):
+
+        z_q = self.first_stage_model.encode(surface, sample_posterior)
+        z_q = self.z_scale_factor * z_q
+
+        return z_q
+
+    @torch.no_grad()
+    def decode_first_stage(self, z_q: torch.FloatTensor, **kwargs):
+
+        z_q = 1. / self.z_scale_factor * z_q
+        latents = self.first_stage_model.decode(z_q, **kwargs)
+        return latents
+
+    @rank_zero_only
+    @torch.no_grad()
+    def on_train_batch_start(self, batch, batch_idx):
+        # only for very first batch
+        if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 \
+                and batch_idx == 0 and self.ckpt_path is None:
+            # set rescale weight to 1./std of encodings
+            print("### USING STD-RESCALING ###")
+
+            z_q = self.encode_first_stage(batch[self.first_stage_key])
+            z = z_q.detach()
+
+            del self.z_scale_factor
+            self.register_buffer("z_scale_factor", 1. / z.flatten().std())
+            print(f"setting self.z_scale_factor to {self.z_scale_factor}")
+
+            print("### USING STD-RESCALING ###")
+
+    def compute_loss(self, model_outputs, split):
+        """
+
+        Args:
+            model_outputs (dict):
+                - x_0:
+                - noise:
+                - noise_prior:
+                - noise_pred:
+                - noise_pred_prior:
+
+            split (str):
+
+        Returns:
+
+        """
+
+        pred = model_outputs["pred"]
+
+        if self.noise_scheduler.prediction_type == "epsilon":
+            target = model_outputs["noise"]
+        elif self.noise_scheduler.prediction_type == "sample":
+            target = model_outputs["x_0"]
+        else:
+            raise NotImplementedError(f"Prediction Type: {self.noise_scheduler.prediction_type} not yet supported.")
+
+        if self.loss_cfg.loss_type == "l1":
+            simple = F.l1_loss(pred, target, reduction="mean")
+        elif self.loss_cfg.loss_type in ["mse", "l2"]:
+            simple = F.mse_loss(pred, target, reduction="mean")
+        else:
+            raise NotImplementedError(f"Loss Type: {self.loss_cfg.loss_type} not yet supported.")
+
+        total_loss = simple
+
+        loss_dict = {
+            f"{split}/total_loss": total_loss.clone().detach(),
+            f"{split}/simple": simple.detach(),
+        }
+
+        return total_loss, loss_dict
+
+    def forward(self, batch):
+        """
+
+        Args:
+            batch:
+
+        Returns:
+
+        """
+
+        latents = self.encode_first_stage(batch[self.first_stage_key])
+        conditions = self.cond_stage_model.encode(batch[self.cond_stage_key])
+
+        # Sample noise that we"ll add to the latents
+        # [batch_size, n_token, latent_dim]
+        noise = torch.randn_like(latents)
+        bs = latents.shape[0]
+        # Sample a random timestep for each motion
+        timesteps = torch.randint(
+            0,
+            self.noise_scheduler.config.num_train_timesteps,
+            (bs,),
+            device=latents.device,
+        )
+        timesteps = timesteps.long()
+        # Add noise to the latents according to the noise magnitude at each timestep
+        noisy_z = self.noise_scheduler.add_noise(latents, noise, timesteps)
+
+        # diffusion model forward
+        noise_pred = self.model(noisy_z, timesteps, conditions)
+
+        diffusion_outputs = {
+            "x_0": noisy_z,
+            "noise": noise,
+            "pred": noise_pred
+        }
+
+        return diffusion_outputs
+
+    def training_step(self, batch: Dict[str, Union[torch.FloatTensor, List[str]]],
+                      batch_idx: int, optimizer_idx: int = 0) -> torch.FloatTensor:
+        """
+
+        Args:
+            batch (dict): the batch sample, and it contains:
+                - surface (torch.FloatTensor):
+                - image (torch.FloatTensor): if provide, [bs, 3, h, w], item range [0, 1]
+                - depth (torch.FloatTensor): if provide, [bs, 1, h, w], item range [-1, 1]
+                - normal (torch.FloatTensor): if provide, [bs, 3, h, w], item range [-1, 1]
+                - text (list of str):
+
+            batch_idx (int):
+
+            optimizer_idx (int):
+
+        Returns:
+            loss (torch.FloatTensor):
+
+        """
+
+        diffusion_outputs = self(batch)
+
+        loss, loss_dict = self.compute_loss(diffusion_outputs, "train")
+        self.log_dict(loss_dict, prog_bar=True, logger=True, sync_dist=False, rank_zero_only=True)
+
+        return loss
+
+    def validation_step(self, batch: Dict[str, torch.FloatTensor],
+                        batch_idx: int, optimizer_idx: int = 0) -> torch.FloatTensor:
+        """
+
+        Args:
+            batch (dict): the batch sample, and it contains:
+                - surface_pc (torch.FloatTensor): [n_pts, 4]
+                - surface_feats (torch.FloatTensor): [n_pts, c]
+                - text (list of str):
+
+            batch_idx (int):
+
+            optimizer_idx (int):
+
+        Returns:
+            loss (torch.FloatTensor):
+
+        """
+
+        diffusion_outputs = self(batch)
+
+        loss, loss_dict = self.compute_loss(diffusion_outputs, "val")
+        self.log_dict(loss_dict, prog_bar=True, logger=True, sync_dist=False, rank_zero_only=True)
+
+        return loss
+
+    @torch.no_grad()
+    def sample(self,
+               batch: Dict[str, Union[torch.FloatTensor, List[str]]],
+               sample_times: int = 1,
+               steps: Optional[int] = None,
+               guidance_scale: Optional[float] = None,
+               eta: float = 0.0,
+               return_intermediates: bool = False, **kwargs):
+
+        if steps is None:
+            steps = self.scheduler_cfg.num_inference_steps
+
+        if guidance_scale is None:
+            guidance_scale = self.scheduler_cfg.guidance_scale
+        do_classifier_free_guidance = guidance_scale > 0
+
+        # conditional encode
+        xc = batch[self.cond_stage_key]
+
+        # print(self.first_stage_model.device, self.cond_stage_model.device, self.device)
+
+        cond = self.cond_stage_model(xc)
+
+        if do_classifier_free_guidance:
+            un_cond = self.cond_stage_model.unconditional_embedding(batch_size=len(xc))
+            cond = torch.cat([un_cond, cond], dim=0)
+
+        outputs = []
+        latents = None
+
+        if not return_intermediates:
+            for _ in range(sample_times):
+                sample_loop = ddim_sample(
+                    self.denoise_scheduler,
+                    self.model,
+                    shape=self.first_stage_model.latent_shape,
+                    cond=cond,
+                    steps=steps,
+                    guidance_scale=guidance_scale,
+                    do_classifier_free_guidance=do_classifier_free_guidance,
+                    device=self.device,
+                    eta=eta,
+                    disable_prog=not self.zero_rank
+                )
+                for sample, t in sample_loop:
+                    latents = sample
+                outputs.append(self.decode_first_stage(latents, **kwargs))
+        else:
+
+            sample_loop = ddim_sample(
+                self.denoise_scheduler,
+                self.model,
+                shape=self.first_stage_model.latent_shape,
+                cond=cond,
+                steps=steps,
+                guidance_scale=guidance_scale,
+                do_classifier_free_guidance=do_classifier_free_guidance,
+                device=self.device,
+                eta=eta,
+                disable_prog=not self.zero_rank
+            )
+
+            iter_size = steps // sample_times
+            i = 0
+            for sample, t in sample_loop:
+                latents = sample
+                if i % iter_size == 0 or i == steps - 1:
+                    outputs.append(self.decode_first_stage(latents, **kwargs))
+                i += 1
+
+        return outputs

MeshAnything/miche/michelangelo/models/asl_diffusion/inference_utils.py

@@ -0,0 +1,80 @@
+# -*- coding: utf-8 -*-
+
+import torch
+from tqdm import tqdm
+from typing import Tuple, List, Union, Optional
+from diffusers.schedulers import DDIMScheduler
+
+
+__all__ = ["ddim_sample"]
+
+
+def ddim_sample(ddim_scheduler: DDIMScheduler,
+                diffusion_model: torch.nn.Module,
+                shape: Union[List[int], Tuple[int]],
+                cond: torch.FloatTensor,
+                steps: int,
+                eta: float = 0.0,
+                guidance_scale: float = 3.0,
+                do_classifier_free_guidance: bool = True,
+                generator: Optional[torch.Generator] = None,
+                device: torch.device = "cuda:0",
+                disable_prog: bool = True):
+
+    assert steps > 0, f"{steps} must > 0."
+
+    # init latents
+    bsz = cond.shape[0]
+    if do_classifier_free_guidance:
+        bsz = bsz // 2
+
+    latents = torch.randn(
+        (bsz, *shape),
+        generator=generator,
+        device=cond.device,
+        dtype=cond.dtype,
+    )
+    # scale the initial noise by the standard deviation required by the scheduler
+    latents = latents * ddim_scheduler.init_noise_sigma
+    # set timesteps
+    ddim_scheduler.set_timesteps(steps)
+    timesteps = ddim_scheduler.timesteps.to(device)
+    # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+    # eta (η) is only used with the DDIMScheduler, and between [0, 1]
+    extra_step_kwargs = {
+        "eta": eta,
+        "generator": generator
+    }
+
+    # reverse
+    for i, t in enumerate(tqdm(timesteps, disable=disable_prog, desc="DDIM Sampling:", leave=False)):
+        # expand the latents if we are doing classifier free guidance
+        latent_model_input = (
+            torch.cat([latents] * 2)
+            if do_classifier_free_guidance
+            else latents
+        )
+        # latent_model_input = scheduler.scale_model_input(latent_model_input, t)
+        # predict the noise residual
+        timestep_tensor = torch.tensor([t], dtype=torch.long, device=device)
+        timestep_tensor = timestep_tensor.expand(latent_model_input.shape[0])
+        noise_pred = diffusion_model.forward(latent_model_input, timestep_tensor, cond)
+
+        # perform guidance
+        if do_classifier_free_guidance:
+            noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+            noise_pred = noise_pred_uncond + guidance_scale * (
+                    noise_pred_text - noise_pred_uncond
+            )
+            # text_embeddings_for_guidance = encoder_hidden_states.chunk(
+            #     2)[1] if do_classifier_free_guidance else encoder_hidden_states
+        # compute the previous noisy sample x_t -> x_t-1
+        latents = ddim_scheduler.step(
+            noise_pred, t, latents, **extra_step_kwargs
+        ).prev_sample
+
+        yield latents, t
+
+
+def karra_sample():
+    pass

MeshAnything/miche/michelangelo/models/conditional_encoders/__init__.py

@@ -0,0 +1,3 @@
+# -*- coding: utf-8 -*-
+
+from .clip import CLIPEncoder

MeshAnything/miche/michelangelo/models/conditional_encoders/clip.py

@@ -0,0 +1,89 @@
+# -*- coding: utf-8 -*-
+
+import torch
+import numpy as np
+from PIL import Image
+from dataclasses import dataclass
+from torchvision.transforms import Normalize
+from transformers import CLIPModel, CLIPTokenizer
+from transformers.utils import ModelOutput
+from typing import Iterable, Optional, Union, List
+
+
+ImageType = Union[np.ndarray, torch.Tensor, Image.Image]
+
+
+@dataclass
+class CLIPEmbedOutput(ModelOutput):
+    last_hidden_state: torch.FloatTensor = None
+    pooler_output: torch.FloatTensor = None
+    embeds: torch.FloatTensor = None
+
+
+class CLIPEncoder(torch.nn.Module):
+
+    def __init__(self, model_path="openai/clip-vit-base-patch32"):
+
+        super().__init__()
+
+        # Load the CLIP model and processor
+        self.model: CLIPModel = CLIPModel.from_pretrained(model_path)
+        self.tokenizer = CLIPTokenizer.from_pretrained(model_path)
+        self.image_preprocess = Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+
+        self.model.training = False
+        for p in self.model.parameters():
+            p.requires_grad = False
+
+    @torch.no_grad()
+    def encode_image(self, images: Iterable[Optional[ImageType]]):
+        pixel_values = self.image_preprocess(images)
+
+        vision_outputs = self.model.vision_model(pixel_values=pixel_values)
+
+        pooler_output = vision_outputs[1]  # pooled_output
+        image_features = self.model.visual_projection(pooler_output)
+
+        visual_embeds = CLIPEmbedOutput(
+            last_hidden_state=vision_outputs.last_hidden_state,
+            pooler_output=pooler_output,
+            embeds=image_features
+        )
+
+        return visual_embeds
+
+    @torch.no_grad()
+    def encode_text(self, texts: List[str]):
+        text_inputs = self.tokenizer(texts, padding=True, return_tensors="pt")
+
+        text_outputs = self.model.text_model(input_ids=text_inputs)
+
+        pooler_output = text_outputs[1]  # pooled_output
+        text_features = self.model.text_projection(pooler_output)
+
+        text_embeds = CLIPEmbedOutput(
+            last_hidden_state=text_outputs.last_hidden_state,
+            pooler_output=pooler_output,
+            embeds=text_features
+        )
+
+        return text_embeds
+
+    def forward(self,
+                images: Iterable[Optional[ImageType]],
+                texts: List[str]):
+
+        visual_embeds = self.encode_image(images)
+        text_embeds = self.encode_text(texts)
+
+        return visual_embeds, text_embeds
+
+
+
+
+
+
+
+
+
+

