repo

.gitattributes

@@ -2,13 +2,11 @@
 *.arrow filter=lfs diff=lfs merge=lfs -text
 *.bin filter=lfs diff=lfs merge=lfs -text
 *.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
 *.ftz filter=lfs diff=lfs merge=lfs -text
 *.gz filter=lfs diff=lfs merge=lfs -text
 *.h5 filter=lfs diff=lfs merge=lfs -text
 *.joblib filter=lfs diff=lfs merge=lfs -text
 *.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
 *.model filter=lfs diff=lfs merge=lfs -text
 *.msgpack filter=lfs diff=lfs merge=lfs -text
 *.npy filter=lfs diff=lfs merge=lfs -text
@@ -16,13 +14,12 @@
 *.onnx filter=lfs diff=lfs merge=lfs -text
 *.ot filter=lfs diff=lfs merge=lfs -text
 *.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
 *.pickle filter=lfs diff=lfs merge=lfs -text
 *.pkl filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
 *.pt filter=lfs diff=lfs merge=lfs -text
 *.pth filter=lfs diff=lfs merge=lfs -text
 *.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
 saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.tar.* filter=lfs diff=lfs merge=lfs -text
 *.tflite filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,14 @@
+.vscode
+__pycache__
+.ipynb_checkpoints
+data
+models
+flagged
+build
+dist
+audiodiffusion.egg-info
+lightning_logs
+taming
+checkpoints
+Pipfile
+Pipfile.lock

LICENSE

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

README.md

@@ -1,13 +1,234 @@
 ---
 title: Audio Diffusion
-emoji: 🔉
-colorFrom: gray
-colorTo: gray
+emoji: 🎵
+colorFrom: pink
+colorTo: blue
 sdk: gradio
-sdk_version: 3.28.0
+sdk_version: 3.1.4
 app_file: app.py
 pinned: false
 license: gpl-3.0
 ---
+# audio-diffusion [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/teticio/audio-diffusion/blob/master/notebooks/gradio_app.ipynb)
 
-taken from teticio's take on audio diffusion
+## Apply diffusion models to synthesize music instead of images using the new Hugging Face [diffusers](https://github.com/huggingface/diffusers) package
+
+---
+#### Sample automatically generated loop
+
+https://user-images.githubusercontent.com/44233095/204103172-27f25d63-5e77-40ca-91ab-d04a45d4726f.mp4
+
+Go to https://soundcloud.com/teticio2/sets/audio-diffusion-loops for more examples.
+
+---
+#### Updates
+
+**25/12/2022**. Now it is possible to train models conditional on an encoding (of text or audio, for example). See the section on Conditional Audio Generation below.
+
+**5/12/2022**. 🤗 Exciting news! `AudioDiffusionPipeline` has been migrated to the Hugging Face `diffusers` package so that it is even easier for others to use and contribute.
+
+**2/12/2022**. Added Mel to pipeline and updated the pretrained models to save Mel config (they are now no longer compatible with previous versions of this repo). It is relatively straightforward to migrate previously trained models to the new format (see https://huggingface.co/teticio/audio-diffusion-256).
+
+**7/11/2022**. Added pre-trained latent audio diffusion models [teticio/latent-audio-diffusion-256](https://huggingface.co/teticio/latent-audio-diffusion-256) and [teticio/latent-audio-diffusion-ddim-256](https://huggingface.co/teticio/latent-audio-diffusion-ddim-256). You can use the pre-trained VAE to train your own latent diffusion models on a different set of audio files.
+
+**22/10/2022**. Added DDIM encoder and ability to interpolate between audios in latent "noise" space. Mel spectrograms no longer have to be square (thanks to Tristan for this one), so you can set the vertical (frequency) and horizontal (time) resolutions independently.
+
+**15/10/2022**. Added latent audio diffusion (see below). Also added the possibility to train a DDIM ([De-noising Diffusion Implicit Models](https://arxiv.org/pdf/2010.02502.pdf)). These have the benefit that samples can be generated with much fewer steps (~50) than used in training.
+
+**4/10/2022**. It is now possible to mask parts of the input audio during generation which means you can stitch several samples together (think "out-painting").
+
+**27/9/2022**. You can now generate an audio based on a previous one. You can use this to generate variations of the same audio or even to "remix" a track (via a sort of "style transfer"). You can find examples of how to do this in the [`test_model.ipynb`](https://colab.research.google.com/github/teticio/audio-diffusion/blob/master/notebooks/test_model.ipynb) notebook.
+
+---
+
+![mel spectrogram](https://user-images.githubusercontent.com/44233095/205305826-8b39c917-26c5-49b4-887c-776f5d69e970.png)
+
+---
+
+## DDPM ([De-noising Diffusion Probabilistic Models](https://arxiv.org/abs/2006.11239))
+
+Audio can be represented as images by transforming to a [mel spectrogram](https://en.wikipedia.org/wiki/Mel-frequency_cepstrum), such as the one shown above. The class `Mel` in `mel.py` can convert a slice of audio into a mel spectrogram of `x_res` x `y_res` and vice versa. The higher the resolution, the less audio information will be lost. You can see how this works in the [`test_mel.ipynb`](https://github.com/teticio/audio-diffusion/blob/main/notebooks/test_mel.ipynb) notebook.
+
+A DDPM is trained on a set of mel spectrograms that have been generated from a directory of audio files. It is then used to synthesize similar mel spectrograms, which are then converted back into audio.
+
+You can play around with some pre-trained models on [Google Colab](https://colab.research.google.com/github/teticio/audio-diffusion/blob/master/notebooks/test_model.ipynb) or [Hugging Face spaces](https://huggingface.co/spaces/teticio/audio-diffusion). Check out some automatically generated loops [here](https://soundcloud.com/teticio2/sets/audio-diffusion-loops).
+
+| Model | Dataset | Description |
+|-------|---------|-------------|
+| [teticio/audio-diffusion-256](https://huggingface.co/teticio/audio-diffusion-256) | [teticio/audio-diffusion-256](https://huggingface.co/datasets/teticio/audio-diffusion-256) | My "liked" Spotify playlist |
+| [teticio/audio-diffusion-breaks-256](https://huggingface.co/teticio/audio-diffusion-breaks-256) | [teticio/audio-diffusion-breaks-256](https://huggingface.co/datasets/teticio/audio-diffusion-breaks-256) | Samples that have been used in music, sourced from [WhoSampled](https://whosampled.com) and [YouTube](https://youtube.com) |
+| [teticio/audio-diffusion-instrumental-hiphop-256](https://huggingface.co/teticio/audio-diffusion-instrumental-hiphop-256) | [teticio/audio-diffusion-instrumental-hiphop-256](https://huggingface.co/datasets/teticio/audio-diffusion-instrumental-hiphop-256) | Instrumental Hip Hop music |
+| [teticio/audio-diffusion-ddim-256](https://huggingface.co/teticio/audio-diffusion-ddim-256) | [teticio/audio-diffusion-256](https://huggingface.co/datasets/teticio/audio-diffusion-256) | De-noising Diffusion Implicit Model |
+| [teticio/latent-audio-diffusion-256](https://huggingface.co/teticio/latent-audio-diffusion-256) | [teticio/audio-diffusion-256](https://huggingface.co/datasets/teticio/audio-diffusion-256) | Latent Audio Diffusion model |
+| [teticio/latent-audio-diffusion-ddim-256](https://huggingface.co/teticio/latent-audio-diffusion-ddim-256) | [teticio/audio-diffusion-256](https://huggingface.co/datasets/teticio/audio-diffusion-256) | Latent Audio Diffusion Implicit Model |
+| [teticio/conditional-latent-audio-diffusion-512](https://huggingface.co/teticio/latent-audio-diffusion-512) | [teticio/audio-diffusion-512](https://huggingface.co/datasets/teticio/audio-diffusion-512) | Conditional Latent Audio Diffusion Model |
+
+---
+
+## Generate Mel spectrogram dataset from directory of audio files
+
+#### Install from GitHub (includes training scripts)
+
+```bash
+git clone https://github.com/teticio/audio-diffusion.git
+cd audio-diffusion
+pip install .
+```
+
+#### Install from PyPI
+
+```bash
+pip install audiodiffusion
+```
+
+#### Training can be run with Mel spectrograms of resolution 64x64 on a single commercial grade GPU (e.g. RTX 2080 Ti). The `hop_length` should be set to 1024 for better results
+
+```bash
+python scripts/audio_to_images.py \
+--resolution 64,64 \
+--hop_length 1024 \
+--input_dir path-to-audio-files \
+--output_dir path-to-output-data
+```
+
+#### Generate dataset of 256x256 Mel spectrograms and push to hub (you will need to be authenticated with `huggingface-cli login`)
+
+```bash
+python scripts/audio_to_images.py \
+--resolution 256 \
+--input_dir path-to-audio-files \
+--output_dir data/audio-diffusion-256 \
+--push_to_hub teticio/audio-diffusion-256
+```
+
+Note that the default `sample_rate` is 22050 and audios will be resampled if they are at a different rate. If you change this value, you may find that the results in the `test_mel.ipynb` notebook are not good (for example, if `sample_rate` is 48000) and that it is necessary to adjust `n_fft` (for example, to 2000 instead of the default value of 2048; alternatively, you can resample to a `sample_rate` of 44100). Make sure you use the same parameters for training and inference. You should also bear in mind that not all resolutions work with the neural network architecture as currently configured - you should be safe if you stick to powers of 2.
+
+## Train model
+
+#### Run training on local machine
+
+```bash
+accelerate launch --config_file config/accelerate_local.yaml \
+scripts/train_unet.py \
+--dataset_name data/audio-diffusion-64 \
+--hop_length 1024 \
+--output_dir models/ddpm-ema-audio-64 \
+--train_batch_size 16 \
+--num_epochs 100 \
+--gradient_accumulation_steps 1 \
+--learning_rate 1e-4 \
+--lr_warmup_steps 500 \
+--mixed_precision no
+```
+
+#### Run training on local machine with `batch_size` of 2 and `gradient_accumulation_steps` 8 to compensate, so that 256x256 resolution model fits on commercial grade GPU and push to hub
+
+```bash
+accelerate launch --config_file config/accelerate_local.yaml \
+scripts/train_unet.py \
+--dataset_name teticio/audio-diffusion-256 \
+--output_dir models/audio-diffusion-256 \
+--num_epochs 100 \
+--train_batch_size 2 \
+--eval_batch_size 2 \
+--gradient_accumulation_steps 8 \
+--learning_rate 1e-4 \
+--lr_warmup_steps 500 \
+--mixed_precision no \
+--push_to_hub True \
+--hub_model_id audio-diffusion-256 \
+--hub_token $(cat $HOME/.huggingface/token)
+```
+
+#### Run training on SageMaker
+
+```bash
+accelerate launch --config_file config/accelerate_sagemaker.yaml \
+scripts/train_unet.py \
+--dataset_name teticio/audio-diffusion-256 \
+--output_dir models/ddpm-ema-audio-256 \
+--train_batch_size 16 \
+--num_epochs 100 \
+--gradient_accumulation_steps 1 \
+--learning_rate 1e-4 \
+--lr_warmup_steps 500 \
+--mixed_precision no
+```
+
+## DDIM ([De-noising Diffusion Implicit Models](https://arxiv.org/pdf/2010.02502.pdf))
+
+#### A DDIM can be trained by adding the parameter
+
+```bash
+--scheduler ddim
+```
+
+Inference can the be run with far fewer steps than the number used for training (e.g., ~50), allowing for much faster generation. Without retraining, the parameter `eta` can be used to replicate a DDPM if it is set to 1 or a DDIM if it is set to 0, with all values in between being valid. When `eta` is 0 (the default value), the de-noising procedure is deterministic, which means that it can be run in reverse as a kind of encoder that recovers the original noise used in generation. A function `encode` has been added to `AudioDiffusionPipeline` for this purpose. It is then possible to interpolate between audios in the latent "noise" space using the function `slerp` (Spherical Linear intERPolation).
+
+## Latent Audio Diffusion
+
+Rather than de-noising images directly, it is interesting to work in the "latent space" after first encoding images using an autoencoder. This has a number of advantages. Firstly, the information in the images is compressed into a latent space of a much lower dimension, so it is much faster to train de-noising diffusion models and run inference with them. Secondly, similar images tend to be clustered together and interpolating between two images in latent space can produce meaningful combinations.
+
+At the time of writing, the Hugging Face `diffusers` library is geared towards inference and lacking in training functionality (rather like its cousin `transformers` in the early days of development). In order to train a VAE (Variational AutoEncoder), I use the [stable-diffusion](https://github.com/CompVis/stable-diffusion) repo from CompVis and convert the checkpoints to `diffusers` format. Note that it uses a perceptual loss function for images; it would be nice to try a perceptual *audio* loss function.
+
+#### Train latent diffusion model using pre-trained VAE
+
+```bash
+accelerate launch ...
+...
+--vae teticio/latent-audio-diffusion-256
+```
+
+#### Install dependencies to train with Stable Diffusion
+
+```bash
+pip install omegaconf pytorch_lightning==1.7.7 torchvision einops
+pip install -e git+https://github.com/CompVis/stable-diffusion.git@main#egg=latent-diffusion
+pip install -e git+https://github.com/CompVis/taming-transformers.git@master#egg=taming-transformers
+```
+
+#### Train an autoencoder
+
+```bash
+python scripts/train_vae.py \
+--dataset_name teticio/audio-diffusion-256 \
+--batch_size 2 \
+--gradient_accumulation_steps 12
+```
+
+#### Train latent diffusion model
+
+```bash
+accelerate launch ...
+...
+--vae models/autoencoder-kl
+```
+
+## Conditional Audio Generation
+
+We can generate audio conditional on a text prompt - or indeed anything which can be encoded into a bunch of numbers - much like DALL-E2, Midjourney and Stable Diffusion. It is generally harder to find good quality datasets of audios together with descriptions, although the people behind the dataset used to train Stable Diffusion are making some very interesting progress [here](https://github.com/LAION-AI/audio-dataset). I have chosen to encode the audio directly instead based on "how it sounds", using a [model which I trained on hundreds of thousands of Spotify playlists](https://github.com/teticio/Deej-AI). To encode an audio into a 100 dimensional vector
+
+```python
+from audiodiffusion.audio_encoder import AudioEncoder
+
+audio_encoder = AudioEncoder.from_pretrained("teticio/audio-encoder")
+audio_encoder.encode(['/home/teticio/Music/liked/Agua Re - Holy Dance - Large Sound Mix.mp3'])
+```
+
+Once you have prepared a dataset, you can encode the audio files with this script
+
+```bash
+python scripts/encode_audio \
+--dataset_name teticio/audio-diffusion-256 \
+--out_file data/encodings.p
+```
+
+Then you can train a model with
+
+```bash
+accelerate launch ...
+...
+--encodings data/encodings.p
+```
+
+When generating audios, you will need to pass an `encodings` Tensor. See the [`conditional_generation.ipynb`](https://colab.research.google.com/github/teticio/audio-diffusion/blob/master/notebooks/conditional_generation.ipynb) notebook for an example that uses encodings of Spotify track previews to influence the generation.

app.py

@@ -0,0 +1,50 @@
+import argparse
+
+import gradio as gr
+
+from audiodiffusion import AudioDiffusion
+
+
+def generate_spectrogram_audio_and_loop(model_id):
+    audio_diffusion = AudioDiffusion(model_id=model_id)
+    image, (sample_rate,
+            audio) = audio_diffusion.generate_spectrogram_and_audio()
+    loop = AudioDiffusion.loop_it(audio, sample_rate)
+    if loop is None:
+        loop = audio
+    return image, (sample_rate, audio), (sample_rate, loop)
+
+
+demo = gr.Interface(
+    fn=generate_spectrogram_audio_and_loop,
+    title="Audio Diffusion",
+    description="Generate audio using Huggingface diffusers.\
+        The models without 'latent' or 'ddim' give better results but take about \
+            20 minutes without a GPU. For GPU, you can use \
+                [colab](https://colab.research.google.com/github/teticio/audio-diffusion/blob/master/notebooks/gradio_app.ipynb) \
+                    to run this app.",
+    inputs=[
+        gr.Dropdown(label="Model",
+                    choices=[
+                        "teticio/audio-diffusion-256",
+                        "teticio/audio-diffusion-breaks-256",
+                        "teticio/audio-diffusion-instrumental-hiphop-256",
+                        "teticio/audio-diffusion-ddim-256",
+                        "teticio/latent-audio-diffusion-256",
+                        "teticio/latent-audio-diffusion-ddim-256"
+                    ],
+                    value="teticio/latent-audio-diffusion-ddim-256")
+    ],
+    outputs=[
+        gr.Image(label="Mel spectrogram", image_mode="L"),
+        gr.Audio(label="Audio"),
+        gr.Audio(label="Loop"),
+    ],
+    allow_flagging="never")
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--port", type=int)
+    parser.add_argument("--server", type=int)
+    args = parser.parse_args()
+    demo.launch(server_name=args.server or "0.0.0.0", server_port=args.port)

