Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,29 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/demo0.gif filter=lfs diff=lfs merge=lfs -text
+assets/demo1.gif filter=lfs diff=lfs merge=lfs -text
+assets/demo2.gif filter=lfs diff=lfs merge=lfs -text
+assets/demo3.gif filter=lfs diff=lfs merge=lfs -text
+assets/demo5.gif filter=lfs diff=lfs merge=lfs -text
+assets/demo6.gif filter=lfs diff=lfs merge=lfs -text
+assets/demo7.gif filter=lfs diff=lfs merge=lfs -text
+assets/demo8.gif filter=lfs diff=lfs merge=lfs -text
+assets/demo9.gif filter=lfs diff=lfs merge=lfs -text
+assets/hg.png filter=lfs diff=lfs merge=lfs -text
+assets/video.png filter=lfs diff=lfs merge=lfs -text
+audio_0_retri_0_watermarked.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/cached_audio/101099-00_18_09-00_18_19.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/cached_audio/1wrQ6Msp7wM_00-00-39.69_00-00-45.68.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/cached_audio/demo1.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/cached_audio/demo3.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/cached_audio/demo4.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/cached_audio/demo5.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/cached_audio/demo7.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/cached_audio/demo9.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/cached_audio/female_test_V1.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/cached_audio/speaker12_10_BVHw8aCPATM_00-01-05.0_00-01-10.0.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/cached_audio/speaker7_iuYlGRnC7J8_00-00-0.00_00-00-3.25.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/cached_audio/speaker9_o7Ik1OB4TaE_00-00-38.15_00-00-42.33.mp4 filter=lfs diff=lfs merge=lfs -text
+datasets/data_json/show-oliver-s40_w128.json filter=lfs diff=lfs merge=lfs -text
+datasets/data_json/show-oliver-s40_w64.json filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,179 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+
+# watermarked videos will be saved at root directory, but we don't want to track them
+demo*watermarked.mp4
+outputs/
+SMPLer-X/common/utils/human_model_files/
+SMPLer-X/pretrained_models/
+Wav2Lip/checkpoints/
+datasets/cached_ckpts/
+datasets/cached_graph/
+emage/smplx_models/
+frame-interpolation-pytorch/*.pt
+
+# submodules
+Wav2Lip/
+frame-interpolation-pytorch/
+SMPLer-X/
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/latest/usage/project/#working-with-version-control
+.pdm.toml
+.pdm-python
+.pdm-build/
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

.idea/.gitignore

@@ -0,0 +1,8 @@
+# 默认忽略的文件
+/shelf/
+/workspace.xml
+# 基于编辑器的 HTTP 客户端请求
+/httpRequests/
+# Datasource local storage ignored files
+/dataSources/
+/dataSources.local.xml

.idea/TANGO.iml

@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<module type="PYTHON_MODULE" version="4">
+  <component name="NewModuleRootManager">
+    <content url="file://$MODULE_DIR$" />
+    <orderEntry type="inheritedJdk" />
+    <orderEntry type="sourceFolder" forTests="false" />
+  </component>
+  <component name="PyDocumentationSettings">
+    <option name="format" value="PLAIN" />
+    <option name="myDocStringFormat" value="Plain" />
+  </component>
+</module>

.idea/inspectionProfiles/profiles_settings.xml

@@ -0,0 +1,6 @@
+<component name="InspectionProjectProfileManager">
+  <settings>
+    <option name="USE_PROJECT_PROFILE" value="false" />
+    <version value="1.0" />
+  </settings>
+</component>

.idea/modules.xml

@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="ProjectModuleManager">
+    <modules>
+      <module fileurl="file://$PROJECT_DIR$/.idea/TANGO.iml" filepath="$PROJECT_DIR$/.idea/TANGO.iml" />
+    </modules>
+  </component>
+</project>

.idea/vcs.xml

@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="VcsDirectoryMappings">
+    <mapping directory="" vcs="Git" />
+    <mapping directory="$PROJECT_DIR$/Wav2Lip" vcs="Git" />
+    <mapping directory="$PROJECT_DIR$/frame-interpolation-pytorch" vcs="Git" />
+  </component>
+</project>

