Rewrite init
commit
bf8cedc9dd
@ -0,0 +1,9 @@
|
||||
__pycache__
|
||||
/data
|
||||
/logs
|
||||
/ckpts
|
||||
/.cache
|
||||
/config
|
||||
/*.egg-info
|
||||
/vall_e/version.py
|
||||
/build
|
||||
@ -0,0 +1,235 @@
|
||||
GNU AFFERO GENERAL PUBLIC LICENSE
|
||||
Version 3, 19 November 2007
|
||||
|
||||
Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
|
||||
|
||||
Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it is not allowed.
|
||||
|
||||
Preamble
|
||||
|
||||
The GNU Affero General Public License is a free, copyleft license for software and other kinds of works, specifically designed to ensure cooperation with the community in the case of network server software.
|
||||
|
||||
The licenses for most software and other practical works are designed to take away your freedom to share and change the works. By contrast, our General Public Licenses are 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.
|
||||
|
||||
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.
|
||||
|
||||
Developers that use our General Public Licenses protect your rights with two steps: (1) assert copyright on the software, and (2) offer you this License which gives you legal permission to copy, distribute and/or modify the software.
|
||||
|
||||
A secondary benefit of defending all users' freedom is that improvements made in alternate versions of the program, if they receive widespread use, become available for other developers to incorporate. Many developers of free software are heartened and encouraged by the resulting cooperation. However, in the case of software used on network servers, this result may fail to come about. The GNU General Public License permits making a modified version and letting the public access it on a server without ever releasing its source code to the public.
|
||||
|
||||
The GNU Affero General Public License is designed specifically to ensure that, in such cases, the modified source code becomes available to the community. It requires the operator of a network server to provide the source code of the modified version running there to the users of that server. Therefore, public use of a modified version, on a publicly accessible server, gives the public access to the source code of the modified version.
|
||||
|
||||
An older license, called the Affero General Public License and published by Affero, was designed to accomplish similar goals. This is a different license, not a version of the Affero GPL, but Affero has released a new version of the Affero GPL which permits relicensing under this license.
|
||||
|
||||
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 Affero 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. Remote Network Interaction; Use with the GNU General Public License.
|
||||
|
||||
Notwithstanding any other provision of this License, if you modify the Program, your modified version must prominently offer all users interacting with it remotely through a computer network (if your version supports such interaction) an opportunity to receive the Corresponding Source of your version by providing access to the Corresponding Source from a network server at no charge, through some standard or customary means of facilitating copying of software. This Corresponding Source shall include the Corresponding Source for any work covered by version 3 of the GNU General Public License that is incorporated pursuant to the following paragraph.
|
||||
|
||||
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 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 work with which it is combined will remain governed by version 3 of the GNU General Public License.
|
||||
|
||||
14. Revised Versions of this License.
|
||||
|
||||
The Free Software Foundation may publish revised and/or new versions of the GNU Affero 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 Affero 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 Affero 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 Affero 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 Affero 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 Affero General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU Affero General Public License along with this program. If not, see <http://www.gnu.org/licenses/>.
|
||||
|
||||
Also add information on how to contact you by electronic and paper mail.
|
||||
|
||||
If your software can interact with users remotely through a computer network, you should also make sure that it provides a way for users to get its source. For example, if your program is a web application, its interface could display a "Source" link that leads users to an archive of the code. There are many ways you could offer source, and different solutions will be better for different programs; see section 13 for the specific requirements.
|
||||
|
||||
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 AGPL, see <http://www.gnu.org/licenses/>.
|
||||
@ -0,0 +1,150 @@
|
||||
<p align="center">
|
||||
<img src="./vall-e.png" width="500px"></img>
|
||||
</p>
|
||||
|
||||
# VALL'Ecker
|
||||
|
||||
An unofficial PyTorch implementation of [VALL-E](https://valle-demo.github.io/), based on the [EnCodec](https://github.com/facebookresearch/encodec) tokenizer.
|
||||
|
||||
> **Note** this is highly experimental. While I've seem to have audited and tighened down as much as I can, I'm still trying to produce a decent model out of it. You're free to train your own model if you happen to have the massive compute for it, but it's quite the beast to properly feed.
|
||||
|
||||
> **Note** This README won't get much love until I truly nail out a quasi-decent model.
|
||||
|
||||
* **Note** Distributed training seems broken? I'm not really sure how to test it, as my two 6800XTs have been redistributed for now, and the last time I tried using them for this, things weren't good.
|
||||
|
||||
* **Note** You can follow along with my pseudo-blog in an issue [here](https://git.ecker.tech/mrq/ai-voice-cloning/issues/152). I currently have a dataset clocking in at 3400+ trimmed hours.
|
||||
|
||||
### Requirements
|
||||
|
||||
Since the trainer is based on [DeepSpeed](https://github.com/microsoft/DeepSpeed#requirements), you will need to have a GPU that DeepSpeed has developed and tested against, as well as a CUDA or ROCm compiler pre-installed to install this package.
|
||||
|
||||
### Install
|
||||
|
||||
```
|
||||
pip install git+https://git.ecker.tech/mrq/vall-e
|
||||
```
|
||||
|
||||
Or you may clone by:
|
||||
|
||||
```
|
||||
git clone --recurse-submodules https://git.ecker.tech/mrq/vall-e.git
|
||||
```
|
||||
|
||||
Note that the code is only tested under `Python 3.10.9`.
|
||||
|
||||
### Train
|
||||
|
||||
Training is very dependent on:
|
||||
* the quality of your dataset.
|
||||
* how much data you have.
|
||||
* the bandwidth you quantized your audio to.
|
||||
|
||||
#### Quick Preparations
|
||||
|
||||
##### Prepared Dataset
|
||||
|
||||
Under `./scripts/download_libritts-small.sh` is a script that will quickly set up an already prepared dataset to train. This leverages a repo I've published to HuggingFace that contains everything processsed, straight from the below method.
|
||||
|
||||
##### Prepare It Yourself
|
||||
|
||||
Under `./scripts/prepare_libri.sh` is a small script to quickly set up a dataset based on LibriSpeech-Finetuning. It'll handle everything from downloading, to extracting, to preparing, to quantizing and phonemizing.
|
||||
|
||||
Afterwards, simply use `./config/libri/config.yaml` as your target YAML.
|
||||
|
||||
However, you'll only train against a small subset of the data with the default settings, due to the maximum phoneme length configured. Increasing this will not only drastically increase VRAM usage, but also reduce iteration rates. It's recommended to further process your files by slicing them down (for example, through [mrq/ai-voice-cloning](https://git.ecker.tech/mrq/ai-voice-cloning)).
|
||||
|
||||
#### Leverage Your Own
|
||||
|
||||
1. Put your data into a folder, e.g. `./data/custom`. Audio files should be named with the suffix `.wav` and text files with `.normalized.txt`.
|
||||
|
||||
2. Quantize the data:
|
||||
|
||||
```
|
||||
python -m vall_e.emb.qnt ./data/custom
|
||||
```
|
||||
|
||||
3. Generate phonemes based on the text:
|
||||
|
||||
```
|
||||
python -m vall_e.emb.g2p data/custom
|
||||
```
|
||||
|
||||
4. Customize your configuration by creating `./config/custom.yml`. Refer to the example configs in `./config/libri-quarter.yaml` and `./vall_e/config.py` for details. If you want to choose between different model presets, check `./vall_e/models/__init__.py`.
|
||||
|
||||
5. Train the AR and NAR models using the following scripts:
|
||||
|
||||
```
|
||||
python -m vall_e.train yaml=config/custom/config.yml
|
||||
```
|
||||
|
||||
You may quit your training any time by just typing `quit` in your CLI. The latest checkpoint will be automatically saved.
|
||||
|
||||
### Dataset Formats
|
||||
|
||||
Two dataset formats are supported:
|
||||
* the standard way:
|
||||
- data is stored under `${speaker}/${id}.phn.txt` and `${speaker}/${id}.qnt.pt`
|
||||
* using an HDF5 dataset:
|
||||
- you can convert from the standard way with the following command: `python3 -m vall_e.data yaml="./path/to/your/config.yaml"`
|
||||
- this will shove everything into a single HDF5 file and store some metadata alongside (for now, the symbol map generated, and text/audio lengths)
|
||||
- be sure to also define `use_hdf5` in your config YAML.
|
||||
|
||||
### Training Tip
|
||||
|
||||
Training a VALL-E model is very, very meticulous. I've fiddled with a lot of """clever""" tricks, but it seems the best is just to pick the highest LR you can get (this heavily depends on your batch size, but hyperparameters of bs=64 * ga=16 on the quarter sized model has an LR of 1.0e-3 stable, while the full size model with hyperparameters of bs=16 * ga=64 needed smaller). Like typical training, it entirely depends on your tradeoff betweeen stability and time.
|
||||
|
||||
### Export
|
||||
|
||||
Both trained models *can* be exported, but is only required if loading them on systems without DeepSpeed for inferencing (Windows systems). To export the models, run:
|
||||
|
||||
```
|
||||
python -m vall_e.export ./models/ yaml=./config/custom.yml
|
||||
```
|
||||
|
||||
This will export the latest checkpoint.
|
||||
|
||||
### Synthesis
|
||||
|
||||
To synthesize speech, invoke either (if exported the models):
|
||||
|
||||
```
|
||||
python -m vall_e <text> <ref_path> <out_path> --ar-ckpt ./models/ar.pt --nar-ckpt ./models/nar.pt
|
||||
```
|
||||
|
||||
or:
|
||||
|
||||
```
|
||||
python -m vall_e <text> <ref_path> <out_path> yaml=<yaml_path>
|
||||
```
|
||||
|
||||
Some additional flags you can pass are:
|
||||
* `--max-ar-steps`: maximum steps for inferencing through the AR model. Each second is 75 steps.
|
||||
* `--ar-temp`: sampling temperature to use for the AR pass.
|
||||
* `--nar-temp`: sampling temperature to use for the NAR pass.
|
||||
* `--device`: device to use (default: `cuda`, examples: `cuda:0`, `cuda:1`, `cpu`)
|
||||
|
||||
## Notice
|
||||
|
||||
- [EnCodec](https://github.com/facebookresearch/encodec) is licensed under CC-BY-NC 4.0. If you use the code to generate audio quantization or perform decoding, it is important to adhere to the terms of their license.
|
||||
|
||||
Unless otherwise credited/noted, this repository is [licensed](LICENSE) under AGPLv3.
|
||||
|
||||
## Citations
|
||||
|
||||
```bibtex
|
||||
@article{wang2023neural,
|
||||
title={Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers},
|
||||
author={Wang, Chengyi and Chen, Sanyuan and Wu, Yu and Zhang, Ziqiang and Zhou, Long and Liu, Shujie and Chen, Zhuo and Liu, Yanqing and Wang, Huaming and Li, Jinyu and others},
|
||||
journal={arXiv preprint arXiv:2301.02111},
|
||||
year={2023}
|
||||
}
|
||||
```
|
||||
|
||||
```bibtex
|
||||
@article{defossez2022highfi,
|
||||
title={High Fidelity Neural Audio Compression},
|
||||
author={Défossez, Alexandre and Copet, Jade and Synnaeve, Gabriel and Adi, Yossi},
|
||||
journal={arXiv preprint arXiv:2210.13438},
|
||||
year={2022}
|
||||
}
|
||||
```
|
||||
@ -0,0 +1,99 @@
|
||||
dataset:
|
||||
training: [
|
||||
|
||||
]
|
||||
|
||||
validation: [
|
||||
|
||||
]
|
||||
|
||||
speaker_name_getter: "lambda p: f'{p.parts[-3]}_{p.parts[-2]}'"
|
||||
|
||||
use_hdf5: True
|
||||
validate: True
|
||||
|
||||
workers: 8
|
||||
cache: True
|
||||
|
||||
phones_range: [4, 192]
|
||||
duration_range: [1.0, 10.0]
|
||||
|
||||
random_utterance: 1.0
|
||||
max_prompts: 3
|
||||
prompt_duration: 3.0
|
||||
|
||||
models:
|
||||
_models:
|
||||
- name: "ar"
|
||||
size: "full"
|
||||
resp_levels: 1
|
||||
use_retnet: True
|
||||
full_retnet: True
|
||||
use_torchscale: True
|
||||
|
||||
- name: "nar"
|
||||
size: "full"
|
||||
resp_levels: 1
|
||||
use_retnet: True
|
||||
full_retnet: True
|
||||
use_torchscale: True
|
||||
|
||||
prom_levels: 2
|
||||
|
||||
hyperparameters:
|
||||
batch_size: 16
|
||||
gradient_accumulation_steps: 8
|
||||
gradient_clipping: 100
|
||||
|
||||
optimizer: Adamw
|
||||
learning_rate: 1.0e-4
|
||||
|
||||
scheduler_type: ""
|
||||
#scheduler_type: OneCycle
|
||||
#scheduler_params:
|
||||
# cycle_first_step_size: 10_000
|
||||
# cycle_first_stair_count: 10_000
|
||||
|
||||
# cycle_second_step_size: 15_000
|
||||
# cycle_second_stair_count: 15_000
|
||||
|
||||
# decay_step_size: 5_000
|
||||
|
||||
# cycle_min_lr: 2.5e-4 # 1.0e-5
|
||||
# cycle_max_lr: 2.5e-4 # 1.0e-4
|
||||
# decay_lr_rate: 0.0
|
||||
|
||||
# cycle_min_mom: 0.90
|
||||
# cycle_max_mom: 0.99
|
||||
# decay_mom_rate: 0.0
|
||||
|
||||
evaluation:
|
||||
batch_size: 64
|
||||
frequency: 250
|
||||
size: 64
|
||||
|
||||
steps: 500
|
||||
temperature: 1.0
|
||||
|
||||
trainer:
|
||||
iterations: 100_000
|
||||
|
||||
save_tag: step
|
||||
save_on_oom: True
|
||||
save_on_quit: True
|
||||
save_frequency: 100
|
||||
|
||||
aggressive_optimizations: False
|
||||
|
||||
#load_tag: "9500"
|
||||
#load_state_dict: True
|
||||
#load_states: False
|
||||
#strict_loading: False
|
||||
#restart_step_count: True
|
||||
|
||||
gc_mode: None # "global_step"
|
||||
|
||||
weight_dtype: bfloat16
|
||||
|
||||
zero_optimization_level: 2
|
||||
use_compression_training: True
|
||||
@ -0,0 +1,9 @@
|
||||
#!/bin/bash
|
||||
|
||||
# do not invoke directly in scripts
|
||||
if [[ ${PWD##*/} == 'scripts' ]]; then
|
||||
cd ..