MeshAnything/miche/michelangelo/models/conditional_encoders/encoder_factory.py

@@ -0,0 +1,562 @@
+# -*- coding: utf-8 -*-
+import os
+
+import torch
+import torch.nn as nn
+from torchvision import transforms
+from transformers import CLIPModel, CLIPTokenizer
+from collections import OrderedDict
+
+from MeshAnything.miche.michelangelo.data.transforms import RandomResize
+
+
+class AbstractEncoder(nn.Module):
+    embedding_dim: int
+
+    def __init__(self):
+        super().__init__()
+
+    def encode(self, *args, **kwargs):
+        raise NotImplementedError
+
+
+class ClassEmbedder(nn.Module):
+    def __init__(self, embed_dim, n_classes=1000, key="class"):
+        super().__init__()
+        self.key = key
+        self.embedding = nn.Embedding(n_classes, embed_dim)
+
+    def forward(self, batch, key=None):
+        if key is None:
+            key = self.key
+        # this is for use in crossattn
+        c = batch[key][:, None]
+        c = self.embedding(c)
+        return c
+
+
+class FrozenCLIPTextEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from Hugging Face)"""
+
+    def __init__(
+        self,
+        version="openai/clip-vit-large-patch14",
+        tokenizer_version=None,
+        device="cuda",
+        max_length=77,
+        zero_embedding_radio: float = 0.1,
+    ):
+        super().__init__()
+        self.tokenizer = CLIPTokenizer.from_pretrained(tokenizer_version or version)
+
+        self.device = device
+        self.max_length = max_length
+        self.zero_embedding_radio = zero_embedding_radio
+
+        self.clip_dict = OrderedDict()
+        self.clip_name = os.path.split(version)[-1]
+
+        transformer = CLIPModel.from_pretrained(version).text_model
+
+        for param in transformer.parameters():
+            param.requires_grad = False
+        self.clip_dict[self.clip_name] = transformer
+
+        self._move_flag = False
+
+    @property
+    def clip(self):
+        return self.clip_dict[self.clip_name]
+
+    def move(self):
+        if self._move_flag:
+            return
+
+        self.clip_dict[self.clip_name] = self.clip_dict[self.clip_name].to(self.device)
+        self._move_flag = True
+
+    def unconditional_embedding(self, batch_size):
+        empty_text = [""] * batch_size
+        empty_z = self.forward(empty_text)
+        return empty_z
+
+    def forward(self, text):
+        self.move()
+
+        batch_encoding = self.tokenizer(
+            text,
+            truncation=True,
+            max_length=self.max_length,
+            return_length=True,
+            return_overflowing_tokens=False,
+            padding="max_length",
+            return_tensors="pt",
+        )
+
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.clip(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        batch_size = len(text)
+        batch_mask = torch.rand((batch_size,))
+        for i in range(batch_size):
+            if batch_mask[i] < self.zero_embedding_radio:
+                text[i] = ""
+
+        return self(text)
+
+class FrozenAlignedCLIPTextEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from Hugging Face)"""
+
+    def __init__(
+        self,
+        version="openai/clip-vit-large-patch14",
+        tokenizer_version=None,
+        device="cuda",
+        max_length=77,
+        zero_embedding_radio: float = 0.1,
+    ):
+        super().__init__()
+        self.tokenizer = CLIPTokenizer.from_pretrained(tokenizer_version or version)
+
+        self.device = device
+        self.max_length = max_length
+        self.zero_embedding_radio = zero_embedding_radio
+
+        self.clip_dict = OrderedDict()
+        self.clip_name = os.path.split(version)[-1]
+
+        transformer = CLIPModel.from_pretrained(version).text_model
+
+        for param in transformer.parameters():
+            param.requires_grad = False
+        self.clip_dict[self.clip_name] = transformer
+
+        self._move_flag = False
+
+    @property
+    def clip(self):
+        return self.clip_dict[self.clip_name]
+
+    def move(self):
+        if self._move_flag:
+            return
+
+        self.clip_dict[self.clip_name] = self.clip_dict[self.clip_name].to(self.device)
+        self._move_flag = True
+
+    def unconditional_embedding(self, batch_size):
+        empty_text = [""] * batch_size
+        empty_z = self.forward(empty_text)
+        return empty_z
+
+    def forward(self, text):
+        self.move()
+
+        batch_encoding = self.tokenizer(
+            text,
+            truncation=True,
+            max_length=self.max_length,
+            return_length=True,
+            return_overflowing_tokens=False,
+            padding="max_length",
+            return_tensors="pt",
+        )
+
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.clip(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        batch_size = len(text)
+        batch_mask = torch.rand((batch_size,))
+        for i in range(batch_size):
+            if batch_mask[i] < self.zero_embedding_radio:
+                text[i] = ""
+
+        return self(text)
+
+
+class FrozenCLIPImageEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from Hugging Face)"""
+
+    def __init__(
+            self,
+            version="openai/clip-vit-large-patch14",
+            device="cuda",
+            zero_embedding_radio=0.1,
+            normalize_embedding=True,
+            num_projection_vector=0,
+            linear_mapping_bias=True,
+            reverse_visual_projection=False,
+    ):
+        super().__init__()
+
+        self.device = device
+
+        self.clip_dict = OrderedDict()
+        self.clip_name = os.path.split(version)[-1]
+
+        clip_model = CLIPModel.from_pretrained(version)
+        clip_model.text_model = None
+        clip_model.text_projection = None
+        clip_model = clip_model.eval()
+        for param in self.parameters():
+            param.requires_grad = False
+        self.clip_dict[self.clip_name] = clip_model
+
+        self.transform = transforms.Compose(
+            [
+                transforms.Resize(224, transforms.InterpolationMode.BICUBIC, antialias=True),
+                transforms.CenterCrop(224),  # crop a (224, 224) square
+                transforms.Normalize(
+                    mean=[0.48145466, 0.4578275, 0.40821073],
+                    std=[0.26862954, 0.26130258, 0.27577711],
+                ),
+            ]
+        )
+        self.zero_embedding_radio = zero_embedding_radio
+
+        self.num_projection_vector = num_projection_vector
+        self.reverse_visual_projection = reverse_visual_projection
+        self.normalize_embedding = normalize_embedding
+
+        embedding_dim = (
+            clip_model.visual_projection.in_features
+            if reverse_visual_projection
+            else clip_model.visual_projection.out_features
+        )
+        self.embedding_dim = embedding_dim
+        if self.num_projection_vector > 0:
+            self.projection = nn.Linear(
+                embedding_dim,
+                clip_model.visual_projection.out_features * num_projection_vector,
+                bias=linear_mapping_bias,
+            )
+            nn.init.normal_(self.projection.weight, std=embedding_dim ** -0.5)
+
+        self._move_flag = False
+
+    @property
+    def clip(self):
+        return self.clip_dict[self.clip_name]
+
+    def unconditional_embedding(self, batch_size):
+        zero = torch.zeros(
+            batch_size,
+            1,
+            self.embedding_dim,
+            device=self.device,
+            dtype=self.clip.visual_projection.weight.dtype,
+        )
+        if self.num_projection_vector > 0:
+            zero = self.projection(zero).view(batch_size, self.num_projection_vector, -1)
+        return zero
+
+    def forward(self, image, value_range=(-1, 1), zero_embedding_radio=0):
+        if value_range is not None:
+            low, high = value_range
+            image = (image - low) / (high - low)
+
+        image = image.to(self.device, dtype=self.clip.visual_projection.weight.dtype)
+
+        if self.reverse_visual_projection:
+            z = self.clip.vision_model(self.transform(image))[1]
+        else:
+            z = self.clip.get_image_features(self.transform(image))
+
+        if self.normalize_embedding:
+            z = z / z.norm(dim=-1, keepdim=True)
+        if z.ndim == 2:
+            z = z.unsqueeze(dim=-2)
+
+        if zero_embedding_radio > 0:
+            mask = torch.rand((len(image), 1, 1), device=z.device, dtype=z.dtype) < zero_embedding_radio
+            z = z * mask.to(z)
+
+        if self.num_projection_vector > 0:
+            z = self.projection(z).view(len(image), self.num_projection_vector, -1)
+
+        return z
+
+    def move(self):
+        if self._move_flag:
+            return
+
+        self.clip_dict[self.clip_name] = self.clip_dict[self.clip_name].to(self.device)
+        self._move_flag = True
+
+    def encode(self, image):
+        self.move()
+        return self(image, zero_embedding_radio=self.zero_embedding_radio)
+
+
+class FrozenCLIPImageGridEmbedder(AbstractEncoder):
+
+    def __init__(
+            self,
+            version="openai/clip-vit-large-patch14",
+            device="cuda",
+            zero_embedding_radio=0.1,
+    ):
+        super().__init__()
+
+        self.device = device
+
+        self.clip_dict = OrderedDict()
+        self.clip_name = os.path.split(version)[-1]
+
+        clip_model: CLIPModel = CLIPModel.from_pretrained(version)
+        clip_model.text_model = None
+        clip_model.text_projection = None
+        clip_model = clip_model.eval()
+        for param in self.parameters():
+            param.requires_grad = False
+        self.clip_dict[self.clip_name] = clip_model
+
+        self.transform = transforms.Compose(
+            [
+                transforms.Resize(224, transforms.InterpolationMode.BILINEAR, antialias=True),
+                transforms.CenterCrop(224),  # crop a (224, 224) square
+                transforms.Normalize(
+                    mean=[0.48145466, 0.4578275, 0.40821073],
+                    std=[0.26862954, 0.26130258, 0.27577711],
+                ),
+            ]
+        )
+        self.zero_embedding_radio = zero_embedding_radio
+        self.embedding_dim = clip_model.vision_embed_dim
+
+        self._move_flag = False
+
+    @property
+    def clip(self):
+        return self.clip_dict[self.clip_name]
+
+    def move(self):
+        if self._move_flag:
+            return
+
+        self.clip_dict[self.clip_name] = self.clip_dict[self.clip_name].to(self.device)
+        self._move_flag = True
+
+    def unconditional_embedding(self, batch_size):
+        zero = torch.zeros(
+            batch_size,
+            self.clip.vision_model.embeddings.num_positions,
+            self.embedding_dim,
+            device=self.device,
+            dtype=self.clip.visual_projection.weight.dtype,
+        )
+        return zero
+
+    def forward(self, image, value_range=(-1, 1), zero_embedding_radio=0):
+        self.move()
+
+        if value_range is not None:
+            low, high = value_range
+            image = (image - low) / (high - low)
+
+        image = image.to(self.device, dtype=self.clip.visual_projection.weight.dtype)
+
+        z = self.clip.vision_model(self.transform(image)).last_hidden_state
+
+        if zero_embedding_radio > 0:
+            mask = torch.rand((len(image), 1, 1), device=z.device, dtype=z.dtype) >= zero_embedding_radio
+            z = z * mask.to(z)
+
+        return z
+
+    def encode(self, image):
+        return self(image, zero_embedding_radio=self.zero_embedding_radio)
+
+
+class MoECLIPImageEncoder(nn.Module):
+    def __init__(
+            self,
+            versions,
+            hidden_state_dim,
+            num_projection_vector=8,
+            zero_embedding_radio=0.1,
+            device="cuda",
+            precision="fp16",
+            normalize=False,
+            clip_max=0,
+            transform_type="base",
+            argument_p=0.2,
+    ):
+        super().__init__()
+
+        self.device = torch.device(device)
+        self.hidden_state_dim = hidden_state_dim
+        self.zero_embedding_radio = zero_embedding_radio
+        self.num_projection_vector = num_projection_vector
+        self.dtype = dict(fp16=torch.float16, fp32=torch.float32, bf16=torch.bfloat16)[precision]
+        self.normalize = normalize
+        self.clip_max = clip_max
+
+        if transform_type == "base":
+            self.transform = transforms.Compose(
+                [
+                    transforms.Resize(224, transforms.InterpolationMode.BICUBIC, antialias=True),
+                    transforms.CenterCrop(224),  # crop a (224, 224) square
+                    transforms.Normalize(
+                        mean=[0.48145466, 0.4578275, 0.40821073],
+                        std=[0.26862954, 0.26130258, 0.27577711],
+                    ),
+                ]
+            )
+        elif transform_type == "crop_blur_resize":
+            self.transform = transforms.Compose(
+                [
+                    transforms.Resize(224, transforms.InterpolationMode.BICUBIC, antialias=True),
+                    transforms.CenterCrop(224),  # crop a (224, 224) square
+                    transforms.RandomApply(
+                        transforms=[
+                            transforms.RandomResizedCrop(
+                                size=224,
+                                scale=(0.8, 1.0),
+                                ratio=(0.99, 1.01),
+                                interpolation=transforms.InterpolationMode.BICUBIC,
+                            ),
+                        ],
+                        p=argument_p,
+                    ),
+                    transforms.RandomApply(
+                        transforms=[
+                            transforms.GaussianBlur(kernel_size=9, sigma=(0.1, 5)),
+                        ],
+                        p=argument_p,
+                    ),
+                    transforms.RandomApply(
+                        transforms=[
+                            RandomResize(size=224, resize_radio=(0.2, 1)),
+                        ],
+                        p=argument_p,
+                    ),
+                    transforms.Normalize(
+                        mean=[0.48145466, 0.4578275, 0.40821073],
+                        std=[0.26862954, 0.26130258, 0.27577711],
+                    ),
+                ]
+            )
+        else:
+            raise ValueError(f"invalid {transform_type=}")
+
+        if isinstance(versions, str):
+            versions = (versions,)
+
+        # 如果直接把clips定位为当前类的子module，1. 会在保存ckp时存无用的多个权重。 2. pl会调用to，导致layer_norm的权重也被转换成fp16
+        clips = OrderedDict()
+
+        for v in versions:
+            # 因为clips不是子module，直接指定device="cuda"会错误地导致clip模型权重都被放到cuda:0上。
+            clips[v], _ = clip.load(name=v, device="cpu", jit=False, download_root=None)
+            delattr(clips[v], "transformer")
+            clips[v].eval()
+            clips[v].requires_grad_(False)
+
+        self.clips_hidden_dim = sum(clips[v].ln_final.weight.size(0) for v in clips)
+
+        if self.num_projection_vector == 0:
+            self.projection = nn.Identity()
+        else:
+            self.projection = nn.Linear(self.clips_hidden_dim, hidden_state_dim * self.num_projection_vector, bias=True)
+            self.projection.to(dtype=self.dtype)
+            nn.init.normal_(self.projection.weight, std=self.clips_hidden_dim ** -0.5)
+
+        self.clips = clips
+
+        self._move_flag = False
+
+    def move(self):
+        if self._move_flag:
+            return
+
+        def convert_weights(model: nn.Module):
+            """Convert applicable model parameters to fp16"""
+
+            def _convert_weights_to_fp16(l):
+                if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+                    l.weight.data = l.weight.data.type(self.dtype)
+                    if l.bias is not None:
+                        l.bias.data = l.bias.data.type(self.dtype)
+
+                if isinstance(l, nn.MultiheadAttention):
+                    for attr in [
+                        *[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]],
+                        "in_proj_bias",
+                        "bias_k",
+                        "bias_v",
+                    ]:
+                        tensor = getattr(l, attr)
+                        if tensor is not None:
+                            tensor.data = tensor.data.type(self.dtype)
+
+                for name in ["text_projection", "proj"]:
+                    if hasattr(l, name):
+                        attr = getattr(l, name)
+                        if attr is not None:
+                            attr.data = attr.data.type(self.dtype)
+
+            model.apply(_convert_weights_to_fp16)
+
+        for k in self.clips:
+            self.clips[k].to(self.device)
+            convert_weights(self.clips[k])  # fp32 -> self.dtype
+        self._move_flag = True
+
+    def unconditional_embedding(self, batch_size=None):
+        zero = torch.zeros(
+            batch_size,
+            self.clips_hidden_dim,
+            device=self.device,
+            dtype=self.dtype,
+        )
+        if self.num_projection_vector > 0:
+            zero = self.projection(zero).view(batch_size, self.num_projection_vector, -1)
+        return zero
+
+    def convert_embedding(self, z):
+        if self.num_projection_vector > 0:
+            z = self.projection(z.type(self.projection.weight.dtype)).view(len(z), self.num_projection_vector, -1)
+        return z
+
+    def forward(self, image, value_range=(-1, 1), zero_embedding_radio=0):
+        if value_range is not None:
+            low, high = value_range
+            image = (image - low) / (high - low)
+
+        image = self.transform(image)
+
+        with torch.no_grad():
+            embs = []
+            for v in self.clips:
+                x = self.clips[v].encode_image(image)
+                if self.normalize:
+                    x = x / x.norm(p=2, dim=-1, keepdim=True) * (x.size(-1) ** 0.5)
+                    # clip_max only works with normalization
+                    if self.clip_max > 0:
+                        x = x.clamp(-self.clip_max, self.clip_max)
+                embs.append(x)
+
+            z = torch.cat(embs, dim=-1)
+            if self.normalize:
+                z /= z.size(-1) ** 0.5
+
+        if zero_embedding_radio > 0:
+            mask = torch.rand((len(image), 1, 1), device=z.device, dtype=z.dtype) >= zero_embedding_radio
+            z = z + mask.to(z)
+
+        if self.num_projection_vector > 0:
+            z = self.projection(z).view(len(image), self.num_projection_vector, -1)
+        return z
+
+    def encode(self, image):
+        self.move()
+        return self(image, zero_embedding_radio=self.zero_embedding_radio)

MeshAnything/miche/michelangelo/models/modules/__init__.py

@@ -0,0 +1,3 @@
+# -*- coding: utf-8 -*-
+
+from .checkpoint import checkpoint

MeshAnything/miche/michelangelo/models/modules/checkpoint.py

@@ -0,0 +1,69 @@
+# -*- coding: utf-8 -*-
+"""
+Adapted from: https://github.com/openai/guided-diffusion/blob/22e0df8183507e13a7813f8d38d51b072ca1e67c/guided_diffusion/nn.py#L124
+"""
+
+import torch
+from typing import Callable, Iterable, Sequence, Union
+
+
+def checkpoint(
+    func: Callable[..., Union[torch.Tensor, Sequence[torch.Tensor]]],
+    inputs: Sequence[torch.Tensor],
+    params: Iterable[torch.Tensor],
+    flag: bool,
+    use_deepspeed: bool = False
+):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    :param use_deepspeed: if True, use deepspeed
+    """
+    if flag:
+        if use_deepspeed:
+            import deepspeed
+            return deepspeed.checkpointing.checkpoint(func, *inputs)
+
+        args = tuple(inputs) + tuple(params)
+        return CheckpointFunction.apply(func, len(inputs), *args)
+    else:
+        return func(*inputs)
+
+
+class CheckpointFunction(torch.autograd.Function):
+    @staticmethod
+    @torch.cuda.amp.custom_fwd
+    def forward(ctx, run_function, length, *args):
+        ctx.run_function = run_function
+        ctx.input_tensors = list(args[:length])
+        ctx.input_params = list(args[length:])
+
+        with torch.no_grad():
+            output_tensors = ctx.run_function(*ctx.input_tensors)
+        return output_tensors
+
+    @staticmethod
+    @torch.cuda.amp.custom_bwd
+    def backward(ctx, *output_grads):
+        ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
+        with torch.enable_grad():
+            # Fixes a bug where the first op in run_function modifies the
+            # Tensor storage in place, which is not allowed for detach()'d
+            # Tensors.
+            shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
+            output_tensors = ctx.run_function(*shallow_copies)
+        input_grads = torch.autograd.grad(
+            output_tensors,
+            ctx.input_tensors + ctx.input_params,
+            output_grads,
+            allow_unused=True,
+        )
+        del ctx.input_tensors
+        del ctx.input_params
+        del output_tensors
+        return (None, None) + input_grads

MeshAnything/miche/michelangelo/models/modules/diffusion_transformer.py

@@ -0,0 +1,218 @@
+# -*- coding: utf-8 -*-
+
+import math
+import torch
+import torch.nn as nn
+from typing import Optional
+
+from MeshAnything.miche.michelangelo.models.modules.checkpoint import checkpoint
+from MeshAnything.miche.michelangelo.models.modules.transformer_blocks import (
+    init_linear,
+    MLP,
+    MultiheadCrossAttention,
+    MultiheadAttention,
+    ResidualAttentionBlock
+)
+
+
+class AdaLayerNorm(nn.Module):
+    def __init__(self,
+                 device: torch.device,
+                 dtype: torch.dtype,
+                 width: int):
+
+        super().__init__()
+
+        self.silu = nn.SiLU(inplace=True)
+        self.linear = nn.Linear(width, width * 2, device=device, dtype=dtype)
+        self.layernorm = nn.LayerNorm(width, elementwise_affine=False, device=device, dtype=dtype)
+
+    def forward(self, x, timestep):
+        emb = self.linear(timestep)
+        scale, shift = torch.chunk(emb, 2, dim=2)
+        x = self.layernorm(x) * (1 + scale) + shift
+        return x
+
+
+class DitBlock(nn.Module):
+    def __init__(
+            self,
+            *,
+            device: torch.device,
+            dtype: torch.dtype,
+            n_ctx: int,
+            width: int,
+            heads: int,
+            context_dim: int,
+            qkv_bias: bool = False,
+            init_scale: float = 1.0,
+            use_checkpoint: bool = False
+    ):
+        super().__init__()
+
+        self.use_checkpoint = use_checkpoint
+
+        self.attn = MultiheadAttention(
+            device=device,
+            dtype=dtype,
+            n_ctx=n_ctx,
+            width=width,
+            heads=heads,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias
+        )
+        self.ln_1 = AdaLayerNorm(device, dtype, width)
+
+        if context_dim is not None:
+            self.ln_2 = AdaLayerNorm(device, dtype, width)
+            self.cross_attn = MultiheadCrossAttention(
+                device=device,
+                dtype=dtype,
+                width=width,
+                heads=heads,
+                data_width=context_dim,
+                init_scale=init_scale,
+                qkv_bias=qkv_bias
+            )
+
+        self.mlp = MLP(device=device, dtype=dtype, width=width, init_scale=init_scale)
+        self.ln_3 = AdaLayerNorm(device, dtype, width)
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor, context: Optional[torch.Tensor] = None):
+        return checkpoint(self._forward, (x, t, context), self.parameters(), self.use_checkpoint)
+
+    def _forward(self, x: torch.Tensor, t: torch.Tensor, context: Optional[torch.Tensor] = None):
+        x = x + self.attn(self.ln_1(x, t))
+        if context is not None:
+            x = x + self.cross_attn(self.ln_2(x, t), context)
+        x = x + self.mlp(self.ln_3(x, t))
+        return x
+
+
+class DiT(nn.Module):
+    def __init__(
+            self,
+            *,
+            device: Optional[torch.device],
+            dtype: Optional[torch.dtype],
+            n_ctx: int,
+            width: int,
+            layers: int,
+            heads: int,
+            context_dim: int,
+            init_scale: float = 0.25,
+            qkv_bias: bool = False,
+            use_checkpoint: bool = False
+    ):
+        super().__init__()
+        self.n_ctx = n_ctx
+        self.width = width
+        self.layers = layers
+
+        self.resblocks = nn.ModuleList(
+            [
+                DitBlock(
+                    device=device,
+                    dtype=dtype,
+                    n_ctx=n_ctx,
+                    width=width,
+                    heads=heads,
+                    context_dim=context_dim,
+                    qkv_bias=qkv_bias,
+                    init_scale=init_scale,
+                    use_checkpoint=use_checkpoint
+                )
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor, context: Optional[torch.Tensor] = None):
+        for block in self.resblocks:
+            x = block(x, t, context)
+        return x
+
+
+class UNetDiffusionTransformer(nn.Module):
+    def __init__(
+            self,
+            *,
+            device: Optional[torch.device],
+            dtype: Optional[torch.dtype],
+            n_ctx: int,
+            width: int,
+            layers: int,
+            heads: int,
+            init_scale: float = 0.25,
+            qkv_bias: bool = False,
+            skip_ln: bool = False,
+            use_checkpoint: bool = False
+    ):
+        super().__init__()
+
+        self.n_ctx = n_ctx
+        self.width = width
+        self.layers = layers
+
+        self.encoder = nn.ModuleList()
+        for _ in range(layers):
+            resblock = ResidualAttentionBlock(
+                device=device,
+                dtype=dtype,
+                n_ctx=n_ctx,
+                width=width,
+                heads=heads,
+                init_scale=init_scale,
+                qkv_bias=qkv_bias,
+                use_checkpoint=use_checkpoint
+            )
+            self.encoder.append(resblock)
+
+        self.middle_block = ResidualAttentionBlock(
+            device=device,
+            dtype=dtype,
+            n_ctx=n_ctx,
+            width=width,
+            heads=heads,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            use_checkpoint=use_checkpoint
+        )
+
+        self.decoder = nn.ModuleList()
+        for _ in range(layers):
+            resblock = ResidualAttentionBlock(
+                device=device,
+                dtype=dtype,
+                n_ctx=n_ctx,
+                width=width,
+                heads=heads,
+                init_scale=init_scale,
+                qkv_bias=qkv_bias,
+                use_checkpoint=use_checkpoint
+            )
+            linear = nn.Linear(width * 2, width, device=device, dtype=dtype)
+            init_linear(linear, init_scale)
+
+            layer_norm = nn.LayerNorm(width, device=device, dtype=dtype) if skip_ln else None
+
+            self.decoder.append(nn.ModuleList([resblock, linear, layer_norm]))
+
+    def forward(self, x: torch.Tensor):
+
+        enc_outputs = []
+        for block in self.encoder:
+            x = block(x)
+            enc_outputs.append(x)
+
+        x = self.middle_block(x)
+
+        for i, (resblock, linear, layer_norm) in enumerate(self.decoder):
+            x = torch.cat([enc_outputs.pop(), x], dim=-1)
+            x = linear(x)
+
+            if layer_norm is not None:
+                x = layer_norm(x)
+
+            x = resblock(x)
+
+        return x

MeshAnything/miche/michelangelo/models/modules/distributions.py

@@ -0,0 +1,100 @@
+import torch
+import numpy as np
+from typing import Union, List
+
+
+class AbstractDistribution(object):
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters: Union[torch.Tensor, List[torch.Tensor]], deterministic=False, feat_dim=1):
+        self.feat_dim = feat_dim
+        self.parameters = parameters
+
+        if isinstance(parameters, list):
+            self.mean = parameters[0]
+            self.logvar = parameters[1]
+        else:
+            self.mean, self.logvar = torch.chunk(parameters, 2, dim=feat_dim)
+
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean)
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn_like(self.mean)
+        return x
+
+    def kl(self, other=None, dims=(1, 2, 3)):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.mean(torch.pow(self.mean, 2)
+                                        + self.var - 1.0 - self.logvar,
+                                        dim=dims)
+            else:
+                return 0.5 * torch.mean(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=dims)
+
+    def nll(self, sample, dims=(1, 2, 3)):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+            -1.0
+            + logvar2
+            - logvar1
+            + torch.exp(logvar1 - logvar2)
+            + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