.idea/workspace.xml

@@ -0,0 +1,114 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="AutoImportSettings">
+    <option name="autoReloadType" value="SELECTIVE" />
+  </component>
+  <component name="ChangeListManager">
+    <list default="true" id="09d685d1-f444-43d5-acad-200263565083" name="更改" comment="">
+      <change beforePath="$PROJECT_DIR$/requirements.txt" beforeDir="false" afterPath="$PROJECT_DIR$/requirements.txt" afterDir="false" />
+    </list>
+    <option name="SHOW_DIALOG" value="false" />
+    <option name="HIGHLIGHT_CONFLICTS" value="true" />
+    <option name="HIGHLIGHT_NON_ACTIVE_CHANGELIST" value="false" />
+    <option name="LAST_RESOLUTION" value="IGNORE" />
+  </component>
+  <component name="Git.Settings">
+    <option name="RECENT_GIT_ROOT_PATH" value="$PROJECT_DIR$" />
+  </component>
+  <component name="ProjectColorInfo">{
+  &quot;associatedIndex&quot;: 5
+}</component>
+  <component name="ProjectId" id="2pVN1AmCm9xKUXNffpEopNvIJF5" />
+  <component name="ProjectViewState">
+    <option name="hideEmptyMiddlePackages" value="true" />
+    <option name="showLibraryContents" value="true" />
+  </component>
+  <component name="PropertiesComponent">{
+  &quot;keyToString&quot;: {
+    &quot;RunOnceActivity.ShowReadmeOnStart&quot;: &quot;true&quot;,
+    &quot;git-widget-placeholder&quot;: &quot;main&quot;,
+    &quot;last_opened_file_path&quot;: &quot;C:/Users/45488/OneDrive/桌面/yzb/TANGO/emage/smplx_models/smplx&quot;,
+    &quot;node.js.detected.package.eslint&quot;: &quot;true&quot;,
+    &quot;node.js.detected.package.tslint&quot;: &quot;true&quot;,
+    &quot;node.js.selected.package.eslint&quot;: &quot;(autodetect)&quot;,
+    &quot;node.js.selected.package.tslint&quot;: &quot;(autodetect)&quot;,
+    &quot;nodejs_package_manager_path&quot;: &quot;npm&quot;,
+    &quot;vue.rearranger.settings.migration&quot;: &quot;true&quot;
+  }
+}</component>
+  <component name="RecentsManager">
+    <key name="CopyFile.RECENT_KEYS">
+      <recent name="C:\Users\45488\OneDrive\桌面\yzb\TANGO\emage\smplx_models\smplx" />
+      <recent name="C:\Users\45488\OneDrive\桌面\yzb\TANGO\datasets\cached_graph\youtube_test" />
+      <recent name="C:\Users\45488\OneDrive\桌面\yzb\TANGO\datasets\cached_graph\show_oliver_test" />
+    </key>
+  </component>
+  <component name="RunManager">
+    <configuration name="create_graph" type="PythonConfigurationType" factoryName="Python" temporary="true" nameIsGenerated="true">
+      <module name="TANGO" />
+      <option name="ENV_FILES" value="" />
+      <option name="INTERPRETER_OPTIONS" value="" />
+      <option name="PARENT_ENVS" value="true" />
+      <envs>
+        <env name="PYTHONUNBUFFERED" value="1" />
+      </envs>
+      <option name="SDK_HOME" value="" />
+      <option name="WORKING_DIRECTORY" value="$PROJECT_DIR$" />
+      <option name="IS_MODULE_SDK" value="true" />
+      <option name="ADD_CONTENT_ROOTS" value="true" />
+      <option name="ADD_SOURCE_ROOTS" value="true" />
+      <EXTENSION ID="PythonCoverageRunConfigurationExtension" runner="coverage.py" />
+      <option name="SCRIPT_NAME" value="$PROJECT_DIR$/create_graph.py" />
+      <option name="PARAMETERS" value="" />
+      <option name="SHOW_COMMAND_LINE" value="false" />
+      <option name="EMULATE_TERMINAL" value="false" />
+      <option name="MODULE_MODE" value="false" />
+      <option name="REDIRECT_INPUT" value="false" />
+      <option name="INPUT_FILE" value="" />
+      <method v="2" />
+    </configuration>
+    <recent_temporary>
+      <list>
+        <item itemvalue="Python.create_graph" />
+      </list>
+    </recent_temporary>
+  </component>
+  <component name="SharedIndexes">
+    <attachedChunks>
+      <set>
+        <option value="bundled-js-predefined-d6986cc7102b-e768b9ed790e-JavaScript-PY-243.21565.199" />
+        <option value="bundled-python-sdk-cab1f2013843-4ae2d6a61b08-com.jetbrains.pycharm.pro.sharedIndexes.bundled-PY-243.21565.199" />
+      </set>
+    </attachedChunks>
+  </component>
+  <component name="SpellCheckerSettings" RuntimeDictionaries="0" Folders="0" CustomDictionaries="0" DefaultDictionary="应用程序级" UseSingleDictionary="true" transferred="true" />
+  <component name="TaskManager">
+    <task active="true" id="Default" summary="默认任务">
+      <changelist id="09d685d1-f444-43d5-acad-200263565083" name="更改" comment="" />
+      <created>1732847065796</created>
+      <option name="number" value="Default" />
+      <option name="presentableId" value="Default" />
+      <updated>1732847065796</updated>
+      <workItem from="1732847068636" duration="14699000" />
+      <workItem from="1733105812677" duration="3426000" />
+      <workItem from="1733110257766" duration="6291000" />
+      <workItem from="1733124449966" duration="296000" />
+      <workItem from="1733304778417" duration="892000" />
+    </task>
+    <servers />
+  </component>
+  <component name="TypeScriptGeneratedFilesManager">
+    <option name="version" value="3" />
+  </component>
+  <component name="XDebuggerManager">
+    <breakpoint-manager>
+      <breakpoints>
+        <line-breakpoint enabled="true" suspend="THREAD" type="python-line">
+          <url>file://$PROJECT_DIR$/create_graph.py</url>
+          <line>460</line>
+          <option name="timeStamp" value="1" />
+        </line-breakpoint>
+      </breakpoints>
+    </breakpoint-manager>
+  </component>
+</project>