|
||||
fi
|
||||
|
||||
# download training data
|
||||
git clone https://huggingface.co/datasets/ecker/libritts-small ./data/libritts-small
|
||||
@ -0,0 +1,106 @@
|
||||
#!/usr/bin/env python3
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import re
|
||||
from pathlib import Path
|
||||
|
||||
import matplotlib.pyplot as plt
|
||||
import pandas as pd
|
||||
|
||||
|
||||
def plot(paths, args):
|
||||
dfs = []
|
||||
|
||||
for path in paths:
|
||||
with open(path, "r") as f:
|
||||
text = f.read()
|
||||
|
||||
rows = []
|
||||
|
||||
pattern = r"(\{.+?\})"
|
||||
|
||||
for row in re.findall(pattern, text, re.DOTALL):
|
||||
try:
|
||||
row = json.loads(row)
|
||||
except Exception as e:
|
||||
continue
|
||||
|
||||
if "global_step" in row:
|
||||
rows.append(row)
|
||||
|
||||
df = pd.DataFrame(rows)
|
||||
|
||||
if "name" in df:
|
||||
df["name"] = df["name"].fillna("train")
|
||||
else:
|
||||
df["name"] = "train"
|
||||
|
||||
df["group"] = str(path.parents[args.group_level])
|
||||
df["group"] = df["group"] + "/" + df["name"]
|
||||
|
||||
dfs.append(df)
|
||||
|
||||
df = pd.concat(dfs)
|
||||
|
||||
if args.max_y is not None:
|
||||
df = df[df["global_step"] < args.max_x]
|
||||
|
||||
for gtag, gdf in sorted(
|
||||
df.groupby("group"),
|
||||
key=lambda p: (p[0].split("/")[-1], p[0]),
|
||||
):
|
||||
for y in args.ys:
|
||||
gdf = gdf.sort_values("global_step")
|
||||
|
||||
if gdf[y].isna().all():
|
||||
continue
|
||||
|
||||
if args.max_y is not None:
|
||||
gdf = gdf[gdf[y] < args.max_y]
|
||||
|
||||
gdf[y] = gdf[y].ewm(10).mean()
|
||||
|
||||
gdf.plot(
|
||||
x="global_step",
|
||||
y=y,
|
||||
label=f"{gtag}/{y}",
|
||||
ax=plt.gca(),
|
||||
marker="x" if len(gdf) < 100 else None,
|
||||
alpha=0.7,
|
||||
)
|
||||
|
||||
plt.gca().legend(
|
||||
loc="center left",
|
||||
fancybox=True,
|
||||
shadow=True,
|
||||
bbox_to_anchor=(1.04, 0.5),
|
||||
)
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("ys", nargs="+")
|
||||
parser.add_argument("--log-dir", default="logs", type=Path)
|
||||
parser.add_argument("--out-dir", default="logs", type=Path)
|
||||
parser.add_argument("--filename", default="log.txt")
|
||||
parser.add_argument("--max-x", type=float, default=float("inf"))
|
||||
parser.add_argument("--max-y", type=float, default=float("inf"))
|
||||
parser.add_argument("--group-level", default=1)
|
||||
parser.add_argument("--filter", default=None)
|
||||
args = parser.parse_args()
|
||||
|
||||
paths = args.log_dir.rglob(f"**/{args.filename}")
|
||||
|
||||
if args.filter:
|
||||
paths = filter(lambda p: re.match(".*" + args.filter + ".*", str(p)), paths)
|
||||
|
||||
plot(paths, args)
|
||||
|
||||
name = "-".join(args.ys)
|
||||
out_path = (args.out_dir / name).with_suffix(".png")
|
||||
plt.savefig(out_path, bbox_inches="tight")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@ -0,0 +1,72 @@
|
||||
import os
|
||||
import json
|
||||
|
||||
for f in os.listdir(f'./data/librispeech_finetuning/1h/'):
|
||||
for j in os.listdir(f'./data/librispeech_finetuning/1h/{f}/clean'):
|
||||
for z in os.listdir(f'./data/librispeech_finetuning/1h/{f}/clean/{j}'):
|
||||
for i in os.listdir(f'./data/librispeech_finetuning/1h/{f}/clean/{j}/{z}'):
|
||||
os.rename(f'./data/librispeech_finetuning/1h/{f}/clean/{j}/{z}/{i}', f'./data/librilight-tts/{i}')
|
||||
|
||||
for j in os.listdir('./data/librispeech_finetuning/9h/clean'):
|
||||
for z in os.listdir(f'./data/librispeech_finetuning/9h/clean/{j}'):
|
||||
for i in os.listdir(f'./data/librispeech_finetuning/9h/clean/{j}/{z}'):
|
||||
os.rename(f'./data/librispeech_finetuning/9h/clean/{j}/{z}/{i}', f'./data/librilight-tts/{i}')
|
||||
|
||||
lst = []
|
||||
for i in os.listdir('./data/librilight-tts/'):
|
||||
try:
|
||||
if 'trans' not in i:
|
||||
continue
|
||||
with open(f'./data/librilight-tts/{i}') as f:
|
||||
for row in f:
|
||||
z = row.split('-')
|
||||
name = z[0]+'-'+z[1]+ '-' + z[2].split(' ')[0]
|
||||
text = " ".join(z[2].split(' ')[1:])
|
||||
lst.append([name, text])
|
||||
except Exception as e:
|
||||
pass
|
||||
|
||||
for i in lst:
|
||||
try:
|
||||
with open(f'./data/librilight-tts/{i[0]}.txt', 'x') as file:
|
||||
file.write(i[1])
|
||||
except:
|
||||
with open(f'./data/librilight-tts/{i[0]}.txt', 'w+') as file:
|
||||
file.write(i[1])
|
||||
|
||||
phoneme_map = {}
|
||||
phoneme_transcript = {}
|
||||
|
||||
with open('./data/librispeech_finetuning/phones/phones_mapping.json', 'r') as f:
|
||||
phoneme_map_rev = json.load(f)
|
||||
for k, v in phoneme_map_rev.items():
|
||||
phoneme_map[f'{v}'] = k
|
||||
|
||||
with open('./data/librispeech_finetuning/phones/10h_phones.txt', 'r') as f:
|
||||
lines = f.readlines()
|
||||
for line in lines:
|
||||
split = line.strip().split(" ")
|
||||
key = split[0]
|
||||
tokens = split[1:]
|
||||
|
||||
phonemes = []
|
||||
for token in tokens:
|
||||
phoneme = phoneme_map[f'{token}']
|
||||
phonemes.append( phoneme )
|
||||
|
||||
phoneme_transcript[key] = " ".join(phonemes)
|
||||
|
||||
for filename in sorted(os.listdir('./data/librilight-tts')):
|
||||
split = filename.split('.')
|
||||
|
||||
key = split[0]
|
||||
extension = split[1] # covers double duty of culling .normalized.txt and .phn.txt
|
||||
|
||||
if extension != 'txt':
|
||||
continue
|
||||
|
||||
os.rename(f'./data/librilight-tts/{filename}', f'./data/librilight-tts/{key}.normalized.txt')
|
||||
|
||||
if key in phoneme_transcript:
|
||||
with open(f'./data/librilight-tts/{key}.phn.txt', 'w', encoding='utf-8') as f:
|
||||
f.write(phoneme_transcript[key])
|
||||
@ -0,0 +1,27 @@
|
||||
#!/bin/bash
|
||||
|
||||
# do not invoke directly in scripts
|
||||
if [[ ${PWD##*/} == 'scripts' ]]; then
|
||||
cd ..
|
||||
fi
|
||||
|
||||
# download training data
|
||||
cd data
|
||||
mkdir librilight-tts
|
||||
if [ ! -e ./librispeech_finetuning.tgz ]; then
|
||||
wget https://dl.fbaipublicfiles.com/librilight/data/librispeech_finetuning.tgz
|
||||
fi
|
||||
tar -xzf librispeech_finetuning.tgz
|
||||
cd ..
|
||||
|
||||
# clean it up
|
||||
python3 ./scripts/prepare_libri.py
|
||||
|
||||
# convert to wav
|
||||
pip3 install AudioConverter
|
||||
audioconvert convert ./data/librilight-tts/ ./data/librilight-tts --output-format .wav
|
||||
|
||||
# process data
|
||||
ulimit -Sn `ulimit -Hn` # ROCm is a bitch
|
||||
python3 -m vall_e.emb.g2p ./data/librilight-tts # phonemizes anything that might have been amiss in the phoneme transcription
|
||||
python3 -m vall_e.emb.qnt ./data/librilight-tts
|
||||
@ -0,0 +1,18 @@
|
||||
import os
|
||||
import json
|
||||
|
||||
for f in os.listdir(f'./LibriTTS/'):
|
||||
if not os.path.isdir(f'./LibriTTS/{f}/'):
|
||||
continue
|
||||
for j in os.listdir(f'./LibriTTS/{f}/'):
|
||||
if not os.path.isdir(f'./LibriTTS/{f}/{j}'):
|
||||
continue
|
||||
for z in os.listdir(f'./LibriTTS/{f}/{j}'):
|
||||
if not os.path.isdir(f'./LibriTTS/{f}/{j}/{z}'):
|
||||
continue
|
||||
for i in os.listdir(f'./LibriTTS/{f}/{j}/{z}'):
|
||||
if i[-4:] != ".wav":
|
||||
continue
|
||||
|
||||
os.makedirs(f'./LibriTTS-Train/{j}/', exist_ok=True)
|
||||
os.rename(f'./LibriTTS/{f}/{j}/{z}/{i}', f'./LibriTTS-Train/{j}/{i}')
|
||||
@ -0,0 +1,3 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
until $@; do echo retrying && pkill python3; done
|
||||
@ -0,0 +1,64 @@
|
||||
import subprocess
|
||||
|
||||
from pathlib import Path
|
||||
from datetime import datetime
|
||||
from setuptools import setup, find_packages
|
||||
|
||||
def shell(*args):
|
||||
out = subprocess.check_output(args)
|
||||
return out.decode("ascii").strip()
|
||||
|
||||
|
||||
def write_version(version_core, pre_release=True):
|
||||
if pre_release:
|
||||
time = shell("git", "log", "-1", "--format=%cd", "--date=iso")
|
||||
time = datetime.strptime(time, "%Y-%m-%d %H:%M:%S %z")
|
||||
time = time.strftime("%Y%m%d%H%M%S")
|
||||
version = f"{version_core}-dev{time}"
|
||||
else:
|
||||
version = version_core
|
||||
|
||||
with open(Path("vall_e", "version.py"), "w") as f:
|
||||
f.write('__version__ = "{}"\n'.format(version))
|
||||
|
||||
return version
|
||||
|
||||
|
||||
with open("README.md", "r") as f:
|
||||
long_description = f.read()
|
||||
|
||||
setup(
|
||||
name="vall-e",
|
||||
python_requires=">=3.10.0",
|
||||
version=write_version("0.0.1"),
|
||||
description="An unofficial implementation of the audio LM VALL-E",
|
||||
author="ecker",
|
||||
author_email="mrq@ecker.tech",
|
||||
long_description=long_description,
|
||||
long_description_content_type="text/markdown",
|
||||
packages=find_packages(),
|
||||
install_requires=[
|
||||
"coloredlogs>=15.0.1",
|
||||
"deepspeed>=0.7.7",
|
||||
"diskcache>=5.4.0",
|
||||
"einops>=0.6.0",
|
||||
"encodec>=0.1.1",
|
||||
"phonemizer>=2.1.0",
|
||||
"matplotlib>=3.6.0",
|
||||
"numpy>=1.23.3",
|
||||
"omegaconf==2.0.6",
|
||||
"tqdm>=4.64.1",
|
||||
|
||||
"pandas>=1.5.0",
|
||||
"torch>=1.13.0",
|
||||
"torchaudio>=0.13.0",
|
||||
"torchmetrics",
|
||||
|
||||
"auraloss[all]",
|
||||
"vocos",
|
||||
"h5py",
|
||||
"git+https://github.com/microsoft/torchscale",
|
||||
"fairseq",
|
||||
],
|
||||
url="https://git.ecker.tech/mrq/vall-e",
|
||||
)
|
||||
Binary file not shown.