audiodiffusion/__init__.py

@@ -0,0 +1,140 @@
+from typing import Iterable, Tuple
+
+import numpy as np
+import torch
+from librosa.beat import beat_track
+from PIL import Image
+from tqdm.auto import tqdm
+
+# from diffusers import AudioDiffusionPipeline
+from .pipeline_audio_diffusion import AudioDiffusionPipeline
+
+VERSION = "1.5.3"
+
+
+class AudioDiffusion:
+    def __init__(
+        self,
+        model_id: str = "teticio/audio-diffusion-256",
+        cuda: bool = torch.cuda.is_available(),
+        progress_bar: Iterable = tqdm,
+    ):
+        """Class for generating audio using De-noising Diffusion Probabilistic Models.
+
+        Args:
+            model_id (String): name of model (local directory or Hugging Face Hub)
+            cuda (bool): use CUDA?
+            progress_bar (iterable): iterable callback for progress updates or None
+        """
+        self.model_id = model_id
+        self.pipe = AudioDiffusionPipeline.from_pretrained(self.model_id)
+        if cuda:
+            self.pipe.to("cuda")
+        self.progress_bar = progress_bar or (lambda _: _)
+
+    def generate_spectrogram_and_audio(
+        self,
+        steps: int = None,
+        generator: torch.Generator = None,
+        step_generator: torch.Generator = None,
+        eta: float = 0,
+        noise: torch.Tensor = None,
+        encoding: torch.Tensor = None,
+    ) -> Tuple[Image.Image, Tuple[int, np.ndarray]]:
+        """Generate random mel spectrogram and convert to audio.
+
+        Args:
+            steps (int): number of de-noising steps (defaults to 50 for DDIM, 1000 for DDPM)
+            generator (torch.Generator): random number generator or None
+            step_generator (torch.Generator): random number generator used to de-noise or None
+            eta (float): parameter between 0 and 1 used with DDIM scheduler
+            noise (torch.Tensor): noisy image or None
+            encoding (`torch.Tensor`): for UNet2DConditionModel shape (batch_size, seq_length, cross_attention_dim)
+
+        Returns:
+            PIL Image: mel spectrogram
+            (float, np.ndarray): sample rate and raw audio
+        """
+        images, (sample_rate, audios) = self.pipe(
+            batch_size=1,
+            steps=steps,
+            generator=generator,
+            step_generator=step_generator,
+            eta=eta,
+            noise=noise,
+            encoding=encoding,
+            return_dict=False,
+        )
+        return images[0], (sample_rate, audios[0])
+
+    def generate_spectrogram_and_audio_from_audio(
+        self,
+        audio_file: str = None,
+        raw_audio: np.ndarray = None,
+        slice: int = 0,
+        start_step: int = 0,
+        steps: int = None,
+        generator: torch.Generator = None,
+        mask_start_secs: float = 0,
+        mask_end_secs: float = 0,
+        step_generator: torch.Generator = None,
+        eta: float = 0,
+        encoding: torch.Tensor = None,
+        noise: torch.Tensor = None,
+    ) -> Tuple[Image.Image, Tuple[int, np.ndarray]]:
+        """Generate random mel spectrogram from audio input and convert to audio.
+
+        Args:
+            audio_file (str): must be a file on disk due to Librosa limitation or
+            raw_audio (np.ndarray): audio as numpy array
+            slice (int): slice number of audio to convert
+            start_step (int): step to start from
+            steps (int): number of de-noising steps (defaults to 50 for DDIM, 1000 for DDPM)
+            generator (torch.Generator): random number generator or None
+            mask_start_secs (float): number of seconds of audio to mask (not generate) at start
+            mask_end_secs (float): number of seconds of audio to mask (not generate) at end
+            step_generator (torch.Generator): random number generator used to de-noise or None
+            eta (float): parameter between 0 and 1 used with DDIM scheduler
+            encoding (`torch.Tensor`): for UNet2DConditionModel shape (batch_size, seq_length, cross_attention_dim)
+            noise (torch.Tensor): noisy image or None
+
+        Returns:
+            PIL Image: mel spectrogram
+            (float, np.ndarray): sample rate and raw audio
+        """
+
+        images, (sample_rate, audios) = self.pipe(
+            batch_size=1,
+            audio_file=audio_file,
+            raw_audio=raw_audio,
+            slice=slice,
+            start_step=start_step,
+            steps=steps,
+            generator=generator,
+            mask_start_secs=mask_start_secs,
+            mask_end_secs=mask_end_secs,
+            step_generator=step_generator,
+            eta=eta,
+            noise=noise,
+            encoding=encoding,
+            return_dict=False,
+        )
+        return images[0], (sample_rate, audios[0])
+
+    @staticmethod
+    def loop_it(audio: np.ndarray, sample_rate: int, loops: int = 12) -> np.ndarray:
+        """Loop audio
+
+        Args:
+            audio (np.ndarray): audio as numpy array
+            sample_rate (int): sample rate of audio
+            loops (int): number of times to loop
+
+        Returns:
+            (float, np.ndarray): sample rate and raw audio or None
+        """
+        _, beats = beat_track(y=audio, sr=sample_rate, units="samples")
+        beats_in_bar = (len(beats) - 1) // 4 * 4
+        if beats_in_bar > 0:
+            return np.tile(audio[beats[0] : beats[beats_in_bar]], loops)
+        return None

audiodiffusion/audio_encoder.py

@@ -0,0 +1,107 @@
+import numpy as np
+import torch
+from diffusers import ConfigMixin, Mel, ModelMixin
+from torch import nn
+
+
+class SeparableConv2d(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size):
+        super(SeparableConv2d, self).__init__()
+        self.depthwise = nn.Conv2d(
+            in_channels,
+            in_channels,
+            kernel_size=kernel_size,
+            groups=in_channels,
+            bias=False,
+            padding=1,
+        )
+        self.pointwise = nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=True)
+
+    def forward(self, x):
+        out = self.depthwise(x)
+        out = self.pointwise(out)
+        return out
+
+
+class ConvBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, dropout_rate):
+        super(ConvBlock, self).__init__()
+        self.sep_conv = SeparableConv2d(in_channels, out_channels, (3, 3))
+        self.leaky_relu = nn.LeakyReLU(0.2)
+        self.batch_norm = nn.BatchNorm2d(out_channels, eps=0.001, momentum=0.01)
+        self.max_pool = nn.MaxPool2d((2, 2))
+        self.dropout = nn.Dropout(dropout_rate)
+
+    def forward(self, x):
+        x = self.sep_conv(x)
+        x = self.leaky_relu(x)
+        x = self.batch_norm(x)
+        x = self.max_pool(x)
+        x = self.dropout(x)
+        return x
+
+
+class DenseBlock(nn.Module):
+    def __init__(self, in_features, out_features, dropout_rate):
+        super(DenseBlock, self).__init__()
+        self.flatten = nn.Flatten()
+        self.dense = nn.Linear(in_features, out_features)
+        self.leaky_relu = nn.LeakyReLU(0.2)
+        self.batch_norm = nn.BatchNorm1d(out_features, eps=0.001, momentum=0.01)
+        self.dropout = nn.Dropout(dropout_rate)
+
+    def forward(self, x):
+        x = self.flatten(x.permute(0, 2, 3, 1))
+        x = self.dense(x)
+        x = self.leaky_relu(x)
+        x = self.batch_norm(x)
+        x = self.dropout(x)
+        return x
+
+
+class AudioEncoder(ModelMixin, ConfigMixin):
+    def __init__(self):
+        super().__init__()
+        self.mel = Mel(
+            x_res=216,
+            y_res=96,
+            sample_rate=22050,
+            n_fft=2048,
+            hop_length=512,
+            top_db=80,
+        )
+        self.conv_blocks = nn.ModuleList([ConvBlock(1, 32, 0.2), ConvBlock(32, 64, 0.3), ConvBlock(64, 128, 0.4)])
+        self.dense_block = DenseBlock(41472, 1024, 0.5)
+        self.embedding = nn.Linear(1024, 100)
+
+    def forward(self, x):
+        for conv_block in self.conv_blocks:
+            x = conv_block(x)
+        x = self.dense_block(x)
+        x = self.embedding(x)
+        return x
+
+    @torch.no_grad()
+    def encode(self, audio_files):
+        self.eval()
+        y = []
+        for audio_file in audio_files:
+            self.mel.load_audio(audio_file)
+            x = [
+                np.expand_dims(
+                    np.frombuffer(self.mel.audio_slice_to_image(slice).tobytes(), dtype="uint8").reshape(
+                        (self.mel.y_res, self.mel.x_res)
+                    )
+                    / 255,
+                    axis=0,
+                )
+                for slice in range(self.mel.get_number_of_slices())
+            ]
+            y += [torch.mean(self(torch.Tensor(x)), dim=0)]
+        return torch.stack(y)
+
+
+# from diffusers import Mel
+# from audiodiffusion.audio_encoder import AudioEncoder
+# audio_encoder = AudioEncoder.from_pretrained("teticio/audio-encoder")
+# audio_encoder.encode(['/home/teticio/Music/liked/Agua Re - Holy Dance - Large Sound Mix.mp3'])

audiodiffusion/mel.py

@@ -0,0 +1,167 @@
+# This code has been migrated to diffusers but can be run locally with
+# pipe = DiffusionPipeline.from_pretrained("teticio/audio-diffusion-256", custom_pipeline="audio-diffusion/audiodiffusion/pipeline_audio_diffusion.py")
+
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import warnings
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+
+warnings.filterwarnings("ignore")
+
+import numpy as np  # noqa: E402
+
+
+try:
+    import librosa  # noqa: E402
+
+    _librosa_can_be_imported = True
+    _import_error = ""
+except Exception as e:
+    _librosa_can_be_imported = False
+    _import_error = (
+        f"Cannot import librosa because {e}. Make sure to correctly install librosa to be able to install it."
+    )
+
+
+from PIL import Image  # noqa: E402
+
+
+class Mel(ConfigMixin, SchedulerMixin):
+    """
+    Parameters:
+        x_res (`int`): x resolution of spectrogram (time)
+        y_res (`int`): y resolution of spectrogram (frequency bins)
+        sample_rate (`int`): sample rate of audio
+        n_fft (`int`): number of Fast Fourier Transforms
+        hop_length (`int`): hop length (a higher number is recommended for lower than 256 y_res)
+        top_db (`int`): loudest in decibels
+        n_iter (`int`): number of iterations for Griffin Linn mel inversion
+    """
+
+    config_name = "mel_config.json"
+
+    @register_to_config
+    def __init__(
+        self,
+        x_res: int = 256,
+        y_res: int = 256,
+        sample_rate: int = 22050,
+        n_fft: int = 2048,
+        hop_length: int = 512,
+        top_db: int = 80,
+        n_iter: int = 32,
+    ):
+        self.hop_length = hop_length
+        self.sr = sample_rate
+        self.n_fft = n_fft
+        self.top_db = top_db
+        self.n_iter = n_iter
+        self.set_resolution(x_res, y_res)
+        self.audio = None
+
+        if not _librosa_can_be_imported:
+            raise ValueError(_import_error)
+
+    def set_resolution(self, x_res: int, y_res: int):
+        """Set resolution.
+
+        Args:
+            x_res (`int`): x resolution of spectrogram (time)
+            y_res (`int`): y resolution of spectrogram (frequency bins)
+        """
+        self.x_res = x_res
+        self.y_res = y_res
+        self.n_mels = self.y_res
+        self.slice_size = self.x_res * self.hop_length - 1
+
+    def load_audio(self, audio_file: str = None, raw_audio: np.ndarray = None):
+        """Load audio.
+
+        Args:
+            audio_file (`str`): must be a file on disk due to Librosa limitation or
+            raw_audio (`np.ndarray`): audio as numpy array
+        """
+        if audio_file is not None:
+            self.audio, _ = librosa.load(audio_file, mono=True, sr=self.sr)
+        else:
+            self.audio = raw_audio
+
+        # Pad with silence if necessary.
+        if len(self.audio) < self.x_res * self.hop_length:
+            self.audio = np.concatenate([self.audio, np.zeros((self.x_res * self.hop_length - len(self.audio),))])
+
+    def get_number_of_slices(self) -> int:
+        """Get number of slices in audio.
+
+        Returns:
+            `int`: number of spectograms audio can be sliced into
+        """
+        return len(self.audio) // self.slice_size
+
+    def get_audio_slice(self, slice: int = 0) -> np.ndarray:
+        """Get slice of audio.
+
+        Args:
+            slice (`int`): slice number of audio (out of get_number_of_slices())
+
+        Returns:
+            `np.ndarray`: audio as numpy array
+        """
+        return self.audio[self.slice_size * slice : self.slice_size * (slice + 1)]
+
+    def get_sample_rate(self) -> int:
+        """Get sample rate:
+
+        Returns:
+            `int`: sample rate of audio
+        """
+        return self.sr
+
+    def audio_slice_to_image(self, slice: int) -> Image.Image:
+        """Convert slice of audio to spectrogram.
+
+        Args:
+            slice (`int`): slice number of audio to convert (out of get_number_of_slices())
+
+        Returns:
+            `PIL Image`: grayscale image of x_res x y_res
+        """
+        S = librosa.feature.melspectrogram(
+            y=self.get_audio_slice(slice), sr=self.sr, n_fft=self.n_fft, hop_length=self.hop_length, n_mels=self.n_mels
+        )
+        log_S = librosa.power_to_db(S, ref=np.max, top_db=self.top_db)
+        bytedata = (((log_S + self.top_db) * 255 / self.top_db).clip(0, 255) + 0.5).astype(np.uint8)
+        image = Image.fromarray(bytedata)
+        return image
+
+    def image_to_audio(self, image: Image.Image) -> np.ndarray:
+        """Converts spectrogram to audio.
+
+        Args:
+            image (`PIL Image`): x_res x y_res grayscale image
+
+        Returns:
+            audio (`np.ndarray`): raw audio
+        """
+        bytedata = np.frombuffer(image.tobytes(), dtype="uint8").reshape((image.height, image.width))
+        log_S = bytedata.astype("float") * self.top_db / 255 - self.top_db
+        S = librosa.db_to_power(log_S)
+        audio = librosa.feature.inverse.mel_to_audio(
+            S, sr=self.sr, n_fft=self.n_fft, hop_length=self.hop_length, n_iter=self.n_iter
+        )
+        return audio

audiodiffusion/pipeline_audio_diffusion.py

@@ -0,0 +1,257 @@
+# This code has been migrated to diffusers but can be run locally with
+# pipe = DiffusionPipeline.from_pretrained("teticio/audio-diffusion-256", custom_pipeline="audio-diffusion/audiodiffusion/pipeline_audio_diffusion.py")
+
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from math import acos, sin
+from typing import List, Tuple, Union
+
+import numpy as np
+import torch
+from diffusers import (
+    AudioPipelineOutput,
+    AutoencoderKL,
+    DDIMScheduler,
+    DDPMScheduler,
+    DiffusionPipeline,
+    ImagePipelineOutput,
+    UNet2DConditionModel,
+)
+from diffusers.utils import BaseOutput
+from PIL import Image
+
+from .mel import Mel
+
+class AudioDiffusionPipeline(DiffusionPipeline):
+    """
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Parameters:
+        vqae ([`AutoencoderKL`]): Variational AutoEncoder for Latent Audio Diffusion or None
+        unet ([`UNet2DConditionModel`]): UNET model
+        mel ([`Mel`]): transform audio <-> spectrogram
+        scheduler ([`DDIMScheduler` or `DDPMScheduler`]): de-noising scheduler
+    """
+
+    _optional_components = ["vqvae"]
+
+    def __init__(
+        self,
+        vqvae: AutoencoderKL,
+        unet: UNet2DConditionModel,
+        mel: Mel,
+        scheduler: Union[DDIMScheduler, DDPMScheduler],
+    ):
+        super().__init__()
+        self.register_modules(unet=unet, scheduler=scheduler, mel=mel, vqvae=vqvae)
+
+    def get_default_steps(self) -> int:
+        """Returns default number of steps recommended for inference
+
+        Returns:
+            `int`: number of steps
+        """
+        return 50 if isinstance(self.scheduler, DDIMScheduler) else 1000
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        batch_size: int = 1,
+        audio_file: str = None,
+        raw_audio: np.ndarray = None,
+        slice: int = 0,
+        start_step: int = 0,
+        steps: int = None,
+        generator: torch.Generator = None,
+        mask_start_secs: float = 0,
+        mask_end_secs: float = 0,
+        step_generator: torch.Generator = None,
+        eta: float = 0,
+        noise: torch.Tensor = None,
+        encoding: torch.Tensor = None,
+        return_dict=True,
+    ) -> Union[
+        Union[AudioPipelineOutput, ImagePipelineOutput],
+        Tuple[List[Image.Image], Tuple[int, List[np.ndarray]]],
+    ]:
+        """Generate random mel spectrogram from audio input and convert to audio.
+
+        Args:
+            batch_size (`int`): number of samples to generate
+            audio_file (`str`): must be a file on disk due to Librosa limitation or
+            raw_audio (`np.ndarray`): audio as numpy array
+            slice (`int`): slice number of audio to convert
+            start_step (int): step to start from
+            steps (`int`): number of de-noising steps (defaults to 50 for DDIM, 1000 for DDPM)
+            generator (`torch.Generator`): random number generator or None
+            mask_start_secs (`float`): number of seconds of audio to mask (not generate) at start
+            mask_end_secs (`float`): number of seconds of audio to mask (not generate) at end
+            step_generator (`torch.Generator`): random number generator used to de-noise or None
+            eta (`float`): parameter between 0 and 1 used with DDIM scheduler
+            noise (`torch.Tensor`): noise tensor of shape (batch_size, 1, height, width) or None
+            encoding (`torch.Tensor`): for UNet2DConditionModel shape (batch_size, seq_length, cross_attention_dim)
+            return_dict (`bool`): if True return AudioPipelineOutput, ImagePipelineOutput else Tuple
+
+        Returns:
+            `List[PIL Image]`: mel spectrograms (`float`, `List[np.ndarray]`): sample rate and raw audios
+        """
+
+        steps = steps or self.get_default_steps()
+        self.scheduler.set_timesteps(steps)
+        step_generator = step_generator or generator
+        # For backwards compatibility
+        if type(self.unet.sample_size) == int:
+            self.unet.sample_size = (self.unet.sample_size, self.unet.sample_size)
+        if noise is None:
+            noise = torch.randn(
+                (
+                    batch_size,
+                    self.unet.in_channels,
+                    self.unet.sample_size[0],
+                    self.unet.sample_size[1],
+                ),
+                generator=generator,
+                device=self.device,
+            )
+        images = noise
+        mask = None
+
+        if audio_file is not None or raw_audio is not None:
+            self.mel.load_audio(audio_file, raw_audio)
+            input_image = self.mel.audio_slice_to_image(slice)
+            input_image = np.frombuffer(input_image.tobytes(), dtype="uint8").reshape(
+                (input_image.height, input_image.width)
+            )
+            input_image = (input_image / 255) * 2 - 1
+            input_images = torch.tensor(input_image[np.newaxis, :, :], dtype=torch.float).to(self.device)
+
+            if self.vqvae is not None:
+                input_images = self.vqvae.encode(torch.unsqueeze(input_images, 0)).latent_dist.sample(
+                    generator=generator
+                )[0]
+                input_images = 0.18215 * input_images
+
+            if start_step > 0:
+                images[0, 0] = self.scheduler.add_noise(input_images, noise, self.scheduler.timesteps[start_step - 1])
+
+            pixels_per_second = (
+                self.unet.sample_size[1] * self.mel.get_sample_rate() / self.mel.x_res / self.mel.hop_length
+            )
+            mask_start = int(mask_start_secs * pixels_per_second)
+            mask_end = int(mask_end_secs * pixels_per_second)
+            mask = self.scheduler.add_noise(input_images, noise, torch.tensor(self.scheduler.timesteps[start_step:]))
+
+        for step, t in enumerate(self.progress_bar(self.scheduler.timesteps[start_step:])):
+            if isinstance(self.unet, UNet2DConditionModel):
+                model_output = self.unet(images, t, encoding)["sample"]
+            else:
+                model_output = self.unet(images, t)["sample"]
+
+            if isinstance(self.scheduler, DDIMScheduler):
+                images = self.scheduler.step(
+                    model_output=model_output,
+                    timestep=t,
+                    sample=images,
+                    eta=eta,
+                    generator=step_generator,
+                )["prev_sample"]
+            else:
+                images = self.scheduler.step(
+                    model_output=model_output,
+                    timestep=t,
+                    sample=images,
+                    generator=step_generator,
+                )["prev_sample"]
+
+            if mask is not None:
+                if mask_start > 0:
+                    images[:, :, :, :mask_start] = mask[:, step, :, :mask_start]
+                if mask_end > 0:
+                    images[:, :, :, -mask_end:] = mask[:, step, :, -mask_end:]
+
+        if self.vqvae is not None:
+            # 0.18215 was scaling factor used in training to ensure unit variance
+            images = 1 / 0.18215 * images
+            images = self.vqvae.decode(images)["sample"]
+
+        images = (images / 2 + 0.5).clamp(0, 1)
+        images = images.cpu().permute(0, 2, 3, 1).numpy()
+        images = (images * 255).round().astype("uint8")
+        images = list(
+            map(lambda _: Image.fromarray(_[:, :, 0]), images)
+            if images.shape[3] == 1
+            else map(lambda _: Image.fromarray(_, mode="RGB").convert("L"), images)
+        )
+
+        audios = list(map(lambda _: self.mel.image_to_audio(_), images))
+        if not return_dict:
+            return images, (self.mel.get_sample_rate(), audios)
+
+        return BaseOutput(**AudioPipelineOutput(np.array(audios)[:, np.newaxis, :]), **ImagePipelineOutput(images))
+
+    @torch.no_grad()
+    def encode(self, images: List[Image.Image], steps: int = 50) -> np.ndarray:
+        """Reverse step process: recover noisy image from generated image.
+
+        Args:
+            images (`List[PIL Image]`): list of images to encode
+            steps (`int`): number of encoding steps to perform (defaults to 50)
+
+        Returns:
+            `np.ndarray`: noise tensor of shape (batch_size, 1, height, width)
+        """
+
+        # Only works with DDIM as this method is deterministic
+        assert isinstance(self.scheduler, DDIMScheduler)
+        self.scheduler.set_timesteps(steps)
+        sample = np.array(
+            [np.frombuffer(image.tobytes(), dtype="uint8").reshape((1, image.height, image.width)) for image in images]
+        )
+        sample = (sample / 255) * 2 - 1
+        sample = torch.Tensor(sample).to(self.device)
+
+        for t in self.progress_bar(torch.flip(self.scheduler.timesteps, (0,))):
+            prev_timestep = t - self.scheduler.config.num_train_timesteps // self.scheduler.num_inference_steps
+            alpha_prod_t = self.scheduler.alphas_cumprod[t]
+            alpha_prod_t_prev = (
+                self.scheduler.alphas_cumprod[prev_timestep]
+                if prev_timestep >= 0
+                else self.scheduler.final_alpha_cumprod
+            )
+            beta_prod_t = 1 - alpha_prod_t
+            model_output = self.unet(sample, t)["sample"]
+            pred_sample_direction = (1 - alpha_prod_t_prev) ** (0.5) * model_output
+            sample = (sample - pred_sample_direction) * alpha_prod_t_prev ** (-0.5)
+            sample = sample * alpha_prod_t ** (0.5) + beta_prod_t ** (0.5) * model_output
+
+        return sample
+
+    @staticmethod
+    def slerp(x0: torch.Tensor, x1: torch.Tensor, alpha: float) -> torch.Tensor:
+        """Spherical Linear intERPolation
+
+        Args:
+            x0 (`torch.Tensor`): first tensor to interpolate between
+            x1 (`torch.Tensor`): seconds tensor to interpolate between
+            alpha (`float`): interpolation between 0 and 1
+
+        Returns:
+            `torch.Tensor`: interpolated tensor
+        """
+
+        theta = acos(torch.dot(torch.flatten(x0), torch.flatten(x1)) / torch.norm(x0) / torch.norm(x1))
+        return sin((1 - alpha) * theta) * x0 / sin(theta) + sin(alpha * theta) * x1 / sin(theta)