MeshAnything/miche/michelangelo/models/modules/embedder.py

@@ -0,0 +1,213 @@
+# -*- coding: utf-8 -*-
+
+import numpy as np
+import torch
+import torch.nn as nn
+import math
+
+VALID_EMBED_TYPES = ["identity", "fourier", "hashgrid", "sphere_harmonic", "triplane_fourier"]
+
+
+class FourierEmbedder(nn.Module):
+    """The sin/cosine positional embedding. Given an input tensor `x` of shape [n_batch, ..., c_dim], it converts
+    each feature dimension of `x[..., i]` into:
+        [
+            sin(x[..., i]),
+            sin(f_1*x[..., i]),
+            sin(f_2*x[..., i]),
+            ...
+            sin(f_N * x[..., i]),
+            cos(x[..., i]),
+            cos(f_1*x[..., i]),
+            cos(f_2*x[..., i]),
+            ...
+            cos(f_N * x[..., i]),
+            x[..., i]     # only present if include_input is True.
+        ], here f_i is the frequency.
+
+    Denote the space is [0 / num_freqs, 1 / num_freqs, 2 / num_freqs, 3 / num_freqs, ..., (num_freqs - 1) / num_freqs].
+    If logspace is True, then the frequency f_i is [2^(0 / num_freqs), ..., 2^(i / num_freqs), ...];
+    Otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)].
+
+    Args:
+        num_freqs (int): the number of frequencies, default is 6;
+        logspace (bool): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+            otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)];
+        input_dim (int): the input dimension, default is 3;
+        include_input (bool): include the input tensor or not, default is True.
+
+    Attributes:
+        frequencies (torch.Tensor): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+                otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1);
+
+        out_dim (int): the embedding size, if include_input is True, it is input_dim * (num_freqs * 2 + 1),
+            otherwise, it is input_dim * num_freqs * 2.
+
+    """
+
+    def __init__(self,
+                 num_freqs: int = 6,
+                 logspace: bool = True,
+                 input_dim: int = 3,
+                 include_input: bool = True,
+                 include_pi: bool = True) -> None:
+
+        """The initialization"""
+
+        super().__init__()
+
+        if logspace:
+            frequencies = 2.0 ** torch.arange(
+                num_freqs,
+                dtype=torch.float32
+            )
+        else:
+            frequencies = torch.linspace(
+                1.0,
+                2.0 ** (num_freqs - 1),
+                num_freqs,
+                dtype=torch.float32
+            )
+
+        if include_pi:
+            frequencies *= torch.pi
+
+        self.register_buffer("frequencies", frequencies, persistent=False)
+        self.include_input = include_input
+        self.num_freqs = num_freqs
+
+        self.out_dim = self.get_dims(input_dim)
+
+    def get_dims(self, input_dim):
+        temp = 1 if self.include_input or self.num_freqs == 0 else 0
+        out_dim = input_dim * (self.num_freqs * 2 + temp)
+
+        return out_dim
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """ Forward process.
+
+        Args:
+            x: tensor of shape [..., dim]
+
+        Returns:
+            embedding: an embedding of `x` of shape [..., dim * (num_freqs * 2 + temp)]
+                where temp is 1 if include_input is True and 0 otherwise.
+        """
+
+        if self.num_freqs > 0:
+            embed = (x[..., None].contiguous() * self.frequencies).view(*x.shape[:-1], -1)
+            if self.include_input:
+                return torch.cat((x, embed.sin(), embed.cos()), dim=-1)
+            else:
+                return torch.cat((embed.sin(), embed.cos()), dim=-1)
+        else:
+            return x
+
+
+class LearnedFourierEmbedder(nn.Module):
+    """ following @crowsonkb "s lead with learned sinusoidal pos emb """
+    """ https://github.com/crowsonkb/v-diffusion-jax/blob/master/diffusion/models/danbooru_128.py#L8 """
+
+    def __init__(self, in_channels, dim):
+        super().__init__()
+        assert (dim % 2) == 0
+        half_dim = dim // 2
+        per_channel_dim = half_dim // in_channels
+        self.weights = nn.Parameter(torch.randn(per_channel_dim))
+
+    def forward(self, x):
+        """
+
+        Args:
+            x (torch.FloatTensor): [..., c]
+
+        Returns:
+            x (torch.FloatTensor): [..., d]
+        """
+
+        # [b, t, c, 1] * [1, d] = [b, t, c, d] -> [b, t, c * d]
+        freqs = (x[..., None] * self.weights[None] * 2 * np.pi).view(*x.shape[:-1], -1)
+        fouriered = torch.cat((x, freqs.sin(), freqs.cos()), dim=-1)
+        return fouriered
+
+
+class TriplaneLearnedFourierEmbedder(nn.Module):
+    def __init__(self, in_channels, dim):
+        super().__init__()
+
+        self.yz_plane_embedder = LearnedFourierEmbedder(in_channels, dim)
+        self.xz_plane_embedder = LearnedFourierEmbedder(in_channels, dim)
+        self.xy_plane_embedder = LearnedFourierEmbedder(in_channels, dim)
+
+        self.out_dim = in_channels + dim
+
+    def forward(self, x):
+
+        yz_embed = self.yz_plane_embedder(x)
+        xz_embed = self.xz_plane_embedder(x)
+        xy_embed = self.xy_plane_embedder(x)
+
+        embed = yz_embed + xz_embed + xy_embed
+
+        return embed
+
+
+def sequential_pos_embed(num_len, embed_dim):
+    assert embed_dim % 2 == 0
+
+    pos = torch.arange(num_len, dtype=torch.float32)
+    omega = torch.arange(embed_dim // 2, dtype=torch.float32)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000 ** omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = torch.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
+
+    embeddings = torch.cat([emb_sin, emb_cos], dim=1)  # (M, D)
+
+    return embeddings
+
+
+def timestep_embedding(timesteps, dim, max_period=10000):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    half = dim // 2
+    freqs = torch.exp(
+        -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+    ).to(device=timesteps.device)
+    args = timesteps[:, None].to(timesteps.dtype) * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+def get_embedder(embed_type="fourier", num_freqs=-1, input_dim=3, degree=4,
+                 num_levels=16, level_dim=2, per_level_scale=2, base_resolution=16,
+                 log2_hashmap_size=19, desired_resolution=None):
+    if embed_type == "identity" or (embed_type == "fourier" and num_freqs == -1):
+        return nn.Identity(), input_dim
+
+    elif embed_type == "fourier":
+        embedder_obj = FourierEmbedder(num_freqs=num_freqs, input_dim=input_dim,
+                                       logspace=True, include_input=True)
+        return embedder_obj, embedder_obj.out_dim
+
+    elif embed_type == "hashgrid":
+        raise NotImplementedError
+
+    elif embed_type == "sphere_harmonic":
+        raise NotImplementedError
+
+    else:
+        raise ValueError(f"{embed_type} is not valid. Currently only supprts {VALID_EMBED_TYPES}")

MeshAnything/miche/michelangelo/models/modules/transformer_blocks.py

@@ -0,0 +1,286 @@
+# -*- coding: utf-8 -*-
+
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Optional
+
+from MeshAnything.miche.michelangelo.models.modules.checkpoint import checkpoint
+
+
+def init_linear(l, stddev):
+    nn.init.normal_(l.weight, std=stddev)
+    if l.bias is not None:
+        nn.init.constant_(l.bias, 0.0)
+
+
+class MultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        n_ctx: int,
+        width: int,
+        heads: int,
+        init_scale: float,
+        qkv_bias: bool,
+        flash: bool = False
+    ):
+        super().__init__()
+        self.n_ctx = n_ctx
+        self.width = width
+        self.heads = heads
+        self.c_qkv = nn.Linear(width, width * 3, bias=qkv_bias, device=device, dtype=dtype)
+        self.c_proj = nn.Linear(width, width, device=device, dtype=dtype)
+        self.attention = QKVMultiheadAttention(device=device, dtype=dtype, heads=heads, n_ctx=n_ctx, flash=flash)
+        init_linear(self.c_qkv, init_scale)
+        init_linear(self.c_proj, init_scale)
+
+    def forward(self, x):
+        x = self.c_qkv(x)
+        x = checkpoint(self.attention, (x,), (), True)
+        x = self.c_proj(x)
+        return x
+
+
+class QKVMultiheadAttention(nn.Module):
+    def __init__(self, *, device: torch.device, dtype: torch.dtype, heads: int, n_ctx: int, flash: bool = False):
+        super().__init__()
+        self.device = device
+        self.dtype = dtype
+        self.heads = heads
+        self.n_ctx = n_ctx
+        self.flash = flash
+
+    def forward(self, qkv):
+        bs, n_ctx, width = qkv.shape
+        attn_ch = width // self.heads // 3
+        scale = 1 / math.sqrt(math.sqrt(attn_ch))
+        qkv = qkv.view(bs, n_ctx, self.heads, -1)
+        q, k, v = torch.split(qkv, attn_ch, dim=-1)
+
+        if self.flash:
+            out = F.scaled_dot_product_attention(q, k, v)
+        else:
+            weight = torch.einsum(
+                "bthc,bshc->bhts", q * scale, k * scale
+            )  # More stable with f16 than dividing afterwards
+            wdtype = weight.dtype
+            weight = torch.softmax(weight.float(), dim=-1).type(wdtype)
+            out = torch.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1)
+
+        return out
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        n_ctx: int,
+        width: int,
+        heads: int,
+        init_scale: float = 1.0,
+        qkv_bias: bool = True,
+        flash: bool = False,
+        use_checkpoint: bool = False
+    ):
+        super().__init__()
+
+        self.use_checkpoint = use_checkpoint
+
+        self.attn = MultiheadAttention(
+            device=device,
+            dtype=dtype,
+            n_ctx=n_ctx,
+            width=width,
+            heads=heads,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            flash=flash
+        )
+        self.ln_1 = nn.LayerNorm(width, device=device, dtype=dtype)
+        self.mlp = MLP(device=device, dtype=dtype, width=width, init_scale=init_scale)
+        self.ln_2 = nn.LayerNorm(width, device=device, dtype=dtype)
+
+    def _forward(self, x: torch.Tensor):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+    def forward(self, x: torch.Tensor):
+        return checkpoint(self._forward, (x,), self.parameters(), self.use_checkpoint)
+
+
+class MultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        width: int,
+        heads: int,
+        init_scale: float,
+        qkv_bias: bool = True,
+        flash: bool = False,
+        n_data: Optional[int] = None,
+        data_width: Optional[int] = None,
+    ):
+        super().__init__()
+        self.n_data = n_data
+        self.width = width
+        self.heads = heads
+        self.data_width = width if data_width is None else data_width
+        self.c_q = nn.Linear(width, width, bias=qkv_bias, device=device, dtype=dtype)
+        self.c_kv = nn.Linear(self.data_width, width * 2, bias=qkv_bias, device=device, dtype=dtype)
+        self.c_proj = nn.Linear(width, width, device=device, dtype=dtype)
+        self.attention = QKVMultiheadCrossAttention(
+            device=device, dtype=dtype, heads=heads, n_data=n_data, flash=flash
+        )
+        init_linear(self.c_q, init_scale)
+        init_linear(self.c_kv, init_scale)
+        init_linear(self.c_proj, init_scale)
+
+    def forward(self, x, data):
+        x = self.c_q(x)
+        data = self.c_kv(data)
+        x = checkpoint(self.attention, (x, data), (), True)
+        x = self.c_proj(x)
+        return x
+
+
+class QKVMultiheadCrossAttention(nn.Module):
+    def __init__(self, *, device: torch.device, dtype: torch.dtype, heads: int,
+                 flash: bool = False, n_data: Optional[int] = None):
+
+        super().__init__()
+        self.device = device
+        self.dtype = dtype
+        self.heads = heads
+        self.n_data = n_data
+        self.flash = flash
+
+    def forward(self, q, kv):
+        _, n_ctx, _ = q.shape
+        bs, n_data, width = kv.shape
+        attn_ch = width // self.heads // 2
+        scale = 1 / math.sqrt(math.sqrt(attn_ch))
+        q = q.view(bs, n_ctx, self.heads, -1)
+        kv = kv.view(bs, n_data, self.heads, -1)
+        k, v = torch.split(kv, attn_ch, dim=-1)
+
+        if self.flash:
+            out = F.scaled_dot_product_attention(q, k, v)
+        else:
+            weight = torch.einsum(
+                "bthc,bshc->bhts", q * scale, k * scale
+            )  # More stable with f16 than dividing afterwards
+            wdtype = weight.dtype
+            weight = torch.softmax(weight.float(), dim=-1).type(wdtype)
+            out = torch.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1)
+
+        return out
+
+
+class ResidualCrossAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: Optional[torch.device],
+        dtype: Optional[torch.dtype],
+        n_data: Optional[int] = None,
+        width: int,
+        heads: int,
+        data_width: Optional[int] = None,
+        init_scale: float = 0.25,
+        qkv_bias: bool = True,
+        flash: bool = False
+    ):
+        super().__init__()
+
+        if data_width is None:
+            data_width = width
+
+        self.attn = MultiheadCrossAttention(
+            device=device,
+            dtype=dtype,
+            n_data=n_data,
+            width=width,
+            heads=heads,
+            data_width=data_width,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            flash=flash,
+        )
+        self.ln_1 = nn.LayerNorm(width, device=device, dtype=dtype)
+        self.ln_2 = nn.LayerNorm(data_width, device=device, dtype=dtype)
+        self.mlp = MLP(device=device, dtype=dtype, width=width, init_scale=init_scale)
+        self.ln_3 = nn.LayerNorm(width, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, data: torch.Tensor):
+        x = x + self.attn(self.ln_1(x), self.ln_2(data))
+        x = x + self.mlp(self.ln_3(x))
+        return x
+
+
+class MLP(nn.Module):
+    def __init__(self, *,
+                 device: Optional[torch.device],
+                 dtype: Optional[torch.dtype],
+                 width: int,
+                 init_scale: float):
+        super().__init__()
+        self.width = width
+        self.c_fc = nn.Linear(width, width * 4, device=device, dtype=dtype)
+        self.c_proj = nn.Linear(width * 4, width, device=device, dtype=dtype)
+        self.gelu = nn.GELU()
+        init_linear(self.c_fc, init_scale)
+        init_linear(self.c_proj, init_scale)
+
+    def forward(self, x):
+        return self.c_proj(self.gelu(self.c_fc(x)))
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: Optional[torch.device],
+        dtype: Optional[torch.dtype],
+        n_ctx: int,
+        width: int,
+        layers: int,
+        heads: int,
+        init_scale: float = 0.25,
+        qkv_bias: bool = True,
+        flash: bool = False,
+        use_checkpoint: bool = False
+    ):
+        super().__init__()
+        self.n_ctx = n_ctx
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(
+                    device=device,
+                    dtype=dtype,
+                    n_ctx=n_ctx,
+                    width=width,
+                    heads=heads,
+                    init_scale=init_scale,
+                    qkv_bias=qkv_bias,
+                    flash=flash,
+                    use_checkpoint=use_checkpoint
+                )
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor):
+        for block in self.resblocks:
+            x = block(x)
+        return x

MeshAnything/miche/michelangelo/models/modules/transformer_vit.py

@@ -0,0 +1,308 @@
+# -*- coding: utf-8 -*-
+
+import math
+import torch
+import torch.nn as nn
+from typing import Optional
+import warnings
+
+from MeshAnything.miche.michelangelo.models.modules.checkpoint import checkpoint
+
+
+def _trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1. + math.erf(x / math.sqrt(2.))) / 2.
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+                      "The distribution of values may be incorrect.",
+                      stacklevel=2)
+
+    # Values are generated by using a truncated uniform distribution and
+    # then using the inverse CDF for the normal distribution.
+    # Get upper and lower cdf values
+    l = norm_cdf((a - mean) / std)
+    u = norm_cdf((b - mean) / std)
+
+    # Uniformly fill tensor with values from [l, u], then translate to
+    # [2l-1, 2u-1].
+    tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+    # Use inverse cdf transform for normal distribution to get truncated
+    # standard normal
+    tensor.erfinv_()
+
+    # Transform to proper mean, std
+    tensor.mul_(std * math.sqrt(2.))
+    tensor.add_(mean)
+
+    # Clamp to ensure it's in the proper range
+    tensor.clamp_(min=a, max=b)
+    return tensor
+
+
+def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
+    # type: (Tensor | nn.Parameter, float, float, float, float) -> Tensor
+    r"""Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \leq \text{mean} \leq b`.
+    NOTE: this impl is similar to the PyTorch trunc_normal_, the bounds [a, b] are
+    applied while sampling the normal with mean/std applied, therefore a, b args
+    should be adjusted to match the range of mean, std args.
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    Examples:
+        >>> w = torch.empty(3, 5)
+        >>> nn.init.trunc_normal_(w)
+    """
+    with torch.no_grad():
+        return _trunc_normal_(tensor, mean, std, a, b)
+
+
+def init_weights(m):
+    if isinstance(m, nn.Linear):
+        trunc_normal_(m.weight, std=.02)
+        if isinstance(m, nn.Linear) and m.bias is not None:
+            nn.init.constant_(m.bias, 0)
+    elif isinstance(m, nn.LayerNorm):
+        nn.init.constant_(m.bias, 0)
+        nn.init.constant_(m.weight, 1.0)
+
+
+class MultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        n_ctx: int,
+        width: int,
+        heads: int,
+        qkv_bias: bool
+    ):
+        super().__init__()
+        self.n_ctx = n_ctx
+        self.width = width
+        self.heads = heads
+        self.c_qkv = nn.Linear(width, width * 3, bias=qkv_bias, device=device, dtype=dtype)
+        self.c_proj = nn.Linear(width, width, device=device, dtype=dtype)
+        self.attention = QKVMultiheadAttention(device=device, dtype=dtype, heads=heads, n_ctx=n_ctx)
+
+    def forward(self, x):
+        x = self.c_qkv(x)
+        x = checkpoint(self.attention, (x,), (), True)
+        x = self.c_proj(x)
+        return x
+
+
+class QKVMultiheadAttention(nn.Module):
+    def __init__(self, *, device: torch.device, dtype: torch.dtype, heads: int, n_ctx: int):
+        super().__init__()
+        self.device = device
+        self.dtype = dtype
+        self.heads = heads
+        self.n_ctx = n_ctx
+
+    def forward(self, qkv):
+        bs, n_ctx, width = qkv.shape
+        attn_ch = width // self.heads // 3
+        scale = 1 / math.sqrt(attn_ch)
+        qkv = qkv.view(bs, n_ctx, self.heads, -1)
+        q, k, v = torch.split(qkv, attn_ch, dim=-1)
+        weight = torch.einsum("bthc,bshc->bhts", q, k) * scale
+        wdtype = weight.dtype
+        weight = torch.softmax(weight.float(), dim=-1).type(wdtype)
+        return torch.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        n_ctx: int,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        use_checkpoint: bool = False
+    ):
+        super().__init__()
+
+        self.use_checkpoint = use_checkpoint
+
+        self.attn = MultiheadAttention(
+            device=device,
+            dtype=dtype,
+            n_ctx=n_ctx,
+            width=width,
+            heads=heads,
+            qkv_bias=qkv_bias
+        )
+        self.ln_1 = nn.LayerNorm(width, device=device, dtype=dtype)
+        self.mlp = MLP(device=device, dtype=dtype, width=width)
+        self.ln_2 = nn.LayerNorm(width, device=device, dtype=dtype)
+
+    def _forward(self, x: torch.Tensor):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+    def forward(self, x: torch.Tensor):
+        return checkpoint(self._forward, (x,), self.parameters(), self.use_checkpoint)
+
+
+class MultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        n_data: Optional[int] = None,
+        data_width: Optional[int] = None,
+    ):
+        super().__init__()
+        self.n_data = n_data
+        self.width = width
+        self.heads = heads
+        self.data_width = width if data_width is None else data_width
+        self.c_q = nn.Linear(width, width, bias=qkv_bias, device=device, dtype=dtype)
+        self.c_kv = nn.Linear(self.data_width, width * 2, bias=qkv_bias, device=device, dtype=dtype)
+        self.c_proj = nn.Linear(width, width, device=device, dtype=dtype)
+        self.attention = QKVMultiheadCrossAttention(
+            device=device, dtype=dtype, heads=heads, n_data=n_data
+        )
+
+    def forward(self, x, data):
+        x = self.c_q(x)
+        data = self.c_kv(data)
+        x = checkpoint(self.attention, (x, data), (), True)
+        x = self.c_proj(x)
+        return x
+
+
+class QKVMultiheadCrossAttention(nn.Module):
+    def __init__(self, *, device: torch.device, dtype: torch.dtype, heads: int, n_data: Optional[int] = None):
+        super().__init__()
+        self.device = device
+        self.dtype = dtype
+        self.heads = heads
+        self.n_data = n_data
+
+    def forward(self, q, kv):
+        _, n_ctx, _ = q.shape
+        bs, n_data, width = kv.shape
+        attn_ch = width // self.heads // 2
+        scale = 1 / math.sqrt(attn_ch)
+        q = q.view(bs, n_ctx, self.heads, -1)
+        kv = kv.view(bs, n_data, self.heads, -1)
+        k, v = torch.split(kv, attn_ch, dim=-1)
+        weight = torch.einsum("bthc,bshc->bhts", q, k) * scale
+        wdtype = weight.dtype
+        weight = torch.softmax(weight.float(), dim=-1).type(wdtype)
+        return torch.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1)
+
+
+class ResidualCrossAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: Optional[torch.device],
+        dtype: Optional[torch.dtype],
+        n_data: Optional[int] = None,
+        width: int,
+        heads: int,
+        data_width: Optional[int] = None,
+        qkv_bias: bool = True
+    ):
+        super().__init__()
+
+        if data_width is None:
+            data_width = width
+
+        self.attn = MultiheadCrossAttention(
+            device=device,
+            dtype=dtype,
+            n_data=n_data,
+            width=width,
+            heads=heads,
+            data_width=data_width,
+            qkv_bias=qkv_bias
+        )
+        self.ln_1 = nn.LayerNorm(width, device=device, dtype=dtype)
+        self.ln_2 = nn.LayerNorm(data_width, device=device, dtype=dtype)
+        self.mlp = MLP(device=device, dtype=dtype, width=width)
+        self.ln_3 = nn.LayerNorm(width, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, data: torch.Tensor):
+        x = x + self.attn(self.ln_1(x), self.ln_2(data))
+        x = x + self.mlp(self.ln_3(x))
+        return x
+
+
+class MLP(nn.Module):
+    def __init__(self, *,
+                 device: Optional[torch.device],
+                 dtype: Optional[torch.dtype],
+                 width: int):
+        super().__init__()
+        self.width = width
+        self.c_fc = nn.Linear(width, width * 4, device=device, dtype=dtype)
+        self.c_proj = nn.Linear(width * 4, width, device=device, dtype=dtype)
+        self.gelu = nn.GELU()
+
+    def forward(self, x):
+        return self.c_proj(self.gelu(self.c_fc(x)))
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: Optional[torch.device],
+        dtype: Optional[torch.dtype],
+        n_ctx: int,
+        width: int,
+        layers: int,
+        heads: int,
+        qkv_bias: bool = True,
+        use_checkpoint: bool = False
+    ):
+        super().__init__()
+        self.n_ctx = n_ctx
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(
+                    device=device,
+                    dtype=dtype,
+                    n_ctx=n_ctx,
+                    width=width,
+                    heads=heads,
+                    qkv_bias=qkv_bias,
+                    use_checkpoint=use_checkpoint
+                )
+                for _ in range(layers)
+            ]
+        )
+
+        self.apply(init_weights)
+
+    def forward(self, x: torch.Tensor):
+        for block in self.resblocks:
+            x = block(x)
+        return x

MeshAnything/miche/michelangelo/models/tsal/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