|
After Width: | Height: | Size: 511 KiB |
@ -0,0 +1,23 @@
|
||||
import argparse
|
||||
from pathlib import Path
|
||||
from .inference import TTS
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser("VALL-E TTS")
|
||||
parser.add_argument("text")
|
||||
parser.add_argument("reference", type=Path)
|
||||
parser.add_argument("out_path", type=Path)
|
||||
parser.add_argument("--yaml", type=Path, default=None)
|
||||
parser.add_argument("--ar-ckpt", type=Path, default=None)
|
||||
parser.add_argument("--nar-ckpt", type=Path, default=None)
|
||||
parser.add_argument("--max-ar-steps", type=int, default=6 * 75)
|
||||
parser.add_argument("--ar-temp", type=float, default=1.0)
|
||||
parser.add_argument("--nar-temp", type=float, default=1.0)
|
||||
parser.add_argument("--device", default="cuda")
|
||||
args = parser.parse_args()
|
||||
|
||||
tts = TTS( config=args.yaml, ar_ckpt=args.ar_ckpt, nar_ckpt=args.nar_ckpt, device=args.device )
|
||||
tts.inference( text=args.text, reference=args.reference, out_path=args.out_path, max_ar_samples=args.max_ar_samples, ar_temp=args.ar_temp, nar_temp=args.nar_temp )
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@ -0,0 +1,450 @@
|
||||
import copy
|
||||
import diskcache
|
||||
import h5py
|
||||
import json
|
||||
import os
|
||||
import subprocess
|
||||
import sys
|
||||
import time
|
||||
|
||||
from dataclasses import asdict, dataclass
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
from functools import cached_property
|
||||
from pathlib import Path
|
||||
|
||||
from omegaconf import OmegaConf
|
||||
|
||||
@dataclass()
|
||||
class _Config:
|
||||
cfg_path: str | None = None
|
||||
|
||||
@property
|
||||
def relpath(self):
|
||||
return Path(self.cfg_path)
|
||||
|
||||
@property
|
||||
def ckpt_dir(self):
|
||||
return self.relpath / "ckpt"
|
||||
|
||||
@property
|
||||
def log_dir(self):
|
||||
return self.relpath / "logs" / str(self.start_time)
|
||||
|
||||
@cached_property
|
||||
def start_time(self):
|
||||
return int(time.time())
|
||||
|
||||
@cached_property
|
||||
def git_commit(self):
|
||||
try:
|
||||
cmd = "git rev-parse HEAD"
|
||||
return subprocess.check_output(cmd.split()).decode("utf8").strip()
|
||||
except:
|
||||
return ""
|
||||
|
||||
@cached_property
|
||||
def git_status(self):
|
||||
try:
|
||||
cmd = "git status"
|
||||
return subprocess.check_output(cmd.split()).decode("utf8").strip()
|
||||
except:
|
||||
return ""
|
||||
|
||||
def dumps(self):
|
||||
data = {k: getattr(self, k) for k in dir(self) if not k.startswith("__")}
|
||||
data = {k: v for k, v in data.items() if not callable(v)}
|
||||
return json.dumps(data, indent=2, default=str)
|
||||
|
||||
def dump(self, path=None):
|
||||
if path is None:
|
||||
path = self.log_dir / "cfg.json"
|
||||
path.parent.mkdir(parents=True, exist_ok=True)
|
||||
with open(path, "w") as f:
|
||||
f.write(self.dumps())
|
||||
|
||||
@staticmethod
|
||||
def _is_cfg_argv(s):
|
||||
return "=" in s and "--" not in s
|
||||
|
||||
@classmethod
|
||||
def from_yaml( cls, yaml_path ):
|
||||
return cls.from_cli( [f'yaml="{yaml_path}"'] )
|
||||
|
||||
@classmethod
|
||||
def from_cli(cls, args=sys.argv):
|
||||
cli_cfg = OmegaConf.from_cli([s for s in args if cls._is_cfg_argv(s)])
|
||||
|
||||
# Replace argv to ensure there are no omegaconf options, for compatibility with argparse.
|
||||
sys.argv = [s for s in sys.argv if not cls._is_cfg_argv(s)]
|
||||
|
||||
if cli_cfg.get("help"):
|
||||
print(f"Configurable hyperparameters with their default values:")
|
||||
print(json.dumps(asdict(cls()), indent=2, default=str))
|
||||
exit()
|
||||
|
||||
if "yaml" in cli_cfg:
|
||||
yaml_cfg = OmegaConf.load(cli_cfg.yaml)
|
||||
yaml_path = Path(cli_cfg.yaml).absolute()
|
||||
cfg_path = Path(*yaml_path.relative_to(Path.cwd()).parts[:-1])
|
||||
cfg_path = cfg_path.with_suffix("")
|
||||
cfg_path = f'./{cfg_path}'
|
||||
|
||||
yaml_cfg.setdefault("cfg_path", cfg_path)
|
||||
cli_cfg.pop("yaml")
|
||||
else:
|
||||
yaml_cfg = {}
|
||||
merged = OmegaConf.merge(yaml_cfg, cli_cfg)
|
||||
return cls(**dict(merged))
|
||||
|
||||
def __repr__(self):
|
||||
return str(self)
|
||||
|
||||
def __str__(self):
|
||||
return self.dumps()
|
||||
|
||||
@dataclass()
|
||||
class Dataset:
|
||||
training: list[Path] = field(default_factory=lambda: [])
|
||||
validation: list[Path] = field(default_factory=lambda: [])
|
||||
|
||||
temp: list[Path] = field(default_factory=lambda: [])
|
||||
|
||||
speaker_name_getter: str = "lambda p: f'{p.parts[-3]}_{p.parts[-2]}'"
|
||||
|
||||
hdf5_name: str = "data.h5"
|
||||
use_hdf5: bool = False
|
||||
validate: bool = True
|
||||
workers: int = 8
|
||||
cache: bool = True
|
||||
|
||||
phones_range: list[int] = field(default_factory=lambda: [4, 256])
|
||||
duration_range: list[float] = field(default_factory=lambda: [1.0, 12.0])
|
||||
|
||||
random_utterance: float = 1.0
|
||||
max_prompts: int = 3
|
||||
prompt_duration: float = 3.0
|
||||
|
||||
@dataclass()
|
||||
class Model:
|
||||
name: str = ""
|
||||
size: str = "full"
|
||||
resp_levels: int = 1
|
||||
arch_type: str = "transformer"
|
||||
|
||||
@property
|
||||
def scale(self):
|
||||
if self.size == "quarter":
|
||||
return 0.25
|
||||
if self.size == "half":
|
||||
return 0.5
|
||||
return 1.0
|
||||
|
||||
@property
|
||||
def full_name(self):
|
||||
name = [ self.name ]
|
||||
|
||||
if self.size != "full":
|
||||
name.append(self.size)
|
||||
|
||||
if self.arch_type != "transformer":
|
||||
name.append(self.arch_type.replace("/", "-"))
|
||||
|
||||
name.append(f'{cfg.models.levels}')
|
||||
|
||||
return "-".join(name)
|
||||
|
||||
@property
|
||||
def tokens(self):
|
||||
return 1024
|
||||
|
||||
@property
|
||||
def dim(self):
|
||||
if self.size == "quarter":
|
||||
return 256
|
||||
if self.size == "half":
|
||||
return 512
|
||||
if self.size == "full":
|
||||
return 1024
|
||||
raise ValueError
|
||||
|
||||
@property
|
||||
def heads(self):
|
||||
if self.size == "quarter":
|
||||
return 4
|
||||
if self.size == "half":
|
||||
return 8
|
||||
if self.size == "full":
|
||||
return 16
|
||||
raise ValueError
|
||||
|
||||
@property
|
||||
def layers(self):
|
||||
return 12
|
||||
|
||||
@dataclass()
|
||||
class Models:
|
||||
_models: list[Model] = field(default_factory=lambda: [
|
||||
Model(name="ar", resp_levels=1),
|
||||
Model(name="nar", resp_levels=7),
|
||||
])
|
||||
|
||||
def get(self, name=None):
|
||||
if not name:
|
||||
return [ Model(**model) for model in self._models ]
|
||||
|
||||
for model in self._models:
|
||||
if model.name == name:
|
||||
return model
|
||||
|
||||
raise ValueError
|
||||
|
||||
@property
|
||||
def ar(self):
|
||||
return self.get("ar")
|
||||
|
||||
@property
|
||||
def nar(self):
|
||||
return self.get("nar")
|
||||
|
||||
@property
|
||||
def levels(self):
|
||||
return self.prom_levels
|
||||
|
||||
prom_levels: int = 8
|
||||
|
||||
@dataclass()
|
||||
class Hyperparameters:
|
||||
batch_size: int = 8
|
||||
gradient_accumulation_steps: int = 32
|
||||
gradient_clipping: int = 100
|
||||
|
||||
optimizer: str = "Adamw"
|
||||
learning_rate: float = 3.25e-4
|
||||
|
||||
scheduler_type: str = ""
|
||||
scheduler_params: dict = field(default_factory=lambda: {})
|
||||
|
||||
@dataclass()
|
||||
class Evaluation:
|
||||
batch_size: int = 64
|
||||
frequency: int = 250
|
||||
size: int = 64
|
||||
|
||||
steps: int = 500
|
||||
temperature: float = 1.0
|
||||
|
||||
@dataclass()
|
||||
class Trainer:
|
||||
iterations: int = 100_000
|
||||
|
||||
save_tag: str = "step"
|
||||
load_tag: str | None = None
|
||||
|
||||
save_on_oom: bool = True
|
||||
save_on_quit: bool = True
|
||||
save_frequency: int = 100
|
||||
|
||||
load_state_dict: bool = False
|
||||
load_states: bool = True
|
||||
strict_loading: bool = True
|
||||
restart_step_count: bool = False
|
||||
|
||||
aggressive_optimizations: bool = False
|
||||
|
||||
gc_mode: str | None = None
|
||||
|
||||
weight_dtype: str = "float16"
|
||||
|
||||
zero_optimization_level: int = 0
|
||||
use_compression_training: bool = False
|
||||
|
||||
@dataclass()
|
||||
class Config(_Config):
|
||||
device: str = "cuda"
|
||||
|
||||
dataset: Dataset = field(default_factory=lambda: Dataset)
|
||||
models: Models = field(default_factory=lambda: Models)
|
||||
hyperparameters: Hyperparameters = field(default_factory=lambda: Hyperparameters)
|
||||
evaluation: Evaluation = field(default_factory=lambda: Evaluation)
|
||||
trainer: Trainer = field(default_factory=lambda: Trainer)
|
||||
|
||||
use_vocos: bool = True
|
||||
|
||||
@property
|
||||
def sample_rate(self):
|
||||
return 24_000
|
||||
|
||||
@cached_property
|
||||
def get_spkr(self):
|
||||
return eval(self.dataset.speaker_name_getter)
|
||||
|
||||
@property
|
||||
def scheduler(self):
|
||||
cfg = {
|
||||
"type": self.hyperparameters.scheduler_type,
|
||||
"params": {},
|
||||
}
|
||||
|
||||
for k in self.hyperparameters.scheduler_params:
|
||||
cfg['params'][k] = self.hyperparameters.scheduler_params[k]
|
||||
|
||||
if self.hyperparameters.scheduler_type == "WarmupDecayLR" and 'total_num_steps' not in cfg['params']:
|
||||
cfg['params']['total_num_steps'] = self.trainer.iterations
|
||||
return cfg
|
||||
|
||||
@property
|
||||
def fp16_cfg(self):
|
||||
if self.trainer.weight_dtype.lower() != "float16":
|
||||
return None
|
||||
return {
|
||||
"enabled": True,
|
||||
"auto_cast": True,
|
||||
}
|
||||
|
||||
@property
|
||||
def bf16_cfg(self):
|
||||
return {
|
||||
"enabled": self.trainer.weight_dtype.lower() == "bfloat16"
|
||||
}
|
||||
|
||||
def get_compression_cfg(self, model):
|
||||
if not self.trainer.use_compression_training:
|
||||
return None
|
||||
|
||||
weights = [ name[0] for name in model.named_parameters() ]
|
||||
bits = 8
|
||||
|
||||
return {
|
||||
"weight_quantization": {
|
||||
"shared_parameters":{
|
||||
"enabled": True,
|
||||
"quantizer_kernel": True,
|
||||
"schedule_offset": 0,
|
||||
"quantize_groups": 64,
|
||||
"quantize_verbose": True,
|
||||
"quantization_type": "symmetric",
|
||||
"rounding": "nearest",
|
||||
"quantize_weight_in_forward": True,
|
||||
"fp16_mixed_quantize":{
|
||||
"enabled": False,
|
||||
"quantize_change_ratio": 1
|
||||
}
|
||||
},
|
||||
"different_groups": {
|
||||
"wq1": {
|
||||
"params": {
|
||||
"start_bits": bits,
|
||||
"target_bits": bits,
|
||||
"quantization_period": 0
|
||||
},
|
||||
"modules": weights
|
||||
}
|
||||
}
|
||||
},
|
||||
"activation_quantization": {
|
||||
"shared_parameters":{
|
||||
"enabled": True,
|
||||
"quantization_type": "symmetric",
|
||||
"range_calibration": "dynamic",
|
||||
"schedule_offset": 0
|
||||
},
|
||||
"different_groups": {
|
||||
"aq1": {
|
||||
"params": {
|
||||
"bits": bits
|
||||
},
|
||||
"modules": weights
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@property
|
||||
def zero_cfg(self):
|
||||
if self.trainer.zero_optimization_level == 0:
|
||||
return None
|
||||
|
||||
return {
|
||||
"stage": self.trainer.zero_optimization_level,
|
||||
"contiguous_gradients": True,
|
||||
"overlap_comm": True,
|
||||
"reduce_scatter": True,
|
||||
"reduce_bucket_size": 5e8,
|
||||
"allgather_bucket_size": 5e8,
|
||||
"sub_group_size": 5e8,
|
||||
"round_robin_gradients": True,
|
||||
"offload_optimizer": {
|
||||
"device": "cpu",
|
||||
"pin_memory": True
|
||||
},
|
||||
"offload_param": {
|
||||
"device": "cpu",
|
||||
"pin_memory": True
|
||||
}
|
||||
}
|
||||
|
||||
def get_ds_cfg(self, model):
|
||||
cfg = {
|
||||
"train_micro_batch_size_per_gpu": self.hyperparameters.batch_size,
|
||||
"gradient_accumulation_steps": self.hyperparameters.gradient_accumulation_steps,
|
||||
"optimizer": {
|
||||
"type": self.hyperparameters.optimizer,
|
||||
"params": {
|
||||
"lr": self.hyperparameters.learning_rate,
|
||||
}
|
||||
},
|
||||
"scheduler": self.hyperparameters.scheduler if self.hyperparameters.scheduler_type != "" else None,
|
||||
"gradient_clipping": self.hyperparameters.gradient_clipping,
|
||||
"fp16": self.fp16_cfg,
|
||||
"bf16": self.bf16_cfg,
|
||||
"compression_training": self.get_compression_cfg(model),
|
||||
"zero_optimization": self.zero_cfg,
|
||||
"comms_logger": {
|
||||
"enabled": False
|
||||
}
|
||||
}
|
||||
|
||||
null_keys = [ k for k in cfg if not cfg[k] ]
|
||||
for k in null_keys:
|
||||
del cfg[k]
|
||||
|
||||
if os.path.exists("./config/ds_config.json"):
|
||||
ds_cfg = json.load(open("./config/ds_config.json", "r", encoding="utf-8"))
|
||||
cfg.update(ds_cfg)
|
||||
|
||||
return cfg
|
||||
|
||||
@property
|
||||
def cache_dir(self):
|
||||
return ".cache" / self.relpath
|
||||
|
||||
@cached_property
|
||||
def diskcache(self):
|
||||
if self.dataset.cache:
|
||||
return diskcache.Cache(self.cache_dir).memoize
|
||||
return lambda: lambda x: x
|
||||
|
||||
def load_yaml( self, config_path ):
|
||||
tmp = Config.from_yaml( config_path )
|
||||
self.__dict__.update(tmp.__dict__)
|
||||
|
||||
|
||||
cfg = Config.from_cli()
|
||||
|
||||
# OmegaConf doesn't actually coerce the dicts into the @dataclass decorated classes, for some god forsaken reason, so we coerce them ourselves
|
||||
cfg.dataset = Dataset(**cfg.dataset)
|
||||
cfg.models = Models(**cfg.models)
|
||||
cfg.hyperparameters = Hyperparameters(**cfg.hyperparameters)
|
||||
cfg.evaluation = Evaluation(**cfg.evaluation)
|
||||
cfg.trainer = Trainer(**cfg.trainer)