audiodiffusion/utils.py

@@ -0,0 +1,303 @@
+# adpated from https://github.com/huggingface/diffusers/blob/main/scripts/convert_original_stable_diffusion_to_diffusers.py
+
+import torch
+from diffusers import AutoencoderKL
+
+
+def shave_segments(path, n_shave_prefix_segments=1):
+    """
+    Removes segments. Positive values shave the first segments, negative shave the last segments.
+    """
+    if n_shave_prefix_segments >= 0:
+        return ".".join(path.split(".")[n_shave_prefix_segments:])
+    else:
+        return ".".join(path.split(".")[:n_shave_prefix_segments])
+
+
+def renew_vae_resnet_paths(old_list, n_shave_prefix_segments=0):
+    """
+    Updates paths inside resnets to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        new_item = old_item
+
+        new_item = new_item.replace("nin_shortcut", "conv_shortcut")
+        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+        mapping.append({"old": old_item, "new": new_item})
+
+    return mapping
+
+
+def renew_vae_attention_paths(old_list, n_shave_prefix_segments=0):
+    """
+    Updates paths inside attentions to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        new_item = old_item
+
+        new_item = new_item.replace("norm.weight", "group_norm.weight")
+        new_item = new_item.replace("norm.bias", "group_norm.bias")
+
+        new_item = new_item.replace("q.weight", "query.weight")
+        new_item = new_item.replace("q.bias", "query.bias")
+
+        new_item = new_item.replace("k.weight", "key.weight")
+        new_item = new_item.replace("k.bias", "key.bias")
+
+        new_item = new_item.replace("v.weight", "value.weight")
+        new_item = new_item.replace("v.bias", "value.bias")
+
+        new_item = new_item.replace("proj_out.weight", "proj_attn.weight")
+        new_item = new_item.replace("proj_out.bias", "proj_attn.bias")
+
+        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+        mapping.append({"old": old_item, "new": new_item})
+
+    return mapping
+
+
+def assign_to_checkpoint(
+    paths,
+    checkpoint,
+    old_checkpoint,
+    attention_paths_to_split=None,
+    additional_replacements=None,
+    config=None,
+):
+    """
+    This does the final conversion step: take locally converted weights and apply a global renaming
+    to them. It splits attention layers, and takes into account additional replacements
+    that may arise.
+
+    Assigns the weights to the new checkpoint.
+    """
+    assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys."
+
+    # Splits the attention layers into three variables.
+    if attention_paths_to_split is not None:
+        for path, path_map in attention_paths_to_split.items():
+            old_tensor = old_checkpoint[path]
+            channels = old_tensor.shape[0] // 3
+
+            target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1)
+
+            num_heads = old_tensor.shape[0] // config["num_head_channels"] // 3
+
+            old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:])
+            query, key, value = old_tensor.split(channels // num_heads, dim=1)
+
+            checkpoint[path_map["query"]] = query.reshape(target_shape)
+            checkpoint[path_map["key"]] = key.reshape(target_shape)
+            checkpoint[path_map["value"]] = value.reshape(target_shape)
+
+    for path in paths:
+        new_path = path["new"]
+
+        # These have already been assigned
+        if attention_paths_to_split is not None and new_path in attention_paths_to_split:
+            continue
+
+        # Global renaming happens here
+        new_path = new_path.replace("middle_block.0", "mid_block.resnets.0")
+        new_path = new_path.replace("middle_block.1", "mid_block.attentions.0")
+        new_path = new_path.replace("middle_block.2", "mid_block.resnets.1")
+
+        if additional_replacements is not None:
+            for replacement in additional_replacements:
+                new_path = new_path.replace(replacement["old"], replacement["new"])
+
+        # proj_attn.weight has to be converted from conv 1D to linear
+        if "proj_attn.weight" in new_path:
+            checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0]
+        else:
+            checkpoint[new_path] = old_checkpoint[path["old"]]
+
+
+def conv_attn_to_linear(checkpoint):
+    keys = list(checkpoint.keys())
+    attn_keys = ["query.weight", "key.weight", "value.weight"]
+    for key in keys:
+        if ".".join(key.split(".")[-2:]) in attn_keys:
+            if checkpoint[key].ndim > 2:
+                checkpoint[key] = checkpoint[key][:, :, 0, 0]
+        elif "proj_attn.weight" in key:
+            if checkpoint[key].ndim > 2:
+                checkpoint[key] = checkpoint[key][:, :, 0]
+
+
+def create_vae_diffusers_config(original_config):
+    """
+    Creates a config for the diffusers based on the config of the LDM model.
+    """
+    vae_params = original_config.model.params.ddconfig
+    _ = original_config.model.params.embed_dim
+
+    block_out_channels = [vae_params.ch * mult for mult in vae_params.ch_mult]
+    down_block_types = ["DownEncoderBlock2D"] * len(block_out_channels)
+    up_block_types = ["UpDecoderBlock2D"] * len(block_out_channels)
+
+    config = dict(
+        sample_size=tuple(vae_params.resolution),
+        in_channels=vae_params.in_channels,
+        out_channels=vae_params.out_ch,
+        down_block_types=tuple(down_block_types),
+        up_block_types=tuple(up_block_types),
+        block_out_channels=tuple(block_out_channels),
+        latent_channels=vae_params.z_channels,
+        layers_per_block=vae_params.num_res_blocks,
+    )
+    return config
+
+
+def convert_ldm_vae_checkpoint(checkpoint, config):
+    # extract state dict for VAE
+    vae_state_dict = checkpoint
+
+    new_checkpoint = {}
+
+    new_checkpoint["encoder.conv_in.weight"] = vae_state_dict["encoder.conv_in.weight"]
+    new_checkpoint["encoder.conv_in.bias"] = vae_state_dict["encoder.conv_in.bias"]
+    new_checkpoint["encoder.conv_out.weight"] = vae_state_dict["encoder.conv_out.weight"]
+    new_checkpoint["encoder.conv_out.bias"] = vae_state_dict["encoder.conv_out.bias"]
+    new_checkpoint["encoder.conv_norm_out.weight"] = vae_state_dict["encoder.norm_out.weight"]
+    new_checkpoint["encoder.conv_norm_out.bias"] = vae_state_dict["encoder.norm_out.bias"]
+
+    new_checkpoint["decoder.conv_in.weight"] = vae_state_dict["decoder.conv_in.weight"]
+    new_checkpoint["decoder.conv_in.bias"] = vae_state_dict["decoder.conv_in.bias"]
+    new_checkpoint["decoder.conv_out.weight"] = vae_state_dict["decoder.conv_out.weight"]
+    new_checkpoint["decoder.conv_out.bias"] = vae_state_dict["decoder.conv_out.bias"]
+    new_checkpoint["decoder.conv_norm_out.weight"] = vae_state_dict["decoder.norm_out.weight"]
+    new_checkpoint["decoder.conv_norm_out.bias"] = vae_state_dict["decoder.norm_out.bias"]
+
+    new_checkpoint["quant_conv.weight"] = vae_state_dict["quant_conv.weight"]
+    new_checkpoint["quant_conv.bias"] = vae_state_dict["quant_conv.bias"]
+    new_checkpoint["post_quant_conv.weight"] = vae_state_dict["post_quant_conv.weight"]
+    new_checkpoint["post_quant_conv.bias"] = vae_state_dict["post_quant_conv.bias"]
+
+    # Retrieves the keys for the encoder down blocks only
+    num_down_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "encoder.down" in layer})
+    down_blocks = {
+        layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks)
+    }
+
+    # Retrieves the keys for the decoder up blocks only
+    num_up_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "decoder.up" in layer})
+    up_blocks = {
+        layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks)
+    }
+
+    for i in range(num_down_blocks):
+        resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key]
+
+        if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict:
+            new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop(
+                f"encoder.down.{i}.downsample.conv.weight"
+            )
+            new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop(
+                f"encoder.down.{i}.downsample.conv.bias"
+            )
+
+        paths = renew_vae_resnet_paths(resnets)
+        meta_path = {"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"}
+        assign_to_checkpoint(
+            paths,
+            new_checkpoint,
+            vae_state_dict,
+            additional_replacements=[meta_path],
+            config=config,
+        )
+
+    mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key]
+    num_mid_res_blocks = 2
+    for i in range(1, num_mid_res_blocks + 1):
+        resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key]
+
+        paths = renew_vae_resnet_paths(resnets)
+        meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"}
+        assign_to_checkpoint(
+            paths,
+            new_checkpoint,
+            vae_state_dict,
+            additional_replacements=[meta_path],
+            config=config,
+        )
+
+    mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key]
+    paths = renew_vae_attention_paths(mid_attentions)
+    meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"}
+    assign_to_checkpoint(
+        paths,
+        new_checkpoint,
+        vae_state_dict,
+        additional_replacements=[meta_path],
+        config=config,
+    )
+    conv_attn_to_linear(new_checkpoint)
+
+    for i in range(num_up_blocks):
+        block_id = num_up_blocks - 1 - i
+        resnets = [
+            key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key
+        ]
+
+        if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict:
+            new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[
+                f"decoder.up.{block_id}.upsample.conv.weight"
+            ]
+            new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[
+                f"decoder.up.{block_id}.upsample.conv.bias"
+            ]
+
+        paths = renew_vae_resnet_paths(resnets)
+        meta_path = {"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"}
+        assign_to_checkpoint(
+            paths,
+            new_checkpoint,
+            vae_state_dict,
+            additional_replacements=[meta_path],
+            config=config,
+        )
+
+    mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key]
+    num_mid_res_blocks = 2
+    for i in range(1, num_mid_res_blocks + 1):
+        resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key]
+
+        paths = renew_vae_resnet_paths(resnets)
+        meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"}
+        assign_to_checkpoint(
+            paths,
+            new_checkpoint,
+            vae_state_dict,
+            additional_replacements=[meta_path],
+            config=config,
+        )
+
+    mid_attentions = [key for key in vae_state_dict if "decoder.mid.attn" in key]
+    paths = renew_vae_attention_paths(mid_attentions)
+    meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"}
+    assign_to_checkpoint(
+        paths,
+        new_checkpoint,
+        vae_state_dict,
+        additional_replacements=[meta_path],
+        config=config,
+    )
+    conv_attn_to_linear(new_checkpoint)
+    return new_checkpoint
+
+
+def convert_ldm_to_hf_vae(ldm_checkpoint, ldm_config, hf_checkpoint, sample_size):
+    checkpoint = torch.load(ldm_checkpoint)["state_dict"]
+
+    # Convert the VAE model.
+    vae_config = create_vae_diffusers_config(ldm_config)
+    converted_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config)
+
+    vae = AutoencoderKL(**vae_config)
+    vae.load_state_dict(converted_vae_checkpoint)
+    vae.save_pretrained(hf_checkpoint)

config/accelerate_deepspeed.yaml

@@ -0,0 +1,18 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config:
+  gradient_accumulation_steps: 1
+  offload_optimizer_device: cpu
+  offload_param_device: cpu
+  zero3_init_flag: false
+  zero_stage: 2
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+fsdp_config: {}
+machine_rank: 0
+main_process_ip: null
+main_process_port: null
+main_training_function: main
+mixed_precision: 'no'
+num_machines: 1
+num_processes: 1
+use_cpu: false

config/accelerate_local.yaml

@@ -0,0 +1,13 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config: {}
+distributed_type: 'NO'
+downcast_bf16: 'no'
+fsdp_config: {}
+machine_rank: 0
+main_process_ip: null
+main_process_port: null
+main_training_function: main
+mixed_precision: 'no'
+num_machines: 1
+num_processes: 1
+use_cpu: false

config/accelerate_sagemaker.yaml

@@ -0,0 +1,16 @@
+base_job_name: accelerate-sagemaker-1
+compute_environment: AMAZON_SAGEMAKER
+distributed_type: 'NO'
+ec2_instance_type: ml.p3.8xlarge
+iam_role_name: accelerate_sagemaker_execution_role
+image_uri: null
+mixed_precision: 'No'
+num_machines: 1
+profile: default
+py_version: py38
+pytorch_version: 1.10.2
+region: eu-west-2
+sagemaker_inputs_file: null
+sagemaker_metrics_file: null
+transformers_version: 4.17.0
+use_cpu: false

config/ldm_autoencoder_kl.yaml

@@ -0,0 +1,34 @@
+
+# based on https://github.com/CompVis/stable-diffusion/blob/main/configs/autoencoder/autoencoder_kl_32x32x4.yaml
+
+model:
+  base_learning_rate: 4.5e-6
+  target: ldm.models.autoencoder.AutoencoderKL
+  params:
+    monitor: "val/rec_loss"
+    embed_dim: 1  # = in_channels
+    lossconfig:
+      target: ldm.modules.losses.LPIPSWithDiscriminator
+      params:
+        disc_start: 50001
+        kl_weight: 0.000001
+        disc_weight: 0.5
+        disc_in_channels: 1  # = out_ch
+
+    ddconfig:
+      double_z: True
+      z_channels: 1  # must = embed_dim due to HF limitation
+      resolution: 256  # overriden by input image size
+      in_channels: 1
+      out_ch: 1
+      ch: 128
+      ch_mult: [ 1,2,4,4 ]  # num_down = len(ch_mult)-1
+      num_res_blocks: 2
+      attn_resolutions: [ ]
+      dropout: 0.0
+
+lightning:
+  trainer:
+    benchmark: True
+    accelerator: gpu
+    devices: 1