MeshAnything/miche/michelangelo/models/tsal/asl_pl_module.py

@@ -0,0 +1,395 @@
+# -*- coding: utf-8 -*-
+
+from typing import List, Tuple, Dict, Optional
+from omegaconf import DictConfig
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.optim import lr_scheduler
+from typing import Union
+from functools import partial
+
+from MeshAnything.miche.michelangelo.utils import instantiate_from_config
+
+from .tsal_base import (
+    AlignedShapeAsLatentModule,
+    ShapeAsLatentModule,
+    Latent2MeshOutput,
+    AlignedMeshOutput
+)
+from MeshAnything.miche.michelangelo.models.tsal.inference_utils import extract_geometry
+import trimesh
+
+class AlignedShapeAsLatentPLModule(nn.Module):
+    def __init__(self, *,
+                 shape_module_cfg,
+                 aligned_module_cfg,
+                 loss_cfg,
+                 optimizer_cfg: Optional[DictConfig] = None,
+                 ckpt_path: Optional[str] = None,
+                 ignore_keys: Union[Tuple[str], List[str]] = ()):
+
+        super().__init__()
+
+        shape_model: ShapeAsLatentModule = instantiate_from_config(
+            shape_module_cfg, device=None, dtype=None
+        )
+        self.model: AlignedShapeAsLatentModule = instantiate_from_config(
+            aligned_module_cfg, shape_model=shape_model
+        )
+
+        self.loss = instantiate_from_config(loss_cfg)
+
+        self.optimizer_cfg = optimizer_cfg
+
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+
+    def set_shape_model_only(self):
+        self.model.set_shape_model_only()
+
+
+
+    @property
+    def latent_shape(self):
+        return self.model.shape_model.latent_shape
+
+    @property
+    def zero_rank(self):
+        if self._trainer:
+            zero_rank = self.trainer.local_rank == 0
+        else:
+            zero_rank = True
+
+        return zero_rank
+
+    def init_from_ckpt(self, path, ignore_keys=()):
+        state_dict = torch.load(path, map_location="cpu")["state_dict"]
+
+        keys = list(state_dict.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del state_dict[k]
+
+        missing, unexpected = self.load_state_dict(state_dict, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+            print(f"Unexpected Keys: {unexpected}")
+
+    def configure_optimizers(self) -> Tuple[List, List]:
+        lr = self.learning_rate
+
+        trainable_parameters = list(self.model.parameters())
+
+        if self.optimizer_cfg is None:
+            optimizers = [torch.optim.AdamW(trainable_parameters, lr=lr, betas=(0.9, 0.99), weight_decay=1e-3)]
+            schedulers = []
+        else:
+            optimizer = instantiate_from_config(self.optimizer_cfg.optimizer, params=trainable_parameters)
+            scheduler_func = instantiate_from_config(
+                self.optimizer_cfg.scheduler,
+                max_decay_steps=self.trainer.max_steps,
+                lr_max=lr
+            )
+            scheduler = {
+                "scheduler": lr_scheduler.LambdaLR(optimizer, lr_lambda=scheduler_func.schedule),
+                "interval": "step",
+                "frequency": 1
+            }
+            optimizers = [optimizer]
+            schedulers = [scheduler]
+
+        return optimizers, schedulers
+
+    def forward(self,
+                surface: torch.FloatTensor,
+                image: torch.FloatTensor,
+                text: torch.FloatTensor,
+                volume_queries: torch.FloatTensor):
+
+        """
+
+        Args:
+            surface (torch.FloatTensor):
+            image (torch.FloatTensor):
+            text (torch.FloatTensor):
+            volume_queries (torch.FloatTensor):
+
+        Returns:
+
+        """
+
+        embed_outputs, shape_z = self.model(surface, image, text)
+
+        shape_zq, posterior = self.model.shape_model.encode_kl_embed(shape_z)
+        latents = self.model.shape_model.decode(shape_zq)
+        logits = self.model.shape_model.query_geometry(volume_queries, latents)
+
+        return embed_outputs, logits, posterior
+
+    def encode(self, surface: torch.FloatTensor, sample_posterior=True):
+
+        pc = surface[..., 0:3]
+        feats = surface[..., 3:6]
+
+        shape_embed, shape_zq, posterior = self.model.shape_model.encode(
+            pc=pc, feats=feats, sample_posterior=sample_posterior
+        )
+
+        return shape_zq
+
+    def encode_latents(self, surface: torch.FloatTensor):
+
+        pc = surface[..., 0:3]
+        feats = surface[..., 3:6]
+
+        shape_embed, shape_latents = self.model.shape_model.encode_latents(
+            pc=pc, feats=feats
+        )
+        shape_embed = shape_embed.unsqueeze(1)
+        assert shape_embed.shape[1] == 1 and shape_latents.shape[1] == 256
+        cat_latents = torch.cat([shape_embed, shape_latents], dim=1)
+
+        return cat_latents
+
+    def recon(self, surface):
+        cat_latents = self.encode_latents(surface)
+        shape_latents = cat_latents[:, 1:]
+        shape_zq, posterior = self.model.shape_model.encode_kl_embed(shape_latents)
+
+        # decoding
+        latents = self.model.shape_model.decode(shape_zq)
+        geometric_func = partial(self.model.shape_model.query_geometry, latents=latents)
+
+        # reconstruction
+        mesh_v_f, has_surface = extract_geometry(
+            geometric_func=geometric_func,
+            device=surface.device,
+            batch_size=surface.shape[0],
+            bounds=(-1.25, -1.25, -1.25, 1.25, 1.25, 1.25),
+            octree_depth=7,
+            num_chunks=10000,
+        )
+        recon_mesh = trimesh.Trimesh(mesh_v_f[0][0], mesh_v_f[0][1])
+
+        return recon_mesh
+
+
+    def to_shape_latents(self, latents):
+
+        shape_zq, posterior = self.model.shape_model.encode_kl_embed(latents, sample_posterior = False)
+        return self.model.shape_model.decode(shape_zq)
+
+    def decode(self,
+               z_q,
+               bounds: Union[Tuple[float], List[float], float] = 1.1,
+               octree_depth: int = 7,
+               num_chunks: int = 10000) -> List[Latent2MeshOutput]:
+
+        latents = self.model.shape_model.decode(z_q)  # latents: [bs, num_latents, dim]
+        outputs = self.latent2mesh(latents, bounds=bounds, octree_depth=octree_depth, num_chunks=num_chunks)
+
+        return outputs
+
+    def training_step(self, batch: Dict[str, torch.FloatTensor],
+                      batch_idx: int, optimizer_idx: int = 0) -> torch.FloatTensor:
+        """
+
+        Args:
+            batch (dict): the batch sample, and it contains:
+                - surface (torch.FloatTensor): [bs, n_surface, (3 + input_dim)]
+                - image (torch.FloatTensor): [bs, 3, 224, 224]
+                - text (torch.FloatTensor): [bs, num_templates, 77]
+                - geo_points (torch.FloatTensor): [bs, n_pts, (3 + 1)]
+
+            batch_idx (int):
+
+            optimizer_idx (int):
+
+        Returns:
+            loss (torch.FloatTensor):
+
+        """
+
+        surface = batch["surface"]
+        image = batch["image"]
+        text = batch["text"]
+
+        volume_queries = batch["geo_points"][..., 0:3]
+        shape_labels = batch["geo_points"][..., -1]
+
+        embed_outputs, shape_logits, posteriors = self(surface, image, text, volume_queries)
+
+        aeloss, log_dict_ae = self.loss(
+            **embed_outputs,
+            posteriors=posteriors,
+            shape_logits=shape_logits,
+            shape_labels=shape_labels,
+            split="train"
+        )
+
+        self.log_dict(log_dict_ae, prog_bar=True, logger=True, batch_size=shape_logits.shape[0],
+                      sync_dist=False, rank_zero_only=True)
+
+        return aeloss
+
+    def validation_step(self, batch: Dict[str, torch.FloatTensor], batch_idx: int) -> torch.FloatTensor:
+
+        surface = batch["surface"]
+        image = batch["image"]
+        text = batch["text"]
+
+        volume_queries = batch["geo_points"][..., 0:3]
+        shape_labels = batch["geo_points"][..., -1]
+
+        embed_outputs, shape_logits, posteriors = self(surface, image, text, volume_queries)
+
+        aeloss, log_dict_ae = self.loss(
+            **embed_outputs,
+            posteriors=posteriors,
+            shape_logits=shape_logits,
+            shape_labels=shape_labels,
+            split="val"
+        )
+        self.log_dict(log_dict_ae, prog_bar=True, logger=True, batch_size=shape_logits.shape[0],
+                      sync_dist=False, rank_zero_only=True)
+
+        return aeloss
+
+    def visual_alignment(self,
+                         surface: torch.FloatTensor,
+                         image: torch.FloatTensor,
+                         text: torch.FloatTensor,
+                         description: Optional[List[str]] = None,
+                         bounds: Union[Tuple[float], List[float]] = (-1.25, -1.25, -1.25, 1.25, 1.25, 1.25),
+                         octree_depth: int = 7,
+                         num_chunks: int = 10000) -> List[AlignedMeshOutput]:
+
+        """
+
+        Args:
+            surface:
+            image:
+            text:
+            description:
+            bounds:
+            octree_depth:
+            num_chunks:
+
+        Returns:
+            mesh_outputs (List[AlignedMeshOutput]): the mesh outputs list.
+
+        """
+
+        outputs = []
+
+        device = surface.device
+        bs = surface.shape[0]
+
+        embed_outputs, shape_z = self.model(surface, image, text)
+
+        # calculate the similarity
+        image_embed = embed_outputs["image_embed"]
+        text_embed = embed_outputs["text_embed"]
+        shape_embed = embed_outputs["shape_embed"]
+
+        # normalized features
+        shape_embed = F.normalize(shape_embed, dim=-1, p=2)
+        text_embed = F.normalize(text_embed, dim=-1, p=2)
+        image_embed = F.normalize(image_embed, dim=-1, p=2)
+
+        # B x B
+        shape_text_similarity = (100.0 * shape_embed @ text_embed.T).softmax(dim=-1)
+
+        # B x B
+        shape_image_similarity = (100.0 * shape_embed @ image_embed.T).softmax(dim=-1)
+
+        # shape reconstruction
+        shape_zq, posterior = self.model.shape_model.encode_kl_embed(shape_z)
+        latents = self.model.shape_model.decode(shape_zq)
+        geometric_func = partial(self.model.shape_model.query_geometry, latents=latents)
+
+        # 2. decode geometry
+        mesh_v_f, has_surface = extract_geometry(
+            geometric_func=geometric_func,
+            device=device,
+            batch_size=bs,
+            bounds=bounds,
+            octree_depth=octree_depth,
+            num_chunks=num_chunks,
+            disable=not self.zero_rank
+        )
+
+        # 3. decode texture
+        for i, ((mesh_v, mesh_f), is_surface) in enumerate(zip(mesh_v_f, has_surface)):
+            if not is_surface:
+                outputs.append(None)
+                continue
+
+            out = AlignedMeshOutput()
+            out.mesh_v = mesh_v
+            out.mesh_f = mesh_f
+            out.surface = surface[i].cpu().numpy()
+            out.image = image[i].cpu().numpy()
+            if description is not None:
+                out.text = description[i]
+            out.shape_text_similarity = shape_text_similarity[i, i]
+            out.shape_image_similarity = shape_image_similarity[i, i]
+
+            outputs.append(out)
+
+        return outputs
+
+    def latent2mesh(self,
+                    latents: torch.FloatTensor,
+                    bounds: Union[Tuple[float], List[float], float] = 1.1,
+                    octree_depth: int = 7,
+                    num_chunks: int = 10000) -> List[Latent2MeshOutput]:
+
+        """
+
+        Args:
+            latents: [bs, num_latents, dim]
+            bounds:
+            octree_depth:
+            num_chunks:
+
+        Returns:
+            mesh_outputs (List[MeshOutput]): the mesh outputs list.
+
+        """
+
+        outputs = []
+
+        geometric_func = partial(self.model.shape_model.query_geometry, latents=latents)
+
+        # 2. decode geometry
+        device = latents.device
+        mesh_v_f, has_surface = extract_geometry(
+            geometric_func=geometric_func,
+            device=device,
+            batch_size=len(latents),
+            bounds=bounds,
+            octree_depth=octree_depth,
+            num_chunks=num_chunks,
+            disable=not self.zero_rank
+        )
+
+        # 3. decode texture
+        for i, ((mesh_v, mesh_f), is_surface) in enumerate(zip(mesh_v_f, has_surface)):
+            if not is_surface:
+                outputs.append(None)
+                continue
+
+            out = Latent2MeshOutput()
+            out.mesh_v = mesh_v
+            out.mesh_f = mesh_f
+
+            outputs.append(out)
+
+        return outputs
+

MeshAnything/miche/michelangelo/models/tsal/clip_asl_module.py

@@ -0,0 +1,118 @@
+# -*- coding: utf-8 -*-
+
+import torch
+from torch import nn
+from einops import rearrange
+from transformers import CLIPModel
+
+from MeshAnything.miche.michelangelo.models.tsal.tsal_base import AlignedShapeAsLatentModule
+
+
+class CLIPAlignedShapeAsLatentModule(AlignedShapeAsLatentModule):
+
+    def __init__(self, *,
+                 shape_model,
+                 clip_model_version: str = "openai/clip-vit-large-patch14"):
+
+        super().__init__()
+
+        # self.clip_model: CLIPModel = CLIPModel.from_pretrained(clip_model_version)
+        # for params in self.clip_model.parameters():
+        #     params.requires_grad = False
+        self.clip_model = None
+        self.shape_model = shape_model
+        self.shape_projection = nn.Parameter(torch.empty(self.shape_model.width, self.shape_model.width))
+        # nn.init.normal_(self.shape_projection, std=self.shape_model.width ** -0.5)
+
+    def set_shape_model_only(self):
+        self.clip_model = None
+
+    def encode_shape_embed(self, surface, return_latents: bool = False):
+        """
+
+        Args:
+            surface (torch.FloatTensor): [bs, n, 3 + c]
+            return_latents (bool):
+
+        Returns:
+            x (torch.FloatTensor): [bs, projection_dim]
+            shape_latents (torch.FloatTensor): [bs, m, d]
+        """
+
+        pc = surface[..., 0:3]
+        feats = surface[..., 3:]
+
+        shape_embed, shape_latents = self.shape_model.encode_latents(pc, feats)
+        x = shape_embed @ self.shape_projection
+
+        if return_latents:
+            return x, shape_latents
+        else:
+            return x
+
+    def encode_image_embed(self, image):
+        """
+
+        Args:
+            image (torch.FloatTensor): [bs, 3, h, w]
+
+        Returns:
+            x (torch.FloatTensor): [bs, projection_dim]
+        """
+
+        x = self.clip_model.get_image_features(image)
+
+        return x
+
+    def encode_text_embed(self, text):
+        x = self.clip_model.get_text_features(text)
+        return x
+
+    def forward(self, surface, image, text):
+        """
+
+        Args:
+            surface (torch.FloatTensor):
+            image (torch.FloatTensor): [bs, 3, 224, 224]
+            text (torch.LongTensor): [bs, num_templates, 77]
+
+        Returns:
+            embed_outputs (dict): the embedding outputs, and it contains:
+                - image_embed (torch.FloatTensor):
+                - text_embed (torch.FloatTensor):
+                - shape_embed (torch.FloatTensor):
+                - logit_scale (float):
+        """
+
+        # # text embedding
+        # text_embed_all = []
+        # for i in range(text.shape[0]):
+        #     text_for_one_sample = text[i]
+        #     text_embed = self.encode_text_embed(text_for_one_sample)
+        #     text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)
+        #     text_embed = text_embed.mean(dim=0)
+        #     text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)
+        #     text_embed_all.append(text_embed)
+        # text_embed_all = torch.stack(text_embed_all)
+
+        b = text.shape[0]
+        text_tokens = rearrange(text, "b t l -> (b t) l")
+        text_embed = self.encode_text_embed(text_tokens)
+        text_embed = rearrange(text_embed, "(b t) d -> b t d", b=b)
+        text_embed = text_embed.mean(dim=1)
+        text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)
+
+        # image embedding
+        image_embed = self.encode_image_embed(image)
+
+        # shape embedding
+        shape_embed, shape_latents = self.encode_shape_embed(surface, return_latents=True)
+
+        embed_outputs = {
+            "image_embed": image_embed,
+            "text_embed": text_embed,
+            "shape_embed": shape_embed,
+            # "logit_scale": self.clip_model.logit_scale.exp()
+        }
+
+        return embed_outputs, shape_latents

MeshAnything/miche/michelangelo/models/tsal/inference_utils.py

@@ -0,0 +1,80 @@
+# -*- coding: utf-8 -*-
+
+import torch
+from tqdm import tqdm
+from einops import repeat
+import numpy as np
+from typing import Callable, Tuple, List, Union, Optional
+from skimage import measure
+
+from MeshAnything.miche.michelangelo.graphics.primitives import generate_dense_grid_points
+
+
+@torch.no_grad()
+def extract_geometry(geometric_func: Callable,
+                     device: torch.device,
+                     batch_size: int = 1,
+                     bounds: Union[Tuple[float], List[float], float] = (-1.25, -1.25, -1.25, 1.25, 1.25, 1.25),
+                     octree_depth: int = 7,
+                     num_chunks: int = 10000,
+                     disable: bool = True):
+    """
+
+    Args:
+        geometric_func:
+        device:
+        bounds:
+        octree_depth:
+        batch_size:
+        num_chunks:
+        disable:
+
+    Returns:
+
+    """
+
+    if isinstance(bounds, float):
+        bounds = [-bounds, -bounds, -bounds, bounds, bounds, bounds]
+
+    bbox_min = np.array(bounds[0:3])
+    bbox_max = np.array(bounds[3:6])
+    bbox_size = bbox_max - bbox_min
+
+    xyz_samples, grid_size, length = generate_dense_grid_points(
+        bbox_min=bbox_min,
+        bbox_max=bbox_max,
+        octree_depth=octree_depth,
+        indexing="ij"
+    )
+    xyz_samples = torch.FloatTensor(xyz_samples)
+
+    batch_logits = []
+    for start in tqdm(range(0, xyz_samples.shape[0], num_chunks),
+                      desc="Implicit Function:", disable=disable, leave=False):
+        queries = xyz_samples[start: start + num_chunks, :].to(device)
+        batch_queries = repeat(queries, "p c -> b p c", b=batch_size)
+
+        logits = geometric_func(batch_queries)
+        batch_logits.append(logits.cpu())
+
+    grid_logits = torch.cat(batch_logits, dim=1).view((batch_size, grid_size[0], grid_size[1], grid_size[2])).numpy()
+
+    mesh_v_f = []
+    has_surface = np.zeros((batch_size,), dtype=np.bool_)
+    for i in range(batch_size):
+        try:
+            vertices, faces, normals, _ = measure.marching_cubes(grid_logits[i], 0, method="lewiner")
+            vertices = vertices / grid_size * bbox_size + bbox_min
+            # vertices[:, [0, 1]] = vertices[:, [1, 0]]
+            mesh_v_f.append((vertices.astype(np.float32), np.ascontiguousarray(faces)))
+            has_surface[i] = True
+
+        except ValueError:
+            mesh_v_f.append((None, None))
+            has_surface[i] = False
+
+        except RuntimeError:
+            mesh_v_f.append((None, None))
+            has_surface[i] = False
+
+    return mesh_v_f, has_surface