|
||||
|
||||
# cached_property stopped working...
|
||||
if cfg.dataset.use_hdf5:
|
||||
try:
|
||||
cfg.hdf5 = h5py.File(f'{cfg.cfg_path}/{cfg.dataset.hdf5_name}', 'a')
|
||||
except Exception as e:
|
||||
print("Error while opening HDF5 file:", f'{cfg.cfg_path}/{cfg.dataset.hdf5_name}', str(e))
|
||||
|
||||
if __name__ == "__main__":
|
||||
print(cfg)
|
||||
@ -0,0 +1,79 @@
|
||||
import argparse
|
||||
import random
|
||||
import string
|
||||
import torch
|
||||
|
||||
from functools import cache
|
||||
from pathlib import Path
|
||||
from phonemizer import phonemize
|
||||
from phonemizer.backend import BACKENDS
|
||||
|
||||
from tqdm import tqdm
|
||||
|
||||
@cache
|
||||
def _get_graphs(path):
|
||||
with open(path, "r") as f:
|
||||
graphs = f.read()
|
||||
return graphs
|
||||
|
||||
cached_backends = {}
|
||||
def _get_backend( language="en-us", backend="espeak" ):
|
||||
key = f'{language}_{backend}'
|
||||
if key in cached_backends:
|
||||
return cached_backends[key]
|
||||
|
||||
if backend == 'espeak':
|
||||
phonemizer = BACKENDS[backend]( language, preserve_punctuation=True, with_stress=True)
|
||||
elif backend == 'espeak-mbrola':
|
||||
phonemizer = BACKENDS[backend]( language )
|
||||
else:
|
||||
phonemizer = BACKENDS[backend]( language, preserve_punctuation=True )
|
||||
|
||||
cached_backends[key] = phonemizer
|
||||
return phonemizer
|
||||
|
||||
|
||||
def encode(text: str, language="en-us", backend="espeak") -> list[str]:
|
||||
if language == "en":
|
||||
language = "en-us"
|
||||
|
||||
text = [ text ]
|
||||
|
||||
backend = _get_backend(language=language, backend=backend)
|
||||
if backend is not None:
|
||||
tokens = backend.phonemize( text, strip=True )
|
||||
else:
|
||||
tokens = phonemize( text, language=language, strip=True, preserve_punctuation=True, with_stress=True )
|
||||
|
||||
tokens = list(tokens[0])
|
||||
tokenized = " ".join( tokens )
|
||||
|
||||
merges = [ "\u02C8", "\u02CC", "\u02D0" ]
|
||||
for merge in merges:
|
||||
tokenized = tokenized.replace( f' {merge}', merge )
|
||||
|
||||
return tokenized.split(" ")
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def main():
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("folder", type=Path)
|
||||
parser.add_argument("--suffix", type=str, default=".normalized.txt")
|
||||
args = parser.parse_args()
|
||||
|
||||
paths = list(args.folder.rglob(f"*{args.suffix}"))
|
||||
random.shuffle(paths)
|
||||
|
||||
for path in tqdm(paths):
|
||||
phone_path = path.with_name(path.stem.split(".")[0] + ".phn.txt")
|
||||
if phone_path.exists():
|
||||
continue
|
||||
graphs = _get_graphs(path)
|
||||
phones = encode(graphs)
|
||||
with open(phone_path, "w") as f:
|
||||
f.write(" ".join(phones))
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@ -0,0 +1,198 @@
|
||||
from ..config import cfg
|
||||
|
||||
import argparse
|
||||
import random
|
||||
import torch
|
||||
import torchaudio
|
||||
|
||||
from functools import cache
|
||||
from pathlib import Path
|
||||
|
||||
|
||||
from encodec import EncodecModel
|
||||
from encodec.utils import convert_audio
|
||||
from einops import rearrange
|
||||
from torch import Tensor
|
||||
from tqdm import tqdm
|
||||
|
||||
USE_VOCOS = False
|
||||
try:
|
||||
from vocos import Vocos
|
||||
USE_VOCOS = True
|
||||
except Exception as e:
|
||||
USE_VOCOS = False
|
||||
|
||||
@cache
|
||||
def _load_encodec_model(device="cuda"):
|
||||
# Instantiate a pretrained EnCodec model
|
||||
assert cfg.sample_rate == 24_000
|
||||
|
||||
# too lazy to un-if ladder this shit
|
||||
if cfg.models.levels == 2:
|
||||
bandwidth_id = 1.5
|
||||
elif cfg.models.levels == 4:
|
||||
bandwidth_id = 3.0
|
||||
elif cfg.models.levels == 8:
|
||||
bandwidth_id = 6.0
|
||||
|
||||
model = EncodecModel.encodec_model_24khz()
|
||||
model.set_target_bandwidth(bandwidth_id)
|
||||
model.to(device)
|
||||
|
||||
return model
|
||||
|
||||
@cache
|
||||
def _load_vocos_model(device="cuda"):
|
||||
assert cfg.sample_rate == 24_000
|
||||
|
||||
model = Vocos.from_pretrained("charactr/vocos-encodec-24khz")
|
||||
model = model.to(device)
|
||||
|
||||
# too lazy to un-if ladder this shit
|
||||
if cfg.models.levels == 2:
|
||||
bandwidth_id = 0
|
||||
elif cfg.models.levels == 4:
|
||||
bandwidth_id = 1
|
||||
elif cfg.models.levels == 8:
|
||||
bandwidth_id = 2
|
||||
|
||||
model.bandwidth_id = torch.tensor([bandwidth_id], device=device)
|
||||
model.sample_rate = cfg.sample_rate
|
||||
|
||||
return model
|
||||
|
||||
@cache
|
||||
def _load_model(device="cuda", vocos=USE_VOCOS):
|
||||
if vocos:
|
||||
model = _load_vocos_model(device)
|
||||
else:
|
||||
model = _load_encodec_model(device)
|
||||
|
||||
return model
|
||||
|
||||
def unload_model():
|
||||
_load_model.cache_clear()
|
||||
_load_encodec_model.cache_clear()
|
||||
|
||||
|
||||
@torch.inference_mode()
|
||||
def decode(codes: Tensor, device="cuda"):
|
||||
"""
|
||||
Args:
|
||||
codes: (b q t)
|
||||
"""
|
||||
|
||||
# expand if we're given a raw 1-RVQ stream
|
||||
if codes.dim() == 1:
|
||||
codes = rearrange(codes, "t -> 1 1 t")
|
||||
# expand to a batch size of one if not passed as a batch
|
||||
# vocos does not do batch decoding, but encodec does, but we don't end up using this anyways *I guess*
|
||||
# to-do, make this logical
|
||||
elif codes.dim() == 2:
|
||||
codes = rearrange(codes, "t q -> 1 q t")
|
||||
|
||||
assert codes.dim() == 3, f'Requires shape (b q t) but got {codes.shape}'
|
||||
model = _load_model(device)
|
||||
|
||||
# upcast so it won't whine
|
||||
if codes.dtype == torch.int8 or codes.dtype == torch.int16 or codes.dtype == torch.uint8:
|
||||
codes = codes.to(torch.int32)
|
||||
|
||||
kwargs = {}
|
||||
if USE_VOCOS:
|
||||
x = model.codes_to_features(codes[0])
|
||||
kwargs['bandwidth_id'] = model.bandwidth_id
|
||||
else:
|
||||
x = [(codes.to(device), None)]
|
||||
|
||||
wav = model.decode(x, **kwargs)
|
||||
|
||||
if not USE_VOCOS:
|
||||
wav = wav[0]
|
||||
|
||||
return wav, model.sample_rate
|
||||
|
||||
# huh
|
||||
def decode_to_wave(resps: Tensor, device="cuda"):
|
||||
return decode(resps, device=device)
|
||||
|
||||
def decode_to_file(resps: Tensor, path: Path, device="cuda"):
|
||||
wavs, sr = decode(resps, device=device)
|
||||
|
||||
torchaudio.save(str(path), wavs.cpu(), sr)
|
||||
return wavs, sr
|
||||
|
||||
def _replace_file_extension(path, suffix):
|
||||
return (path.parent / path.name.split(".")[0]).with_suffix(suffix)
|
||||
|
||||
|
||||
@torch.inference_mode()
|
||||
def encode(wav: Tensor, sr: int, device="cuda"):
|
||||
"""
|
||||
Args:
|
||||
wav: (t)
|
||||
sr: int
|
||||
"""
|
||||
|
||||
model = _load_encodec_model(device)
|
||||
wav = wav.unsqueeze(0)
|
||||
wav = convert_audio(wav, sr, model.sample_rate, model.channels)
|
||||
wav = wav.to(device)
|
||||
|
||||
encoded_frames = model.encode(wav)
|
||||
qnt = torch.cat([encoded[0] for encoded in encoded_frames], dim=-1) # (b q t)
|
||||
|
||||
# duration = qnt.shape[-1] / 75
|
||||
|
||||
return qnt
|
||||
|
||||
|
||||
def encode_from_files(paths, device="cuda"):
|
||||
tuples = [ torchaudio.load(str(path)) for path in paths ]
|
||||
|
||||
wavs = []
|
||||
main_sr = tuples[0][1]
|
||||
for wav, sr in tuples:
|
||||
assert sr == main_sr, "Mismatching sample rates"
|
||||
|
||||
if wav.shape[0] == 2:
|
||||
wav = wav[:1]
|
||||
|
||||
wavs.append(wav)
|
||||
|
||||
wav = torch.cat(wavs, dim=-1)
|
||||
|
||||
return encode(wav, sr, "cpu")
|
||||
|
||||
def encode_from_file(path, device="cuda"):
|
||||
if isinstance( path, list ):
|
||||
return encode_from_files( path, device )
|
||||
else:
|
||||
wav, sr = torchaudio.load(str(path), format=path[-3:])
|
||||
|
||||
if wav.shape[0] == 2:
|
||||
wav = wav[:1]
|
||||
|
||||
qnt = encode(wav, sr, device)
|
||||
|
||||
return qnt
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("folder", type=Path)
|
||||
parser.add_argument("--suffix", default=".wav")
|
||||
args = parser.parse_args()
|
||||
|
||||
paths = [*args.folder.rglob(f"*{args.suffix}")]
|
||||
|
||||
for path in tqdm(paths):
|
||||
out_path = _replace_file_extension(path, ".qnt.pt")
|
||||
if out_path.exists():
|
||||
continue
|
||||
qnt = encode_from_file(path)
|
||||
torch.save(qnt.cpu(), out_path)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@ -0,0 +1,34 @@
|
||||
import argparse
|
||||
|
||||
import torch
|
||||
|
||||
from .data import get_phone_symmap
|
||||
from .train import load_engines
|
||||
|
||||
def load_models():
|
||||
models = {}
|
||||
engines = load_engines()
|
||||
for name in engines:
|
||||
model = engines[name].module.cpu()
|
||||
|
||||
model.phone_symmap = get_phone_symmap()
|
||||
|
||||
models[name] = model
|
||||
|
||||
return models
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser("Save trained model to path.")