notebooks/audio_diffusion_pipeline.ipynb

@@ -0,0 +1,687 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "fef7e1fb",
+   "metadata": {},
+   "source": [
+    "<a href=\"https://colab.research.google.com/github/teticio/audio-diffusion/blob/master/notebooks/audio_diffusion_pipeline.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "2ada074b",
+   "metadata": {},
+   "source": [
+    "# Audio Diffusion\n",
+    "For training scripts and notebooks visit https://github.com/teticio/audio-diffusion"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6c7800a6",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "try:\n",
+    "    # are we running on Google Colab?\n",
+    "    import google.colab\n",
+    "    %pip install -q diffusers torch librosa datasets\n",
+    "except:\n",
+    "    pass"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c2fc0e7a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import random\n",
+    "import librosa\n",
+    "import numpy as np\n",
+    "from datasets import load_dataset\n",
+    "from IPython.display import Audio\n",
+    "from librosa.beat import beat_track\n",
+    "from diffusers import DiffusionPipeline"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b294a94a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
+    "generator = torch.Generator(device=device)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "f3feb265",
+   "metadata": {},
+   "source": [
+    "## DDPM (De-noising Diffusion Probabilistic Models)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "7fd945bb",
+   "metadata": {},
+   "source": [
+    "### Select model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "97f24046",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#@markdown teticio/audio-diffusion-256                     - trained on my Spotify \"liked\" playlist\n",
+    "\n",
+    "#@markdown teticio/audio-diffusion-breaks-256              - trained on samples used in music\n",
+    "\n",
+    "#@markdown teticio/audio-diffusion-instrumental-hiphop-256 - trained on instrumental hiphop\n",
+    "\n",
+    "model_id = \"teticio/audio-diffusion-256\"  #@param [\"teticio/audio-diffusion-256\", \"teticio/audio-diffusion-breaks-256\", \"audio-diffusion-instrumenal-hiphop-256\", \"teticio/audio-diffusion-ddim-256\"]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a3d45c36",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "audio_diffusion = DiffusionPipeline.from_pretrained(model_id).to(device)\n",
+    "mel = audio_diffusion.mel\n",
+    "sample_rate = mel.get_sample_rate()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ab0d705c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def loop_it(audio: np.ndarray,\n",
+    "        sample_rate: int,\n",
+    "        loops: int = 12) -> np.ndarray:\n",
+    "    \"\"\"Loop audio\n",
+    "\n",
+    "    Args:\n",
+    "        audio (np.ndarray): audio as numpy array\n",
+    "        sample_rate (int): sample rate of audio\n",
+    "        loops (int): number of times to loop\n",
+    "\n",
+    "    Returns:\n",
+    "        (float, np.ndarray): sample rate and raw audio or None\n",
+    "    \"\"\"\n",
+    "    _, beats = beat_track(y=audio, sr=sample_rate, units='samples')\n",
+    "    for beats_in_bar in [16, 12, 8, 4]:\n",
+    "        if len(beats) > beats_in_bar:\n",
+    "            return np.tile(audio[beats[0]:beats[beats_in_bar]], loops)\n",
+    "    return None"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "011fb5a1",
+   "metadata": {},
+   "source": [
+    "### Run model inference to generate mel spectrogram, audios and loops"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b809fed5",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for _ in range(10):\n",
+    "    seed = generator.seed()\n",
+    "    print(f'Seed = {seed}')\n",
+    "    generator.manual_seed(seed)\n",
+    "    output = audio_diffusion(generator=generator)\n",
+    "    image = output.images[0]\n",
+    "    audio = output.audios[0, 0]\n",
+    "    display(image)\n",
+    "    display(Audio(audio, rate=sample_rate))\n",
+    "    loop = loop_it(audio, sample_rate)\n",
+    "    if loop is not None:\n",
+    "        display(Audio(loop, rate=sample_rate))\n",
+    "    else:\n",
+    "        print(\"Unable to determine loop points\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "0bb03e33",
+   "metadata": {},
+   "source": [
+    "### Generate variations of audios"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "80e5b5fa",
+   "metadata": {},
+   "source": [
+    "Try playing around with `start_steps`. Values closer to zero will produce new samples, while values closer to 1,000 will produce samples more faithful to the original."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5074ec11",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "seed = 2391504374279719  #@param {type:\"integer\"}\n",
+    "generator.manual_seed(seed)\n",
+    "output = audio_diffusion(generator=generator)\n",
+    "image = output.images[0]\n",
+    "audio = output.audios[0, 0]\n",
+    "display(image)\n",
+    "display(Audio(audio, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a0fefe28",
+   "metadata": {
+    "scrolled": false
+   },
+   "outputs": [],
+   "source": [
+    "start_step = 500  #@param {type:\"slider\", min:0, max:1000, step:10}\n",
+    "track = loop_it(audio, sample_rate, loops=1)\n",
+    "for variation in range(12):\n",
+    "    output = audio_diffusion(raw_audio=audio, start_step=start_step)\n",
+    "    image2 = output.images[0]\n",
+    "    audio2 = output.audios[0, 0]\n",
+    "    display(image2)\n",
+    "    display(Audio(audio2, rate=sample_rate))\n",
+    "    track = np.concatenate([track, loop_it(audio2, sample_rate, loops=1)])\n",
+    "display(Audio(track, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "58a876c1",
+   "metadata": {},
+   "source": [
+    "### Generate continuations (\"out-painting\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b95d5780",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "overlap_secs = 2  #@param {type:\"integer\"}\n",
+    "start_step = 0  #@param {type:\"slider\", min:0, max:1000, step:10}\n",
+    "overlap_samples = overlap_secs * sample_rate\n",
+    "track = audio\n",
+    "for variation in range(12):\n",
+    "    output = audio_diffusion(raw_audio=audio[-overlap_samples:],\n",
+    "                             start_step=start_step,\n",
+    "                             mask_start_secs=overlap_secs)\n",
+    "    image2 = output.images[0]\n",
+    "    audio2 = output.audios[0, 0]\n",
+    "    display(image2)\n",
+    "    display(Audio(audio2, rate=sample_rate))\n",
+    "    track = np.concatenate([track, audio2[overlap_samples:]])\n",
+    "    audio = audio2\n",
+    "display(Audio(track, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b6434d3f",
+   "metadata": {},
+   "source": [
+    "### Remix (style transfer)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "0da030b2",
+   "metadata": {},
+   "source": [
+    "Alternatively, you can start from another audio altogether, resulting in a kind of style transfer. Maintaining the same seed during generation fixes the style, while masking helps stitch consecutive segments together more smoothly."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "fc620a80",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "try:\n",
+    "    # are we running on Google Colab?\n",
+    "    from google.colab import files\n",
+    "    audio_file = list(files.upload().keys())[0]\n",
+    "except:\n",
+    "    audio_file = \"/home/teticio/Music/liked/El Michels Affair - Glaciers Of Ice.mp3\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5a257e69",
+   "metadata": {
+    "scrolled": false
+   },
+   "outputs": [],
+   "source": [
+    "start_step = 500  #@param {type:\"slider\", min:0, max:1000, step:10}\n",
+    "overlap_secs = 2  #@param {type:\"integer\"}\n",
+    "track_audio, _ = librosa.load(audio_file, mono=True, sr=sample_rate)\n",
+    "overlap_samples = overlap_secs * sample_rate\n",
+    "slice_size = mel.x_res * mel.hop_length\n",
+    "stride = slice_size - overlap_samples\n",
+    "generator = torch.Generator(device=device)\n",
+    "seed = generator.seed()\n",
+    "print(f'Seed = {seed}')\n",
+    "track = np.array([])\n",
+    "not_first = 0\n",
+    "for sample in range(len(track_audio) // stride):\n",
+    "    generator.manual_seed(seed)\n",
+    "    audio = np.array(track_audio[sample * stride:sample * stride + slice_size])\n",
+    "    if not_first:\n",
+    "        # Normalize and re-insert generated audio\n",
+    "        audio[:overlap_samples] = audio2[-overlap_samples:] * np.max(\n",
+    "            audio[:overlap_samples]) / np.max(audio2[-overlap_samples:])\n",
+    "    output = audio_diffusion(raw_audio=audio,\n",
+    "                             start_step=start_step,\n",
+    "                             generator=generator,\n",
+    "                             mask_start_secs=overlap_secs * not_first)\n",
+    "    audio2 = output.audios[0, 0]\n",
+    "    track = np.concatenate([track, audio2[overlap_samples * not_first:]])\n",
+    "    not_first = 1\n",
+    "    display(Audio(track, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "924ff9d5",
+   "metadata": {},
+   "source": [
+    "### Fill the gap (\"in-painting\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "0200264c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "sample = 3  #@param {type:\"integer\"}\n",
+    "raw_audio = track_audio[sample * stride:sample * stride + slice_size]\n",
+    "output = audio_diffusion(raw_audio=raw_audio,\n",
+    "                         mask_start_secs=1,\n",
+    "                         mask_end_secs=1,\n",
+    "                         step_generator=torch.Generator(device=device))\n",
+    "audio2 = output.audios[0, 0]\n",
+    "display(Audio(audio, rate=sample_rate))\n",
+    "display(Audio(audio2, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "efc32dae",
+   "metadata": {},
+   "source": [
+    "## DDIM (De-noising Diffusion Implicit Models)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a021f78a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "audio_diffusion = DiffusionPipeline.from_pretrained('teticio/audio-diffusion-ddim-256').to(device)\n",
+    "mel = audio_diffusion.mel\n",
+    "sample_rate = mel.get_sample_rate()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "deb23339",
+   "metadata": {},
+   "source": [
+    "### Generation can be done in many fewer steps with DDIMs"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c105a497",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for _ in range(10):\n",
+    "    seed = generator.seed()\n",
+    "    print(f'Seed = {seed}')\n",
+    "    generator.manual_seed(seed)\n",
+    "    output = audio_diffusion(generator=generator)\n",
+    "    image = output.images[0]\n",
+    "    audio = output.audios[0, 0]\n",
+    "    display(image)\n",
+    "    display(Audio(audio, rate=sample_rate))\n",
+    "    loop = loop_it(audio, sample_rate)\n",
+    "    if loop is not None:\n",
+    "        display(Audio(loop, rate=sample_rate))\n",
+    "    else:\n",
+    "        print(\"Unable to determine loop points\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "cab4692c",
+   "metadata": {},
+   "source": [
+    "The parameter eta controls the variance:\n",
+    "* 0 - DDIM (deterministic)\n",
+    "* 1 - DDPM (De-noising Diffusion Probabilistic Model)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "72bdd207",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "output = audio_diffusion(steps=1000, generator=generator, eta=1)\n",
+    "image = output.images[0]\n",
+    "audio = output.audios[0, 0]\n",
+    "display(image)\n",
+    "display(Audio(audio, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b8d5442c",
+   "metadata": {},
+   "source": [
+    "### DDIMs can be used as encoders..."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "269ee816",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Doesn't have to be an audio from the train dataset, this is just for convenience\n",
+    "ds = load_dataset('teticio/audio-diffusion-256')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "278d1d80",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image = ds['train'][264]['image']\n",
+    "display(Audio(mel.image_to_audio(image), rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "912b54e4",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "noise = audio_diffusion.encode([image])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c7b31f97",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Reconstruct original audio from noise\n",
+    "output = audio_diffusion(noise=noise, generator=generator)\n",
+    "image = output.images[0]\n",
+    "audio = output.audios[0, 0]\n",
+    "display(Audio(audio, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "998c776b",
+   "metadata": {},
+   "source": [
+    "### ...or to interpolate between audios"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "33f82367",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image2 = ds['train'][15978]['image']\n",
+    "display(Audio(mel.image_to_audio(image2), rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f93fb6c0",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "noise2 = audio_diffusion.encode([image2])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a4190563",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "alpha = 0.5  #@param {type:\"slider\", min:0, max:1, step:0.1}\n",
+    "output = audio_diffusion(\n",
+    "    noise=audio_diffusion.slerp(noise, noise2, alpha),\n",
+    "    generator=generator)\n",
+    "audio = output.audios[0, 0]\n",
+    "display(Audio(mel.image_to_audio(image), rate=sample_rate))\n",
+    "display(Audio(mel.image_to_audio(image2), rate=sample_rate))\n",
+    "display(Audio(audio, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "9b244547",
+   "metadata": {},
+   "source": [
+    "## Latent Audio Diffusion\n",
+    "Instead of de-noising images directly in the pixel space, we can work in the latent space of a pre-trained VAE (Variational AutoEncoder). This is much faster to train and run inference on, although the quality suffers as there are now three stages involved in encoding / decoding: mel spectrogram, VAE and de-noising."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a88b3fbb",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "model_id = \"teticio/latent-audio-diffusion-ddim-256\"  #@param [\"teticio/latent-audio-diffusion-256\", \"teticio/latent-audio-diffusion-ddim-256\"]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "15e353ee",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "audio_diffusion = DiffusionPipeline.from_pretrained(model_id).to(device)\n",
+    "mel = audio_diffusion.mel\n",
+    "sample_rate = mel.get_sample_rate()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "fa0f0c8c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "seed = 3412253600050855  #@param {type:\"integer\"}\n",
+    "generator.manual_seed(seed)\n",
+    "output = audio_diffusion(generator=generator)\n",
+    "image = output.images[0]\n",
+    "audio = output.audios[0, 0]\n",
+    "display(image)\n",
+    "display(Audio(audio, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "73dc575d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "seed2 = 7016114633369557  #@param {type:\"integer\"}\n",
+    "generator.manual_seed(seed2)\n",
+    "output = audio_diffusion(generator=generator)\n",
+    "image2 = output.images[0]\n",
+    "audio2 = output.audios[0, 0]\n",
+    "display(image2)\n",
+    "display(Audio(audio2, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "428d2d67",
+   "metadata": {},
+   "source": [
+    "### Interpolation in latent space\n",
+    "As the VAE forces a more compact, lower dimensional representation for the spectrograms, interpolation in latent space can lead to meaningful combinations of audios. In combination with the (deterministic) DDIM from the previous section, the model can be used as an encoder / decoder to a lower dimensional space."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "72211c2b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "generator.manual_seed(seed)\n",
+    "latents = torch.randn(\n",
+    "    (1, audio_diffusion.unet.in_channels, audio_diffusion.unet.sample_size[0],\n",
+    "     audio_diffusion.unet.sample_size[1]),\n",
+    "    generator=generator, device=device)\n",
+    "latents.shape"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6c732dbe",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "generator.manual_seed(seed2)\n",
+    "latents2 = torch.randn(\n",
+    "    (1, audio_diffusion.unet.in_channels, audio_diffusion.unet.sample_size[0],\n",
+    "     audio_diffusion.unet.sample_size[1]),\n",
+    "    generator=generator,\n",
+    "    device=device)\n",
+    "latents2.shape"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "159bcfc4",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "alpha = 0.5  #@param {type:\"slider\", min:0, max:1, step:0.1}\n",
+    "output = audio_diffusion(\n",
+    "    noise=audio_diffusion.slerp(latents, latents2, alpha),\n",
+    "    generator=generator)\n",
+    "audio3 = output.audios[0, 0]\n",
+    "display(Audio(audio, rate=mel.get_sample_rate()))\n",
+    "display(Audio(audio2, rate=mel.get_sample_rate()))\n",
+    "display(Audio(audio3, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ce6c9cc1",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "accelerator": "GPU",
+  "colab": {
+   "provenance": []
+  },
+  "gpuClass": "standard",
+  "kernelspec": {
+   "display_name": "huggingface",
+   "language": "python",
+   "name": "huggingface"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.6"
+  },
+  "toc": {
+   "base_numbering": 1,
+   "nav_menu": {},
+   "number_sections": true,
+   "sideBar": true,
+   "skip_h1_title": false,
+   "title_cell": "Table of Contents",
+   "title_sidebar": "Contents",
+   "toc_cell": false,
+   "toc_position": {},
+   "toc_section_display": true,
+   "toc_window_display": false
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

notebooks/audio_encoder.ipynb

@@ -0,0 +1,68 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "592fff30",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from audiodiffusion.audio_encoder import AudioEncoder"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d99ef523",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "audio_encoder = AudioEncoder.from_pretrained(\"teticio/audio-encoder\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4eb3bbd7",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "audio_encoder.encode(['/home/teticio/Music/liked/Agua Re - Holy Dance - Large Sound Mix.mp3'])"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "huggingface",
+   "language": "python",
+   "name": "huggingface"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.6"
+  },
+  "toc": {
+   "base_numbering": 1,
+   "nav_menu": {},
+   "number_sections": true,
+   "sideBar": true,
+   "skip_h1_title": false,
+   "title_cell": "Table of Contents",
+   "title_sidebar": "Contents",
+   "toc_cell": false,
+   "toc_position": {},
+   "toc_section_display": true,
+   "toc_window_display": false
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