MeshAnything/miche/michelangelo/models/tsal/loss.py

@@ -0,0 +1,303 @@
+# -*- coding: utf-8 -*-
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from typing import Optional, Tuple, Dict
+
+from MeshAnything.miche.michelangelo.models.modules.distributions import DiagonalGaussianDistribution
+from MeshAnything.miche.michelangelo.utils.eval import compute_psnr
+from MeshAnything.miche.michelangelo.utils import misc
+
+
+class KLNearFar(nn.Module):
+    def __init__(self,
+                 near_weight: float = 0.1,
+                 kl_weight: float = 1.0,
+                 num_near_samples: Optional[int] = None):
+
+        super().__init__()
+
+        self.near_weight = near_weight
+        self.kl_weight = kl_weight
+        self.num_near_samples = num_near_samples
+        self.geo_criterion = nn.BCEWithLogitsLoss()
+
+    def forward(self,
+                posteriors: Optional[DiagonalGaussianDistribution],
+                logits: torch.FloatTensor,
+                labels: torch.FloatTensor,
+                split: Optional[str] = "train", **kwargs) -> Tuple[torch.FloatTensor, Dict[str, float]]:
+
+        """
+
+        Args:
+            posteriors (DiagonalGaussianDistribution or torch.distributions.Normal):
+            logits (torch.FloatTensor): [B, 2*N], logits[:, 0:N] is the volume points; logits[:, N:2N] is the near points;
+            labels (torch.FloatTensor): [B, 2*N], labels[:, 0:N] is the volume points; labels[:, N:2N] is the near points;
+            split (str):
+            **kwargs:
+
+        Returns:
+            loss (torch.Tensor): (,)
+            log (dict):
+
+        """
+
+        if self.num_near_samples is None:
+            num_vol = logits.shape[1] // 2
+        else:
+            num_vol = logits.shape[1] - self.num_near_samples
+
+        vol_logits = logits[:, 0:num_vol]
+        vol_labels = labels[:, 0:num_vol]
+
+        near_logits = logits[:, num_vol:]
+        near_labels = labels[:, num_vol:]
+
+        # occupancy loss
+        # vol_bce = self.geo_criterion(vol_logits, vol_labels)
+        # near_bce = self.geo_criterion(near_logits, near_labels)
+        vol_bce = self.geo_criterion(vol_logits.float(), vol_labels.float())
+        near_bce = self.geo_criterion(near_logits.float(), near_labels.float())
+
+        if posteriors is None:
+            kl_loss = torch.tensor(0.0, dtype=vol_logits.dtype, device=vol_logits.device)
+        else:
+            kl_loss = posteriors.kl(dims=(1, 2))
+            kl_loss = torch.mean(kl_loss)
+
+        loss = vol_bce + near_bce * self.near_weight + kl_loss * self.kl_weight
+
+        with torch.no_grad():
+            preds = logits >= 0
+            accuracy = (preds == labels).float()
+            accuracy = accuracy.mean()
+            pos_ratio = torch.mean(labels)
+
+        log = {
+            "{}/total_loss".format(split): loss.clone().detach(),
+            "{}/near".format(split): near_bce.detach(),
+            "{}/far".format(split): vol_bce.detach(),
+            "{}/kl".format(split): kl_loss.detach(),
+            "{}/accuracy".format(split): accuracy,
+            "{}/pos_ratio".format(split): pos_ratio
+        }
+
+        if posteriors is not None:
+            log[f"{split}/mean"] = posteriors.mean.mean().detach()
+            log[f"{split}/std_mean"] = posteriors.std.mean().detach()
+            log[f"{split}/std_max"] = posteriors.std.max().detach()
+
+        return loss, log
+
+
+class KLNearFarColor(nn.Module):
+    def __init__(self,
+                 near_weight: float = 0.1,
+                 kl_weight: float = 1.0,
+                 color_weight: float = 1.0,
+                 color_criterion: str = "mse",
+                 num_near_samples: Optional[int] = None):
+
+        super().__init__()
+
+        self.color_weight = color_weight
+        self.near_weight = near_weight
+        self.kl_weight = kl_weight
+        self.num_near_samples = num_near_samples
+
+        if color_criterion == "mse":
+            self.color_criterion = nn.MSELoss()
+
+        elif color_criterion == "l1":
+            self.color_criterion = nn.L1Loss()
+
+        else:
+            raise ValueError(f"{color_criterion} must be [`mse`, `l1`].")
+
+        self.geo_criterion = nn.BCEWithLogitsLoss()
+
+    def forward(self,
+                posteriors: Optional[DiagonalGaussianDistribution],
+                logits: torch.FloatTensor,
+                labels: torch.FloatTensor,
+                pred_colors: torch.FloatTensor,
+                gt_colors: torch.FloatTensor,
+                split: Optional[str] = "train", **kwargs) -> Tuple[torch.FloatTensor, Dict[str, float]]:
+
+        """
+
+        Args:
+            posteriors (DiagonalGaussianDistribution or torch.distributions.Normal):
+            logits (torch.FloatTensor): [B, 2*N], logits[:, 0:N] is the volume points; logits[:, N:2N] is the near points;
+            labels (torch.FloatTensor): [B, 2*N], labels[:, 0:N] is the volume points; labels[:, N:2N] is the near points;
+            pred_colors (torch.FloatTensor): [B, M, 3]
+            gt_colors (torch.FloatTensor): [B, M, 3]
+            split (str):
+            **kwargs:
+
+        Returns:
+            loss (torch.Tensor): (,)
+            log (dict):
+
+        """
+
+        if self.num_near_samples is None:
+            num_vol = logits.shape[1] // 2
+        else:
+            num_vol = logits.shape[1] - self.num_near_samples
+
+        vol_logits = logits[:, 0:num_vol]
+        vol_labels = labels[:, 0:num_vol]
+
+        near_logits = logits[:, num_vol:]
+        near_labels = labels[:, num_vol:]
+
+        # occupancy loss
+        # vol_bce = self.geo_criterion(vol_logits, vol_labels)
+        # near_bce = self.geo_criterion(near_logits, near_labels)
+        vol_bce = self.geo_criterion(vol_logits.float(), vol_labels.float())
+        near_bce = self.geo_criterion(near_logits.float(), near_labels.float())
+
+        # surface color loss
+        color = self.color_criterion(pred_colors, gt_colors)
+
+        if posteriors is None:
+            kl_loss = torch.tensor(0.0, dtype=pred_colors.dtype, device=pred_colors.device)
+        else:
+            kl_loss = posteriors.kl(dims=(1, 2))
+            kl_loss = torch.mean(kl_loss)
+
+        loss = vol_bce + near_bce * self.near_weight + color * self.color_weight + kl_loss * self.kl_weight
+
+        with torch.no_grad():
+            preds = logits >= 0
+            accuracy = (preds == labels).float()
+            accuracy = accuracy.mean()
+            psnr = compute_psnr(pred_colors, gt_colors)
+
+        log = {
+            "{}/total_loss".format(split): loss.clone().detach(),
+            "{}/near".format(split): near_bce.detach(),
+            "{}/far".format(split): vol_bce.detach(),
+            "{}/color".format(split): color.detach(),
+            "{}/kl".format(split): kl_loss.detach(),
+            "{}/psnr".format(split): psnr.detach(),
+            "{}/accuracy".format(split): accuracy
+        }
+
+        return loss, log
+
+
+class ContrastKLNearFar(nn.Module):
+    def __init__(self,
+                 contrast_weight: float = 1.0,
+                 near_weight: float = 0.1,
+                 kl_weight: float = 1.0,
+                 num_near_samples: Optional[int] = None):
+
+        super().__init__()
+
+        self.labels = None
+        self.last_local_batch_size = None
+
+        self.contrast_weight = contrast_weight
+        self.near_weight = near_weight
+        self.kl_weight = kl_weight
+        self.num_near_samples = num_near_samples
+        self.geo_criterion = nn.BCEWithLogitsLoss()
+
+    def forward(self,
+                shape_embed: torch.FloatTensor,
+                text_embed: torch.FloatTensor,
+                image_embed: torch.FloatTensor,
+                logit_scale: torch.FloatTensor,
+                posteriors: Optional[DiagonalGaussianDistribution],
+                shape_logits: torch.FloatTensor,
+                shape_labels: torch.FloatTensor,
+                split: Optional[str] = "train", **kwargs):
+
+        local_batch_size = shape_embed.size(0)
+
+        if local_batch_size != self.last_local_batch_size:
+            self.labels = local_batch_size * misc.get_rank() + torch.arange(
+                local_batch_size, device=shape_embed.device
+            ).long()
+            self.last_local_batch_size = local_batch_size
+
+        # normalized features
+        shape_embed = F.normalize(shape_embed, dim=-1, p=2)
+        text_embed = F.normalize(text_embed, dim=-1, p=2)
+        image_embed = F.normalize(image_embed, dim=-1, p=2)
+
+        # gather features from all GPUs
+        shape_embed_all, text_embed_all, image_embed_all = misc.all_gather_batch(
+            [shape_embed, text_embed, image_embed]
+        )
+
+        # cosine similarity as logits
+        logits_per_shape_text = logit_scale * shape_embed @ text_embed_all.t()
+        logits_per_text_shape = logit_scale * text_embed @ shape_embed_all.t()
+        logits_per_shape_image = logit_scale * shape_embed @ image_embed_all.t()
+        logits_per_image_shape = logit_scale * image_embed @ shape_embed_all.t()
+        contrast_loss = (F.cross_entropy(logits_per_shape_text, self.labels) +
+                         F.cross_entropy(logits_per_text_shape, self.labels)) / 2 + \
+                        (F.cross_entropy(logits_per_shape_image, self.labels) +
+                         F.cross_entropy(logits_per_image_shape, self.labels)) / 2
+
+        # shape reconstruction
+        if self.num_near_samples is None:
+            num_vol = shape_logits.shape[1] // 2
+        else:
+            num_vol = shape_logits.shape[1] - self.num_near_samples
+
+        vol_logits = shape_logits[:, 0:num_vol]
+        vol_labels = shape_labels[:, 0:num_vol]
+
+        near_logits = shape_logits[:, num_vol:]
+        near_labels = shape_labels[:, num_vol:]
+
+        # occupancy loss
+        vol_bce = self.geo_criterion(vol_logits.float(), vol_labels.float())
+        near_bce = self.geo_criterion(near_logits.float(), near_labels.float())
+
+        if posteriors is None:
+            kl_loss = torch.tensor(0.0, dtype=vol_logits.dtype, device=vol_logits.device)
+        else:
+            kl_loss = posteriors.kl(dims=(1, 2))
+            kl_loss = torch.mean(kl_loss)
+
+        loss = vol_bce + near_bce * self.near_weight + kl_loss * self.kl_weight + contrast_loss * self.contrast_weight
+
+        # compute accuracy
+        with torch.no_grad():
+            pred = torch.argmax(logits_per_shape_text, dim=-1)
+            correct = pred.eq(self.labels).sum()
+            shape_text_acc = 100 * correct / local_batch_size
+
+            pred = torch.argmax(logits_per_shape_image, dim=-1)
+            correct = pred.eq(self.labels).sum()
+            shape_image_acc = 100 * correct / local_batch_size
+
+            preds = shape_logits >= 0
+            accuracy = (preds == shape_labels).float()
+            accuracy = accuracy.mean()
+
+            log = {
+                "{}/contrast".format(split): contrast_loss.clone().detach(),
+                "{}/near".format(split): near_bce.detach(),
+                "{}/far".format(split): vol_bce.detach(),
+                "{}/kl".format(split): kl_loss.detach(),
+                "{}/shape_text_acc".format(split): shape_text_acc,
+                "{}/shape_image_acc".format(split): shape_image_acc,
+                "{}/total_loss".format(split): loss.clone().detach(),
+                "{}/accuracy".format(split): accuracy,
+            }
+
+            if posteriors is not None:
+                log[f"{split}/mean"] = posteriors.mean.mean().detach()
+                log[f"{split}/std_mean"] = posteriors.std.mean().detach()
+                log[f"{split}/std_max"] = posteriors.std.max().detach()
+
+        return loss, log

MeshAnything/miche/michelangelo/models/tsal/sal_perceiver.py

@@ -0,0 +1,423 @@
+# -*- coding: utf-8 -*-
+
+import torch
+import torch.nn as nn
+from typing import Optional
+from einops import repeat
+import math
+
+from MeshAnything.miche.michelangelo.models.modules import checkpoint
+from MeshAnything.miche.michelangelo.models.modules.embedder import FourierEmbedder
+from MeshAnything.miche.michelangelo.models.modules.distributions import DiagonalGaussianDistribution
+from MeshAnything.miche.michelangelo.models.modules.transformer_blocks import (
+    ResidualCrossAttentionBlock,
+    Transformer
+)
+
+from .tsal_base import ShapeAsLatentModule
+
+
+class CrossAttentionEncoder(nn.Module):
+
+    def __init__(self, *,
+                 device: Optional[torch.device],
+                 dtype: Optional[torch.dtype],
+                 num_latents: int,
+                 fourier_embedder: FourierEmbedder,
+                 point_feats: int,
+                 width: int,
+                 heads: int,
+                 layers: int,
+                 init_scale: float = 0.25,
+                 qkv_bias: bool = True,
+                 flash: bool = False,
+                 use_ln_post: bool = False,
+                 use_checkpoint: bool = False):
+
+        super().__init__()
+
+        self.use_checkpoint = use_checkpoint
+        self.num_latents = num_latents
+
+        self.query = nn.Parameter(torch.randn((num_latents, width), device=device, dtype=dtype) * 0.02)
+
+        self.fourier_embedder = fourier_embedder
+        self.input_proj = nn.Linear(self.fourier_embedder.out_dim + point_feats, width, device=device, dtype=dtype)
+        self.cross_attn = ResidualCrossAttentionBlock(
+            device=device,
+            dtype=dtype,
+            width=width,
+            heads=heads,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            flash=flash,
+        )
+
+        self.self_attn = Transformer(
+            device=device,
+            dtype=dtype,
+            n_ctx=num_latents,
+            width=width,
+            layers=layers,
+            heads=heads,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            flash=flash,
+            use_checkpoint=False
+        )
+
+        if use_ln_post:
+            self.ln_post = nn.LayerNorm(width, dtype=dtype, device=device)
+        else:
+            self.ln_post = None
+
+    def _forward(self, pc, feats):
+        """
+
+        Args:
+            pc (torch.FloatTensor): [B, N, 3]
+            feats (torch.FloatTensor or None): [B, N, C]
+
+        Returns:
+
+        """
+
+        bs = pc.shape[0]
+
+        data = self.fourier_embedder(pc)
+        if feats is not None:
+            data = torch.cat([data, feats], dim=-1)
+        data = self.input_proj(data)
+
+        query = repeat(self.query, "m c -> b m c", b=bs)
+        latents = self.cross_attn(query, data)
+        latents = self.self_attn(latents)
+
+        if self.ln_post is not None:
+            latents = self.ln_post(latents)
+
+        return latents, pc
+
+    def forward(self, pc: torch.FloatTensor, feats: Optional[torch.FloatTensor] = None):
+        """
+
+        Args:
+            pc (torch.FloatTensor): [B, N, 3]
+            feats (torch.FloatTensor or None): [B, N, C]
+
+        Returns:
+            dict
+        """
+
+        return checkpoint(self._forward, (pc, feats), self.parameters(), self.use_checkpoint)
+
+
+class CrossAttentionDecoder(nn.Module):
+
+    def __init__(self, *,
+                 device: Optional[torch.device],
+                 dtype: Optional[torch.dtype],
+                 num_latents: int,
+                 out_channels: int,
+                 fourier_embedder: FourierEmbedder,
+                 width: int,
+                 heads: int,
+                 init_scale: float = 0.25,
+                 qkv_bias: bool = True,
+                 flash: bool = False,
+                 use_checkpoint: bool = False):
+
+        super().__init__()
+
+        self.use_checkpoint = use_checkpoint
+        self.fourier_embedder = fourier_embedder
+
+        self.query_proj = nn.Linear(self.fourier_embedder.out_dim, width, device=device, dtype=dtype)
+
+        self.cross_attn_decoder = ResidualCrossAttentionBlock(
+            device=device,
+            dtype=dtype,
+            n_data=num_latents,
+            width=width,
+            heads=heads,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            flash=flash
+        )
+
+        self.ln_post = nn.LayerNorm(width, device=device, dtype=dtype)
+        self.output_proj = nn.Linear(width, out_channels, device=device, dtype=dtype)
+
+    def _forward(self, queries: torch.FloatTensor, latents: torch.FloatTensor):
+        queries = self.query_proj(self.fourier_embedder(queries))
+        x = self.cross_attn_decoder(queries, latents)
+        x = self.ln_post(x)
+        x = self.output_proj(x)
+        return x
+
+    def forward(self, queries: torch.FloatTensor, latents: torch.FloatTensor):
+        return checkpoint(self._forward, (queries, latents), self.parameters(), self.use_checkpoint)
+
+
+class ShapeAsLatentPerceiver(ShapeAsLatentModule):
+    def __init__(self, *,
+                 device: Optional[torch.device],
+                 dtype: Optional[torch.dtype],
+                 num_latents: int,
+                 point_feats: int = 0,
+                 embed_dim: int = 0,
+                 num_freqs: int = 8,
+                 include_pi: bool = True,
+                 width: int,
+                 heads: int,
+                 num_encoder_layers: int,
+                 num_decoder_layers: int,
+                 init_scale: float = 0.25,
+                 qkv_bias: bool = True,
+                 flash: bool = False,
+                 use_ln_post: bool = False,
+                 use_checkpoint: bool = False):
+
+        super().__init__()
+
+        self.use_checkpoint = use_checkpoint
+
+        self.num_latents = num_latents
+        self.fourier_embedder = FourierEmbedder(num_freqs=num_freqs, include_pi=include_pi)
+
+        init_scale = init_scale * math.sqrt(1.0 / width)
+        self.encoder = CrossAttentionEncoder(
+            device=device,
+            dtype=dtype,
+            fourier_embedder=self.fourier_embedder,
+            num_latents=num_latents,
+            point_feats=point_feats,
+            width=width,
+            heads=heads,
+            layers=num_encoder_layers,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            flash=flash,
+            use_ln_post=use_ln_post,
+            use_checkpoint=use_checkpoint
+        )
+
+        self.embed_dim = embed_dim
+        if embed_dim > 0:
+            # VAE embed
+            self.pre_kl = nn.Linear(width, embed_dim * 2, device=device, dtype=dtype)
+            self.post_kl = nn.Linear(embed_dim, width, device=device, dtype=dtype)
+            self.latent_shape = (num_latents, embed_dim)
+        else:
+            self.latent_shape = (num_latents, width)
+
+        self.transformer = Transformer(
+            device=device,
+            dtype=dtype,
+            n_ctx=num_latents,
+            width=width,
+            layers=num_decoder_layers,
+            heads=heads,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            flash=flash,
+            use_checkpoint=use_checkpoint
+        )
+
+        # geometry decoder
+        self.geo_decoder = CrossAttentionDecoder(
+            device=device,
+            dtype=dtype,
+            fourier_embedder=self.fourier_embedder,
+            out_channels=1,
+            num_latents=num_latents,
+            width=width,
+            heads=heads,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            flash=flash,
+            use_checkpoint=use_checkpoint
+        )
+
+    def encode(self,
+               pc: torch.FloatTensor,
+               feats: Optional[torch.FloatTensor] = None,
+               sample_posterior: bool = True):
+        """
+
+        Args:
+            pc (torch.FloatTensor): [B, N, 3]
+            feats (torch.FloatTensor or None): [B, N, C]
+            sample_posterior (bool):
+
+        Returns:
+            latents (torch.FloatTensor)
+            center_pos (torch.FloatTensor or None):
+            posterior (DiagonalGaussianDistribution or None):
+        """
+
+        latents, center_pos = self.encoder(pc, feats)
+
+        posterior = None
+        if self.embed_dim > 0:
+            moments = self.pre_kl(latents)
+            posterior = DiagonalGaussianDistribution(moments, feat_dim=-1)
+
+            if sample_posterior:
+                latents = posterior.sample()
+            else:
+                latents = posterior.mode()
+
+        return latents, center_pos, posterior
+
+    def decode(self, latents: torch.FloatTensor):
+        latents = self.post_kl(latents)
+        return self.transformer(latents)
+
+    def query_geometry(self, queries: torch.FloatTensor, latents: torch.FloatTensor):
+        logits = self.geo_decoder(queries, latents).squeeze(-1)
+        return logits
+
+    def forward(self,
+                pc: torch.FloatTensor,
+                feats: torch.FloatTensor,
+                volume_queries: torch.FloatTensor,
+                sample_posterior: bool = True):
+        """
+
+        Args:
+            pc (torch.FloatTensor): [B, N, 3]
+            feats (torch.FloatTensor or None): [B, N, C]
+            volume_queries (torch.FloatTensor): [B, P, 3]
+            sample_posterior (bool):
+
+        Returns:
+            logits (torch.FloatTensor): [B, P]
+            center_pos (torch.FloatTensor): [B, M, 3]
+            posterior (DiagonalGaussianDistribution or None).
+
+        """
+
+        latents, center_pos, posterior = self.encode(pc, feats, sample_posterior=sample_posterior)
+
+        latents = self.decode(latents)
+        logits = self.query_geometry(volume_queries, latents)
+
+        return logits, center_pos, posterior
+
+
+class AlignedShapeLatentPerceiver(ShapeAsLatentPerceiver):
+
+    def __init__(self, *,
+                 device: Optional[torch.device],
+                 dtype: Optional[torch.dtype],
+                 num_latents: int,
+                 point_feats: int = 0,
+                 embed_dim: int = 0,
+                 num_freqs: int = 8,
+                 include_pi: bool = True,
+                 width: int,
+                 heads: int,
+                 num_encoder_layers: int,
+                 num_decoder_layers: int,
+                 init_scale: float = 0.25,
+                 qkv_bias: bool = True,
+                 flash: bool = False,
+                 use_ln_post: bool = False,
+                 use_checkpoint: bool = False):
+
+        super().__init__(
+            device=device,
+            dtype=dtype,
+            num_latents=1 + num_latents,
+            point_feats=point_feats,
+            embed_dim=embed_dim,
+            num_freqs=num_freqs,
+            include_pi=include_pi,
+            width=width,
+            heads=heads,
+            num_encoder_layers=num_encoder_layers,
+            num_decoder_layers=num_decoder_layers,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            flash=flash,
+            use_ln_post=use_ln_post,
+            use_checkpoint=use_checkpoint
+        )
+
+        self.width = width
+
+    def encode(self,
+               pc: torch.FloatTensor,
+               feats: Optional[torch.FloatTensor] = None,
+               sample_posterior: bool = True):
+        """
+
+        Args:
+            pc (torch.FloatTensor): [B, N, 3]
+            feats (torch.FloatTensor or None): [B, N, c]
+            sample_posterior (bool):
+
+        Returns:
+            shape_embed (torch.FloatTensor)
+            kl_embed (torch.FloatTensor):
+            posterior (DiagonalGaussianDistribution or None):
+        """
+
+        shape_embed, latents = self.encode_latents(pc, feats)
+        kl_embed, posterior = self.encode_kl_embed(latents, sample_posterior)
+
+        return shape_embed, kl_embed, posterior
+
+    def encode_latents(self,
+                       pc: torch.FloatTensor,
+                       feats: Optional[torch.FloatTensor] = None):
+
+        x, _ = self.encoder(pc, feats)
+
+        shape_embed = x[:, 0]
+        latents = x[:, 1:]
+
+        return shape_embed, latents
+
+    def encode_kl_embed(self, latents: torch.FloatTensor, sample_posterior: bool = True):
+        posterior = None
+        if self.embed_dim > 0:
+            moments = self.pre_kl(latents)
+            posterior = DiagonalGaussianDistribution(moments, feat_dim=-1)
+
+            if sample_posterior:
+                kl_embed = posterior.sample()
+            else:
+                kl_embed = posterior.mode()
+        else:
+            kl_embed = latents
+
+        return kl_embed, posterior
+
+    def forward(self,
+                pc: torch.FloatTensor,
+                feats: torch.FloatTensor,
+                volume_queries: torch.FloatTensor,
+                sample_posterior: bool = True):
+        """
+
+        Args:
+            pc (torch.FloatTensor): [B, N, 3]
+            feats (torch.FloatTensor or None): [B, N, C]
+            volume_queries (torch.FloatTensor): [B, P, 3]
+            sample_posterior (bool):
+
+        Returns:
+            shape_embed (torch.FloatTensor): [B, projection_dim]
+            logits (torch.FloatTensor): [B, M]
+            posterior (DiagonalGaussianDistribution or None).
+
+        """
+
+        shape_embed, kl_embed, posterior = self.encode(pc, feats, sample_posterior=sample_posterior)
+
+        latents = self.decode(kl_embed)
+        logits = self.query_geometry(volume_queries, latents)
+
+        return shape_embed, logits, posterior