|
||||
parser.add_argument("path")
|
||||
args = parser.parse_args()
|
||||
|
||||
models = load_models()
|
||||
for name in models:
|
||||
model = models[name]
|
||||
|
||||
outpath = f'{args.path}/{name}.pt'
|
||||
torch.save(model, outpath)
|
||||
print(f"Exported {name} to {outpath}")
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@ -0,0 +1,82 @@
|
||||
import torch
|
||||
import torchaudio
|
||||
import soundfile
|
||||
|
||||
from einops import rearrange
|
||||
|
||||
from .emb import g2p, qnt
|
||||
from .utils import to_device
|
||||
|
||||
from .config import cfg
|
||||
from .export import load_models
|
||||
|
||||
class TTS():
|
||||
def __init__( self, config=None, ar_ckpt=None, nar_ckpt=None, device="cuda" ):
|
||||
self.loading = True
|
||||
self.device = device
|
||||
|
||||
self.input_sample_rate = 24000
|
||||
self.output_sample_rate = 24000
|
||||
|
||||
if ar_ckpt and nar_ckpt:
|
||||
self.load_ar( ar_ckpt )
|
||||
self.load_nar( nar_ckpt )
|
||||
else:
|
||||
self.load_models( config )
|
||||
|
||||
self.loading = False
|
||||
|
||||
def load_models( self, config_path ):
|
||||
if config_path:
|
||||
cfg.load_yaml( config_path )
|
||||
|
||||
print("Loading models...")
|
||||
models = load_models()
|
||||
print("Loaded models")
|
||||
for name in models:
|
||||
model = models[name]
|
||||
if name[:2] == "ar":
|
||||
self.ar = model.to(self.device)
|
||||
self.symmap = self.ar.phone_symmap
|
||||
elif name[:3] == "nar":
|
||||
self.nar = model.to(self.device)
|
||||
else:
|
||||
print("Unknown:", name)
|
||||
|
||||
def load_ar( self, ckpt ):
|
||||
self.ar_ckpt = ckpt
|
||||
|
||||
self.ar = torch.load(self.ar_ckpt).to(self.device)
|
||||
self.symmap = self.ar.phone_symmap
|
||||
|
||||
def load_nar( self, ckpt ):
|
||||
self.nar_ckpt = nar_ckpt
|
||||
|
||||
self.nar = torch.load(self.nar_ckpt).to(self.device)
|
||||
|
||||
def encode_text( self, text, lang_marker="en" ):
|
||||
text = g2p.encode(text)
|
||||
phones = [f"<{lang_marker}>"] + [ " " if not p else p for p in text ] + [f"</{lang_marker}>"]
|
||||
mapped = [self.symmap[p] for p in phones if p in self.symmap]
|
||||
return torch.tensor( mapped )
|
||||
|
||||
def encode_audio( self, path ):
|
||||
enc = qnt.encode_from_file( path )
|
||||
return enc[0].t().to(torch.int16)
|
||||
|
||||
|
||||
def inference( self, text, reference, mode="both", max_ar_steps=6 * 75, ar_temp=1.0, nar_temp=1.0, out_path="./.tmp.wav" ):
|
||||
prom = self.encode_audio( reference )
|
||||
phns = self.encode_text(text)
|
||||
|
||||
prom = to_device(prom, self.device).to(torch.int16)
|
||||
phns = to_device(phns, self.device).to(torch.uint8 if len(self.symmap) < 256 else torch.int16)
|
||||
|
||||
resp_list = self.ar(text_list=[phns], proms_list=[prom], max_steps=max_ar_steps, sampling_temperature=ar_temp)
|
||||
resps_list = [r.unsqueeze(-1) for r in resp_list]
|
||||
resps_list = self.nar(text_list=[phns], proms_list=[prom], resps_list=resps_list, sampling_temperature=nar_temp)
|
||||
|
||||
wav, sr = qnt.decode_to_file(resps_list[0], out_path)
|
||||
|
||||
return (wav, sr)
|
||||
|
||||
@ -0,0 +1,267 @@
|
||||
# todo: clean this mess up
|
||||
# todo: yank deepspeed dependent code out into its own thing
|
||||
|
||||
from .config import cfg
|
||||
from .data import create_train_val_dataloader
|
||||
from .emb import qnt
|
||||
|
||||
from .utils import setup_logging, to_device, trainer, flatten_dict, do_gc
|
||||
from .utils import wrapper as ml
|
||||
|
||||
from .models import get_models
|
||||
|
||||
import auraloss
|
||||
import deepspeed
|
||||
import json
|
||||
import logging
|
||||
import random
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
import traceback
|
||||
|
||||
from collections import defaultdict
|
||||
|
||||
from deepspeed import comm as dist
|
||||
from deepspeed import DeepSpeedConfig
|
||||
from deepspeed.accelerator import get_accelerator
|
||||
|
||||
from tqdm import tqdm
|
||||
|
||||
mel_stft_loss = auraloss.freq.MelSTFTLoss(24_000, device="cuda")
|
||||
|
||||
def center_crop(x, len):
|
||||
start = (x.shape[-1] - len) // 2
|
||||
stop = start + len
|
||||
return x[..., start:stop]
|
||||
|
||||
def left_crop(x, len):
|
||||
return x[..., 0:len]
|
||||
|
||||
_logger = logging.getLogger(__name__)
|
||||
deepspeed._initialized_dist = False
|
||||
|
||||
def load_engines():
|
||||
if not deepspeed._initialized_dist:
|
||||
deepspeed._initialized_dist = True
|
||||
deepspeed.init_distributed()
|
||||
|
||||
models = get_models(cfg.models.get())
|
||||
engines = dict()
|
||||
|
||||
for name in models:
|
||||
model = models[name]
|
||||
|
||||
optimizer = None
|
||||
lr_scheduler = None
|
||||
|
||||
Adam = ml.Adam
|
||||
AdamW = ml.AdamW
|
||||
|
||||
if cfg.hyperparameters.optimizer.lower() == "adamw-torch":
|
||||
optimizer = AdamW(
|
||||
model.parameters(),
|
||||
lr=cfg.hyperparameters.learning_rate,
|
||||
betas=(0.9, 0.96),
|
||||
eps=1e-07,
|
||||
weight_decay=0.01,
|
||||
)
|
||||
|
||||
if cfg.trainer.load_state_dict:
|
||||
load_dir = cfg.ckpt_dir / name / "fp32.pth"
|
||||
model.load_state_dict(torch.load(load_dir))
|
||||
|
||||
ds_cfg=cfg.get_ds_cfg(model=model)
|
||||
config_class=DeepSpeedConfig(ds_cfg)
|
||||
engines[name] = trainer.Engine(
|
||||
model=model,
|
||||
config=ds_cfg,
|
||||
config_class=config_class,
|
||||
optimizer=optimizer,
|
||||
lr_scheduler=lr_scheduler,
|
||||
)
|
||||
|
||||
return trainer.load_engines(engines, cfg)
|
||||
|
||||
def main():
|
||||
#dist.init_distributed(dist_backend=get_accelerator().communication_backend_name())
|
||||
if not deepspeed._initialized_dist:
|
||||
deepspeed._initialized_dist = True
|
||||
deepspeed.init_distributed()
|
||||
|
||||
setup_logging(cfg.log_dir)
|
||||
|
||||
train_dl, subtrain_dl, val_dl = create_train_val_dataloader()
|
||||
|
||||
def train_feeder(engines, batch, name):
|
||||
stats = {}
|
||||
model = engines[name]
|
||||
if name.startswith("ar"):
|
||||
_ = model(
|
||||
text_list=batch["text"],
|
||||
proms_list=batch["proms"],
|
||||
resp_list=[r[..., 0] for r in batch["resps"]],
|
||||
)
|
||||
elif name.startswith("nar"):
|
||||
_ = model(
|
||||
text_list=batch["text"],
|
||||
proms_list=batch["proms"],
|
||||
resps_list=batch["resps"],
|
||||
)
|
||||
else:
|
||||
raise NotImplementedError(name)
|
||||
|
||||
losses = model.gather_attribute("loss")
|
||||
|
||||
loss = torch.stack([*losses.values()]).sum()
|
||||
|
||||
stats = {}
|
||||
stats |= {k: v.item() for k, v in losses.items()}
|
||||
stats |= engines.gather_attribute("scalar")
|
||||
|
||||
return loss, stats
|
||||
|
||||
@torch.inference_mode()
|
||||
def run_eval(engines, eval_name, dl):
|
||||
engines_stats = {
|
||||
'eval': eval_name
|
||||
}
|
||||
|
||||
AR = None
|
||||
NAR = None
|
||||
|
||||
names = []
|
||||
for name in engines:
|
||||
model = engines[name]
|
||||
names.append(name)
|
||||
if name[:2] == "ar":
|
||||
AR = model
|
||||
elif name[:3] == "nar":
|
||||
NAR = model
|
||||
|
||||
stats = defaultdict(list)
|
||||
stats['loss'] = []
|
||||
|
||||
for batch in tqdm(dl):
|
||||
batch: dict = to_device(batch, cfg.device)
|
||||
|
||||
# if we're training both models, provide output for both
|
||||
if AR is not None and NAR is not None:
|
||||
name = "+".join(names)
|
||||
|
||||
resp_list = AR(text_list=batch["text"], proms_list=batch["proms"], max_steps=cfg.evaluation.steps, sampling_temperature=cfg.evaluation.temperature)
|
||||
resps_list = [ r.unsqueeze(-1) for r in resp_list ]
|
||||
resps_list = NAR(text_list=batch["text"], proms_list=batch["proms"], resps_list=resps_list, sampling_temperature=cfg.evaluation.temperature)
|
||||
|
||||
for speaker, path, ref, hyp, prom in zip(batch["spkr_name"], batch["path"], batch["resps"], resps_list, batch["proms"]):
|
||||
if len(hyp) == 0:
|
||||
continue
|
||||
|
||||
filename = f'{speaker}_{path.parts[-1]}'
|
||||
|
||||
# to-do, refine the output dir to be sane-er
|
||||
ref_path = (cfg.log_dir / str(engines.global_step) / "ref" / filename).with_suffix(".wav")
|
||||
hyp_path = (cfg.log_dir / str(engines.global_step) / name / eval_name / filename).with_suffix(".wav")
|
||||
prom_path = (cfg.log_dir / str(engines.global_step) / name / "prom" / filename).with_suffix(".wav")
|
||||
|
||||
hyp_path.parent.mkdir(parents=True, exist_ok=True)
|
||||
ref_path.parent.mkdir(parents=True, exist_ok=True)
|
||||
prom_path.parent.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
ref_audio, sr = qnt.decode_to_file(ref, ref_path)
|
||||
hyp_audio, sr = qnt.decode_to_file(hyp, hyp_path)
|
||||
prom_audio, sr = qnt.decode_to_file(prom, prom_path)
|
||||
|
||||
min_length = min( ref_audio.shape[-1], hyp_audio.shape[-1] )
|
||||
ref_audio = ref_audio[..., 0:min_length]
|
||||
hyp_audio = hyp_audio[..., 0:min_length]
|
||||
|
||||
stats['loss'].append(mel_stft_loss(hyp_audio, ref_audio).item())
|
||||
else:
|
||||
for name in engines:
|
||||
model = engines[name]
|
||||
|
||||
if name.startswith("ar"):
|
||||
resp_list = model(
|
||||
text_list=batch["text"],
|
||||
proms_list=batch["proms"],
|
||||
max_steps=cfg.evaluation.steps,
|
||||
sampling_temperature=cfg.evaluation.temperature,
|
||||
)
|
||||
resps_list = [r.unsqueeze(-1) for r in resp_list]
|
||||
elif name.startswith("nar"):
|
||||
resps_list = model(
|
||||
text_list=batch["text"],
|
||||
proms_list=batch["proms"],
|
||||
resps_list=[r[..., 0].unsqueeze(-1) for r in batch["resps"]],
|
||||
sampling_temperature=cfg.evaluation.temperature,
|
||||
)
|
||||
else:
|
||||
raise NotImplementedError(name)
|
||||
|
||||
losses = model.gather_attribute("loss")
|
||||
|
||||
batch_stats = {}
|
||||
batch_stats |= {k: v.item() for k, v in losses.items()}
|
||||
batch_stats |= engines.gather_attribute("scalar")
|
||||
|
||||
for k, v in batch_stats.items():
|
||||
stats[k].append(v)
|
||||
|
||||
for speaker, path, ref, hyp, prom in zip(batch["spkr_name"], batch["path"], batch["resps"], resps_list, batch["proms"]):
|
||||
if len(hyp) == 0:
|
||||
continue
|
||||
|
||||
filename = f'{speaker}_{path.parts[-1]}'
|
||||
|
||||
# to-do, refine the output dir to be sane-er
|
||||
ref_path = (cfg.log_dir / str(engines.global_step) / "ref" / filename).with_suffix(".wav")
|
||||
hyp_path = (cfg.log_dir / str(engines.global_step) / name / eval_name / filename).with_suffix(".wav")
|
||||
prom_path = (cfg.log_dir / str(engines.global_step) / name / "prom" / filename).with_suffix(".wav")
|
||||
|
||||
hyp_path.parent.mkdir(parents=True, exist_ok=True)
|
||||
ref_path.parent.mkdir(parents=True, exist_ok=True)
|
||||
prom_path.parent.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
ref_audio, sr = qnt.decode_to_file(ref, ref_path)
|
||||
hyp_audio, sr = qnt.decode_to_file(hyp, hyp_path)
|
||||
prom_audio, sr = qnt.decode_to_file(prom, prom_path)
|
||||
|
||||
# pseudo loss calculation since we don't get the logits during eval
|
||||
min_length = min( ref_audio.shape[-1], hyp_audio.shape[-1] )
|
||||
ref_audio = ref_audio[..., 0:min_length]
|
||||
hyp_audio = hyp_audio[..., 0:min_length]
|
||||
|
||||
stats['loss'].append(mel_stft_loss(hyp_audio, ref_audio).item())
|
||||
|
||||
stats = {k: sum(v) / len(v) for k, v in stats.items()}
|
||||
engines_stats.update(flatten_dict({ name: stats }))
|
||||
|
||||
iteration = engines.global_step
|
||||
engines_stats['it'] = iteration
|
||||
engines_stats['epoch'] = iteration * cfg.hyperparameters.gradient_accumulation_steps / len(train_dl)
|
||||
|
||||
_logger.info(f"Validation Metrics: {json.dumps(engines_stats)}.")