LICENSE

@@ -0,0 +1,407 @@
+Attribution-NonCommercial 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+    wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More considerations
+     for the public:
+    wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-NonCommercial 4.0 International Public
+License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial 4.0 International Public License ("Public
+License"). To the extent this Public License may be interpreted as a
+contract, You are granted the Licensed Rights in consideration of Your
+acceptance of these terms and conditions, and the Licensor grants You
+such rights in consideration of benefits the Licensor receives from
+making the Licensed Material available under these terms and
+conditions.
+
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+  d. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  e. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  f. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  g. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  h. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  i. NonCommercial means not primarily intended for or directed towards
+     commercial advantage or monetary compensation. For purposes of
+     this Public License, the exchange of the Licensed Material for
+     other material subject to Copyright and Similar Rights by digital
+     file-sharing or similar means is NonCommercial provided there is
+     no payment of monetary compensation in connection with the
+     exchange.
+
+  j. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  k. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  l. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part, for NonCommercial purposes only; and
+
+            b. produce, reproduce, and Share Adapted Material for
+               NonCommercial purposes only.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties, including when
+          the Licensed Material is used other than for NonCommercial
+          purposes.
+
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+       4. If You Share Adapted Material You produce, the Adapter's
+          License You apply must not prevent recipients of the Adapted
+          Material from complying with this Public License.
+
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database for NonCommercial purposes
+     only;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material; and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