MeshAnything/miche/michelangelo/models/tsal/sal_pl_module.py

@@ -0,0 +1,290 @@
+# -*- coding: utf-8 -*-
+
+from typing import List, Tuple, Dict, Optional
+from omegaconf import DictConfig
+
+import torch
+from torch.optim import lr_scheduler
+import pytorch_lightning as pl
+from typing import Union
+from functools import partial
+
+from MeshAnything.miche.michelangelo.utils import instantiate_from_config
+
+from .inference_utils import extract_geometry
+from .tsal_base import (
+    ShapeAsLatentModule,
+    Latent2MeshOutput,
+    Point2MeshOutput
+)
+
+
+class ShapeAsLatentPLModule(pl.LightningModule):
+
+    def __init__(self, *,
+                 module_cfg,
+                 loss_cfg,
+                 optimizer_cfg: Optional[DictConfig] = None,
+                 ckpt_path: Optional[str] = None,
+                 ignore_keys: Union[Tuple[str], List[str]] = ()):
+
+        super().__init__()
+
+        self.sal: ShapeAsLatentModule = instantiate_from_config(module_cfg, device=None, dtype=None)
+
+        self.loss = instantiate_from_config(loss_cfg)
+
+        self.optimizer_cfg = optimizer_cfg
+
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+
+        self.save_hyperparameters()
+
+    @property
+    def latent_shape(self):
+        return self.sal.latent_shape
+
+    @property
+    def zero_rank(self):
+        if self._trainer:
+            zero_rank = self.trainer.local_rank == 0
+        else:
+            zero_rank = True
+
+        return zero_rank
+
+    def init_from_ckpt(self, path, ignore_keys=()):
+        state_dict = torch.load(path, map_location="cpu")["state_dict"]
+
+        keys = list(state_dict.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del state_dict[k]
+
+        missing, unexpected = self.load_state_dict(state_dict, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+            print(f"Unexpected Keys: {unexpected}")
+
+    def configure_optimizers(self) -> Tuple[List, List]:
+        lr = self.learning_rate
+
+        # optimizers = [torch.optim.AdamW(self.sal.parameters(), lr=lr, betas=(0.9, 0.99), weight_decay=1e-4)]
+        # optimizers = [torch.optim.AdamW(self.sal.parameters(), lr=lr, betas=(0.9, 0.99), weight_decay=1e-3)]
+
+        if self.optimizer_cfg is None:
+            optimizers = [torch.optim.AdamW(self.sal.parameters(), lr=lr, betas=(0.9, 0.99), weight_decay=1e-3)]
+            schedulers = []
+        else:
+            optimizer = instantiate_from_config(self.optimizer_cfg.optimizer, params=self.sal.parameters())
+            scheduler_func = instantiate_from_config(
+                self.optimizer_cfg.scheduler,
+                max_decay_steps=self.trainer.max_steps,
+                lr_max=lr
+            )
+            scheduler = {
+                "scheduler": lr_scheduler.LambdaLR(optimizer, lr_lambda=scheduler_func.schedule),
+                "interval": "step",
+                "frequency": 1
+            }
+            optimizers = [optimizer]
+            schedulers = [scheduler]
+
+        return optimizers, schedulers
+
+    def forward(self,
+                pc: torch.FloatTensor,
+                feats: torch.FloatTensor,
+                volume_queries: torch.FloatTensor):
+
+        logits, center_pos, posterior = self.sal(pc, feats, volume_queries)
+
+        return posterior, logits
+
+    def encode(self, surface: torch.FloatTensor, sample_posterior=True):
+
+        pc = surface[..., 0:3]
+        feats = surface[..., 3:6]
+
+        latents, center_pos, posterior = self.sal.encode(
+            pc=pc, feats=feats, sample_posterior=sample_posterior
+        )
+
+        return latents
+
+    def decode(self,
+               z_q,
+               bounds: Union[Tuple[float], List[float], float] = 1.1,
+               octree_depth: int = 7,
+               num_chunks: int = 10000) -> List[Latent2MeshOutput]:
+
+        latents = self.sal.decode(z_q)  # latents: [bs, num_latents, dim]
+        outputs = self.latent2mesh(latents, bounds=bounds, octree_depth=octree_depth, num_chunks=num_chunks)
+
+        return outputs
+
+    def training_step(self, batch: Dict[str, torch.FloatTensor],
+                      batch_idx: int, optimizer_idx: int = 0) -> torch.FloatTensor:
+        """
+
+        Args:
+            batch (dict): the batch sample, and it contains:
+                - surface (torch.FloatTensor): [bs, n_surface, (3 + input_dim)]
+                - geo_points (torch.FloatTensor): [bs, n_pts, (3 + 1)]
+
+            batch_idx (int):
+
+            optimizer_idx (int):
+
+        Returns:
+            loss (torch.FloatTensor):
+
+        """
+
+        pc = batch["surface"][..., 0:3]
+        feats = batch["surface"][..., 3:]
+
+        volume_queries = batch["geo_points"][..., 0:3]
+        volume_labels = batch["geo_points"][..., -1]
+
+        posterior, logits = self(
+            pc=pc, feats=feats, volume_queries=volume_queries
+        )
+        aeloss, log_dict_ae = self.loss(posterior, logits, volume_labels, split="train")
+
+        self.log_dict(log_dict_ae, prog_bar=True, logger=True, batch_size=logits.shape[0],
+                      sync_dist=False, rank_zero_only=True)
+
+        return aeloss
+
+    def validation_step(self, batch: Dict[str, torch.FloatTensor], batch_idx: int) -> torch.FloatTensor:
+
+        pc = batch["surface"][..., 0:3]
+        feats = batch["surface"][..., 3:]
+
+        volume_queries = batch["geo_points"][..., 0:3]
+        volume_labels = batch["geo_points"][..., -1]
+
+        posterior, logits = self(
+            pc=pc, feats=feats, volume_queries=volume_queries,
+        )
+        aeloss, log_dict_ae = self.loss(posterior, logits, volume_labels, split="val")
+
+        self.log_dict(log_dict_ae, prog_bar=True, logger=True, batch_size=logits.shape[0],
+                      sync_dist=False, rank_zero_only=True)
+
+        return aeloss
+
+    def point2mesh(self,
+                   pc: torch.FloatTensor,
+                   feats: torch.FloatTensor,
+                   bounds: Union[Tuple[float], List[float]] = (-1.25, -1.25, -1.25, 1.25, 1.25, 1.25),
+                   octree_depth: int = 7,
+                   num_chunks: int = 10000) -> List[Point2MeshOutput]:
+
+        """
+
+        Args:
+            pc:
+            feats:
+            bounds:
+            octree_depth:
+            num_chunks:
+
+        Returns:
+            mesh_outputs (List[MeshOutput]): the mesh outputs list.
+
+        """
+
+        outputs = []
+
+        device = pc.device
+        bs = pc.shape[0]
+
+        # 1. point encoder + latents transformer
+        latents, center_pos, posterior = self.sal.encode(pc, feats)
+        latents = self.sal.decode(latents)  # latents: [bs, num_latents, dim]
+
+        geometric_func = partial(self.sal.query_geometry, latents=latents)
+
+        # 2. decode geometry
+        mesh_v_f, has_surface = extract_geometry(
+            geometric_func=geometric_func,
+            device=device,
+            batch_size=bs,
+            bounds=bounds,
+            octree_depth=octree_depth,
+            num_chunks=num_chunks,
+            disable=not self.zero_rank
+        )
+
+        # 3. decode texture
+        for i, ((mesh_v, mesh_f), is_surface) in enumerate(zip(mesh_v_f, has_surface)):
+            if not is_surface:
+                outputs.append(None)
+                continue
+
+            out = Point2MeshOutput()
+            out.mesh_v = mesh_v
+            out.mesh_f = mesh_f
+            out.pc = torch.cat([pc[i], feats[i]], dim=-1).cpu().numpy()
+
+            if center_pos is not None:
+                out.center = center_pos[i].cpu().numpy()
+
+            outputs.append(out)
+
+        return outputs
+
+    def latent2mesh(self,
+                    latents: torch.FloatTensor,
+                    bounds: Union[Tuple[float], List[float], float] = 1.1,
+                    octree_depth: int = 7,
+                    num_chunks: int = 10000) -> List[Latent2MeshOutput]:
+
+        """
+
+        Args:
+            latents: [bs, num_latents, dim]
+            bounds:
+            octree_depth:
+            num_chunks:
+
+        Returns:
+            mesh_outputs (List[MeshOutput]): the mesh outputs list.
+
+        """
+
+        outputs = []
+
+        geometric_func = partial(self.sal.query_geometry, latents=latents)
+
+        # 2. decode geometry
+        device = latents.device
+        mesh_v_f, has_surface = extract_geometry(
+            geometric_func=geometric_func,
+            device=device,
+            batch_size=len(latents),
+            bounds=bounds,
+            octree_depth=octree_depth,
+            num_chunks=num_chunks,
+            disable=not self.zero_rank
+        )
+
+        # 3. decode texture
+        for i, ((mesh_v, mesh_f), is_surface) in enumerate(zip(mesh_v_f, has_surface)):
+            if not is_surface:
+                outputs.append(None)
+                continue
+
+            out = Latent2MeshOutput()
+            out.mesh_v = mesh_v
+            out.mesh_f = mesh_f
+
+            outputs.append(out)
+
+        return outputs

MeshAnything/miche/michelangelo/models/tsal/tsal_base.py

@@ -0,0 +1,120 @@
+# -*- coding: utf-8 -*-
+
+import torch.nn as nn
+from typing import Tuple, List, Optional
+
+
+class Point2MeshOutput(object):
+    def __init__(self):
+        self.mesh_v = None
+        self.mesh_f = None
+        self.center = None
+        self.pc = None
+
+
+class Latent2MeshOutput(object):
+
+    def __init__(self):
+        self.mesh_v = None
+        self.mesh_f = None
+
+
+class AlignedMeshOutput(object):
+
+    def __init__(self):
+        self.mesh_v = None
+        self.mesh_f = None
+        self.surface = None
+        self.image = None
+        self.text: Optional[str] = None
+        self.shape_text_similarity: Optional[float] = None
+        self.shape_image_similarity: Optional[float] = None
+
+
+class ShapeAsLatentPLModule(nn.Module):
+    latent_shape: Tuple[int]
+
+    def encode(self, surface, *args, **kwargs):
+        raise NotImplementedError
+
+    def decode(self, z_q, *args, **kwargs):
+        raise NotImplementedError
+
+    def latent2mesh(self, latents, *args, **kwargs) -> List[Latent2MeshOutput]:
+        raise NotImplementedError
+
+    def point2mesh(self, *args, **kwargs) -> List[Point2MeshOutput]:
+        raise NotImplementedError
+
+
+class ShapeAsLatentModule(nn.Module):
+    latent_shape: Tuple[int, int]
+
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+
+    def encode(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def decode(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def query_geometry(self, *args, **kwargs):
+        raise NotImplementedError
+
+
+class AlignedShapeAsLatentPLModule(nn.Module):
+    latent_shape: Tuple[int]
+
+    def set_shape_model_only(self):
+        raise NotImplementedError
+
+    def encode(self, surface, *args, **kwargs):
+        raise NotImplementedError
+
+    def decode(self, z_q, *args, **kwargs):
+        raise NotImplementedError
+
+    def latent2mesh(self, latents, *args, **kwargs) -> List[Latent2MeshOutput]:
+        raise NotImplementedError
+
+    def point2mesh(self, *args, **kwargs) -> List[Point2MeshOutput]:
+        raise NotImplementedError
+
+
+class AlignedShapeAsLatentModule(nn.Module):
+    shape_model: ShapeAsLatentModule
+    latent_shape: Tuple[int, int]
+
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+
+    def set_shape_model_only(self):
+        raise NotImplementedError
+
+    def encode_image_embed(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def encode_text_embed(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def encode_shape_embed(self, *args, **kwargs):
+        raise NotImplementedError
+
+
+class TexturedShapeAsLatentModule(nn.Module):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+
+    def encode(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def decode(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def query_geometry(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def query_color(self, *args, **kwargs):
+        raise NotImplementedError

MeshAnything/miche/michelangelo/utils/__init__.py

@@ -0,0 +1,3 @@
+# -*- coding: utf-8 -*-
+
+from .misc import instantiate_from_config

MeshAnything/miche/michelangelo/utils/eval.py

@@ -0,0 +1,12 @@
+# -*- coding: utf-8 -*-
+
+import torch
+
+
+def compute_psnr(x, y, data_range: float = 2, eps: float = 1e-7):
+
+    mse = torch.mean((x - y) ** 2)
+    psnr = 10 * torch.log10(data_range / (mse + eps))
+
+    return psnr
+

MeshAnything/miche/michelangelo/utils/io.py

@@ -0,0 +1,47 @@
+# -*- coding: utf-8 -*-
+
+import os
+import io
+import tarfile
+import json
+import numpy as np
+import numpy.lib.format
+
+
+def mkdir(path):
+    os.makedirs(path, exist_ok=True)
+    return path
+
+
+def npy_loads(data):
+    stream = io.BytesIO(data)
+    return np.lib.format.read_array(stream)
+
+
+def npz_loads(data):
+    return np.load(io.BytesIO(data))
+
+
+def json_loads(data):
+    return json.loads(data)
+
+
+def load_json(filepath):
+    with open(filepath, "r") as f:
+        data = json.load(f)
+        return data
+
+
+def write_json(filepath, data):
+    with open(filepath, "w") as f:
+        json.dump(data, f, indent=2)
+
+
+def extract_tar(tar_path, tar_cache_folder):
+
+    with tarfile.open(tar_path, "r") as tar:
+        tar.extractall(path=tar_cache_folder)
+
+    tar_uids = sorted(os.listdir(tar_cache_folder))
+    print(f"extract tar: {tar_path} to {tar_cache_folder}")
+    return tar_uids

MeshAnything/miche/michelangelo/utils/misc.py

@@ -0,0 +1,83 @@
+# -*- coding: utf-8 -*-
+
+import importlib
+
+import torch
+import torch.distributed as dist
+
+
+
+def get_obj_from_str(string, reload=False):
+    module, cls = string.rsplit(".", 1)
+    if reload:
+        module_imp = importlib.import_module(module)
+        importlib.reload(module_imp)
+    return getattr(importlib.import_module(module, package=None), cls)
+
+
+def get_obj_from_config(config):
+    if "target" not in config:
+        raise KeyError("Expected key `target` to instantiate.")
+
+    return get_obj_from_str(config["target"])
+
+
+def instantiate_from_config(config, **kwargs):
+    if "target" not in config:
+        raise KeyError("Expected key `target` to instantiate.")
+
+    cls = get_obj_from_str(config["target"])
+
+    params = config.get("params", dict())
+    # params.update(kwargs)
+    # instance = cls(**params)
+    kwargs.update(params)
+    instance = cls(**kwargs)
+
+    return instance
+
+
+def is_dist_avail_and_initialized():
+    if not dist.is_available():
+        return False
+    if not dist.is_initialized():
+        return False
+    return True
+
+
+def get_rank():
+    if not is_dist_avail_and_initialized():
+        return 0
+    return dist.get_rank()
+
+
+def get_world_size():
+    if not is_dist_avail_and_initialized():
+        return 1
+    return dist.get_world_size()
+
+
+def all_gather_batch(tensors):
+    """
+    Performs all_gather operation on the provided tensors.
+    """
+    # Queue the gathered tensors
+    world_size = get_world_size()
+    # There is no need for reduction in the single-proc case
+    if world_size == 1:
+        return tensors
+    tensor_list = []
+    output_tensor = []
+    for tensor in tensors:
+        tensor_all = [torch.ones_like(tensor) for _ in range(world_size)]
+        dist.all_gather(
+            tensor_all,
+            tensor,
+            async_op=False  # performance opt
+        )
+
+        tensor_list.append(tensor_all)
+
+    for tensor_all in tensor_list:
+        output_tensor.append(torch.cat(tensor_all, dim=0))
+    return output_tensor

MeshAnything/miche/michelangelo/utils/visualizers/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

MeshAnything/miche/michelangelo/utils/visualizers/color_util.py

@@ -0,0 +1,43 @@
+import numpy as np
+import matplotlib.pyplot as plt
+
+
+# Helper functions
+def get_colors(inp, colormap="viridis", normalize=True, vmin=None, vmax=None):
+    colormap = plt.cm.get_cmap(colormap)
+    if normalize:
+        vmin = np.min(inp)
+        vmax = np.max(inp)
+
+    norm = plt.Normalize(vmin, vmax)
+    return colormap(norm(inp))[:, :3]
+
+
+def gen_checkers(n_checkers_x, n_checkers_y, width=256, height=256):
+    # tex dims need to be power of two.
+    array = np.ones((width, height, 3), dtype='float32')
+
+    # width in texels of each checker
+    checker_w = width / n_checkers_x
+    checker_h = height / n_checkers_y
+
+    for y in range(height):
+        for x in range(width):
+            color_key = int(x / checker_w) + int(y / checker_h)
+            if color_key % 2 == 0:
+                array[x, y, :] = [1., 0.874, 0.0]
+            else:
+                array[x, y, :] = [0., 0., 0.]
+    return array
+
+
+def gen_circle(width=256, height=256):
+    xx, yy = np.mgrid[:width, :height]
+    circle = (xx - width / 2 + 0.5) ** 2 + (yy - height / 2 + 0.5) ** 2
+    array = np.ones((width, height, 4), dtype='float32')
+    array[:, :, 0] = (circle <= width)
+    array[:, :, 1] = (circle <= width)
+    array[:, :, 2] = (circle <= width)
+    array[:, :, 3] = circle <= width
+    return array
+

MeshAnything/miche/michelangelo/utils/visualizers/html_util.py

@@ -0,0 +1,49 @@
+# -*- coding: utf-8 -*-
+import io
+import base64
+import numpy as np
+from PIL import Image
+
+
+def to_html_frame(content):
+
+    html_frame = f"""
+    <html>
+      <body>
+        {content}
+      </body>
+    </html>
+    """
+
+    return html_frame
+
+
+def to_single_row_table(caption: str, content: str):
+
+    table_html = f"""
+    <table border = "1">
+        <caption>{caption}</caption>
+        <tr>
+            <td>{content}</td>
+        </tr>
+    </table>
+    """
+
+    return table_html
+
+
+def to_image_embed_tag(image: np.ndarray):
+
+    # Convert np.ndarray to bytes
+    img = Image.fromarray(image)
+    raw_bytes = io.BytesIO()
+    img.save(raw_bytes, "PNG")
+
+    # Encode bytes to base64
+    image_base64 = base64.b64encode(raw_bytes.getvalue()).decode("utf-8")
+
+    image_tag = f"""
+    <img src="data:image/png;base64,{image_base64}" alt="Embedded Image">
+    """
+
+    return image_tag