|
||||
|
||||
def eval_fn(engines):
|
||||
try:
|
||||
run_eval(engines, "subtrain", subtrain_dl)
|
||||
run_eval(engines, "val", val_dl)
|
||||
except Exception as e:
|
||||
print("Error occurred while performing eval:", str(e))
|
||||
print(traceback.format_exc())
|
||||
|
||||
qnt.unload_model()
|
||||
do_gc()
|
||||
|
||||
qnt.unload_model()
|
||||
|
||||
trainer.train(
|
||||
engines_loader=load_engines,
|
||||
train_dl=train_dl,
|
||||
train_feeder=train_feeder,
|
||||
eval_fn=eval_fn,
|
||||
)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@ -0,0 +1,10 @@
|
||||
from .utils import (
|
||||
dispatch_attribute,
|
||||
flatten_dict,
|
||||
gather_attribute,
|
||||
load_state_dict_non_strict,
|
||||
setup_logging,
|
||||
to_device,
|
||||
tree_map,
|
||||
do_gc,
|
||||
)
|
||||
@ -0,0 +1,81 @@
|
||||
"""
|
||||
# https://github.com/enhuiz/pytorch-training-utilities
|
||||
"""
|
||||
|
||||
import os
|
||||
import socket
|
||||
|
||||
from functools import cache, wraps
|
||||
from typing import Callable
|
||||
|
||||
def get_free_port():
|
||||
sock = socket.socket()
|
||||
sock.bind(("", 0))
|
||||
return sock.getsockname()[1]
|
||||
|
||||
|
||||
@cache
|
||||
def fix_unset_envs():
|
||||
envs = dict(
|
||||
RANK="0",
|
||||
WORLD_SIZE="1",
|
||||
MASTER_ADDR="localhost",
|
||||
MASTER_PORT=str(get_free_port()),
|
||||
LOCAL_RANK="0",
|
||||
)
|
||||
|
||||
for key in envs:
|
||||
value = os.getenv(key)
|
||||
if value is not None:
|
||||
return
|
||||
|
||||
for key, value in envs.items():
|
||||
os.environ[key] = value
|
||||
|
||||
|
||||
def local_rank():
|
||||
return int(os.getenv("LOCAL_RANK", 0))
|
||||
|
||||
|
||||
def global_rank():
|
||||
return int(os.getenv("RANK", 0))
|
||||
|
||||
|
||||
def is_local_leader():
|
||||
return local_rank() == 0
|
||||
|
||||
|
||||
def is_global_leader():
|
||||
return global_rank() == 0
|
||||
|
||||
|
||||
def local_leader_only(fn=None, *, default=None) -> Callable:
|
||||
def wrapper(fn):
|
||||
@wraps(fn)
|
||||
def wrapped(*args, **kwargs):
|
||||
if is_local_leader():
|
||||
return fn(*args, **kwargs)
|
||||
return default
|
||||
|
||||
return wrapped
|
||||
|
||||
if fn is None:
|
||||
return wrapper
|
||||
|
||||
return wrapper(fn)
|
||||
|
||||
|
||||
def global_leader_only(fn: Callable | None = None, *, default=None) -> Callable:
|
||||
def wrapper(fn):
|
||||
@wraps(fn)
|
||||
def wrapped(*args, **kwargs):
|
||||
if is_global_leader():
|
||||
return fn(*args, **kwargs)
|
||||
return default
|
||||
|
||||
return wrapped
|
||||
|
||||
if fn is None:
|
||||
return wrapper
|
||||
|
||||
return wrapper(fn)
|
||||
@ -0,0 +1,252 @@
|
||||
"""
|
||||
# https://github.com/enhuiz/pytorch-training-utilities
|
||||
"""
|
||||
|
||||
# to-do: replace this
|
||||
# to-do: swap out deepspeed
|
||||
|
||||
from .config import Config
|
||||
from .distributed import fix_unset_envs
|
||||
from .utils import dispatch_attribute, flatten_dict, gather_attribute, do_gc, to_device
|
||||
|
||||
import logging
|
||||
import time
|
||||
import torch
|
||||
import torch.distributed
|
||||
|
||||
from deepspeed import DeepSpeedEngine
|
||||
from torch import Tensor
|
||||
from torch.distributed import all_reduce
|
||||
from typing import Any, Protocol
|
||||
|
||||
Stats = dict[str, float]
|
||||
|
||||
_logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class Engine(DeepSpeedEngine):
|
||||
def __init__(self, *args, **kwargs):
|
||||
fix_unset_envs()
|
||||
super().__init__(None, *args, **kwargs)
|
||||
self._frozen_params = set()
|
||||
|
||||
def freeze(self):
|
||||
for p in self.module.parameters():
|
||||
if p.requires_grad:
|
||||
p.requires_grad_(False)
|
||||
self._frozen_params.add(p)
|
||||
|
||||
def unfreeze(self):
|
||||
for p in self._frozen_params:
|
||||
p.requires_grad_(True)
|
||||
self._frozen_params.clear()
|
||||
|
||||
@property
|
||||
def global_step(self):
|
||||
return self.global_steps
|
||||
|
||||
def gather_attribute(self, *args, **kwargs):
|
||||
return gather_attribute(self.module, *args, **kwargs)
|
||||
|
||||
def dispatch_attribute(self, *args, **kwargs):
|
||||
return dispatch_attribute(self.module, *args, **kwargs)
|
||||
|
||||
|
||||
class TrainFeeder(Protocol):
|
||||
def __call__(
|
||||
self, *, engines: "Engines", batch: Any, name: str
|
||||
) -> None | tuple[Tensor, Stats]:
|
||||
...
|
||||
|
||||
|
||||
class Engines(dict[str, Engine]):
|
||||
def setup(self, cfg: Config):
|
||||
self._cfg = cfg
|
||||
self._global_step = 0
|
||||
|
||||
@property
|
||||
def cfg(self) -> Config:
|
||||
return self._cfg
|
||||
|
||||
@property
|
||||
def config(self):
|
||||
return self._cfg
|
||||
|
||||
@property
|
||||
def global_step(self):
|
||||
return self._global_step
|
||||
|
||||
def gather_attribute(self, *args, **kwargs):
|
||||
ret = {}
|
||||
for engine in self.values():
|
||||
ret |= engine.gather_attribute(*args, **kwargs)
|
||||
return ret
|
||||
|
||||
def dispatch_attribute(self, *args, **kwargs):
|
||||
for engine in self.values():
|
||||
engine.dispatch_attribute(*args, **kwargs)
|
||||
|
||||
def save_checkpoint(self, tag=None):
|
||||
if not tag:
|
||||
tag = self.cfg.trainer.save_tag
|
||||
tag = tag.lower()
|
||||
if tag[:2] == "it" or tag[:4] == "step":
|
||||
tag = self.global_step
|
||||
|
||||
self.cfg.ckpt_dir.mkdir(parents=True, exist_ok=True)
|
||||
for name, engine in self.items():
|
||||
engine.save_checkpoint(self.cfg.ckpt_dir / name, tag=tag)
|
||||
|
||||
def load_checkpoint(self, tag=None):
|
||||
if not tag:
|
||||
tag = self.cfg.trainer.load_tag
|
||||
|
||||
for name, engine in self.items():
|
||||
load_dir = self.cfg.ckpt_dir / name
|
||||
engine.load_checkpoint(
|
||||
tag=tag,
|
||||
load_dir=load_dir,
|
||||
load_module_strict=self.cfg.trainer.strict_loading,
|
||||
load_optimizer_states=self.cfg.trainer.load_states,
|
||||
load_lr_scheduler_states=self.cfg.trainer.load_states,
|
||||
load_module_only=False, # not self.cfg.trainer.load_states,
|
||||
)
|
||||
if self.cfg.trainer.restart_step_count:
|
||||
engine.global_steps = 0
|
||||
|
||||
# update the LR because for some god awful reason it gets overwritten when loading from a checkpoint but only when it's not using a scheduler
|
||||
if self.cfg.hyperparameters.scheduler_type == "":
|
||||
self.set_lr(self.cfg.hyperparameters.learning_rate)
|
||||
|
||||
self._update_global_step()
|
||||
|
||||
def set_lr(self, lr):
|
||||
try:
|
||||
for engine in self.values():
|
||||
if hasattr(engine.optimizer, 'param_groups'):
|
||||
print(engine.optimizer.param_groups)
|
||||
for param_group in engine.optimizer.param_groups:
|
||||
param_group['lr'] = lr
|
||||
else:
|
||||
engine.optimizer.set_lr(lr)
|
||||
except Exception as e:
|
||||
print(str(e))
|
||||
|
||||
def _update_global_step(self):
|
||||
for engine in self.values():
|
||||
self._global_step = max(self._global_step, engine.global_step)
|
||||
|
||||
def eval(self):
|
||||
for engine in self.values():
|
||||
engine.eval()
|
||||
|
||||
def train(self):
|
||||
for engine in self.values():
|
||||
engine.train()
|
||||
|
||||
def step(self, feeder: TrainFeeder, batch):
|
||||
total_elapsed_time = 0
|
||||
|
||||
stats: Any = dict()
|
||||
|
||||
if self.cfg.trainer.gc_mode == 'step':
|
||||
do_gc()
|
||||
|
||||
batch = to_device(batch, torch.cuda.current_device())
|
||||
|
||||
for name, engine in self.items():
|
||||
torch.cuda.synchronize()
|
||||
if self.cfg.trainer.gc_mode == 'substep':
|
||||
do_gc()
|
||||
|
||||
start_time = time.time()
|
||||
|
||||
tries = 4
|
||||
n_ooms = torch.zeros([], device=self.cfg.device)
|
||||
if self.cfg.trainer.aggressive_optimizations:
|
||||
batch = to_device(batch, torch.cuda.current_device())
|
||||
# engine = engine.to(torch.cuda.current_device())
|
||||
|
||||
while tries >= 0:
|
||||
try:
|
||||
res = feeder( engines=self, batch=batch, name=name )
|
||||
break
|
||||
except RuntimeError as e:
|
||||
print("Forward", str(e))
|
||||
|
||||
if "out of memory" not in str(e):
|
||||
self.save_checkpoint()
|
||||
raise e
|
||||
|
||||
# shrink batch size until it's happy
|
||||
for k in batch:
|
||||
batch[k] = batch[k][:-1]
|
||||
|
||||
if tries <= 0:
|
||||
# trigger OOM
|
||||
n_ooms += 1
|
||||
else:
|
||||
# also do GC
|
||||
do_gc()
|
||||
continue
|
||||
|
||||
all_reduce(n_ooms)
|
||||
if n_ooms.item() > 0:
|
||||
self.save_checkpoint()
|
||||
raise RuntimeError("Out of memory during forward pass!")
|
||||
|
||||
if res is None:
|
||||
continue
|
||||
|
||||
loss, engine_stats = res
|
||||
|
||||
n_ooms = torch.zeros([], device=self.cfg.device)
|
||||
|
||||
if self.cfg.trainer.aggressive_optimizations:
|
||||
batch = to_device(batch, 'cpu')
|
||||
|
||||
try:
|
||||
engine.backward(loss)
|
||||
except RuntimeError as e:
|
||||
print("Backwards:", str(e))
|
||||
|
||||
if "out of memory" not in str(e):
|
||||
self.save_checkpoint()
|
||||
raise e
|
||||
|
||||
n_ooms += 1
|
||||
|
||||
all_reduce(n_ooms)
|
||||
if n_ooms.item() > 0:
|
||||
self.save_checkpoint()
|
||||
raise RuntimeError("Out of memory during backwards pass!")
|
||||
|
||||
engine.step()
|
||||
torch.cuda.synchronize()
|
||||
elapsed_time = time.time() - start_time
|
||||
total_elapsed_time += elapsed_time
|
||||
|
||||
stats.update(
|
||||
flatten_dict(
|
||||
{
|
||||
name.split("-")[0]: dict(
|
||||
loss=loss.item(),
|
||||
lr=engine.get_lr()[0],
|
||||
grad_norm=engine.get_global_grad_norm(), # This norm is delayed but global and avoids extra computation
|
||||
elapsed_time=elapsed_time,
|
||||
engine_step=engine.global_step,
|
||||
**engine_stats,
|
||||
)
|
||||
}
|
||||
),
|
||||
)
|
||||
del loss
|
||||
# engine = engine.to('cpu')
|
||||
|
||||
self._update_global_step()
|
||||
stats["batch_size"] = len(batch["text"])
|
||||
stats["elapsed_time"] = total_elapsed_time
|
||||
stats["wall_time"] = time.time()
|
||||
stats["global_step"] = self.global_step
|
||||
|
||||
return stats
|
||||
@ -0,0 +1,48 @@
|
||||
"""
|
||||
A sampler that balances data by key_fns.