README.md

@@ -1,12 +1,111 @@
----
-title: My Tango
-emoji: 🦀
-colorFrom: blue
-colorTo: blue
-sdk: gradio
-sdk_version: 5.7.1
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+---
+title: my_tango
+app_file: app.py
+sdk: gradio
+sdk_version: 4.44.1
+---
+<div align="center">
+<!-- <p align="center"> <img src="./assets/EMAGE_2024/1711449143651.jpg" width="100px"> </p> -->
+<h2>TANGO: Co-Speech Gesture Video Reenactment with Hierarchical Audio-Motion Embedding and Diffusion Interpolation</h2>
+
+<a href='https://pantomatrix.github.io/TANGO/'><img src='https://img.shields.io/badge/Project-TANGO-blue' alt='Project'></a>
+<a href='https://www.youtube.com/watch?v=_DfsA11puBc'><img src='https://img.shields.io/badge/YouTube-TANGO-rgb(255, 0, 0)' alt='Youtube'></a>
+<a href='https://huggingface.co/spaces/H-Liu1997/TANGO'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue'></a>
+<a href='https://arxiv.org/abs/2410.04221'><img src='https://img.shields.io/badge/Paper-ArXiv-yellow' alt='Project'></a>
+
+</div>
+
+# News
+
+Welcome contributors! Feel free to submit the pull requests!
+
+- **[2024/10]** Welcome to try our [TANGO](<(https://huggingface.co/spaces/H-Liu1997/TANGO)!>) on Hugging face space !
+- **[2024/10]** Code for create gesture graph is available.
+
+<p align=center>
+<img src ="./assets/hg.png" width="60%" >
+</p>
+
+# Results Videos
+
+<p align="center">
+  <img src="./assets/demo8.gif" width="32%" alt="demo0">
+  <img src="./assets/demo1.gif" width="32%" alt="demo1">
+  <img src="./assets/demo2.gif" width="32%" alt="demo2">
+</p>
+<p align="center">
+  <img src="./assets/demo3.gif" width="32%" alt="demo3">
+  <img src="./assets/demo5.gif" width="32%" alt="demo5">
+  <img src="./assets/demo0.gif" width="32%" alt="demo6">
+</p>
+<p align="center">
+  <img src="./assets/demo7.gif" width="32%" alt="demo7">
+  <img src="./assets/demo6.gif" width="32%" alt="demo8">
+  <img src="./assets/demo9.gif" width="32%" alt="demo9">
+</p>
+
+# Demo Video (on Youtube)
+
+<p align=center>
+    <a href="https://youtu.be/xuhD_-tMH1w?si=Tr6jHAhOR1fxWIjb">
+    <img  width="68%" src="./assets/video.png">
+    </a>
+</p>
+
+# 📝 Release Plans
+
+- [ ] Training codes for AuMoClip and ACInterp
+- [ ] Inference codes for ACInterp
+- [ ] Processed Youtube Buiness Video data (very small, around 15 mins)
+- [x] Scripts for creating gesture graph
+- [x] Inference codes with AuMoClip and pretrained weights
+
+# ⚒️ Installation
+
+## Clone the repository
+
+```shell
+git clone https://github.com/CyberAgentAILab/TANGO.git
+cd TANGO
+git clone https://github.com/justinjohn0306/Wav2Lip.git
+git clone https://github.com/dajes/frame-interpolation-pytorch.git
+```
+
+## Build Environtment
+
+We Recommend a python version `==3.9.20` and cuda version `==11.8`. Then build environment as follows:
+
+```shell
+# [Optional] Create a virtual env
+conda create -n tango python==3.9.20
+conda activate tango
+# Install with pip:
+pip install -r ./pre-requirements.txt
+pip install -r ./requirements.txt
+```
+
+# 🚀 Training and Inference
+
+## Inference
+
+Here is the command for running inference scripts under the path `<your root>/TANGO/`, it will take around 3 min to generate two 8s vidoes. You can visualize by directly check the video or check the result .npz files via blender using our blender addon in [EMAGE](https://github.com/PantoMatrix/PantoMatrix).
+
+_Necessary checkpoints and pre-computed graphs will be automatically downloaded during the first run. Please ensure that at least 35GB of disk space is available._
+
+```shell
+python app.py
+```
+
+### Create the graph for custom character
+
+```shell
+python create_graph.py
+```
+
+# Copyright Information
+
+We thanks the open-source project [Wav2Lip](https://github.com/Rudrabha/Wav2Lip), [FiLM](https://github.com/caffeinism/FiLM-pytorch), [SMPLerX](https://github.com/caizhongang/SMPLer-X).
+
+Check out our previous works for Co-Speech 3D motion Generation <a href="https://github.com/PantoMatrix/PantoMatrix">DisCo, BEAT, EMAGE</a>.
+
+This project is only for research or education purposes, and not freely available for commercial use or redistribution. The srcipt is available only under the terms of the [Attribution-NonCommercial 4.0 International](https://creativecommons.org/licenses/by-nc/4.0/legalcode) (CC BY-NC 4.0) license.