notebooks/conditional_generation.ipynb

@@ -0,0 +1,226 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "5f6c6cc2",
+   "metadata": {},
+   "source": [
+    "<a href=\"https://colab.research.google.com/github/teticio/audio-diffusion/blob/master/notebooks/conditional_generation.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f1935544",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "try:\n",
+    "    # are we running on Google Colab?\n",
+    "    import google.colab\n",
+    "    !git clone -q https://github.com/teticio/audio-diffusion.git\n",
+    "    %cd audio-diffusion\n",
+    "    %pip install -q -r requirements.txt\n",
+    "except:\n",
+    "    pass"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b0e656c9",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import os\n",
+    "import sys\n",
+    "sys.path.insert(0, os.path.dirname(os.path.abspath(\"\")))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d448b299",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import urllib\n",
+    "import requests\n",
+    "from IPython.display import Audio\n",
+    "from audiodiffusion import AudioDiffusion\n",
+    "from audiodiffusion.audio_encoder import AudioEncoder"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f1548971",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
+    "generator = torch.Generator(device=device)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "056f179c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "audio_diffusion = AudioDiffusion(model_id=\"teticio/conditional-latent-audio-diffusion-512\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b4a08500",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "audio_encoder = AudioEncoder.from_pretrained(\"teticio/audio-encoder\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "387550ac",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Uncomment for faster (but slightly lower quality) generation\n",
+    "#from diffusers import DDIMScheduler\n",
+    "#audio_diffusion.pipe.scheduler = DDIMScheduler()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "9936a72f",
+   "metadata": {},
+   "source": [
+    "## Download and encode preview track from Spotify"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "57a9b134",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Get temporary API credentials\n",
+    "credentials = requests.get(\n",
+    "    \"https://open.spotify.com/get_access_token?reason=transport&productType=embed\"\n",
+    ").json()\n",
+    "headers = {\n",
+    "    \"Accept\": \"application/json\",\n",
+    "    \"Content-Type\": \"application/json\",\n",
+    "    \"Authorization\": \"Bearer \" + credentials[\"accessToken\"]\n",
+    "}\n",
+    "\n",
+    "# Search for tracks\n",
+    "search_string = input(\"Search: \")\n",
+    "response = requests.get(\n",
+    "    f\"https://api.spotify.com/v1/search?q={urllib.parse.quote(search_string)}&type=track\",\n",
+    "    headers=headers).json()\n",
+    "\n",
+    "# List results\n",
+    "for _, track in enumerate(response[\"tracks\"][\"items\"]):\n",
+    "    print(f\"{_ + 1}. {track['artists'][0]['name']} - {track['name']}\")\n",
+    "selection = input(\"Select a track: \")\n",
+    "\n",
+    "# Download and encode selection\n",
+    "r = requests.get(response[\"tracks\"][\"items\"][int(selection) -\n",
+    "                                             1][\"preview_url\"],\n",
+    "                 stream=True)\n",
+    "with open(\"temp.mp3\", \"wb\") as f:\n",
+    "    for chunk in r:\n",
+    "        f.write(chunk)\n",
+    "encoding = torch.unsqueeze(audio_encoder.encode([\"temp.mp3\"]),\n",
+    "                           axis=1).to(device)\n",
+    "os.remove(\"temp.mp3\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "8af863f5",
+   "metadata": {},
+   "source": [
+    "## Conditional Generation\n",
+    "Bear in mind that the generative model can only generate music similar to that on which it was trained. The audio encoding will influence the generation within those limitations."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "8f119ddd",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for _ in range(10):\n",
+    "    seed = generator.seed()\n",
+    "    print(f'Seed = {seed}')\n",
+    "    generator.manual_seed(seed)\n",
+    "    image, (sample_rate,\n",
+    "            audio) = audio_diffusion.generate_spectrogram_and_audio(\n",
+    "                generator=generator, encoding=encoding)\n",
+    "    display(image)\n",
+    "    display(Audio(audio, rate=sample_rate))\n",
+    "    loop = AudioDiffusion.loop_it(audio, sample_rate)\n",
+    "    if loop is not None:\n",
+    "        display(Audio(loop, rate=sample_rate))\n",
+    "    else:\n",
+    "        print(\"Unable to determine loop points\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d0bd18c0",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "accelerator": "GPU",
+  "colab": {
+   "provenance": []
+  },
+  "gpuClass": "standard",
+  "kernelspec": {
+   "display_name": "huggingface",
+   "language": "python",
+   "name": "huggingface"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.6"
+  },
+  "toc": {
+   "base_numbering": 1,
+   "nav_menu": {},
+   "number_sections": true,
+   "sideBar": true,
+   "skip_h1_title": false,
+   "title_cell": "Table of Contents",
+   "title_sidebar": "Contents",
+   "toc_cell": false,
+   "toc_position": {},
+   "toc_section_display": true,
+   "toc_window_display": false
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

notebooks/gradio_app.ipynb

@@ -0,0 +1,101 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "a489aa44",
+   "metadata": {},
+   "source": [
+    "<a href=\"https://colab.research.google.com/github/teticio/audio-diffusion/blob/master/notebooks/gradio_app.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "9502ffa7",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "try:\n",
+    "    # are we running on Google Colab?\n",
+    "    import google.colab\n",
+    "    !git clone -q https://github.com/teticio/audio-diffusion.git\n",
+    "    %cd audio-diffusion\n",
+    "    %pip install -q -r requirements.txt\n",
+    "except:\n",
+    "    pass"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "8f8b6e43",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import os\n",
+    "import sys\n",
+    "sys.path.insert(0, os.path.dirname(os.path.abspath(\"\")))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "2d948967",
+   "metadata": {
+    "scrolled": false
+   },
+   "outputs": [],
+   "source": [
+    "import app\n",
+    "app.demo.launch(share=True);"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "46f03607",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "accelerator": "GPU",
+  "colab": {
+   "provenance": []
+  },
+  "gpuClass": "standard",
+  "kernelspec": {
+   "display_name": "huggingface",
+   "language": "python",
+   "name": "huggingface"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.6"
+  },
+  "toc": {
+   "base_numbering": 1,
+   "nav_menu": {},
+   "number_sections": true,
+   "sideBar": true,
+   "skip_h1_title": false,
+   "title_cell": "Table of Contents",
+   "title_sidebar": "Contents",
+   "toc_cell": false,
+   "toc_position": {},
+   "toc_section_display": true,
+   "toc_window_display": false
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

notebooks/test_mel.ipynb

@@ -0,0 +1,220 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "3de63898",
+   "metadata": {},
+   "source": [
+    "<a href=\"https://colab.research.google.com/github/teticio/audio-diffusion/blob/master/notebooks/test_mel.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "81fbd495",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "try:\n",
+    "    # are we running on Google Colab?\n",
+    "    import google.colab\n",
+    "    !git clone -q https://github.com/teticio/audio-diffusion.git\n",
+    "    %cd audio-diffusion\n",
+    "    %pip install -q -r requirements.txt\n",
+    "except:\n",
+    "    pass"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "218fcdf1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from datasets import load_dataset\n",
+    "from IPython.display import Audio\n",
+    "from diffusers import Mel"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5e4f8ee5",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# These are the default parameters. If you change any of them, you may have to adjust the others.\n",
+    "mel = Mel(x_res=256,\n",
+    "          y_res=256,\n",
+    "          hop_length=512,\n",
+    "          sample_rate=22050,\n",
+    "          n_fft=2048,\n",
+    "          n_iter=32)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b2178c3f",
+   "metadata": {},
+   "source": [
+    "### Transform slice of audio to mel spectrogram"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f32bb35e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "try:\n",
+    "    # are we running on Google Colab?\n",
+    "    from google.colab import files\n",
+    "    audio_file = list(files.upload().keys())[0]\n",
+    "except:\n",
+    "    audio_file = \"/home/teticio/Music/Music/A Tribe Called Quest/The Anthology/08 Can I Kick It_.mp3\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "61dbcd2e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "mel.load_audio(audio_file)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ccadcc0f",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image = mel.audio_slice_to_image(15)\n",
+    "image"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "8cec79c6",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image.width, image.height"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "fe112fef",
+   "metadata": {},
+   "source": [
+    "### Transform mel spectrogram back to audio"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "0b268a54",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "audio = mel.image_to_audio(image)\n",
+    "Audio(data=audio, rate=mel.get_sample_rate())"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "0f1f2006",
+   "metadata": {},
+   "source": [
+    "### Select a random image from the training set"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1f29f025",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "ds = load_dataset('teticio/audio-diffusion-256')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e002482d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image = ds['train'].shuffle().select(range(1))['image'][0]\n",
+    "image"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "6a801fc5",
+   "metadata": {},
+   "source": [
+    "### Convert to audio"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a2421827",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "audio = mel.image_to_audio(image)\n",
+    "Audio(data=audio, rate=mel.get_sample_rate())"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "2281fb55",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "huggingface",
+   "language": "python",
+   "name": "huggingface"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.6"
+  },
+  "toc": {
+   "base_numbering": 1,
+   "nav_menu": {},
+   "number_sections": true,
+   "sideBar": true,
+   "skip_h1_title": false,
+   "title_cell": "Table of Contents",
+   "title_sidebar": "Contents",
+   "toc_cell": false,
+   "toc_position": {},
+   "toc_section_display": true,
+   "toc_window_display": false
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

notebooks/test_model.ipynb

@@ -0,0 +1,660 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "62c5865f",
+   "metadata": {},
+   "source": [
+    "<a href=\"https://colab.research.google.com/github/teticio/audio-diffusion/blob/master/notebooks/test_model.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6c7800a6",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "try:\n",
+    "    # are we running on Google Colab?\n",
+    "    import google.colab\n",
+    "    !git clone -q https://github.com/teticio/audio-diffusion.git\n",
+    "    %cd audio-diffusion\n",
+    "    %pip install -q -r requirements.txt\n",
+    "except:\n",
+    "    pass"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b447e2c4",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import os\n",
+    "import sys\n",
+    "sys.path.insert(0, os.path.dirname(os.path.abspath(\"\")))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c2fc0e7a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import random\n",
+    "import librosa\n",
+    "import numpy as np\n",
+    "from datasets import load_dataset\n",
+    "from IPython.display import Audio\n",
+    "from audiodiffusion import AudioDiffusion"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b294a94a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
+    "generator = torch.Generator(device=device)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "f3feb265",
+   "metadata": {},
+   "source": [
+    "## DDPM (De-noising Diffusion Probabilistic Models)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "7fd945bb",
+   "metadata": {},
+   "source": [
+    "### Select model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "97f24046",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#@markdown teticio/audio-diffusion-256                     - trained on my Spotify \"liked\" playlist\n",
+    "\n",
+    "#@markdown teticio/audio-diffusion-breaks-256              - trained on samples used in music\n",
+    "\n",
+    "#@markdown teticio/audio-diffusion-instrumental-hiphop-256 - trained on instrumental hiphop\n",
+    "\n",
+    "model_id = \"teticio/audio-diffusion-256\"  #@param [\"teticio/audio-diffusion-256\", \"teticio/audio-diffusion-breaks-256\", \"audio-diffusion-instrumenal-hiphop-256\", \"teticio/audio-diffusion-ddim-256\"]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a3d45c36",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "audio_diffusion = AudioDiffusion(model_id=model_id)\n",
+    "mel = audio_diffusion.pipe.mel"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "011fb5a1",
+   "metadata": {},
+   "source": [
+    "### Run model inference to generate mel spectrogram, audios and loops"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b809fed5",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for _ in range(10):\n",
+    "    seed = generator.seed()\n",
+    "    print(f'Seed = {seed}')\n",
+    "    generator.manual_seed(seed)\n",
+    "    image, (sample_rate,\n",
+    "            audio) = audio_diffusion.generate_spectrogram_and_audio(\n",
+    "                generator=generator)\n",
+    "    display(image)\n",
+    "    display(Audio(audio, rate=sample_rate))\n",
+    "    loop = AudioDiffusion.loop_it(audio, sample_rate)\n",
+    "    if loop is not None:\n",
+    "        display(Audio(loop, rate=sample_rate))\n",
+    "    else:\n",
+    "        print(\"Unable to determine loop points\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "0bb03e33",
+   "metadata": {},
+   "source": [
+    "### Generate variations of audios"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "80e5b5fa",
+   "metadata": {},
+   "source": [
+    "Try playing around with `start_steps`. Values closer to zero will produce new samples, while values closer to 1,000 will produce samples more faithful to the original."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5074ec11",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "seed = 2391504374279719  #@param {type:\"integer\"}\n",
+    "generator.manual_seed(seed)\n",
+    "image, (sample_rate, audio) = audio_diffusion.generate_spectrogram_and_audio(\n",
+    "    generator=generator)\n",
+    "display(image)\n",
+    "display(Audio(audio, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a0fefe28",
+   "metadata": {
+    "scrolled": false
+   },
+   "outputs": [],
+   "source": [
+    "start_step = 500  #@param {type:\"slider\", min:0, max:1000, step:10}\n",
+    "track = AudioDiffusion.loop_it(audio, sample_rate, loops=1)\n",
+    "for variation in range(12):\n",
+    "    image2, (\n",
+    "        sample_rate,\n",
+    "        audio2) = audio_diffusion.generate_spectrogram_and_audio_from_audio(\n",
+    "            raw_audio=audio, start_step=start_step)\n",
+    "    display(image2)\n",
+    "    display(Audio(audio2, rate=sample_rate))\n",
+    "    track = np.concatenate(\n",
+    "        [track, AudioDiffusion.loop_it(audio2, sample_rate, loops=1)])\n",
+    "display(Audio(track, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "58a876c1",
+   "metadata": {},
+   "source": [
+    "### Generate continuations (\"out-painting\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b95d5780",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "overlap_secs = 2  #@param {type:\"integer\"}\n",
+    "start_step = 0  #@param {type:\"slider\", min:0, max:1000, step:10}\n",
+    "overlap_samples = overlap_secs * sample_rate\n",
+    "track = audio\n",
+    "for variation in range(12):\n",
+    "    image2, (\n",
+    "        sample_rate,\n",
+    "        audio2) = audio_diffusion.generate_spectrogram_and_audio_from_audio(\n",
+    "            raw_audio=audio[-overlap_samples:],\n",
+    "            start_step=start_step,\n",
+    "            mask_start_secs=overlap_secs)\n",
+    "    display(image2)\n",
+    "    display(Audio(audio2, rate=sample_rate))\n",
+    "    track = np.concatenate([track, audio2[overlap_samples:]])\n",
+    "    audio = audio2\n",
+    "display(Audio(track, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b6434d3f",
+   "metadata": {},
+   "source": [
+    "### Remix (style transfer)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "0da030b2",
+   "metadata": {},
+   "source": [
+    "Alternatively, you can start from another audio altogether, resulting in a kind of style transfer. Maintaining the same seed during generation fixes the style, while masking helps stitch consecutive segments together more smoothly."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "fc620a80",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "try:\n",
+    "    # are we running on Google Colab?\n",
+    "    from google.colab import files\n",
+    "    audio_file = list(files.upload().keys())[0]\n",
+    "except:\n",
+    "    audio_file = \"/home/teticio/Music/liked/El Michels Affair - Glaciers Of Ice.mp3\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5a257e69",
+   "metadata": {
+    "scrolled": false
+   },
+   "outputs": [],
+   "source": [
+    "start_step = 500  #@param {type:\"slider\", min:0, max:1000, step:10}\n",
+    "overlap_secs = 2  #@param {type:\"integer\"}\n",
+    "track_audio, _ = librosa.load(audio_file, mono=True, sr=mel.get_sample_rate())\n",
+    "overlap_samples = overlap_secs * sample_rate\n",
+    "slice_size = mel.x_res * mel.hop_length\n",
+    "stride = slice_size - overlap_samples\n",
+    "generator = torch.Generator(device=device)\n",
+    "seed = generator.seed()\n",
+    "print(f'Seed = {seed}')\n",
+    "track = np.array([])\n",
+    "not_first = 0\n",
+    "for sample in range(len(track_audio) // stride):\n",
+    "    generator.manual_seed(seed)\n",
+    "    audio = np.array(track_audio[sample * stride:sample * stride + slice_size])\n",
+    "    if not_first:\n",
+    "        # Normalize and re-insert generated audio\n",
+    "        audio[:overlap_samples] = audio2[-overlap_samples:] * np.max(\n",
+    "            audio[:overlap_samples]) / np.max(audio2[-overlap_samples:])\n",
+    "    _, (sample_rate,\n",
+    "        audio2) = audio_diffusion.generate_spectrogram_and_audio_from_audio(\n",
+    "            raw_audio=audio,\n",
+    "            start_step=start_step,\n",
+    "            generator=generator,\n",
+    "            mask_start_secs=overlap_secs * not_first)\n",
+    "    track = np.concatenate([track, audio2[overlap_samples * not_first:]])\n",
+    "    not_first = 1\n",
+    "    display(Audio(track, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "924ff9d5",
+   "metadata": {},
+   "source": [
+    "### Fill the gap (\"in-painting\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "0200264c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "slice = 3  #@param {type:\"integer\"}\n",
+    "raw_audio = track_audio[sample * stride:sample * stride + slice_size]\n",
+    "_, (sample_rate,\n",
+    "    audio2) = audio_diffusion.generate_spectrogram_and_audio_from_audio(\n",
+    "        raw_audio=raw_audio,\n",
+    "        mask_start_secs=1,\n",
+    "        mask_end_secs=1,\n",
+    "        step_generator=torch.Generator(device=device))\n",
+    "display(Audio(audio, rate=sample_rate))\n",
+    "display(Audio(audio2, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "efc32dae",
+   "metadata": {},
+   "source": [
+    "## DDIM (De-noising Diffusion Implicit Models)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a021f78a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "audio_diffusion = AudioDiffusion(model_id='teticio/audio-diffusion-ddim-256')\n",
+    "mel = audio_diffusion.pipe.mel"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "deb23339",
+   "metadata": {},
+   "source": [
+    "### Generation can be done in many fewer steps with DDIMs"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c105a497",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for _ in range(10):\n",
+    "    seed = generator.seed()\n",
+    "    print(f'Seed = {seed}')\n",
+    "    generator.manual_seed(seed)\n",
+    "    image, (sample_rate,\n",
+    "            audio) = audio_diffusion.generate_spectrogram_and_audio(\n",
+    "                generator=generator)\n",
+    "    display(image)\n",
+    "    display(Audio(audio, rate=sample_rate))\n",
+    "    loop = AudioDiffusion.loop_it(audio, sample_rate)\n",
+    "    if loop is not None:\n",
+    "        display(Audio(loop, rate=sample_rate))\n",
+    "    else:\n",
+    "        print(\"Unable to determine loop points\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "cab4692c",
+   "metadata": {},
+   "source": [
+    "The parameter eta controls the variance:\n",
+    "* 0 - DDIM (deterministic)\n",
+    "* 1 - DDPM (De-noising Diffusion Probabilistic Model)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "72bdd207",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image, (sample_rate, audio) = audio_diffusion.generate_spectrogram_and_audio(\n",
+    "    steps=1000, generator=generator, eta=1)\n",
+    "display(image)\n",
+    "display(Audio(audio, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b8d5442c",
+   "metadata": {},
+   "source": [
+    "### DDIMs can be used as encoders..."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "269ee816",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Doesn't have to be an audio from the train dataset, this is just for convenience\n",
+    "ds = load_dataset('teticio/audio-diffusion-256')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "278d1d80",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image = ds['train'][264]['image']\n",
+    "display(Audio(mel.image_to_audio(image), rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "912b54e4",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "noise = audio_diffusion.pipe.encode([image])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c7b31f97",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Reconstruct original audio from noise\n",
+    "_, (sample_rate, audio) = audio_diffusion.generate_spectrogram_and_audio(\n",
+    "    noise=noise, generator=generator)\n",
+    "display(Audio(audio, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "998c776b",
+   "metadata": {},
+   "source": [
+    "### ...or to interpolate between audios"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "33f82367",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image2 = ds['train'][15978]['image']\n",
+    "display(Audio(mel.image_to_audio(image2), rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f93fb6c0",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "noise2 = audio_diffusion.pipe.encode([image2])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a4190563",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "alpha = 0.5  #@param {type:\"slider\", min:0, max:1, step:0.1}\n",
+    "_, (sample_rate, audio) = audio_diffusion.generate_spectrogram_and_audio(\n",
+    "    noise=audio_diffusion.pipe.slerp(noise, noise2, alpha),\n",
+    "    generator=generator)\n",
+    "display(Audio(mel.image_to_audio(image), rate=sample_rate))\n",
+    "display(Audio(mel.image_to_audio(image2), rate=sample_rate))\n",
+    "display(Audio(audio, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "9b244547",
+   "metadata": {},
+   "source": [
+    "## Latent Audio Diffusion\n",
+    "Instead of de-noising images directly in the pixel space, we can work in the latent space of a pre-trained VAE (Variational AutoEncoder). This is much faster to train and run inference on, although the quality suffers as there are now three stages involved in encoding / decoding: mel spectrogram, VAE and de-noising."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a88b3fbb",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "model_id = \"teticio/latent-audio-diffusion-ddim-256\"  #@param [\"teticio/latent-audio-diffusion-256\", \"teticio/latent-audio-diffusion-ddim-256\"]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "15e353ee",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "audio_diffusion = AudioDiffusion(model_id=model_id)\n",
+    "mel = audio_diffusion.pipe.mel"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "fa0f0c8c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "seed = 3412253600050855  #@param {type:\"integer\"}\n",
+    "generator.manual_seed(seed)\n",
+    "image, (sample_rate, audio) = audio_diffusion.generate_spectrogram_and_audio(\n",
+    "    generator=generator)\n",
+    "display(image)\n",
+    "display(Audio(audio, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "73dc575d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "seed2 = 7016114633369557  #@param {type:\"integer\"}\n",
+    "generator.manual_seed(seed2)\n",
+    "image2, (sample_rate, audio2) = audio_diffusion.generate_spectrogram_and_audio(\n",
+    "    generator=generator)\n",
+    "display(image2)\n",
+    "display(Audio(audio2, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "428d2d67",
+   "metadata": {},
+   "source": [
+    "### Interpolation in latent space\n",
+    "As the VAE forces a more compact, lower dimensional representation for the spectrograms, interpolation in latent space can lead to meaningful combinations of audios. In combination with the (deterministic) DDIM from the previous section, the model can be used as an encoder / decoder to a lower dimensional space."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "72211c2b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "generator.manual_seed(seed)\n",
+    "latents = torch.randn((1, audio_diffusion.pipe.unet.in_channels,\n",
+    "                       audio_diffusion.pipe.unet.sample_size[0],\n",
+    "                       audio_diffusion.pipe.unet.sample_size[1]),\n",
+    "                      device=device,\n",
+    "                      generator=generator)\n",
+    "latents.shape"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6c732dbe",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "generator.manual_seed(seed2)\n",
+    "latents2 = torch.randn((1, audio_diffusion.pipe.unet.in_channels,\n",
+    "                        audio_diffusion.pipe.unet.sample_size[0],\n",
+    "                        audio_diffusion.pipe.unet.sample_size[1]),\n",
+    "                       device=device,\n",
+    "                       generator=generator)\n",
+    "latents2.shape"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "159bcfc4",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "alpha = 0.5  #@param {type:\"slider\", min:0, max:1, step:0.1}\n",
+    "_, (sample_rate, audio3) = audio_diffusion.generate_spectrogram_and_audio(\n",
+    "    noise=audio_diffusion.pipe.slerp(latents, latents2, alpha),\n",
+    "    generator=generator)\n",
+    "display(Audio(audio, rate=sample_rate))\n",
+    "display(Audio(audio2, rate=sample_rate))\n",
+    "display(Audio(audio3, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ce6c9cc1",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "accelerator": "GPU",
+  "colab": {
+   "provenance": []
+  },
+  "gpuClass": "standard",
+  "kernelspec": {
+   "display_name": "huggingface",
+   "language": "python",
+   "name": "huggingface"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.6"
+  },
+  "toc": {
+   "base_numbering": 1,
+   "nav_menu": {},
+   "number_sections": true,
+   "sideBar": true,
+   "skip_h1_title": false,
+   "title_cell": "Table of Contents",
+   "title_sidebar": "Contents",
+   "toc_cell": false,
+   "toc_position": {},
+   "toc_section_display": true,
+   "toc_window_display": false
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