|
||||
|
||||
MIT License
|
||||
|
||||
Copyright (c) 2023 Zhe Niu
|
||||
|
||||
niuzhe.nz@outlook.com
|
||||
"""
|
||||
|
||||
import random
|
||||
|
||||
|
||||
class Sampler:
|
||||
def __init__(self, l, key_fns):
|
||||
self.tree = self._build(l, key_fns)
|
||||
|
||||
def _build(self, l, key_fns) -> dict[dict, list]:
|
||||
if not key_fns:
|
||||
return l
|
||||
|
||||
tree = {}
|
||||
|
||||
key_fn, *key_fns = key_fns
|
||||
|
||||
for x in l:
|
||||
k = key_fn(x)
|
||||
|
||||
if k in tree:
|
||||
tree[k].append(x)
|
||||
else:
|
||||
tree[k] = [x]
|
||||
|
||||
for k in tree:
|
||||
tree[k] = self._build(tree[k], key_fns)
|
||||
|
||||
return tree
|
||||
|
||||
def _sample(self, tree: dict | list):
|
||||
if isinstance(tree, list):
|
||||
ret = random.choice(tree)
|
||||
else:
|
||||
key = random.choice([*tree.keys()])
|
||||
ret = self._sample(tree[key])
|
||||
return ret
|
||||
|
||||
def sample(self):
|
||||
return self._sample(self.tree)
|
||||
@ -0,0 +1,253 @@
|
||||
"""
|
||||
# https://github.com/enhuiz/pytorch-training-utilities
|
||||
"""
|
||||
|
||||
# todo: replace this
|
||||
|
||||
import logging
|
||||
import time
|
||||
from typing import Any, Protocol
|
||||
|
||||
import torch
|
||||
import torch.distributed
|
||||
from deepspeed import DeepSpeedEngine
|
||||
from torch import Tensor
|
||||
from torch.distributed import all_reduce
|
||||
|
||||
from .config import Config
|
||||
from .distributed import fix_unset_envs
|
||||
from .utils import dispatch_attribute, flatten_dict, gather_attribute, do_gc, to_device
|
||||
|
||||
Stats = dict[str, float]
|
||||
|
||||
_logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class Engine(DeepSpeedEngine):
|
||||
def __init__(self, *args, **kwargs):
|
||||
fix_unset_envs()
|
||||
super().__init__(None, *args, **kwargs)
|
||||
self._frozen_params = set()
|
||||
|
||||
def freeze(self):
|
||||
for p in self.module.parameters():
|
||||
if p.requires_grad:
|
||||
p.requires_grad_(False)
|
||||
self._frozen_params.add(p)
|
||||
|
||||
def unfreeze(self):
|
||||
for p in self._frozen_params:
|
||||
p.requires_grad_(True)
|
||||
self._frozen_params.clear()
|
||||
|
||||
@property
|
||||
def global_step(self):
|
||||
return self.global_steps
|
||||
|
||||
def gather_attribute(self, *args, **kwargs):
|
||||
return gather_attribute(self.module, *args, **kwargs)
|
||||
|
||||
def dispatch_attribute(self, *args, **kwargs):
|
||||
return dispatch_attribute(self.module, *args, **kwargs)
|
||||
|
||||
|
||||
class TrainFeeder(Protocol):
|
||||
def __call__(
|
||||
self, *, engines: "Engines", batch: Any, name: str
|
||||
) -> None | tuple[Tensor, Stats]:
|
||||
...
|
||||
|
||||
|
||||
class Engines(dict[str, Engine]):
|
||||
def setup(self, cfg: Config):
|
||||
self._cfg = cfg
|
||||
self._global_step = 0
|
||||
|
||||
@property
|
||||
def cfg(self) -> Config:
|
||||
return self._cfg
|
||||
|
||||
@property
|
||||
def config(self):
|
||||
return self._cfg
|
||||
|
||||
@property
|
||||
def global_step(self):
|
||||
return self._global_step
|
||||
|
||||
def gather_attribute(self, *args, **kwargs):
|
||||
ret = {}
|
||||
for engine in self.values():
|
||||
ret |= engine.gather_attribute(*args, **kwargs)
|
||||
return ret
|
||||
|
||||
def dispatch_attribute(self, *args, **kwargs):
|
||||
for engine in self.values():
|
||||
engine.dispatch_attribute(*args, **kwargs)
|
||||
|
||||
def save_checkpoint(self, tag=None):
|
||||
if not tag:
|
||||
tag = self.cfg.trainer.save_tag
|
||||
tag = tag.lower()
|
||||
if tag[:2] == "it" or tag[:4] == "step":
|
||||
tag = self.global_step
|
||||
|
||||
self.cfg.ckpt_dir.mkdir(parents=True, exist_ok=True)
|
||||
for name, engine in self.items():
|
||||
engine.save_checkpoint(self.cfg.ckpt_dir / name, tag=tag)
|
||||
|
||||
def load_checkpoint(self, tag=None):
|
||||
if not tag:
|
||||
tag = self.cfg.trainer.load_tag
|
||||
|
||||
for name, engine in self.items():
|
||||
load_dir = self.cfg.ckpt_dir / name
|
||||
engine.load_checkpoint(
|
||||
tag=tag,
|
||||
load_dir=load_dir,
|
||||
load_module_strict=self.cfg.trainer.strict_loading,
|
||||
load_optimizer_states=self.cfg.trainer.load_states,
|
||||
load_lr_scheduler_states=self.cfg.trainer.load_states,
|
||||
load_module_only=False, # not self.cfg.trainer.load_states,
|
||||
)
|
||||
if self.cfg.trainer.restart_step_count:
|
||||
engine.global_steps = 0
|
||||
|
||||
# update the LR because for some god awful reason it gets overwritten when loading from a checkpoint but only when it's not using a scheduler
|
||||
if self.cfg.hyperparameters.scheduler_type == "":
|
||||
self.set_lr(self.cfg.hyperparameters.learning_rate)
|
||||
|
||||
self._update_global_step()
|
||||
|
||||
def set_lr(self, lr):
|
||||
try:
|
||||
for engine in self.values():
|
||||
if hasattr(engine.optimizer, 'param_groups'):
|
||||
print(engine.optimizer.param_groups)
|
||||
for param_group in engine.optimizer.param_groups:
|
||||
param_group['lr'] = lr
|
||||
else:
|
||||
engine.optimizer.set_lr(lr)
|
||||
except Exception as e:
|
||||
print(str(e))
|
||||
|
||||
def _update_global_step(self):
|
||||
for engine in self.values():
|
||||
self._global_step = max(self._global_step, engine.global_step)
|
||||
|
||||
def eval(self):
|
||||
for engine in self.values():
|
||||
engine.eval()
|
||||
|
||||
def train(self):
|
||||
for engine in self.values():
|
||||
engine.train()
|
||||
|
||||
def step(self, feeder: TrainFeeder, batch):
|
||||
total_elapsed_time = 0
|
||||
|
||||
stats: Any = dict()
|
||||
|
||||
if self.cfg.trainer.gc_mode == 'step':
|
||||
do_gc()
|
||||
|
||||
batch = to_device(batch, torch.cuda.current_device())
|
||||
|
||||
for name, engine in self.items():
|
||||
torch.cuda.synchronize()
|
||||
if self.cfg.trainer.gc_mode == 'substep':
|
||||
do_gc()
|
||||
|
||||
start_time = time.time()
|
||||
|
||||
tries = 4
|
||||
n_ooms = torch.zeros([], device=self.cfg.device)
|
||||
if self.cfg.trainer.aggressive_optimizations:
|
||||
batch = to_device(batch, torch.cuda.current_device())
|
||||
# engine = engine.to(torch.cuda.current_device())
|
||||
|
||||
while tries >= 0:
|
||||
try:
|
||||
maybe_loss_and_engine_stats = feeder( engines=self, batch=batch, name=name )
|
||||
break
|
||||
except RuntimeError as e:
|
||||
print("Forward", str(e))
|
||||
|
||||
if "out of memory" not in str(e):
|
||||
self.save_checkpoint()
|
||||
raise e
|
||||
|
||||
# shrink batch size until it's happy
|
||||
for k in batch:
|
||||
batch[k] = batch[k][:-1]
|
||||
|
||||
if tries <= 0:
|
||||
# trigger OOM
|
||||
n_ooms += 1
|
||||
else:
|
||||
# also do GC
|
||||
do_gc()
|
||||
continue
|
||||
|
||||
all_reduce(n_ooms)
|
||||
if n_ooms.item() > 0:
|
||||
self.save_checkpoint()
|
||||
raise RuntimeError("Out of memory during forward pass!")
|
||||
|
||||
# Here we allow skip optimizers. It's useful when, for example,
|
||||
# skipping discriminators in the begining of GAN training.
|
||||
if maybe_loss_and_engine_stats is None:
|
||||
continue
|
||||
|
||||
loss, engine_stats = maybe_loss_and_engine_stats
|
||||
|
||||
n_ooms = torch.zeros([], device=self.cfg.device)
|
||||
|
||||
if self.cfg.trainer.aggressive_optimizations:
|
||||
batch = to_device(batch, 'cpu')
|
||||
|
||||
try:
|
||||
engine.backward(loss)
|
||||
except RuntimeError as e:
|
||||
print("Backwards:", str(e))
|
||||
|
||||
if "out of memory" not in str(e):
|
||||
self.save_checkpoint()
|
||||
raise e
|
||||
|
||||
n_ooms += 1
|
||||
|
||||
all_reduce(n_ooms)
|
||||
if n_ooms.item() > 0:
|
||||
self.save_checkpoint()
|
||||
raise RuntimeError("Out of memory during backwards pass!")
|
||||
|
||||
engine.step()
|
||||
torch.cuda.synchronize()
|
||||
elapsed_time = time.time() - start_time
|
||||
total_elapsed_time += elapsed_time
|
||||
|
||||
stats.update(
|
||||
flatten_dict(
|
||||
{
|
||||
name.split("-")[0]: dict(
|
||||
loss=loss.item(),
|
||||
lr=engine.get_lr()[0],
|
||||
grad_norm=engine.get_global_grad_norm(), # This norm is delayed but global and avoids extra computation
|
||||
elapsed_time=elapsed_time,
|
||||
engine_step=engine.global_step,
|
||||
**engine_stats,
|
||||
)
|
||||
}
|
||||
),
|
||||
)
|
||||
del loss
|
||||
# engine = engine.to('cpu')
|
||||
|
||||
self._update_global_step()
|
||||
stats["batch_size"] = len(batch["text"])
|
||||
stats["elapsed_time"] = total_elapsed_time
|
||||
stats["wall_time"] = time.time()
|
||||
stats["global_step"] = self.global_step
|
||||
|
||||
return stats
|
||||
@ -0,0 +1,159 @@
|
||||
"""
|
||||
# https://github.com/enhuiz/pytorch-training-utilities
|
||||
"""
|
||||
|
||||
from .distributed import global_rank, local_rank, global_leader_only
|
||||
|
||||
import gc
|
||||
import logging
|
||||
import pandas as pd
|
||||
import re
|
||||
import torch
|
||||
|
||||
from coloredlogs import ColoredFormatter
|
||||
from logging import StreamHandler
|
||||
from pathlib import Path
|
||||
from torch import Tensor, nn
|
||||
from tqdm.auto import tqdm
|
||||
from typing import Callable, TypeVar, overload
|
||||
|
||||
T = TypeVar("T")
|
||||
|
||||
def do_gc():
|
||||
gc.collect()
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
def flatten_dict(d):
|
||||
records = pd.json_normalize(d).to_dict(orient="records")
|
||||
return records[0] if records else {}
|
||||
|
||||
|
||||
def _get_named_modules(module, attrname):
|
||||
for name, module in module.named_modules():
|
||||
if hasattr(module, attrname):
|
||||
yield name, module
|
||||
|
||||
|
||||
def gather_attribute(module, attrname, delete=True, prefix=True):
|
||||
ret = {}
|
||||
for name, module in _get_named_modules(module, attrname):
|
||||
ret[name] = getattr(module, attrname)
|
||||
if delete:
|
||||
try:
|
||||
delattr(module, attrname)
|
||||
except Exception as e:
|
||||
raise RuntimeError(f"{name} {module} {attrname}") from e
|
||||
if prefix:
|
||||
ret = {attrname: ret}
|
||||
ret = flatten_dict(ret)
|
||||
# remove consecutive dots
|
||||
ret = {re.sub(r"\.+", ".", k): v for k, v in ret.items()}
|
||||
return ret
|
||||
|
||||
|
||||
def dispatch_attribute(
|
||||
module,
|
||||
attrname,
|
||||
value,
|
||||
filter_fn: Callable[[nn.Module], bool] | None = None,
|
||||
):
|
||||
for _, module in _get_named_modules(module, attrname):
|
||||
if filter_fn is None or filter_fn(module):
|
||||
setattr(module, attrname, value)
|
||||
|
||||
|
||||
def load_state_dict_non_strict(model, state_dict, logger=None):
|
||||
model_state_dict = model.state_dict()
|
||||
provided = set(state_dict)
|
||||
required = set(model_state_dict)
|
||||
agreed = provided & required
|
||||
for k in list(agreed):
|
||||
if model_state_dict[k].shape != state_dict[k].shape:
|
||||
agreed.remove(k)
|
||||
provided.remove(k)
|
||||
state_dict = {k: state_dict[k] for k in agreed}
|
||||
if logger is not None and (diff := provided - required):
|
||||
logger.warning(
|
||||
f"Extra parameters are found. "
|
||||
f"Provided but not required parameters: \n{diff}."
|
||||
)
|
||||
if logger is not None and (diff := required - provided):
|
||||
logger.warning(
|
||||
f"Some parameters are missing. "
|
||||
f"Required but not provided parameters: \n{diff}."