MeshAnything/miche/michelangelo/utils/visualizers/pythreejs_viewer.py

@@ -0,0 +1,534 @@
+import numpy as np
+from ipywidgets import embed
+import pythreejs as p3s
+import uuid
+
+from .color_util import get_colors, gen_circle, gen_checkers
+
+
+EMBED_URL = "https://cdn.jsdelivr.net/npm/@jupyter-widgets/[email protected]/dist/embed-amd.js"
+
+
+class PyThreeJSViewer(object):
+
+    def __init__(self, settings, render_mode="WEBSITE"):
+        self.render_mode = render_mode
+        self.__update_settings(settings)
+        self._light = p3s.DirectionalLight(color='white', position=[0, 0, 1], intensity=0.6)
+        self._light2 = p3s.AmbientLight(intensity=0.5)
+        self._cam = p3s.PerspectiveCamera(position=[0, 0, 1], lookAt=[0, 0, 0], fov=self.__s["fov"],
+                                          aspect=self.__s["width"] / self.__s["height"], children=[self._light])
+        self._orbit = p3s.OrbitControls(controlling=self._cam)
+        self._scene = p3s.Scene(children=[self._cam, self._light2], background=self.__s["background"])  # "#4c4c80"
+        self._renderer = p3s.Renderer(camera=self._cam, scene=self._scene, controls=[self._orbit],
+                                      width=self.__s["width"], height=self.__s["height"],
+                                      antialias=self.__s["antialias"])
+
+        self.__objects = {}
+        self.__cnt = 0
+
+    def jupyter_mode(self):
+        self.render_mode = "JUPYTER"
+
+    def offline(self):
+        self.render_mode = "OFFLINE"
+
+    def website(self):
+        self.render_mode = "WEBSITE"
+
+    def __get_shading(self, shading):
+        shad = {"flat": True, "wireframe": False, "wire_width": 0.03, "wire_color": "black",
+                "side": 'DoubleSide', "colormap": "viridis", "normalize": [None, None],
+                "bbox": False, "roughness": 0.5, "metalness": 0.25, "reflectivity": 1.0,
+                "line_width": 1.0, "line_color": "black",
+                "point_color": "red", "point_size": 0.01, "point_shape": "circle",
+                "text_color": "red"
+                }
+        for k in shading:
+            shad[k] = shading[k]
+        return shad
+
+    def __update_settings(self, settings={}):
+        sett = {"width": 600, "height": 600, "antialias": True, "scale": 1.5, "background": "#ffffff",
+                "fov": 30}
+        for k in settings:
+            sett[k] = settings[k]
+        self.__s = sett
+
+    def __add_object(self, obj, parent=None):
+        if not parent:  # Object is added to global scene and objects dict
+            self.__objects[self.__cnt] = obj
+            self.__cnt += 1
+            self._scene.add(obj["mesh"])
+        else:  # Object is added to parent object and NOT to objects dict
+            parent.add(obj["mesh"])
+
+        self.__update_view()
+
+        if self.render_mode == "JUPYTER":
+            return self.__cnt - 1
+        elif self.render_mode == "WEBSITE":
+            return self
+
+    def __add_line_geometry(self, lines, shading, obj=None):
+        lines = lines.astype("float32", copy=False)
+        mi = np.min(lines, axis=0)
+        ma = np.max(lines, axis=0)
+
+        geometry = p3s.LineSegmentsGeometry(positions=lines.reshape((-1, 2, 3)))
+        material = p3s.LineMaterial(linewidth=shading["line_width"], color=shading["line_color"])
+        # , vertexColors='VertexColors'),
+        lines = p3s.LineSegments2(geometry=geometry, material=material)  # type='LinePieces')
+        line_obj = {"geometry": geometry, "mesh": lines, "material": material,
+                    "max": ma, "min": mi, "type": "Lines", "wireframe": None}
+
+        if obj:
+            return self.__add_object(line_obj, obj), line_obj
+        else:
+            return self.__add_object(line_obj)
+
+    def __update_view(self):
+        if len(self.__objects) == 0:
+            return
+        ma = np.zeros((len(self.__objects), 3))
+        mi = np.zeros((len(self.__objects), 3))
+        for r, obj in enumerate(self.__objects):
+            ma[r] = self.__objects[obj]["max"]
+            mi[r] = self.__objects[obj]["min"]
+        ma = np.max(ma, axis=0)
+        mi = np.min(mi, axis=0)
+        diag = np.linalg.norm(ma - mi)
+        mean = ((ma - mi) / 2 + mi).tolist()
+        scale = self.__s["scale"] * (diag)
+        self._orbit.target = mean
+        self._cam.lookAt(mean)
+        self._cam.position = [mean[0], mean[1], mean[2] + scale]
+        self._light.position = [mean[0], mean[1], mean[2] + scale]
+
+        self._orbit.exec_three_obj_method('update')
+        self._cam.exec_three_obj_method('updateProjectionMatrix')
+
+    def __get_bbox(self, v):
+        m = np.min(v, axis=0)
+        M = np.max(v, axis=0)
+
+        # Corners of the bounding box
+        v_box = np.array([[m[0], m[1], m[2]], [M[0], m[1], m[2]], [M[0], M[1], m[2]], [m[0], M[1], m[2]],
+                          [m[0], m[1], M[2]], [M[0], m[1], M[2]], [M[0], M[1], M[2]], [m[0], M[1], M[2]]])
+
+        f_box = np.array([[0, 1], [1, 2], [2, 3], [3, 0], [4, 5], [5, 6], [6, 7], [7, 4],
+                          [0, 4], [1, 5], [2, 6], [7, 3]], dtype=np.uint32)
+        return v_box, f_box
+
+    def __get_colors(self, v, f, c, sh):
+        coloring = "VertexColors"
+        if type(c) == np.ndarray and c.size == 3:  # Single color
+            colors = np.ones_like(v)
+            colors[:, 0] = c[0]
+            colors[:, 1] = c[1]
+            colors[:, 2] = c[2]
+            # print("Single colors")
+        elif type(c) == np.ndarray and len(c.shape) == 2 and c.shape[1] == 3:  # Color values for
+            if c.shape[0] == f.shape[0]:  # faces
+                colors = np.hstack([c, c, c]).reshape((-1, 3))
+                coloring = "FaceColors"
+                # print("Face color values")
+            elif c.shape[0] == v.shape[0]:  # vertices
+                colors = c
+                # print("Vertex color values")
+            else:  # Wrong size, fallback
+                print("Invalid color array given! Supported are numpy arrays.", type(c))
+                colors = np.ones_like(v)
+                colors[:, 0] = 1.0
+                colors[:, 1] = 0.874
+                colors[:, 2] = 0.0
+        elif type(c) == np.ndarray and c.size == f.shape[0]:  # Function values for faces
+            normalize = sh["normalize"][0] != None and sh["normalize"][1] != None
+            cc = get_colors(c, sh["colormap"], normalize=normalize,
+                            vmin=sh["normalize"][0], vmax=sh["normalize"][1])
+            # print(cc.shape)
+            colors = np.hstack([cc, cc, cc]).reshape((-1, 3))
+            coloring = "FaceColors"
+            # print("Face function values")
+        elif type(c) == np.ndarray and c.size == v.shape[0]:  # Function values for vertices
+            normalize = sh["normalize"][0] != None and sh["normalize"][1] != None
+            colors = get_colors(c, sh["colormap"], normalize=normalize,
+                                vmin=sh["normalize"][0], vmax=sh["normalize"][1])
+            # print("Vertex function values")
+
+        else:
+            colors = np.ones_like(v)
+            colors[:, 0] = 1.0
+            colors[:, 1] = 0.874
+            colors[:, 2] = 0.0
+
+            # No color
+            if c is not None:
+                print("Invalid color array given! Supported are numpy arrays.", type(c))
+
+        return colors, coloring
+
+    def __get_point_colors(self, v, c, sh):
+        v_color = True
+        if c is None:  # No color given, use global color
+            # conv = mpl.colors.ColorConverter()
+            colors = sh["point_color"]  # np.array(conv.to_rgb(sh["point_color"]))
+            v_color = False
+        elif isinstance(c, str):  # No color given, use global color
+            # conv = mpl.colors.ColorConverter()
+            colors = c  # np.array(conv.to_rgb(c))
+            v_color = False
+        elif type(c) == np.ndarray and len(c.shape) == 2 and c.shape[0] == v.shape[0] and c.shape[1] == 3:
+            # Point color
+            colors = c.astype("float32", copy=False)
+
+        elif isinstance(c, np.ndarray) and len(c.shape) == 2 and c.shape[0] == v.shape[0] and c.shape[1] != 3:
+            # Function values for vertices, but the colors are features
+            c_norm = np.linalg.norm(c, ord=2, axis=-1)
+            normalize = sh["normalize"][0] != None and sh["normalize"][1] != None
+            colors = get_colors(c_norm, sh["colormap"], normalize=normalize,
+                                vmin=sh["normalize"][0], vmax=sh["normalize"][1])
+            colors = colors.astype("float32", copy=False)
+
+        elif type(c) == np.ndarray and c.size == v.shape[0]:  # Function color
+            normalize = sh["normalize"][0] != None and sh["normalize"][1] != None
+            colors = get_colors(c, sh["colormap"], normalize=normalize,
+                                vmin=sh["normalize"][0], vmax=sh["normalize"][1])
+            colors = colors.astype("float32", copy=False)
+            # print("Vertex function values")
+
+        else:
+            print("Invalid color array given! Supported are numpy arrays.", type(c))
+            colors = sh["point_color"]
+            v_color = False
+
+        return colors, v_color
+
+    def add_mesh(self, v, f, c=None, uv=None, n=None, shading={}, texture_data=None, **kwargs):
+        shading.update(kwargs)
+        sh = self.__get_shading(shading)
+        mesh_obj = {}
+
+        # it is a tet
+        if v.shape[1] == 3 and f.shape[1] == 4:
+            f_tmp = np.ndarray([f.shape[0] * 4, 3], dtype=f.dtype)
+            for i in range(f.shape[0]):
+                f_tmp[i * 4 + 0] = np.array([f[i][1], f[i][0], f[i][2]])
+                f_tmp[i * 4 + 1] = np.array([f[i][0], f[i][1], f[i][3]])
+                f_tmp[i * 4 + 2] = np.array([f[i][1], f[i][2], f[i][3]])
+                f_tmp[i * 4 + 3] = np.array([f[i][2], f[i][0], f[i][3]])
+            f = f_tmp
+
+        if v.shape[1] == 2:
+            v = np.append(v, np.zeros([v.shape[0], 1]), 1)
+
+        # Type adjustment vertices
+        v = v.astype("float32", copy=False)
+
+        # Color setup
+        colors, coloring = self.__get_colors(v, f, c, sh)
+
+        # Type adjustment faces and colors
+        c = colors.astype("float32", copy=False)
+
+        # Material and geometry setup
+        ba_dict = {"color": p3s.BufferAttribute(c)}
+        if coloring == "FaceColors":
+            verts = np.zeros((f.shape[0] * 3, 3), dtype="float32")
+            for ii in range(f.shape[0]):
+                # print(ii*3, f[ii])
+                verts[ii * 3] = v[f[ii, 0]]
+                verts[ii * 3 + 1] = v[f[ii, 1]]
+                verts[ii * 3 + 2] = v[f[ii, 2]]
+            v = verts
+        else:
+            f = f.astype("uint32", copy=False).ravel()
+            ba_dict["index"] = p3s.BufferAttribute(f, normalized=False)
+
+        ba_dict["position"] = p3s.BufferAttribute(v, normalized=False)
+
+        if uv is not None:
+            uv = (uv - np.min(uv)) / (np.max(uv) - np.min(uv))
+            if texture_data is None:
+                texture_data = gen_checkers(20, 20)
+            tex = p3s.DataTexture(data=texture_data, format="RGBFormat", type="FloatType")
+            material = p3s.MeshStandardMaterial(map=tex, reflectivity=sh["reflectivity"], side=sh["side"],
+                                                roughness=sh["roughness"], metalness=sh["metalness"],
+                                                flatShading=sh["flat"],
+                                                polygonOffset=True, polygonOffsetFactor=1, polygonOffsetUnits=5)
+            ba_dict["uv"] = p3s.BufferAttribute(uv.astype("float32", copy=False))
+        else:
+            material = p3s.MeshStandardMaterial(vertexColors=coloring, reflectivity=sh["reflectivity"],
+                                                side=sh["side"], roughness=sh["roughness"], metalness=sh["metalness"],
+                                                flatShading=sh["flat"],
+                                                polygonOffset=True, polygonOffsetFactor=1, polygonOffsetUnits=5)
+
+        if type(n) != type(None) and coloring == "VertexColors":  # TODO: properly handle normals for FaceColors as well
+            ba_dict["normal"] = p3s.BufferAttribute(n.astype("float32", copy=False), normalized=True)
+
+        geometry = p3s.BufferGeometry(attributes=ba_dict)
+
+        if coloring == "VertexColors" and type(n) == type(None):
+            geometry.exec_three_obj_method('computeVertexNormals')
+        elif coloring == "FaceColors" and type(n) == type(None):
+            geometry.exec_three_obj_method('computeFaceNormals')
+
+        # Mesh setup
+        mesh = p3s.Mesh(geometry=geometry, material=material)
+
+        # Wireframe setup
+        mesh_obj["wireframe"] = None
+        if sh["wireframe"]:
+            wf_geometry = p3s.WireframeGeometry(mesh.geometry)  # WireframeGeometry
+            wf_material = p3s.LineBasicMaterial(color=sh["wire_color"], linewidth=sh["wire_width"])
+            wireframe = p3s.LineSegments(wf_geometry, wf_material)
+            mesh.add(wireframe)
+            mesh_obj["wireframe"] = wireframe
+
+        # Bounding box setup
+        if sh["bbox"]:
+            v_box, f_box = self.__get_bbox(v)
+            _, bbox = self.add_edges(v_box, f_box, sh, mesh)
+            mesh_obj["bbox"] = [bbox, v_box, f_box]
+
+        # Object setup
+        mesh_obj["max"] = np.max(v, axis=0)
+        mesh_obj["min"] = np.min(v, axis=0)
+        mesh_obj["geometry"] = geometry
+        mesh_obj["mesh"] = mesh
+        mesh_obj["material"] = material
+        mesh_obj["type"] = "Mesh"
+        mesh_obj["shading"] = sh
+        mesh_obj["coloring"] = coloring
+        mesh_obj["arrays"] = [v, f, c]  # TODO replays with proper storage or remove if not needed
+
+        return self.__add_object(mesh_obj)
+
+    def add_lines(self, beginning, ending, shading={}, obj=None, **kwargs):
+        shading.update(kwargs)
+        if len(beginning.shape) == 1:
+            if len(beginning) == 2:
+                beginning = np.array([[beginning[0], beginning[1], 0]])
+        else:
+            if beginning.shape[1] == 2:
+                beginning = np.append(
+                    beginning, np.zeros([beginning.shape[0], 1]), 1)
+        if len(ending.shape) == 1:
+            if len(ending) == 2:
+                ending = np.array([[ending[0], ending[1], 0]])
+        else:
+            if ending.shape[1] == 2:
+                ending = np.append(
+                    ending, np.zeros([ending.shape[0], 1]), 1)
+
+        sh = self.__get_shading(shading)
+        lines = np.hstack([beginning, ending])
+        lines = lines.reshape((-1, 3))
+        return self.__add_line_geometry(lines, sh, obj)
+
+    def add_edges(self, vertices, edges, shading={}, obj=None, **kwargs):
+        shading.update(kwargs)
+        if vertices.shape[1] == 2:
+            vertices = np.append(
+                vertices, np.zeros([vertices.shape[0], 1]), 1)
+        sh = self.__get_shading(shading)
+        lines = np.zeros((edges.size, 3))
+        cnt = 0
+        for e in edges:
+            lines[cnt, :] = vertices[e[0]]
+            lines[cnt + 1, :] = vertices[e[1]]
+            cnt += 2
+        return self.__add_line_geometry(lines, sh, obj)
+
+    def add_points(self, points, c=None, shading={}, obj=None, **kwargs):
+        shading.update(kwargs)
+        if len(points.shape) == 1:
+            if len(points) == 2:
+                points = np.array([[points[0], points[1], 0]])
+        else:
+            if points.shape[1] == 2:
+                points = np.append(
+                    points, np.zeros([points.shape[0], 1]), 1)
+        sh = self.__get_shading(shading)
+        points = points.astype("float32", copy=False)
+        mi = np.min(points, axis=0)
+        ma = np.max(points, axis=0)
+
+        g_attributes = {"position": p3s.BufferAttribute(points, normalized=False)}
+        m_attributes = {"size": sh["point_size"]}
+
+        if sh["point_shape"] == "circle":  # Plot circles
+            tex = p3s.DataTexture(data=gen_circle(16, 16), format="RGBAFormat", type="FloatType")
+            m_attributes["map"] = tex
+            m_attributes["alphaTest"] = 0.5
+            m_attributes["transparency"] = True
+        else:  # Plot squares
+            pass
+
+        colors, v_colors = self.__get_point_colors(points, c, sh)
+        if v_colors:  # Colors per point
+            m_attributes["vertexColors"] = 'VertexColors'
+            g_attributes["color"] = p3s.BufferAttribute(colors, normalized=False)
+
+        else:  # Colors for all points
+            m_attributes["color"] = colors
+
+        material = p3s.PointsMaterial(**m_attributes)
+        geometry = p3s.BufferGeometry(attributes=g_attributes)
+        points = p3s.Points(geometry=geometry, material=material)
+        point_obj = {"geometry": geometry, "mesh": points, "material": material,
+                     "max": ma, "min": mi, "type": "Points", "wireframe": None}
+
+        if obj:
+            return self.__add_object(point_obj, obj), point_obj
+        else:
+            return self.__add_object(point_obj)
+
+    def remove_object(self, obj_id):
+        if obj_id not in self.__objects:
+            print("Invalid object id. Valid ids are: ", list(self.__objects.keys()))
+            return
+        self._scene.remove(self.__objects[obj_id]["mesh"])
+        del self.__objects[obj_id]
+        self.__update_view()
+
+    def reset(self):
+        for obj_id in list(self.__objects.keys()).copy():
+            self._scene.remove(self.__objects[obj_id]["mesh"])
+            del self.__objects[obj_id]
+        self.__update_view()
+
+    def update_object(self, oid=0, vertices=None, colors=None, faces=None):
+        obj = self.__objects[oid]
+        if type(vertices) != type(None):
+            if obj["coloring"] == "FaceColors":
+                f = obj["arrays"][1]
+                verts = np.zeros((f.shape[0] * 3, 3), dtype="float32")
+                for ii in range(f.shape[0]):
+                    # print(ii*3, f[ii])
+                    verts[ii * 3] = vertices[f[ii, 0]]
+                    verts[ii * 3 + 1] = vertices[f[ii, 1]]
+                    verts[ii * 3 + 2] = vertices[f[ii, 2]]
+                v = verts
+
+            else:
+                v = vertices.astype("float32", copy=False)
+            obj["geometry"].attributes["position"].array = v
+            # self.wireframe.attributes["position"].array = v # Wireframe updates?
+            obj["geometry"].attributes["position"].needsUpdate = True
+        #           obj["geometry"].exec_three_obj_method('computeVertexNormals')
+        if type(colors) != type(None):
+            colors, coloring = self.__get_colors(obj["arrays"][0], obj["arrays"][1], colors, obj["shading"])
+            colors = colors.astype("float32", copy=False)
+            obj["geometry"].attributes["color"].array = colors
+            obj["geometry"].attributes["color"].needsUpdate = True
+        if type(faces) != type(None):
+            if obj["coloring"] == "FaceColors":
+                print("Face updates are currently only possible in vertex color mode.")
+                return
+            f = faces.astype("uint32", copy=False).ravel()
+            print(obj["geometry"].attributes)
+            obj["geometry"].attributes["index"].array = f
+            # self.wireframe.attributes["position"].array = v # Wireframe updates?
+            obj["geometry"].attributes["index"].needsUpdate = True
+        #            obj["geometry"].exec_three_obj_method('computeVertexNormals')
+        # self.mesh.geometry.verticesNeedUpdate = True
+        # self.mesh.geometry.elementsNeedUpdate = True
+        # self.update()
+        if self.render_mode == "WEBSITE":
+            return self
+
+    #    def update(self):
+    #        self.mesh.exec_three_obj_method('update')
+    #        self.orbit.exec_three_obj_method('update')
+    #        self.cam.exec_three_obj_method('updateProjectionMatrix')
+    #        self.scene.exec_three_obj_method('update')
+
+    def add_text(self, text, shading={}, **kwargs):
+        shading.update(kwargs)
+        sh = self.__get_shading(shading)
+        tt = p3s.TextTexture(string=text, color=sh["text_color"])
+        sm = p3s.SpriteMaterial(map=tt)
+        text = p3s.Sprite(material=sm, scaleToTexture=True)
+        self._scene.add(text)
+
+    # def add_widget(self, widget, callback):
+    #    self.widgets.append(widget)
+    #    widget.observe(callback, names='value')
+
+    #    def add_dropdown(self, options, default, desc, cb):
+    #        widget = widgets.Dropdown(options=options, value=default, description=desc)
+    #        self.__widgets.append(widget)
+    #        widget.observe(cb, names="value")
+    #        display(widget)
+
+    #    def add_button(self, text, cb):
+    #        button = widgets.Button(description=text)
+    #        self.__widgets.append(button)
+    #        button.on_click(cb)
+    #        display(button)
+
+    def to_html(self, imports=True, html_frame=True):
+        # Bake positions (fixes centering bug in offline rendering)
+        if len(self.__objects) == 0:
+            return
+        ma = np.zeros((len(self.__objects), 3))
+        mi = np.zeros((len(self.__objects), 3))
+        for r, obj in enumerate(self.__objects):
+            ma[r] = self.__objects[obj]["max"]
+            mi[r] = self.__objects[obj]["min"]
+        ma = np.max(ma, axis=0)
+        mi = np.min(mi, axis=0)
+        diag = np.linalg.norm(ma - mi)
+        mean = (ma - mi) / 2 + mi
+        for r, obj in enumerate(self.__objects):
+            v = self.__objects[obj]["geometry"].attributes["position"].array
+            v -= mean
+            v += np.array([0.0, .9, 0.0]) #! to move the obj to the center of window
+
+        scale = self.__s["scale"] * (diag)
+        self._orbit.target = [0.0, 0.0, 0.0]
+        self._cam.lookAt([0.0, 0.0, 0.0])
+        # self._cam.position = [0.0, 0.0, scale]
+        self._cam.position = [0.0, 0.5, scale * 1.3] #! show four complete meshes in the window
+        self._light.position = [0.0, 0.0, scale]
+
+        state = embed.dependency_state(self._renderer)
+
+        # Somehow these entries are missing when the state is exported in python.
+        # Exporting from the GUI works, so we are inserting the missing entries.
+        for k in state:
+            if state[k]["model_name"] == "OrbitControlsModel":
+                state[k]["state"]["maxAzimuthAngle"] = "inf"
+                state[k]["state"]["maxDistance"] = "inf"
+                state[k]["state"]["maxZoom"] = "inf"
+                state[k]["state"]["minAzimuthAngle"] = "-inf"
+
+        tpl = embed.load_requirejs_template
+        if not imports:
+            embed.load_requirejs_template = ""
+
+        s = embed.embed_snippet(self._renderer, state=state, embed_url=EMBED_URL)
+        # s = embed.embed_snippet(self.__w, state=state)
+        embed.load_requirejs_template = tpl
+
+        if html_frame:
+            s = "<html>\n<body>\n" + s + "\n</body>\n</html>"
+
+        # Revert changes
+        for r, obj in enumerate(self.__objects):
+            v = self.__objects[obj]["geometry"].attributes["position"].array
+            v += mean
+        self.__update_view()
+
+        return s
+
+    def save(self, filename=""):
+        if filename == "":
+            uid = str(uuid.uuid4()) + ".html"
+        else:
+            filename = filename.replace(".html", "")
+            uid = filename + '.html'
+        with open(uid, "w") as f:
+            f.write(self.to_html())
+        print("Plot saved to file %s." % uid)