notebooks/test_vae.ipynb

@@ -0,0 +1,233 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b451ab22",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import random\n",
+    "import numpy as np\n",
+    "from PIL import Image\n",
+    "from datasets import load_dataset\n",
+    "from IPython.display import Audio\n",
+    "from diffusers import AutoencoderKL, AudioDiffusionPipeline, Mel"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "324cef44",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "mel = Mel()\n",
+    "vae = AutoencoderKL.from_pretrained('../models/autoencoder-kl')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "da55ce79",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "vae.config"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5fea99ff",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "ds = load_dataset('teticio/audio-diffusion-256')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "3a65ec4d",
+   "metadata": {},
+   "source": [
+    "### Reconstruct audio"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "426c6edd",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image = random.choice(ds['train'])['image']\n",
+    "display(image)\n",
+    "Audio(data=mel.image_to_audio(image), rate=mel.get_sample_rate())"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "29c9011d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# encode\n",
+    "input_image = np.frombuffer(image.tobytes(), dtype=\"uint8\").reshape(\n",
+    "    (image.height, image.width, 1))\n",
+    "input_image = ((input_image / 255) * 2 - 1).transpose(2, 0, 1)\n",
+    "posterior = vae.encode(torch.tensor([input_image],\n",
+    "                                    dtype=torch.float32)).latent_dist\n",
+    "latents = posterior.sample()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "323ba46d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# reconstruct\n",
+    "output_image = vae.decode(latents)['sample']\n",
+    "output_image = torch.clamp(output_image, -1., 1.)\n",
+    "output_image = (output_image + 1.0) / 2.0  # -1,1 -> 0,1; c,h,w\n",
+    "output_image = (output_image.detach().cpu().numpy() *\n",
+    "                255).round().astype(\"uint8\").transpose(0, 2, 3, 1)[0, :, :, 0]\n",
+    "output_image = Image.fromarray(output_image)\n",
+    "display(output_image)\n",
+    "Audio(data=mel.image_to_audio(output_image), rate=mel.get_sample_rate())"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "00ff2ffa",
+   "metadata": {},
+   "source": [
+    "### Random sample from latent space\n",
+    "(Don't expect interesting results!)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "156a06a2",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# sample\n",
+    "output_image = vae.decode(torch.randn_like(latents))['sample']\n",
+    "output_image = torch.clamp(output_image, -1., 1.)\n",
+    "output_image = (output_image + 1.0) / 2.0  # -1,1 -> 0,1; c,h,w\n",
+    "output_image = (output_image.detach().cpu().numpy() *\n",
+    "                255).round().astype(\"uint8\").transpose(0, 2, 3, 1)[0, :, :, 0]\n",
+    "output_image = Image.fromarray(output_image)\n",
+    "display(output_image)\n",
+    "Audio(data=mel.image_to_audio(output_image), rate=mel.get_sample_rate())"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "ee3997cf",
+   "metadata": {},
+   "source": [
+    "### Interpolate between two audios in latent space"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "46019770",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image2 = random.choice(ds['train'])['image']\n",
+    "display(image2)\n",
+    "Audio(data=mel.image_to_audio(image2), rate=mel.get_sample_rate())"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e6552b19",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# encode\n",
+    "input_image2 = np.frombuffer(image2.tobytes(), dtype=\"uint8\").reshape(\n",
+    "    (image2.height, image2.width, 1))\n",
+    "input_image2 = ((input_image2 / 255) * 2 - 1).transpose(2, 0, 1)\n",
+    "posterior2 = vae.encode(torch.tensor([input_image2],\n",
+    "                                     dtype=torch.float32)).latent_dist\n",
+    "latents2 = posterior2.sample()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "060a0b25",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# interpolate\n",
+    "alpha = 0.5  #@param {type:\"slider\", min:0, max:1, step:0.1}\n",
+    "output_image = vae.decode(\n",
+    "    AudioDiffusionPipeline.slerp(latents, latents2, alpha))['sample']\n",
+    "output_image = torch.clamp(output_image, -1., 1.)\n",
+    "output_image = (output_image + 1.0) / 2.0  # -1,1 -> 0,1; c,h,w\n",
+    "output_image = (output_image.detach().cpu().numpy() *\n",
+    "                255).round().astype(\"uint8\").transpose(0, 2, 3, 1)[0, :, :, 0]\n",
+    "output_image = Image.fromarray(output_image)\n",
+    "display(output_image)\n",
+    "display(Audio(data=mel.image_to_audio(image), rate=mel.get_sample_rate()))\n",
+    "display(Audio(data=mel.image_to_audio(image2), rate=mel.get_sample_rate()))\n",
+    "display(\n",
+    "    Audio(data=mel.image_to_audio(output_image), rate=mel.get_sample_rate()))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d6c74105",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "huggingface",
+   "language": "python",
+   "name": "huggingface"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.6"
+  },
+  "toc": {
+   "base_numbering": 1,
+   "nav_menu": {},
+   "number_sections": true,
+   "sideBar": true,
+   "skip_h1_title": false,
+   "title_cell": "Table of Contents",
+   "title_sidebar": "Contents",
+   "toc_cell": false,
+   "toc_position": {},
+   "toc_section_display": true,
+   "toc_window_display": false
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

notebooks/train_model.ipynb

@@ -0,0 +1,600 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "62c5865f",
+   "metadata": {
+    "id": "62c5865f"
+   },
+   "source": [
+    "<a href=\"https://colab.research.google.com/github/teticio/audio-diffusion/blob/master/notebooks/train_model.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6c7800a6",
+   "metadata": {
+    "colab": {
+     "base_uri": "https://localhost:8080/"
+    },
+    "id": "6c7800a6",
+    "outputId": "ed18f4a9-ccea-4d7c-c82b-1749f1041f6c"
+   },
+   "outputs": [],
+   "source": [
+    "try:\n",
+    "    # are we running on Google Colab?\n",
+    "    import google.colab\n",
+    "    !git clone -q https://github.com/teticio/audio-diffusion.git\n",
+    "    %cd audio-diffusion\n",
+    "    %pip install -q -r requirements.txt .\n",
+    "except:\n",
+    "    pass"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c2fc0e7a",
+   "metadata": {
+    "id": "c2fc0e7a"
+   },
+   "outputs": [],
+   "source": [
+    "import sys\n",
+    "from IPython.display import Audio\n",
+    "from audiodiffusion import AudioDiffusion"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "MqlpL75_mDVv",
+   "metadata": {
+    "id": "MqlpL75_mDVv"
+   },
+   "source": [
+    "### Upload / specify audio files to train on\n",
+    "Provide some MP3 or WAV files that will be split into samples and converted to Mel spectrograms. For a resolution of 256, the samples will be about 5 seconds long."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "jg1zAHVsmCBG",
+   "metadata": {
+    "colab": {
+     "base_uri": "https://localhost:8080/",
+     "height": 73
+    },
+    "id": "jg1zAHVsmCBG",
+    "outputId": "414244c9-02b6-4ccf-cbfd-83f9022a0fc1"
+   },
+   "outputs": [],
+   "source": [
+    "try:\n",
+    "    # are we running on Google Colab?\n",
+    "    from google.colab import files\n",
+    "    input_dir = '.'\n",
+    "    files.upload();\n",
+    "except:\n",
+    "    input_dir = \".\""
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "10v0RCSUu75P",
+   "metadata": {
+    "id": "10v0RCSUu75P"
+   },
+   "source": [
+    "### Prepare dataset"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "NJNeEU6ftaTM",
+   "metadata": {
+    "colab": {
+     "base_uri": "https://localhost:8080/"
+    },
+    "id": "NJNeEU6ftaTM",
+    "outputId": "6c5bed15-c821-4def-eb90-3ab1a17b3c3d"
+   },
+   "outputs": [],
+   "source": [
+    "!{sys.executable} scripts/audio_to_images.py \\\n",
+    "  --resolution 256,256 \\\n",
+    "  --input_dir {input_dir} \\\n",
+    "  --output_dir data"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "5mGeXyJFvQCO",
+   "metadata": {
+    "id": "5mGeXyJFvQCO"
+   },
+   "source": [
+    "### Train model\n",
+    "The DDIM scheduler generates samples much faster."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "JGnlePbLvTOH",
+   "metadata": {
+    "colab": {
+     "base_uri": "https://localhost:8080/"
+    },
+    "id": "JGnlePbLvTOH",
+    "outputId": "69b6f53e-25a3-4c59-e205-2eab42889cd8"
+   },
+   "outputs": [],
+   "source": [
+    "!{sys.executable} scripts/train_unet.py \\\n",
+    "  --dataset_name data \\\n",
+    "  --output_dir models/model \\\n",
+    "  --num_epochs 10 \\\n",
+    "  --train_batch_size 2 \\\n",
+    "  --eval_batch_size 2 \\\n",
+    "  --gradient_accumulation_steps 8 \\\n",
+    "  --save_images_epochs 100 \\\n",
+    "  --save_model_epochs 1 \\\n",
+    "  --scheduler ddim"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "nTMAYEtMxtt0",
+   "metadata": {
+    "id": "nTMAYEtMxtt0"
+   },
+   "source": [
+    "### Generate samples with model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b294a94a",
+   "metadata": {
+    "id": "b294a94a"
+   },
+   "outputs": [],
+   "source": [
+    "audio_diffusion = AudioDiffusion(\"models/model\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "k2bKq3aqyAIM",
+   "metadata": {
+    "colab": {
+     "base_uri": "https://localhost:8080/",
+     "height": 363,
+     "referenced_widgets": [
+      "474d4db933d54e0497da4076a7fe135b",
+      "a849a3a1b46947db830a6a087411ec68",
+      "378f819239274ac88d913714bc27bf06",
+      "cc3b33e508744206955b26a417fbbdec",
+      "6015e5a9e6774e9abf7db273bca57363",
+      "629c21c68d22447185bb961e22bce4a6",
+      "2d5abefbc2ed4b72aed8c4f8ddc7a00c",
+      "11d1dbae00764a1c9dcc899c0b0f67dc",
+      "acdb5ddc7bda411a948689787b18b21e",
+      "9c4955f9d0f443a7b28ed827c5cdb37f",
+      "f9a1a976d82148f8961e80c357bc2764"
+     ]
+    },
+    "id": "k2bKq3aqyAIM",
+    "outputId": "d48238fe-ae36-4736-e67b-b69e3729304a"
+   },
+   "outputs": [],
+   "source": [
+    "image, (sample_rate, audio) = audio_diffusion.generate_spectrogram_and_audio()\n",
+    "display(image)\n",
+    "display(Audio(audio, rate=sample_rate))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "K2qAIJzg2DNK",
+   "metadata": {
+    "id": "K2qAIJzg2DNK"
+   },
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "accelerator": "GPU",
+  "colab": {
+   "collapsed_sections": [],
+   "provenance": []
+  },
+  "gpuClass": "standard",
+  "kernelspec": {
+   "display_name": "tmp",
+   "language": "python",
+   "name": "tmp"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.6"
+  },
+  "toc": {
+   "base_numbering": 1,
+   "nav_menu": {},
+   "number_sections": true,
+   "sideBar": true,
+   "skip_h1_title": false,
+   "title_cell": "Table of Contents",
+   "title_sidebar": "Contents",
+   "toc_cell": false,
+   "toc_position": {},
+   "toc_section_display": true,
+   "toc_window_display": false
+  },
+  "widgets": {
+   "application/vnd.jupyter.widget-state+json": {
+    "11d1dbae00764a1c9dcc899c0b0f67dc": {
+     "model_module": "@jupyter-widgets/base",
+     "model_module_version": "1.2.0",
+     "model_name": "LayoutModel",
+     "state": {
+      "_model_module": "@jupyter-widgets/base",
+      "_model_module_version": "1.2.0",
+      "_model_name": "LayoutModel",
+      "_view_count": null,
+      "_view_module": "@jupyter-widgets/base",
+      "_view_module_version": "1.2.0",
+      "_view_name": "LayoutView",
+      "align_content": null,
+      "align_items": null,
+      "align_self": null,
+      "border": null,
+      "bottom": null,
+      "display": null,
+      "flex": null,
+      "flex_flow": null,
+      "grid_area": null,
+      "grid_auto_columns": null,
+      "grid_auto_flow": null,
+      "grid_auto_rows": null,
+      "grid_column": null,
+      "grid_gap": null,
+      "grid_row": null,
+      "grid_template_areas": null,
+      "grid_template_columns": null,
+      "grid_template_rows": null,
+      "height": null,
+      "justify_content": null,
+      "justify_items": null,
+      "left": null,
+      "margin": null,
+      "max_height": null,
+      "max_width": null,
+      "min_height": null,
+      "min_width": null,
+      "object_fit": null,
+      "object_position": null,
+      "order": null,
+      "overflow": null,
+      "overflow_x": null,
+      "overflow_y": null,
+      "padding": null,
+      "right": null,
+      "top": null,
+      "visibility": null,
+      "width": null
+     }
+    },
+    "2d5abefbc2ed4b72aed8c4f8ddc7a00c": {
+     "model_module": "@jupyter-widgets/controls",
+     "model_module_version": "1.5.0",
+     "model_name": "DescriptionStyleModel",
+     "state": {
+      "_model_module": "@jupyter-widgets/controls",
+      "_model_module_version": "1.5.0",
+      "_model_name": "DescriptionStyleModel",
+      "_view_count": null,
+      "_view_module": "@jupyter-widgets/base",
+      "_view_module_version": "1.2.0",
+      "_view_name": "StyleView",
+      "description_width": ""
+     }
+    },
+    "378f819239274ac88d913714bc27bf06": {
+     "model_module": "@jupyter-widgets/controls",
+     "model_module_version": "1.5.0",
+     "model_name": "FloatProgressModel",
+     "state": {
+      "_dom_classes": [],
+      "_model_module": "@jupyter-widgets/controls",
+      "_model_module_version": "1.5.0",
+      "_model_name": "FloatProgressModel",
+      "_view_count": null,
+      "_view_module": "@jupyter-widgets/controls",
+      "_view_module_version": "1.5.0",
+      "_view_name": "ProgressView",
+      "bar_style": "success",
+      "description": "",
+      "description_tooltip": null,
+      "layout": "IPY_MODEL_11d1dbae00764a1c9dcc899c0b0f67dc",
+      "max": 50,
+      "min": 0,
+      "orientation": "horizontal",
+      "style": "IPY_MODEL_acdb5ddc7bda411a948689787b18b21e",
+      "value": 50
+     }
+    },
+    "474d4db933d54e0497da4076a7fe135b": {
+     "model_module": "@jupyter-widgets/controls",
+     "model_module_version": "1.5.0",
+     "model_name": "HBoxModel",
+     "state": {
+      "_dom_classes": [],
+      "_model_module": "@jupyter-widgets/controls",
+      "_model_module_version": "1.5.0",
+      "_model_name": "HBoxModel",
+      "_view_count": null,
+      "_view_module": "@jupyter-widgets/controls",
+      "_view_module_version": "1.5.0",
+      "_view_name": "HBoxView",
+      "box_style": "",
+      "children": [
+       "IPY_MODEL_a849a3a1b46947db830a6a087411ec68",
+       "IPY_MODEL_378f819239274ac88d913714bc27bf06",
+       "IPY_MODEL_cc3b33e508744206955b26a417fbbdec"
+      ],
+      "layout": "IPY_MODEL_6015e5a9e6774e9abf7db273bca57363"
+     }
+    },
+    "6015e5a9e6774e9abf7db273bca57363": {
+     "model_module": "@jupyter-widgets/base",
+     "model_module_version": "1.2.0",
+     "model_name": "LayoutModel",
+     "state": {
+      "_model_module": "@jupyter-widgets/base",
+      "_model_module_version": "1.2.0",
+      "_model_name": "LayoutModel",
+      "_view_count": null,
+      "_view_module": "@jupyter-widgets/base",
+      "_view_module_version": "1.2.0",
+      "_view_name": "LayoutView",
+      "align_content": null,
+      "align_items": null,
+      "align_self": null,
+      "border": null,
+      "bottom": null,
+      "display": null,
+      "flex": null,
+      "flex_flow": null,
+      "grid_area": null,
+      "grid_auto_columns": null,
+      "grid_auto_flow": null,
+      "grid_auto_rows": null,
+      "grid_column": null,
+      "grid_gap": null,
+      "grid_row": null,
+      "grid_template_areas": null,
+      "grid_template_columns": null,
+      "grid_template_rows": null,
+      "height": null,
+      "justify_content": null,
+      "justify_items": null,
+      "left": null,
+      "margin": null,
+      "max_height": null,
+      "max_width": null,
+      "min_height": null,
+      "min_width": null,
+      "object_fit": null,
+      "object_position": null,
+      "order": null,
+      "overflow": null,
+      "overflow_x": null,
+      "overflow_y": null,
+      "padding": null,
+      "right": null,
+      "top": null,
+      "visibility": null,
+      "width": null
+     }
+    },
+    "629c21c68d22447185bb961e22bce4a6": {
+     "model_module": "@jupyter-widgets/base",
+     "model_module_version": "1.2.0",
+     "model_name": "LayoutModel",
+     "state": {
+      "_model_module": "@jupyter-widgets/base",
+      "_model_module_version": "1.2.0",
+      "_model_name": "LayoutModel",
+      "_view_count": null,
+      "_view_module": "@jupyter-widgets/base",
+      "_view_module_version": "1.2.0",
+      "_view_name": "LayoutView",
+      "align_content": null,
+      "align_items": null,
+      "align_self": null,
+      "border": null,
+      "bottom": null,
+      "display": null,
+      "flex": null,
+      "flex_flow": null,
+      "grid_area": null,
+      "grid_auto_columns": null,
+      "grid_auto_flow": null,
+      "grid_auto_rows": null,
+      "grid_column": null,
+      "grid_gap": null,
+      "grid_row": null,
+      "grid_template_areas": null,
+      "grid_template_columns": null,
+      "grid_template_rows": null,
+      "height": null,
+      "justify_content": null,
+      "justify_items": null,
+      "left": null,
+      "margin": null,
+      "max_height": null,
+      "max_width": null,
+      "min_height": null,
+      "min_width": null,
+      "object_fit": null,
+      "object_position": null,
+      "order": null,
+      "overflow": null,
+      "overflow_x": null,
+      "overflow_y": null,
+      "padding": null,
+      "right": null,
+      "top": null,
+      "visibility": null,
+      "width": null
+     }
+    },
+    "9c4955f9d0f443a7b28ed827c5cdb37f": {
+     "model_module": "@jupyter-widgets/base",
+     "model_module_version": "1.2.0",
+     "model_name": "LayoutModel",
+     "state": {
+      "_model_module": "@jupyter-widgets/base",
+      "_model_module_version": "1.2.0",
+      "_model_name": "LayoutModel",
+      "_view_count": null,
+      "_view_module": "@jupyter-widgets/base",
+      "_view_module_version": "1.2.0",
+      "_view_name": "LayoutView",
+      "align_content": null,
+      "align_items": null,
+      "align_self": null,
+      "border": null,
+      "bottom": null,
+      "display": null,
+      "flex": null,
+      "flex_flow": null,
+      "grid_area": null,
+      "grid_auto_columns": null,
+      "grid_auto_flow": null,
+      "grid_auto_rows": null,
+      "grid_column": null,
+      "grid_gap": null,
+      "grid_row": null,
+      "grid_template_areas": null,
+      "grid_template_columns": null,
+      "grid_template_rows": null,
+      "height": null,
+      "justify_content": null,
+      "justify_items": null,
+      "left": null,
+      "margin": null,
+      "max_height": null,
+      "max_width": null,
+      "min_height": null,
+      "min_width": null,
+      "object_fit": null,
+      "object_position": null,
+      "order": null,
+      "overflow": null,
+      "overflow_x": null,
+      "overflow_y": null,
+      "padding": null,
+      "right": null,
+      "top": null,
+      "visibility": null,
+      "width": null
+     }
+    },
+    "a849a3a1b46947db830a6a087411ec68": {
+     "model_module": "@jupyter-widgets/controls",
+     "model_module_version": "1.5.0",
+     "model_name": "HTMLModel",
+     "state": {
+      "_dom_classes": [],
+      "_model_module": "@jupyter-widgets/controls",
+      "_model_module_version": "1.5.0",
+      "_model_name": "HTMLModel",
+      "_view_count": null,
+      "_view_module": "@jupyter-widgets/controls",
+      "_view_module_version": "1.5.0",
+      "_view_name": "HTMLView",
+      "description": "",
+      "description_tooltip": null,
+      "layout": "IPY_MODEL_629c21c68d22447185bb961e22bce4a6",
+      "placeholder": "​",
+      "style": "IPY_MODEL_2d5abefbc2ed4b72aed8c4f8ddc7a00c",
+      "value": "100%"
+     }
+    },
+    "acdb5ddc7bda411a948689787b18b21e": {
+     "model_module": "@jupyter-widgets/controls",
+     "model_module_version": "1.5.0",
+     "model_name": "ProgressStyleModel",
+     "state": {
+      "_model_module": "@jupyter-widgets/controls",
+      "_model_module_version": "1.5.0",
+      "_model_name": "ProgressStyleModel",
+      "_view_count": null,
+      "_view_module": "@jupyter-widgets/base",
+      "_view_module_version": "1.2.0",
+      "_view_name": "StyleView",
+      "bar_color": null,
+      "description_width": ""
+     }
+    },
+    "cc3b33e508744206955b26a417fbbdec": {
+     "model_module": "@jupyter-widgets/controls",
+     "model_module_version": "1.5.0",
+     "model_name": "HTMLModel",
+     "state": {
+      "_dom_classes": [],
+      "_model_module": "@jupyter-widgets/controls",
+      "_model_module_version": "1.5.0",
+      "_model_name": "HTMLModel",
+      "_view_count": null,
+      "_view_module": "@jupyter-widgets/controls",
+      "_view_module_version": "1.5.0",
+      "_view_name": "HTMLView",
+      "description": "",
+      "description_tooltip": null,
+      "layout": "IPY_MODEL_9c4955f9d0f443a7b28ed827c5cdb37f",
+      "placeholder": "​",
+      "style": "IPY_MODEL_f9a1a976d82148f8961e80c357bc2764",
+      "value": " 50/50 [00:07&lt;00:00,  8.13it/s]"
+     }
+    },
+    "f9a1a976d82148f8961e80c357bc2764": {
+     "model_module": "@jupyter-widgets/controls",
+     "model_module_version": "1.5.0",
+     "model_name": "DescriptionStyleModel",
+     "state": {
+      "_model_module": "@jupyter-widgets/controls",
+      "_model_module_version": "1.5.0",
+      "_model_name": "DescriptionStyleModel",
+      "_view_count": null,
+      "_view_module": "@jupyter-widgets/base",
+      "_view_module_version": "1.2.0",
+      "_view_name": "StyleView",
+      "description_width": ""
+     }
+    }
+   }
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