app.py

@@ -0,0 +1,796 @@
+import os
+import gradio as gr
+import gc
+import soundfile as sf
+import shutil
+import argparse
+from omegaconf import OmegaConf
+import random
+import numpy as np
+import librosa
+import emage.mertic  # noqa: F401 # somehow this must be imported, even though it is not used directly
+from decord import VideoReader
+from PIL import Image
+import cv2
+import subprocess
+import importlib
+import torch
+import torch.nn.functional as F
+import smplx
+import igraph
+
+# import emage
+from utils.video_io import save_videos_from_pil
+from utils.genextend_inference_utils import adjust_statistics_to_match_reference
+from create_graph import path_visualization, graph_pruning, get_motion_reps_tensor, path_visualization_v2
+from utils.download_utils import download_files_from_repo
+
+SCRIPT_PATH = os.path.dirname(os.path.realpath(__file__))
+
+download_files_from_repo()
+
+shutil.copyfile("./assets/app.py", "./SMPLer-X/app.py")
+shutil.copyfile("./assets/transforms.py", "./SMPLer-X/common/utils/transforms.py")
+shutil.copyfile("./assets/inference.py", "./SMPLer-X/main/inference.py")
+
+
+def search_path_dp(graph, audio_low_np, audio_high_np, loop_penalty=0.1, top_k=1, search_mode="both", continue_penalty=0.1):
+    T = audio_low_np.shape[0]  # Total time steps
+    # N = len(graph.vs)  # Total number of nodes in the graph
+
+    # Initialize DP tables
+    min_cost = [
+        {} for _ in range(T)
+    ]  # min_cost[t][node_index] = list of tuples: (cost, prev_node_index, prev_tuple_index, non_continue_count, visited_nodes)
+
+    # Initialize the first time step
+    start_nodes = [v for v in graph.vs if v["previous"] is None or v["previous"] == -1]
+    for node in start_nodes:
+        node_index = node.index
+        motion_low = node["motion_low"]  # Shape: [C]
+        motion_high = node["motion_high"]  # Shape: [C]
+
+        # Cost using cosine similarity
+        if search_mode == "both":
+            cost = 2 - (np.dot(audio_low_np[0], motion_low.T) + np.dot(audio_high_np[0], motion_high.T))
+        elif search_mode == "high_level":
+            cost = 1 - np.dot(audio_high_np[0], motion_high.T)
+        elif search_mode == "low_level":
+            cost = 1 - np.dot(audio_low_np[0], motion_low.T)
+
+        visited_nodes = {node_index: 1}  # Initialize visit count as a dictionary
+
+        min_cost[0][node_index] = [(cost, None, None, 0, visited_nodes)]  # Initialize with no predecessor and 0 non-continue count
+
+    # DP over time steps
+    for t in range(1, T):
+        for node in graph.vs:
+            node_index = node.index
+            candidates = []
+
+            # Incoming edges to the current node
+            incoming_edges = graph.es.select(_to=node_index)
+            for edge in incoming_edges:
+                prev_node_index = edge.source
+                edge_id = edge.index
+                is_continue_edge = graph.es[edge_id]["is_continue"]
+                # prev_node = graph.vs[prev_node_index]
+                if prev_node_index in min_cost[t - 1]:
+                    for tuple_index, (prev_cost, _, _, prev_non_continue_count, prev_visited) in enumerate(min_cost[t - 1][prev_node_index]):
+                        # Loop punishment
+                        if node_index in prev_visited:
+                            loop_time = prev_visited[node_index]  # Get the count of previous visits
+                            loop_cost = prev_cost + loop_penalty * np.exp(loop_time)  # Apply exponential penalty
+                            new_visited = prev_visited.copy()
+                            new_visited[node_index] = loop_time + 1  # Increment visit count