MeshAnything/miche/shapevae-256.yaml

@@ -0,0 +1,46 @@
+model:
+  target: MeshAnything.miche.michelangelo.models.tsal.asl_pl_module.AlignedShapeAsLatentPLModule
+  params:
+    shape_module_cfg:
+      target: MeshAnything.miche.michelangelo.models.tsal.sal_perceiver.AlignedShapeLatentPerceiver
+      params:
+        num_latents: 256
+        embed_dim: 64
+        point_feats: 3   # normal
+        num_freqs: 8
+        include_pi: false
+        heads: 12
+        width: 768
+        num_encoder_layers: 8
+        num_decoder_layers: 16
+        use_ln_post: true
+        init_scale: 0.25
+        qkv_bias: false
+        use_checkpoint: true
+    aligned_module_cfg:
+      target: MeshAnything.miche.michelangelo.models.tsal.clip_asl_module.CLIPAlignedShapeAsLatentModule
+      params:
+        clip_model_version: "./checkpoints/clip/clip-vit-large-patch14"
+
+    loss_cfg:
+      target: MeshAnything.miche.michelangelo.models.tsal.loss.ContrastKLNearFar
+      params:
+        contrast_weight: 0.1
+        near_weight: 0.1
+        kl_weight: 0.001
+
+    optimizer_cfg:
+      optimizer:
+        target: torch.optim.AdamW
+        params:
+          betas: [0.9, 0.99]
+          eps: 1.e-6
+          weight_decay: 1.e-2
+
+      scheduler:
+        target: MeshAnything.miche.michelangelo.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler
+        params:
+          warm_up_steps: 5000
+          f_start: 1.e-6
+          f_min: 1.e-3
+          f_max: 1.0

MeshAnything/models/meshanything.py

@@ -0,0 +1,223 @@
+import torch
+from torch import nn, Tensor
+from transformers import AutoModelForCausalLM, AutoConfig, AutoModel
+from MeshAnything.miche.encode import load_model
+from MeshAnything.models.shape_opt import ShapeOPTConfig
+from einops.layers.torch import Rearrange
+from einops import rearrange, repeat, reduce, pack, unpack
+import torch.nn.functional as F
+
+class NoiseResistantDecoder(nn.Module):
+
+    def __init__(self, args):
+        super().__init__()
+        self.args = args
+        self.pad_id = -1
+        self.num_quantizers = 3
+
+        self.discrete_num = 128
+        self.codebook_size = args.codebook_size
+        self.codebook_dim = args.codebook_dim
+
+        config = AutoConfig.from_pretrained("bert-base-uncased")
+        config.num_hidden_layers = 6
+        self.decoder= AutoModel.from_config(config=config).to_bettertransformer().encoder
+        self.n_embd = self.decoder.config.hidden_size
+
+        self.pos_embedding = nn.Embedding(18000, self.n_embd)
+        self.layernorm = nn.LayerNorm(self.n_embd)
+        self.point_layernorm = nn.LayerNorm(self.n_embd)
+
+        self.cond_length = 257
+        self.cond_dim = 768
+        self.point_pe = nn.Embedding(self.cond_length, self.n_embd)
+        self.cond_proj = nn.Linear(self.cond_dim, self.n_embd)
+        self.cond_head_proj = nn.Linear(self.cond_dim, self.n_embd)
+
+        self.project_down_codebook = nn.Linear(self.codebook_dim * 3, self.n_embd)
+        self.to_coor_logits = nn.Sequential(
+            nn.Linear(self.n_embd, self.discrete_num * 9),
+            Rearrange('... (v c) -> ... v c', v = 9)
+        )
+    def process_point_feature(self, encode_feature):
+        point_feature = torch.zeros(encode_feature.shape[0], self.cond_length, self.n_embd, device=self.cond_head_proj.weight.device, dtype=self.cond_head_proj.weight.dtype)
+        point_feature[:, 0] = self.cond_head_proj(encode_feature[:, 0])
+        point_feature[:, 1:] = self.cond_proj(encode_feature[:, 1:])
+
+        point_feature = self.point_layernorm(point_feature + self.point_pe.weight[None, :point_feature.shape[1]])
+        return point_feature
+
+    def forward(self, input_ids, input_embeds, point_feature = None):
+        input_ids = input_ids.reshape(input_ids.shape[0], -1)
+        point_feature = self.process_point_feature(point_feature)
+
+        face_embeds = rearrange(input_embeds, 'b (nf nv) d -> b nf (nv d)', nv = 3)
+        face_embeds = self.project_down_codebook(face_embeds)
+
+        face_mask = reduce(input_ids != self.pad_id, 'b (nf nv q) -> b nf', 'all', nv = 3, q = self.num_quantizers)
+        face_embeds[~face_mask] = 0
+
+        face_embeds = self.layernorm(face_embeds + self.pos_embedding.weight[None, :face_embeds.shape[1]])
+
+        outputs = self.decoder(
+            hidden_states=torch.concatenate([point_feature, face_embeds], dim=1),
+        )
+        decoded = outputs.last_hidden_state[:, self.cond_length:]       # batch x nfaces x dim
+        decoded = decoded.masked_fill(~face_mask.unsqueeze(-1), 0.)
+
+        # batch x nfaces x 9 -> batch x nfaces x 3 x 3
+        pred_face_logits = self.to_coor_logits(decoded)     # batch x nfaces x 9 x ndiscrete
+        pred_face_coords = rearrange(pred_face_logits.argmax(dim = -1), '... (v c) -> ... v c', v = 3)
+
+        continuous_coors = undiscretize(
+            pred_face_coords,
+            num_discrete = self.discrete_num,
+            low = -0.5,
+            high = 0.5
+        )
+        continuous_coors = continuous_coors.masked_fill(~rearrange(face_mask, 'b nf -> b nf 1 1'), float('nan'))
+
+        return continuous_coors
+
+class MeshAnything(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.args = args
+        self.point_encoder = load_model(ckpt_path=None)
+        self.tokenizer = NoiseResistantDecoder(args)
+
+        self.num_quantizers = 3
+        self.face_per_token = self.num_quantizers * 3
+        self.cond_length = 257
+        self.cond_dim = 768
+        self.max_length = args.n_max_triangles * self.face_per_token + 2 + self.cond_length
+
+        self.config = ShapeOPTConfig.from_pretrained(
+            args.llm,
+            n_positions=18259,
+            max_position_embeddings=18259,
+            vocab_size=self.tokenizer.codebook_size + 3,
+            _attn_implementation="flash_attention_2"
+        )
+        self.bos_token_id = 0
+        self.eos_token_id = 1
+        self.pad_token_id = 2
+        self.config.bos_token_id = self.bos_token_id
+        self.config.eos_token_id = self.eos_token_id
+        self.config.pad_token_id = self.pad_token_id
+        self.config.quantize_codebook_dim = self.tokenizer.codebook_dim
+        self.config.face_per_token = self.face_per_token
+        self.config._attn_implementation="flash_attention_2"
+        self.config.cond_length = self.cond_length
+        if self.config.word_embed_proj_dim != self.config.hidden_size:
+            self.config.word_embed_proj_dim = self.config.hidden_size
+        self.transformer = AutoModelForCausalLM.from_config(
+            config=self.config, use_flash_attention_2 = True
+        )
+        self.transformer.to_bettertransformer()
+        self.transformer.model.decoder.quantize_codebooks = nn.Parameter(torch.zeros(1, self.tokenizer.codebook_size, self.tokenizer.codebook_dim))
+
+        self.cond_head_proj = nn.Linear(self.cond_dim, self.config.word_embed_proj_dim)
+        self.cond_proj = nn.Linear(self.cond_dim * 2, self.config.word_embed_proj_dim)
+
+        self.eval()
+
+    def process_point_feature(self, point_feature):
+        encode_feature = torch.zeros(point_feature.shape[0], self.cond_length, self.config.word_embed_proj_dim,
+                                    device=self.cond_head_proj.weight.device, dtype=self.cond_head_proj.weight.dtype)
+        encode_feature[:, 0] = self.cond_head_proj(point_feature[:, 0])
+        shape_latents = self.point_encoder.to_shape_latents(point_feature[:, 1:])
+        encode_feature[:, 1:] = self.cond_proj(torch.cat([point_feature[:, 1:], shape_latents], dim=-1))
+
+        return encode_feature
+
+    @torch.no_grad()
+    def forward(self, pc_normal, sampling=False) -> dict:
+        batch_size = pc_normal.shape[0]
+        point_feature = self.point_encoder.encode_latents(pc_normal)
+        processed_point_feature = self.process_point_feature(point_feature)
+
+        generate_length = self.max_length - self.cond_length
+        net_device = next(self.parameters()).device
+        outputs = torch.ones(batch_size, generate_length).long().to(net_device) * self.eos_token_id
+        if not sampling:
+            results = self.transformer.generate(
+                inputs_embeds=processed_point_feature,
+                max_new_tokens=generate_length,  # all faces plus two
+                num_beams=1,
+                bos_token_id=self.bos_token_id,
+                eos_token_id=self.eos_token_id,
+                pad_token_id=self.pad_token_id,
+            )
+        else:
+            results = self.transformer.generate(
+                inputs_embeds = processed_point_feature,
+                max_new_tokens=generate_length, # all faces plus two
+                do_sample=True,
+                top_k=50,
+                top_p=0.95,
+                bos_token_id = self.bos_token_id,
+                eos_token_id = self.eos_token_id,
+                pad_token_id = self.pad_token_id,
+            )
+        assert results.shape[1] <= generate_length # B x ID  bos is not included since it's predicted
+        outputs[:, :results.shape[1]] = results
+        # batch x ntokens ====> batch x ntokens x D
+        outputs = outputs[:, 1: -1]
+
+        outputs[outputs == self.bos_token_id] = self.tokenizer.pad_id
+        outputs[outputs == self.eos_token_id] = self.tokenizer.pad_id
+        outputs[outputs == self.pad_token_id] = self.tokenizer.pad_id
+
+        outputs[outputs != self.tokenizer.pad_id] -= 3
+        code_embed = self.get_codes(outputs)
+        decoder_output = self.tokenizer(outputs, code_embed, point_feature=point_feature)
+
+        return decoder_output
+
+    def get_codes(self, indices):
+        indices = indices.reshape(indices.shape[0], -1)
+
+        indices = rearrange(indices, 'b (n q) -> b n q', q=self.num_quantizers)
+
+        batch, quantize_dim = indices.shape[0], indices.shape[-1]
+        # may also receive indices in the shape of 'b h w q' (accept_image_fmap)
+
+        indices, ps = pack([indices], 'b * q')
+
+        # because of quantize dropout, one can pass in indices that are coarse
+        # and the network should be able to reconstruct
+
+        if quantize_dim < self.num_quantizers:
+            indices = F.pad(indices, (0, self.num_quantizers - quantize_dim), value = -1)
+
+        # take care of quantizer dropout
+
+        mask = indices == -1.
+        indices = indices.masked_fill(mask, 0) # have it fetch a dummy code to be masked out later
+
+        # dummy implementation for shared codebook
+        all_codes = self.transformer.model.decoder.quantize_codebooks[0][indices]
+        all_codes = all_codes.permute(2, 0, 1, 3)
+
+        # mask out any codes that were dropout-ed
+
+        all_codes = all_codes.masked_fill(rearrange(mask, 'b n q -> q b n 1'), 0.)
+
+        # if (accept_image_fmap = True) then return shape (quantize, batch, height, width, dimension)
+
+        codes, = unpack(all_codes, ps, 'q b * d')
+
+        codes_summed = reduce(codes, 'q ... -> ...', 'sum')
+        return codes_summed
+
+def undiscretize(
+    t,
+    low,
+    high,
+    num_discrete
+) -> Tensor:
+    t = t.float()
+
+    t /= num_discrete
+    return t * (high - low) + low

MeshAnything/models/shape_opt.py

@@ -0,0 +1,464 @@
+from transformers import AutoModelForCausalLM, AutoConfig, OPTConfig
+from transformers.models.opt.modeling_opt import OPTForCausalLM, OPTModel, OPTDecoder, OPTLearnedPositionalEmbedding, OPTDecoderLayer
+from typing import List, Optional, Tuple, Union
+from einops import repeat
+from transformers.modeling_outputs import (
+    CausalLMOutputWithPast,
+)
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from transformers.utils import replace_return_docstrings, logging
+from transformers.modeling_outputs import BaseModelOutputWithPast
+# from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
+
+class ShapeOPTConfig(OPTConfig):
+    model_type = "shape_opt"
+
+class ShapeOPT(OPTForCausalLM):
+    config_class = ShapeOPTConfig
+    def __init__(self, config: ShapeOPTConfig):
+        super(OPTForCausalLM, self).__init__(config)
+        self.model = ShapeOPTModel(config)
+
+        self.lm_head = nn.Linear(config.word_embed_proj_dim, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def tie_weights(self):
+        """
+        Tie the weights between the input embeddings and the output embeddings.
+
+        If the `torchscript` flag is set in the configuration, can't handle parameter sharing so we are cloning the
+        weights instead.
+        """
+        if getattr(self.config, "is_encoder_decoder", False) and getattr(self.config, "tie_encoder_decoder", False):
+            if hasattr(self, self.base_model_prefix):
+                self = getattr(self, self.base_model_prefix)
+            self._tie_encoder_decoder_weights(self.encoder, self.decoder, self.base_model_prefix)
+
+        for module in self.modules():
+            if hasattr(module, "_tie_weights"):
+                module._tie_weights()
+
+
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class="OPTConfig")
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        face_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+                [`PreTrainedTokenizer.__call__`] for details.
+
+                [What are input IDs?](../glossary#input-ids)
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            head_mask (`torch.Tensor` of shape `(num_hidden_layers, num_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+                shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. The two additional
+                tensors are only required when the model is used as a decoder in a Sequence to Sequence model.
+
+                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
+                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
+                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, OPTForCausalLM
+
+        >>> model = OPTForCausalLM.from_pretrained("facebook/opt-350m")
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious. I'm just a little bit of a weirdo."
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model.decoder(
+            input_ids=input_ids,
+            face_ids = face_ids,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        logits = self.lm_head(outputs[0]).contiguous()
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+class ShapeOPTModel(OPTModel):
+    config_class = ShapeOPTConfig
+    def __init__(self, config: ShapeOPTConfig):
+        super(OPTModel,self).__init__(config)
+        self.decoder = ShapeOPTDecoder(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+class ShapeOPTDecoder(OPTDecoder):
+    config_class = ShapeOPTConfig
+    def __init__(self, config: ShapeOPTConfig):
+        super(OPTDecoder,self).__init__(config)
+        self.config = config
+        self.dropout = config.dropout
+        self.layerdrop = config.layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_position_embeddings
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.word_embed_proj_dim, self.padding_idx) # not used
+        self.hidden_size = config.hidden_size
+        self.word_embed_proj_dim = config.word_embed_proj_dim
+        self.extra_embeds = nn.Embedding(3, config.word_embed_proj_dim) #padding_idx=self.padding_idx)
+        self.input_layer = nn.Linear(config.quantize_codebook_dim, config.word_embed_proj_dim)
+
+        self.embed_positions = OPTLearnedPositionalEmbedding(config.max_position_embeddings, config.hidden_size)
+        self.token_embed_positions =  OPTFacePositionalEmbedding(config.face_per_token + 3, config.word_embed_proj_dim) #padding_idx=self.padding_idx)
+        self.face_per_token = config.face_per_token
+        self.cond_length = config.cond_length
+        self.cond_embed = nn.Embedding(2, config.word_embed_proj_dim)
+
+        if config.word_embed_proj_dim != config.hidden_size:
+            self.project_out = nn.Linear(config.hidden_size, config.word_embed_proj_dim, bias=False)
+        else:
+            self.project_out = None
+
+        if config.word_embed_proj_dim != config.hidden_size:
+            self.project_in = nn.Linear(config.word_embed_proj_dim, config.hidden_size, bias=False)
+        else:
+            self.project_in = None
+        # Note that the only purpose of `config._remove_final_layer_norm` is to keep backward compatibility
+        # with checkpoints that have been fine-tuned before transformers v4.20.1
+        # see https://github.com/facebookresearch/metaseq/pull/164
+        if config.do_layer_norm_before and not config._remove_final_layer_norm:
+            self.final_layer_norm = nn.LayerNorm(
+                config.hidden_size, elementwise_affine=config.layer_norm_elementwise_affine
+            )
+        else:
+            self.final_layer_norm = None
+
+        self.layers = nn.ModuleList([OPTDecoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def embed_with_vae(self, input_ids):
+        inputs_embeds = repeat(torch.zeros(input_ids.shape, device=input_ids.device), 'b n -> b n d',
+                               d=self.word_embed_proj_dim).clone().detach()
+        idx_in_extra = torch.isin(input_ids, torch.LongTensor([0, 1, 2]).to(input_ids.device))
+        inputs_embeds[idx_in_extra] += self.extra_embeds(input_ids[idx_in_extra])
+        self.quantize_codebooks = self.quantize_codebooks.to(input_ids.device)
+        inputs_embeds[~idx_in_extra] += self.input_layer(self.quantize_codebooks[0][input_ids[~idx_in_extra] - 3])
+
+        return inputs_embeds
+
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        face_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        r"""
+        Args:
+            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+                [`PreTrainedTokenizer.__call__`] for details.
+
+                [What are input IDs?](../glossary#input-ids)
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            head_mask (`torch.Tensor` of shape `(num_hidden_layers, num_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+
+                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
+                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
+                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        # OPT Decoder
+        # print("used my Trans")
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        # Transformer Decoder
+        if input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            inputs_embeds = self.embed_with_vae(input_ids) # nothing to do with position
+
+            face_embeds = self.token_embed_positions(attention_mask[:, self.cond_length:], face_ids, input_ids,
+                                                     self.face_per_token)
+            inputs_embeds += face_embeds
+            cond_embed_query = torch.ones((inputs_embeds.shape[0], inputs_embeds.shape[1]), device=inputs_embeds.device,
+                                            dtype=inputs_embeds.dtype).long()
+            inputs_embeds = inputs_embeds + self.cond_embed(cond_embed_query)
+
+        elif inputs_embeds is not None:
+            # assert self.cond and not self.training
+
+            total_length = inputs_embeds.shape[1] # B x length x embeding
+            cond_embed_query = torch.zeros((inputs_embeds.shape[0], total_length), device=inputs_embeds.device,
+                                            dtype=inputs_embeds.dtype).long()
+            inputs_embeds = inputs_embeds + self.cond_embed(cond_embed_query)
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        batch_size, seq_length = inputs_embeds.shape[:2] # seq_length not used since mask_seq_length is not used
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+        # required mask seq length can be calculated via length of past
+        mask_seq_length = past_key_values_length + seq_length # not used since attention mask is input
+
+        # embed positions
+        if self._use_flash_attention_2:
+            # 2d mask is passed through the layers
+            assert attention_mask is not None
+            causal_attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+            attention_mask = (
+                torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)
+                if attention_mask is None
+                else attention_mask
+            )
+        else:
+            raise ValueError("Only flash_attention_2 is supported in MeshAnything")
+
+        pos_embeds = self.embed_positions(attention_mask, past_key_values_length)
+
+        if self.project_in is not None:
+            inputs_embeds = self.project_in(inputs_embeds)
+
+        hidden_states = inputs_embeds + pos_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        # check if head_mask has a correct number of layers specified if desired
+        for attn_mask, mask_name in zip([head_mask], ["head_mask"]):
+            if attn_mask is not None:
+                if attn_mask.size()[0] != (len(self.layers)):
+                    raise ValueError(
+                        f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
+                        f" {head_mask.size()[0]}."
+                    )
+
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.training:
+                dropout_probability = torch.rand([])
+                if dropout_probability < self.layerdrop:
+                    continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_attention_mask,
+                    head_mask[idx] if head_mask is not None else None,
+                    None,
+                    output_attentions,
+                    use_cache,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_attention_mask,
+                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        if self.final_layer_norm is not None:
+            hidden_states = self.final_layer_norm(hidden_states)
+
+        if self.project_out is not None:
+            hidden_states = self.project_out(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+class OPTFacePositionalEmbedding(nn.Embedding):
+    """
+    This module learns positional embeddings up to a fixed maximum size.
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int):
+        super().__init__(num_embeddings, embedding_dim)
+
+    def forward(self, attention_mask=None, face_ids = None, input_ids = None, face_per_token = None):
+        """`input_ids_shape` is expected to be [bsz x seqlen]."""
+        if face_ids is not None:
+            return super().forward(face_ids)
+
+        assert input_ids.shape[1] == 1
+        idx_in_extra = torch.isin(input_ids, torch.LongTensor([0, 1, 2]).to(input_ids.device))
+        cur_ids = input_ids.clone().detach()
+
+        cur_index = (attention_mask.sum(dim=1, keepdim=True) - 2) % face_per_token + 3
+        cur_ids[~idx_in_extra]=cur_index[~idx_in_extra]
+
+        return super().forward(cur_ids)
+
+
+AutoConfig.register("shape_opt", ShapeOPTConfig)
+AutoModelForCausalLM.register(ShapeOPTConfig, ShapeOPT)
+