pyproject.toml

@@ -0,0 +1,3 @@
+[tool.black]
+line-length = 119
+target-version = ['py36']

requirements.txt

@@ -0,0 +1,12 @@
+torch
+numpy
+Pillow
+diffusers>=0.12.0
+librosa
+datasets>=2.9.0
+gradio
+streamlit
+tensorboard
+accelerate
+torchvision
+transformers

scripts/audio_to_images.py

@@ -0,0 +1,113 @@
+import argparse
+import io
+import logging
+import os
+import re
+
+import numpy as np
+import pandas as pd
+from datasets import Dataset, DatasetDict, Features, Image, Value
+from diffusers.pipelines.audio_diffusion import Mel
+from tqdm.auto import tqdm
+
+logging.basicConfig(level=logging.WARN)
+logger = logging.getLogger("audio_to_images")
+
+
+def main(args):
+    mel = Mel(
+        x_res=args.resolution[0],
+        y_res=args.resolution[1],
+        hop_length=args.hop_length,
+        sample_rate=args.sample_rate,
+        n_fft=args.n_fft,
+    )
+    os.makedirs(args.output_dir, exist_ok=True)
+    audio_files = [
+        os.path.join(root, file)
+        for root, _, files in os.walk(args.input_dir)
+        for file in files
+        if re.search("\.(mp3|wav|m4a)$", file, re.IGNORECASE)
+    ]
+    examples = []
+    try:
+        for audio_file in tqdm(audio_files):
+            try:
+                mel.load_audio(audio_file)
+            except KeyboardInterrupt:
+                raise
+            except:
+                continue
+            for slice in range(mel.get_number_of_slices()):
+                image = mel.audio_slice_to_image(slice)
+                assert image.width == args.resolution[0] and image.height == args.resolution[1], "Wrong resolution"
+                # skip completely silent slices
+                if all(np.frombuffer(image.tobytes(), dtype=np.uint8) == 255):
+                    logger.warn("File %s slice %d is completely silent", audio_file, slice)
+                    continue
+                with io.BytesIO() as output:
+                    image.save(output, format="PNG")
+                    bytes = output.getvalue()
+                examples.extend(
+                    [
+                        {
+                            "image": {"bytes": bytes},
+                            "audio_file": audio_file,
+                            "slice": slice,
+                        }
+                    ]
+                )
+    except Exception as e:
+        print(e)
+    finally:
+        if len(examples) == 0:
+            logger.warn("No valid audio files were found.")
+            return
+        ds = Dataset.from_pandas(
+            pd.DataFrame(examples),
+            features=Features(
+                {
+                    "image": Image(),
+                    "audio_file": Value(dtype="string"),
+                    "slice": Value(dtype="int16"),
+                }
+            ),
+        )
+        dsd = DatasetDict({"train": ds})
+        dsd.save_to_disk(os.path.join(args.output_dir))
+        if args.push_to_hub:
+            dsd.push_to_hub(args.push_to_hub)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Create dataset of Mel spectrograms from directory of audio files.")
+    parser.add_argument("--input_dir", type=str)
+    parser.add_argument("--output_dir", type=str, default="data")
+    parser.add_argument(
+        "--resolution",
+        type=str,
+        default="256",
+        help="Either square resolution or width,height.",
+    )
+    parser.add_argument("--hop_length", type=int, default=512)
+    parser.add_argument("--push_to_hub", type=str, default=None)
+    parser.add_argument("--sample_rate", type=int, default=22050)
+    parser.add_argument("--n_fft", type=int, default=2048)
+    args = parser.parse_args()
+
+    if args.input_dir is None:
+        raise ValueError("You must specify an input directory for the audio files.")
+
+    # Handle the resolutions.
+    try:
+        args.resolution = (int(args.resolution), int(args.resolution))
+    except ValueError:
+        try:
+            args.resolution = tuple(int(x) for x in args.resolution.split(","))
+            if len(args.resolution) != 2:
+                raise ValueError
+        except ValueError:
+            raise ValueError("Resolution must be a tuple of two integers or a single integer.")
+    assert isinstance(args.resolution, tuple)
+
+    main(args)

scripts/encode_audio.py

@@ -0,0 +1,38 @@
+import argparse
+import os
+import pickle
+
+from datasets import load_dataset, load_from_disk
+from tqdm.auto import tqdm
+
+from audiodiffusion.audio_encoder import AudioEncoder
+
+
+def main(args):
+    audio_encoder = AudioEncoder.from_pretrained("teticio/audio-encoder")
+
+    if args.dataset_name is not None:
+        if os.path.exists(args.dataset_name):
+            dataset = load_from_disk(args.dataset_name)["train"]
+        else:
+            dataset = load_dataset(
+                args.dataset_name,
+                args.dataset_config_name,
+                cache_dir=args.cache_dir,
+                use_auth_token=True if args.use_auth_token else None,
+                split="train",
+            )
+
+    encodings = {}
+    for audio_file in tqdm(dataset.to_pandas()["audio_file"].unique()):
+        encodings[audio_file] = audio_encoder.encode([audio_file])
+    pickle.dump(encodings, open(args.output_file, "wb"))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Create pickled audio encodings for dataset of audio files.")
+    parser.add_argument("--dataset_name", type=str, default=None)
+    parser.add_argument("--output_file", type=str, default="data/encodings.p")
+    parser.add_argument("--use_auth_token", type=bool, default=False)
+    args = parser.parse_args()
+    main(args)