|
||||
)
|
||||
model.load_state_dict(state_dict, strict=False)
|
||||
|
||||
class TqdmLoggingHandler(logging.Handler):
|
||||
def __init__(self, level=logging.NOTSET):
|
||||
super().__init__(level)
|
||||
|
||||
def emit(self, record):
|
||||
try:
|
||||
msg = self.format(record)
|
||||
tqdm.write(msg)
|
||||
self.flush()
|
||||
except Exception:
|
||||
self.handleError(record)
|
||||
|
||||
@global_leader_only
|
||||
def setup_logging(log_dir: str | Path | None = "log", log_level="info"):
|
||||
handlers = []
|
||||
|
||||
#stdout_handler = StreamHandler()
|
||||
stdout_handler = TqdmLoggingHandler()
|
||||
stdout_handler.setLevel(logging.INFO)
|
||||
formatter = ColoredFormatter(
|
||||
f"%(asctime)s - %(name)s - %(levelname)s - GR={global_rank()};LR={local_rank()} - \n%(message)s"
|
||||
)
|
||||
stdout_handler.setFormatter(formatter)
|
||||
handlers.append(stdout_handler)
|
||||
|
||||
if log_dir is not None:
|
||||
filename = Path(log_dir) / f"log.txt"
|
||||
filename.parent.mkdir(parents=True, exist_ok=True)
|
||||
file_handler = logging.FileHandler(filename, mode="a")
|
||||
file_handler.setLevel(logging.DEBUG)
|
||||
handlers.append(file_handler)
|
||||
|
||||
logging.basicConfig(
|
||||
level=logging.getLevelName(log_level.upper()),
|
||||
format="%(asctime)s - %(name)s - %(levelname)s - \n%(message)s",
|
||||
handlers=handlers,
|
||||
)
|
||||
|
||||
|
||||
@overload
|
||||
def tree_map(fn: Callable, x: list[T]) -> list[T]:
|
||||
...
|
||||
|
||||
|
||||
@overload
|
||||
def tree_map(fn: Callable, x: tuple[T]) -> tuple[T]:
|
||||
...
|
||||
|
||||
|
||||
@overload
|
||||
def tree_map(fn: Callable, x: dict[str, T]) -> dict[str, T]:
|
||||
...
|
||||
|
||||
|
||||
@overload
|
||||
def tree_map(fn: Callable, x: T) -> T:
|
||||
...
|
||||
|
||||
|
||||
def tree_map(fn: Callable, x):
|
||||
if isinstance(x, list):
|
||||
x = [tree_map(fn, xi) for xi in x]
|
||||
elif isinstance(x, tuple):
|
||||
x = (tree_map(fn, xi) for xi in x)
|
||||
elif isinstance(x, dict):
|
||||
x = {k: tree_map(fn, v) for k, v in x.items()}
|
||||
elif isinstance(x, Tensor):
|
||||
x = fn(x)
|
||||
return x
|
||||
|
||||
|
||||
def to_device(x: T, device) -> T:
|
||||
return tree_map(lambda t: t.to(device), x)
|
||||
@ -0,0 +1,60 @@
|
||||
# to-do: re-introduce bitsandbytes support
|
||||
|
||||
from contextlib import contextmanager
|
||||
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
|
||||
Embedding = torch.nn.Embedding
|
||||
Linear = torch.nn.Linear
|
||||
|
||||
"""
|
||||
if cfg.bitsandbytes:
|
||||
import bitsandbytes as bnb
|
||||
|
||||
if cfg.bitsandbytes_linear:
|
||||
Linear = bnb.nn.Linear8bitLt
|
||||
|
||||
if cfg.bitsandbytes_embedding:
|
||||
Embedding = bnb.nn.StableEmbedding
|
||||
Embedding.forward = lambda self, input: ( self.norm(F.embedding(
|
||||
input,
|
||||
self.weight,
|
||||
self.padding_idx,
|
||||
self.max_norm,
|
||||
self.norm_type,
|
||||
self.scale_grad_by_freq,
|
||||
self.sparse,
|
||||
)).to(self.weight.dtype) )
|
||||
"""
|
||||
|
||||
Adam = torch.optim.Adam
|
||||
AdamW = torch.optim.AdamW
|
||||
|
||||
"""
|
||||
if cfg.bitsandbytes:
|
||||
import bitsandbytes as bnb
|
||||
|
||||
Adam = bnb.optim.Adam
|
||||
AdamW = bnb.optim.AdamW
|
||||
"""
|
||||
|
||||
# handles temporarily upcasting 'index tensors' so torch will stop bitching
|
||||
def autocast_forward( func ):
|
||||
def wrapper( self, input, *args, **kwargs ):
|
||||
if input.dtype == torch.int16 or input.dtype == torch.int8 or input.dtype == torch.uint8:
|
||||
input = input.to(torch.int32)
|
||||
|
||||
return func( self, input, *args, **kwargs )
|
||||
return wrapper
|
||||
Embedding.forward = autocast_forward(Embedding.forward)
|
||||
|
||||
# handles generically converting to a specific tensor type and converting back (implemented solely for bfloat16)
|
||||
@contextmanager
|
||||
def autocast(input, from_dtype, to_dtype):
|
||||
if input.dtype == from_dtype:
|
||||
input = input.to(to_dtype)
|
||||
yield input
|
||||
input = input.to(from_dtype)
|
||||
else:
|
||||
yield input
|
||||
@ -0,0 +1,24 @@
|
||||
from .ar import AR
|
||||
from .nar import NAR
|
||||
|
||||
def get_model(model):
|
||||
if model.name == "ar":
|
||||
Model = AR
|
||||
elif model.name == "nar":
|
||||
Model = NAR
|
||||
else:
|
||||
raise f"invalid model name: {model.name}"
|
||||
name = model.name
|
||||
model = Model(
|
||||
n_tokens=model.tokens,
|
||||
d_model=model.dim,
|
||||
n_heads=model.heads,
|
||||
n_layers=model.layers,
|
||||
)
|
||||
|
||||
print(f"{name} parameter count: {sum(p.numel() for p in model.parameters() if p.requires_grad)}")
|
||||
|
||||
return model
|
||||
|
||||
def get_models(models):
|
||||
return { model.full_name: get_model(model) for model in models }
|
||||
@ -0,0 +1,19 @@
|
||||
from fairseq.models import FairseqIncrementalDecoder
|
||||
from fairseq.incremental_decoding_utils import with_incremental_state
|
||||
|
||||
from torchscale.architecture.config import RetNetConfig
|
||||
from torchscale.architecture.retnet import RetNetDecoder as Decoder
|
||||
|
||||
@with_incremental_state
|
||||
class RetNetDecoder(Decoder):
|
||||
def forward(self, src_tokens, **kwargs):
|
||||
return super().forward(src_tokens, **kwargs)
|
||||
|
||||
def max_positions(self):
|
||||
return self.args.max_token_positions
|
||||
|
||||
def reorder_incremental_state( self, incremental_state, new_order ):
|
||||
for module in incremental_state:
|
||||
for key in incremental_state[module]:
|
||||
result = incremental_state[module][key].index_select(0, new_order)
|
||||
incremental_state[module][key] = result
|
||||
@ -0,0 +1,195 @@
|
||||
"""
|
||||
# https://github.com/enhuiz/vall-e/
|
||||
"""
|
||||
|
||||
import math
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
import traceback
|
||||
|
||||
from typing import Literal, overload
|
||||
from functools import partial
|
||||
from einops import rearrange
|
||||
|
||||
from torch import Tensor, einsum, nn
|
||||
from torch.utils.checkpoint import checkpoint
|
||||
|
||||
from ..utils import wrapper as ml
|
||||
from .adaln import AdaLN
|
||||
|
||||
class SinusoidalEmbedding(nn.Module):
|
||||
def __init__(self, d_model):
|
||||
super().__init__()
|
||||
self.d_model = d_model
|
||||
exponent = torch.arange(self.d_half, dtype=torch.float32)
|
||||
exponent = exponent / self.d_half
|
||||
omega = torch.exp(-math.log(1e4) * exponent)
|
||||
self.omega: torch.Tensor
|
||||
self.register_buffer("omega", omega, persistent=False)
|
||||
|
||||
@property
|
||||
def d_half(self):
|
||||
assert self.d_model % 2 == 0, "Only support even d_model."
|
||||
return self.d_model // 2
|
||||
|
||||
def forward(self, x):
|
||||
"""
|
||||
Args:
|
||||
x: (...)
|
||||
Returns:
|
||||
pe: (... d)
|
||||
"""
|
||||
omega = self.omega
|
||||
|
||||
while omega.dim() <= x.dim():
|
||||
omega = omega.unsqueeze(0) # (... d)
|
||||
|
||||
x = x.unsqueeze(-1) # (... 1)
|
||||
x = omega * x
|
||||
x = torch.cat([x.sin(), x.cos()], dim=-1)
|
||||
|
||||
return x
|
||||
|
||||
def get_pe(self, n: int):
|
||||
"""
|
||||
Args:
|
||||
n: int
|
||||
Returns:
|
||||
pe: (n d)
|
||||
"""
|
||||
device = self.omega.device
|
||||
return self.forward(torch.arange(n, device=device))
|
||||
|
||||
def add_pe(self, x):
|
||||
"""
|
||||
Args:
|
||||
x: (b t c)
|
||||
"""
|
||||
e = self.get_pe(x.shape[1]) # t d
|
||||
e = e[None] # b t d
|
||||
x = x + e
|
||||
return x
|
||||
|
||||
|
||||
class Attention(nn.Module):
|
||||
def __init__(self, d_model, n_heads, causal):
|
||||
super().__init__()
|
||||
assert d_model % n_heads == 0
|
||||
dim_head = d_model // n_heads
|
||||
self.causal = causal
|
||||
self.n_heads = n_heads
|
||||
self.scale = dim_head**-0.5
|
||||
|
||||
self.to_qkv = nn.Linear(d_model, d_model * 3, bias=False)
|
||||
self.to_out = nn.Linear(d_model, d_model)
|
||||
|
||||
def forward(self, x, m):
|
||||
"""
|
||||
Args:
|
||||
x: (b t c)
|
||||
m: (b t c), 1 is data, 0 is padding
|
||||
Returns:
|
||||
x: (b t c)
|
||||
"""
|
||||
h = self.n_heads
|
||||
|
||||
q, k, v = self.to_qkv(x).chunk(3, dim=-1)
|
||||
q, k, v = map(lambda t: rearrange(t, "b t (h d) -> b t h d", h=h), (q, k, v))
|
||||
|
||||
e = einsum("b i h d, b j h d -> b i j h", q, k)
|
||||
e = e * self.scale
|
||||
|
||||
kpm = m.unsqueeze(1) * m.unsqueeze(2) # b i j 1
|
||||
|
||||
if self.causal:
|
||||
with ml.autocast(kpm, torch.bfloat16, torch.float16) as k:
|
||||
kpm = k.squeeze(-1).tril().unsqueeze(-1) # b i j 1
|
||||
|
||||
e = e.masked_fill(kpm == 0, -torch.finfo(e.dtype).max)
|
||||
a = e.softmax(dim=2) # Normalize on j, i.e. key
|
||||
|
||||
o = einsum("b i j h, b j h d -> b i h d", a, v)
|
||||
o = o.flatten(-2)
|
||||
o = self.to_out(o) # b t c
|
||||
|
||||
o = o * m
|
||||
|
||||
return o
|
||||
|
||||
class PrenormResidual(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
block,
|
||||
d_model,
|
||||
p_dropout,
|
||||
requires_mask=False,
|
||||
norm_type="ln",
|
||||
n_levels: int | None = None,
|
||||
):
|
||||
super().__init__()
|
||||
self.block = block
|
||||
self.requires_mask = requires_mask
|
||||
self.norm_type = norm_type
|
||||
if norm_type == "ln":
|
||||
self.norm = nn.LayerNorm(d_model)
|
||||
elif norm_type == "adaln":
|
||||
assert n_levels is not None
|
||||
self.norm = AdaLN(d_model, n_levels)
|
||||
else:
|
||||
raise NotImplementedError(norm_type)
|
||||
self.dropout = nn.Dropout(p_dropout)
|
||||
|
||||
def forward(self, x, m, l):
|
||||
"""
|
||||
Args:
|
||||
x: input (b t d)
|
||||
m: mask (b t 1), 1 is valuable and 0 is padding
|
||||
l: level to use, required only for AdaLN
|
||||
"""
|
||||
nopts = {"l": l} if self.norm_type == "adaln" else {}
|
||||
bopts = {"m": m} if self.requires_mask else {}
|
||||
x = x + self.dropout(self.block(self.norm(x, **nopts) * m, **bopts))
|
||||
return x * m
|
||||
|
||||
|
||||
class Block(nn.Sequential):
|
||||
def __init__(self, d_model, n_heads, p_dropout, causal, norm_type, n_levels):
|
||||
super().__init__()
|
||||
|
||||
self.attn = PrenormResidual(
|
||||
Attention(d_model, n_heads, causal),
|
||||
d_model=d_model,
|
||||
p_dropout=p_dropout,
|
||||
requires_mask=True,
|
||||
norm_type=norm_type,
|
||||
n_levels=n_levels,
|
||||
)
|
||||
|
||||
n_ff = d_model * 4 # 1024 * 4 = 4096 feed-forwards
|
||||
self.ffn = PrenormResidual(
|
||||
nn.Sequential(
|
||||
nn.Linear(d_model, n_ff),
|
||||
nn.GELU(),
|
||||
nn.Dropout(p_dropout),
|
||||
nn.Linear(n_ff, d_model),
|
||||
),
|
||||
d_model=d_model,
|
||||
p_dropout=p_dropout,
|
||||
norm_type=norm_type,
|
||||
n_levels=n_levels,
|
||||
)
|
||||
|
||||
def forward(self, x, m, l):
|
||||
"""
|
||||
Args:
|
||||
x: (b t c)
|
||||
m: (b t 1)
|
||||
l: (b)
|
||||
"""
|
||||
poor_in_vram = True
|
||||
if x.requires_grad and poor_in_vram:
|
||||
x = checkpoint(self.attn, x, m, l, use_reentrant=False)
|
||||
else:
|
||||
x = self.attn(x, m, l)
|
||||
x = self.ffn(x, m, l)
|
||||
return x
|
||||
Loading…
Reference in New Issue