scripts/train_unet.py

@@ -0,0 +1,453 @@
+# based on https://github.com/huggingface/diffusers/blob/main/examples/train_unconditional.py
+
+import argparse
+import os
+import pickle
+import random
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from accelerate import Accelerator
+from accelerate.logging import get_logger
+from datasets import load_dataset, load_from_disk
+from diffusers import (AutoencoderKL, DDIMScheduler, DDPMScheduler,
+                       UNet2DConditionModel, UNet2DModel)
+from diffusers.optimization import get_scheduler
+from diffusers.pipelines.audio_diffusion import Mel
+from diffusers.training_utils import EMAModel
+from huggingface_hub import HfFolder, Repository, whoami
+from librosa.util import normalize
+from torchvision.transforms import Compose, Normalize, ToTensor
+from tqdm.auto import tqdm
+
+from audiodiffusion.pipeline_audio_diffusion import AudioDiffusionPipeline
+
+logger = get_logger(__name__)
+
+
+def get_full_repo_name(model_id: str,
+                       organization: Optional[str] = None,
+                       token: Optional[str] = None):
+    if token is None:
+        token = HfFolder.get_token()
+    if organization is None:
+        username = whoami(token)["name"]
+        return f"{username}/{model_id}"
+    else:
+        return f"{organization}/{model_id}"
+
+
+def main(args):
+    output_dir = os.environ.get("SM_MODEL_DIR", None) or args.output_dir
+    logging_dir = os.path.join(output_dir, args.logging_dir)
+    accelerator = Accelerator(
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        mixed_precision=args.mixed_precision,
+        log_with="tensorboard",
+        logging_dir=logging_dir,
+    )
+
+    if args.dataset_name is not None:
+        if os.path.exists(args.dataset_name):
+            dataset = load_from_disk(
+                args.dataset_name,
+                storage_options=args.dataset_config_name)["train"]
+        else:
+            dataset = load_dataset(
+                args.dataset_name,
+                args.dataset_config_name,
+                cache_dir=args.cache_dir,
+                use_auth_token=True if args.use_auth_token else None,
+                split="train",
+            )
+    else:
+        dataset = load_dataset(
+            "imagefolder",
+            data_dir=args.train_data_dir,
+            cache_dir=args.cache_dir,
+            split="train",
+        )
+    # Determine image resolution
+    resolution = dataset[0]["image"].height, dataset[0]["image"].width
+
+    augmentations = Compose([
+        ToTensor(),
+        Normalize([0.5], [0.5]),
+    ])
+
+    def transforms(examples):
+        if args.vae is not None and vqvae.config["in_channels"] == 3:
+            images = [
+                augmentations(image.convert("RGB"))
+                for image in examples["image"]
+            ]
+        else:
+            images = [augmentations(image) for image in examples["image"]]
+        if args.encodings is not None:
+            encoding = [encodings[file] for file in examples["audio_file"]]
+            return {"input": images, "encoding": encoding}
+        return {"input": images}
+
+    dataset.set_transform(transforms)
+    train_dataloader = torch.utils.data.DataLoader(
+        dataset, batch_size=args.train_batch_size, shuffle=True)
+
+    if args.encodings is not None:
+        encodings = pickle.load(open(args.encodings, "rb"))
+
+    vqvae = None
+    if args.vae is not None:
+        try:
+            vqvae = AutoencoderKL.from_pretrained(args.vae)
+        except EnvironmentError:
+            vqvae = AudioDiffusionPipeline.from_pretrained(args.vae).vqvae
+        # Determine latent resolution
+        with torch.no_grad():
+            latent_resolution = vqvae.encode(
+                torch.zeros((1, 1) +
+                            resolution)).latent_dist.sample().shape[2:]
+
+    if args.from_pretrained is not None:
+        pipeline = AudioDiffusionPipeline.from_pretrained(args.from_pretrained)
+        mel = pipeline.mel
+        model = pipeline.unet
+        if hasattr(pipeline, "vqvae"):
+            vqvae = pipeline.vqvae
+
+    else:
+        if args.encodings is None:
+            model = UNet2DModel(
+                sample_size=resolution if vqvae is None else latent_resolution,
+                in_channels=1
+                if vqvae is None else vqvae.config["latent_channels"],
+                out_channels=1
+                if vqvae is None else vqvae.config["latent_channels"],
+                layers_per_block=2,
+                block_out_channels=(128, 128, 256, 256, 512, 512),
+                down_block_types=(
+                    "DownBlock2D",
+                    "DownBlock2D",
+                    "DownBlock2D",
+                    "DownBlock2D",
+                    "AttnDownBlock2D",
+                    "DownBlock2D",
+                ),
+                up_block_types=(
+                    "UpBlock2D",
+                    "AttnUpBlock2D",
+                    "UpBlock2D",
+                    "UpBlock2D",
+                    "UpBlock2D",
+                    "UpBlock2D",
+                ),
+            )
+
+        else:
+            model = UNet2DConditionModel(
+                sample_size=resolution if vqvae is None else latent_resolution,
+                in_channels=1
+                if vqvae is None else vqvae.config["latent_channels"],
+                out_channels=1
+                if vqvae is None else vqvae.config["latent_channels"],
+                layers_per_block=2,
+                block_out_channels=(128, 256, 512, 512),
+                down_block_types=(
+                    "CrossAttnDownBlock2D",
+                    "CrossAttnDownBlock2D",
+                    "CrossAttnDownBlock2D",
+                    "DownBlock2D",
+                ),
+                up_block_types=(
+                    "UpBlock2D",
+                    "CrossAttnUpBlock2D",
+                    "CrossAttnUpBlock2D",
+                    "CrossAttnUpBlock2D",
+                ),
+                cross_attention_dim=list(encodings.values())[0].shape[-1],
+            )
+
+    if args.scheduler == "ddpm":
+        noise_scheduler = DDPMScheduler(
+            num_train_timesteps=args.num_train_steps)
+    else:
+        noise_scheduler = DDIMScheduler(
+            num_train_timesteps=args.num_train_steps)
+
+    optimizer = torch.optim.AdamW(
+        model.parameters(),
+        lr=args.learning_rate,
+        betas=(args.adam_beta1, args.adam_beta2),
+        weight_decay=args.adam_weight_decay,
+        eps=args.adam_epsilon,
+    )
+
+    lr_scheduler = get_scheduler(
+        args.lr_scheduler,
+        optimizer=optimizer,
+        num_warmup_steps=args.lr_warmup_steps,
+        num_training_steps=(len(train_dataloader) * args.num_epochs) //
+        args.gradient_accumulation_steps,
+    )
+
+    model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        model, optimizer, train_dataloader, lr_scheduler)
+
+    ema_model = EMAModel(
+        getattr(model, "module", model),
+        inv_gamma=args.ema_inv_gamma,
+        power=args.ema_power,
+        max_value=args.ema_max_decay,
+    )
+
+    if args.push_to_hub:
+        if args.hub_model_id is None:
+            repo_name = get_full_repo_name(Path(output_dir).name,
+                                           token=args.hub_token)
+        else:
+            repo_name = args.hub_model_id
+        repo = Repository(output_dir, clone_from=repo_name)
+
+    if accelerator.is_main_process:
+        run = os.path.split(__file__)[-1].split(".")[0]
+        accelerator.init_trackers(run)
+
+    mel = Mel(
+        x_res=resolution[1],
+        y_res=resolution[0],
+        hop_length=args.hop_length,
+        sample_rate=args.sample_rate,
+        n_fft=args.n_fft,
+    )
+
+    global_step = 0
+    for epoch in range(args.num_epochs):
+        progress_bar = tqdm(total=len(train_dataloader),
+                            disable=not accelerator.is_local_main_process)
+        progress_bar.set_description(f"Epoch {epoch}")
+
+        if epoch < args.start_epoch:
+            for step in range(len(train_dataloader)):
+                optimizer.step()
+                lr_scheduler.step()
+                progress_bar.update(1)
+                global_step += 1
+            if epoch == args.start_epoch - 1 and args.use_ema:
+                ema_model.optimization_step = global_step
+            continue
+
+        model.train()
+        for step, batch in enumerate(train_dataloader):
+            clean_images = batch["input"]
+
+            if vqvae is not None:
+                vqvae.to(clean_images.device)
+                with torch.no_grad():
+                    clean_images = vqvae.encode(
+                        clean_images).latent_dist.sample()
+                # Scale latent images to ensure approximately unit variance
+                clean_images = clean_images * 0.18215
+
+            # Sample noise that we'll add to the images
+            noise = torch.randn(clean_images.shape).to(clean_images.device)
+            bsz = clean_images.shape[0]
+            # Sample a random timestep for each image
+            timesteps = torch.randint(
+                0,
+                noise_scheduler.config.num_train_timesteps,
+                (bsz, ),
+                device=clean_images.device,
+            ).long()
+
+            # Add noise to the clean images according to the noise magnitude at each timestep
+            # (this is the forward diffusion process)
+            noisy_images = noise_scheduler.add_noise(clean_images, noise,
+                                                     timesteps)
+
+            with accelerator.accumulate(model):
+                # Predict the noise residual
+                if args.encodings is not None:
+                    noise_pred = model(noisy_images, timesteps,
+                                       batch["encoding"])["sample"]
+                else:
+                    noise_pred = model(noisy_images, timesteps)["sample"]
+                loss = F.mse_loss(noise_pred, noise)
+                accelerator.backward(loss)
+
+                if accelerator.sync_gradients:
+                    accelerator.clip_grad_norm_(model.parameters(), 1.0)
+                optimizer.step()
+                lr_scheduler.step()
+                if args.use_ema:
+                    ema_model.step(model)
+                optimizer.zero_grad()
+
+            progress_bar.update(1)
+            global_step += 1
+
+            logs = {
+                "loss": loss.detach().item(),
+                "lr": lr_scheduler.get_last_lr()[0],
+                "step": global_step,
+            }
+            if args.use_ema:
+                logs["ema_decay"] = ema_model.decay
+            progress_bar.set_postfix(**logs)
+            accelerator.log(logs, step=global_step)
+        progress_bar.close()
+
+        accelerator.wait_for_everyone()
+
+        # Generate sample images for visual inspection
+        if accelerator.is_main_process:
+            if ((epoch + 1) % args.save_model_epochs == 0
+                    or (epoch + 1) % args.save_images_epochs == 0
+                    or epoch == args.num_epochs - 1):
+                unet = accelerator.unwrap_model(model)
+                if args.use_ema:
+                    ema_model.copy_to(unet.parameters())
+                pipeline = AudioDiffusionPipeline(
+                    vqvae=vqvae,
+                    unet=unet,
+                    mel=mel,
+                    scheduler=noise_scheduler,
+                )
+
+            if (
+                    epoch + 1
+            ) % args.save_model_epochs == 0 or epoch == args.num_epochs - 1:
+                pipeline.save_pretrained(output_dir)
+
+                # save the model
+                if args.push_to_hub:
+                    repo.push_to_hub(
+                        commit_message=f"Epoch {epoch}",
+                        blocking=False,
+                        auto_lfs_prune=True,
+                    )
+
+            if (epoch + 1) % args.save_images_epochs == 0:
+                generator = torch.Generator(
+                    device=clean_images.device).manual_seed(42)
+
+                if args.encodings is not None:
+                    random.seed(42)
+                    encoding = torch.stack(
+                        random.sample(list(encodings.values()),
+                                      args.eval_batch_size)).to(
+                                          clean_images.device)
+                else:
+                    encoding = None
+
+                # run pipeline in inference (sample random noise and denoise)
+                images, (sample_rate, audios) = pipeline(
+                    generator=generator,
+                    batch_size=args.eval_batch_size,
+                    return_dict=False,
+                    encoding=encoding,
+                )
+
+                # denormalize the images and save to tensorboard
+                images = np.array([
+                    np.frombuffer(image.tobytes(), dtype="uint8").reshape(
+                        (len(image.getbands()), image.height, image.width))
+                    for image in images
+                ])
+                accelerator.trackers[0].writer.add_images(
+                    "test_samples", images, epoch)
+                for _, audio in enumerate(audios):
+                    accelerator.trackers[0].writer.add_audio(
+                        f"test_audio_{_}",
+                        normalize(audio),
+                        epoch,
+                        sample_rate=sample_rate,
+                    )
+        accelerator.wait_for_everyone()
+
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Simple example of a training script.")
+    parser.add_argument("--local_rank", type=int, default=-1)
+    parser.add_argument("--dataset_name", type=str, default=None)
+    parser.add_argument("--dataset_config_name", type=str, default=None)
+    parser.add_argument(
+        "--train_data_dir",
+        type=str,
+        default=None,
+        help="A folder containing the training data.",
+    )
+    parser.add_argument("--output_dir", type=str, default="ddpm-model-64")
+    parser.add_argument("--overwrite_output_dir", type=bool, default=False)
+    parser.add_argument("--cache_dir", type=str, default=None)
+    parser.add_argument("--train_batch_size", type=int, default=16)
+    parser.add_argument("--eval_batch_size", type=int, default=16)
+    parser.add_argument("--num_epochs", type=int, default=100)
+    parser.add_argument("--save_images_epochs", type=int, default=10)
+    parser.add_argument("--save_model_epochs", type=int, default=10)
+    parser.add_argument("--gradient_accumulation_steps", type=int, default=1)
+    parser.add_argument("--learning_rate", type=float, default=1e-4)
+    parser.add_argument("--lr_scheduler", type=str, default="cosine")
+    parser.add_argument("--lr_warmup_steps", type=int, default=500)
+    parser.add_argument("--adam_beta1", type=float, default=0.95)
+    parser.add_argument("--adam_beta2", type=float, default=0.999)
+    parser.add_argument("--adam_weight_decay", type=float, default=1e-6)
+    parser.add_argument("--adam_epsilon", type=float, default=1e-08)
+    parser.add_argument("--use_ema", type=bool, default=True)
+    parser.add_argument("--ema_inv_gamma", type=float, default=1.0)
+    parser.add_argument("--ema_power", type=float, default=3 / 4)
+    parser.add_argument("--ema_max_decay", type=float, default=0.9999)
+    parser.add_argument("--push_to_hub", type=bool, default=False)
+    parser.add_argument("--use_auth_token", type=bool, default=False)
+    parser.add_argument("--hub_token", type=str, default=None)
+    parser.add_argument("--hub_model_id", type=str, default=None)
+    parser.add_argument("--hub_private_repo", type=bool, default=False)
+    parser.add_argument("--logging_dir", type=str, default="logs")
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default="no",
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose"
+            "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10."
+            "and an Nvidia Ampere GPU."),
+    )
+    parser.add_argument("--hop_length", type=int, default=512)
+    parser.add_argument("--sample_rate", type=int, default=22050)
+    parser.add_argument("--n_fft", type=int, default=2048)
+    parser.add_argument("--from_pretrained", type=str, default=None)
+    parser.add_argument("--start_epoch", type=int, default=0)
+    parser.add_argument("--num_train_steps", type=int, default=1000)
+    parser.add_argument("--scheduler",
+                        type=str,
+                        default="ddpm",
+                        help="ddpm or ddim")
+    parser.add_argument(
+        "--vae",
+        type=str,
+        default=None,
+        help="pretrained VAE model for latent diffusion",
+    )
+    parser.add_argument(
+        "--encodings",
+        type=str,
+        default=None,
+        help="picked dictionary mapping audio_file to encoding",
+    )
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    if args.dataset_name is None and args.train_data_dir is None:
+        raise ValueError(
+            "You must specify either a dataset name from the hub or a train data directory."
+        )
+
+    main(args)

scripts/train_vae.py

@@ -0,0 +1,177 @@
+# based on https://github.com/CompVis/stable-diffusion/blob/main/main.py
+
+import argparse
+import os
+
+import numpy as np
+import pytorch_lightning as pl
+import torch
+import torchvision
+from datasets import load_dataset, load_from_disk
+from diffusers.pipelines.audio_diffusion import Mel
+from ldm.util import instantiate_from_config
+from librosa.util import normalize
+from omegaconf import OmegaConf
+from PIL import Image
+from pytorch_lightning.callbacks import Callback, ModelCheckpoint
+from pytorch_lightning.trainer import Trainer
+from pytorch_lightning.utilities.distributed import rank_zero_only
+from torch.utils.data import DataLoader, Dataset
+
+from audiodiffusion.utils import convert_ldm_to_hf_vae
+
+
+class AudioDiffusion(Dataset):
+    def __init__(self, model_id, channels=3):
+        super().__init__()
+        self.channels = channels
+        if os.path.exists(model_id):
+            self.hf_dataset = load_from_disk(model_id)["train"]
+        else:
+            self.hf_dataset = load_dataset(model_id)["train"]
+
+    def __len__(self):
+        return len(self.hf_dataset)
+
+    def __getitem__(self, idx):
+        image = self.hf_dataset[idx]["image"]
+        if self.channels == 3:
+            image = image.convert("RGB")
+        image = np.frombuffer(image.tobytes(), dtype="uint8").reshape((image.height, image.width, self.channels))
+        image = (image / 255) * 2 - 1
+        return {"image": image}
+
+
+class AudioDiffusionDataModule(pl.LightningDataModule):
+    def __init__(self, model_id, batch_size, channels):
+        super().__init__()
+        self.batch_size = batch_size
+        self.dataset = AudioDiffusion(model_id=model_id, channels=channels)
+        self.num_workers = 1
+
+    def train_dataloader(self):
+        return DataLoader(self.dataset, batch_size=self.batch_size, num_workers=self.num_workers)
+
+
+class ImageLogger(Callback):
+    def __init__(self, every=1000, hop_length=512, sample_rate=22050, n_fft=2048):
+        super().__init__()
+        self.every = every
+        self.hop_length = hop_length
+        self.sample_rate = sample_rate
+        self.n_fft = n_fft
+
+    @rank_zero_only
+    def log_images_and_audios(self, pl_module, batch):
+        pl_module.eval()
+        with torch.no_grad():
+            images = pl_module.log_images(batch, split="train")
+        pl_module.train()
+
+        image_shape = next(iter(images.values())).shape
+        channels = image_shape[1]
+        mel = Mel(
+            x_res=image_shape[2],
+            y_res=image_shape[3],
+            hop_length=self.hop_length,
+            sample_rate=self.sample_rate,
+            n_fft=self.n_fft,
+        )
+
+        for k in images:
+            images[k] = images[k].detach().cpu()
+            images[k] = torch.clamp(images[k], -1.0, 1.0)
+            images[k] = (images[k] + 1.0) / 2.0  # -1,1 -> 0,1; c,h,w
+            grid = torchvision.utils.make_grid(images[k])
+
+            tag = f"train/{k}"
+            pl_module.logger.experiment.add_image(tag, grid, global_step=pl_module.global_step)
+
+            images[k] = (images[k].numpy() * 255).round().astype("uint8").transpose(0, 2, 3, 1)
+            for _, image in enumerate(images[k]):
+                audio = mel.image_to_audio(
+                    Image.fromarray(image, mode="RGB").convert("L")
+                    if channels == 3
+                    else Image.fromarray(image[:, :, 0])
+                )
+                pl_module.logger.experiment.add_audio(
+                    tag + f"/{_}",
+                    normalize(audio),
+                    global_step=pl_module.global_step,
+                    sample_rate=mel.get_sample_rate(),
+                )
+
+    def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx):
+        if (batch_idx + 1) % self.every != 0:
+            return
+        self.log_images_and_audios(pl_module, batch)
+
+
+class HFModelCheckpoint(ModelCheckpoint):
+    def __init__(self, ldm_config, hf_checkpoint, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.ldm_config = ldm_config
+        self.hf_checkpoint = hf_checkpoint
+        self.sample_size = None
+
+    def on_train_batch_start(self, trainer, pl_module, batch, batch_idx):
+        if self.sample_size is None:
+            self.sample_size = list(batch["image"].shape[1:3])
+
+    def on_train_epoch_end(self, trainer, pl_module):
+        ldm_checkpoint = self._get_metric_interpolated_filepath_name({"epoch": trainer.current_epoch}, trainer)
+        super().on_train_epoch_end(trainer, pl_module)
+        self.ldm_config.model.params.ddconfig.resolution = self.sample_size
+        convert_ldm_to_hf_vae(ldm_checkpoint, self.ldm_config, self.hf_checkpoint, self.sample_size)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Train VAE using ldm.")
+    parser.add_argument("-d", "--dataset_name", type=str, default=None)
+    parser.add_argument("-b", "--batch_size", type=int, default=1)
+    parser.add_argument("-c", "--ldm_config_file", type=str, default="config/ldm_autoencoder_kl.yaml")
+    parser.add_argument("--ldm_checkpoint_dir", type=str, default="models/ldm-autoencoder-kl")
+    parser.add_argument("--hf_checkpoint_dir", type=str, default="models/autoencoder-kl")
+    parser.add_argument("-r", "--resume_from_checkpoint", type=str, default=None)
+    parser.add_argument("-g", "--gradient_accumulation_steps", type=int, default=1)
+    parser.add_argument("--hop_length", type=int, default=512)
+    parser.add_argument("--sample_rate", type=int, default=22050)
+    parser.add_argument("--n_fft", type=int, default=2048)
+    parser.add_argument("--save_images_batches", type=int, default=1000)
+    parser.add_argument("--max_epochs", type=int, default=100)
+    args = parser.parse_args()
+
+    config = OmegaConf.load(args.ldm_config_file)
+    model = instantiate_from_config(config.model)
+    model.learning_rate = config.model.base_learning_rate
+    data = AudioDiffusionDataModule(
+        model_id=args.dataset_name,
+        batch_size=args.batch_size,
+        channels=config.model.params.ddconfig.in_channels,
+    )
+    lightning_config = config.pop("lightning", OmegaConf.create())
+    trainer_config = lightning_config.get("trainer", OmegaConf.create())
+    trainer_config.accumulate_grad_batches = args.gradient_accumulation_steps
+    trainer_opt = argparse.Namespace(**trainer_config)
+    trainer = Trainer.from_argparse_args(
+        trainer_opt,
+        max_epochs=args.max_epochs,
+        resume_from_checkpoint=args.resume_from_checkpoint,
+        callbacks=[
+            ImageLogger(
+                every=args.save_images_batches,
+                hop_length=args.hop_length,
+                sample_rate=args.sample_rate,
+                n_fft=args.n_fft,
+            ),
+            HFModelCheckpoint(
+                ldm_config=config,
+                hf_checkpoint=args.hf_checkpoint_dir,
+                dirpath=args.ldm_checkpoint_dir,
+                filename="{epoch:06}",
+                verbose=True,
+                save_last=True,
+            ),
+        ],
+    )
+    trainer.fit(model, data)

setup.cfg

@@ -0,0 +1,20 @@
+[metadata]
+name = audiodiffusion
+version = attr: audiodiffusion.VERSION
+description = Generate Mel spectrogram dataset from directory of audio files.
+long_description = file: README.md
+long_description_content_type = text/markdown
+license = GPL3
+classifiers =
+    Programming Language :: Python :: 3
+
+[options]
+zip_safe = False
+packages = audiodiffusion
+install_requires =
+    torch
+    numpy
+    Pillow
+    diffusers>=0.12.0
+    librosa
+    datasets>=2.9.0

setup.py

@@ -0,0 +1,6 @@
+#!/usr/bin/env python
+
+from setuptools import setup
+
+if __name__ == "__main__":
+    setup()