+                        else:
+                            loop_cost = prev_cost
+                            new_visited = prev_visited.copy()
+                            new_visited[node_index] = 1  # Initialize visit count for the new node
+
+                        motion_low = node["motion_low"]  # Shape: [C]
+                        motion_high = node["motion_high"]  # Shape: [C]
+
+                        if search_mode == "both":
+                            cost_increment = 2 - (np.dot(audio_low_np[t], motion_low.T) + np.dot(audio_high_np[t], motion_high.T))
+                        elif search_mode == "high_level":
+                            cost_increment = 1 - np.dot(audio_high_np[t], motion_high.T)
+                        elif search_mode == "low_level":
+                            cost_increment = 1 - np.dot(audio_low_np[t], motion_low.T)
+
+                        # Check if the edge is "is_continue"
+                        if not is_continue_edge:
+                            non_continue_count = prev_non_continue_count + 1  # Increment the count of non-continue edges
+                        else:
+                            non_continue_count = prev_non_continue_count
+
+                        # Apply the penalty based on the square of the number of non-continuous edges
+                        continue_penalty_cost = continue_penalty * non_continue_count
+
+                        total_cost = loop_cost + cost_increment + continue_penalty_cost
+
+                        candidates.append((total_cost, prev_node_index, tuple_index, non_continue_count, new_visited))
+
+            # Keep the top k candidates
+            if candidates:
+                # Sort candidates by total_cost
+                candidates.sort(key=lambda x: x[0])
+                # Keep top k
+                min_cost[t][node_index] = candidates[:top_k]
+            else:
+                # No candidates, do nothing
+                pass
+
+    # Collect all possible end paths at time T-1
+    end_candidates = []
+    for node_index, tuples in min_cost[T - 1].items():
+        for tuple_index, (cost, _, _, _, _) in enumerate(tuples):
+            end_candidates.append((cost, node_index, tuple_index))
+
+    if not end_candidates:
+        print("No valid path found.")
+        return [], []
+
+    # Sort end candidates by cost
+    end_candidates.sort(key=lambda x: x[0])
+
+    # Keep top k paths
+    top_k_paths_info = end_candidates[:top_k]
+
+    # Reconstruct the paths
+    optimal_paths = []
+    is_continue_lists = []
+    for final_cost, node_index, tuple_index in top_k_paths_info:
+        optimal_path_indices = []
+        current_node_index = node_index
+        current_tuple_index = tuple_index
+        for t in range(T - 1, -1, -1):
+            optimal_path_indices.append(current_node_index)
+            tuple_data = min_cost[t][current_node_index][current_tuple_index]
+            _, prev_node_index, prev_tuple_index, _, _ = tuple_data
+            current_node_index = prev_node_index
+            current_tuple_index = prev_tuple_index
+            if current_node_index is None:
+                break  # Reached the start node
+        optimal_path_indices = optimal_path_indices[::-1]  # Reverse to get correct order
+        optimal_path = [graph.vs[idx] for idx in optimal_path_indices]
+        optimal_paths.append(optimal_path)
+
+        # Extract continuity information
+        is_continue = []
+        for i in range(len(optimal_path) - 1):
+            edge_id = graph.get_eid(optimal_path[i].index, optimal_path[i + 1].index)
+            is_cont = graph.es[edge_id]["is_continue"]
+            is_continue.append(is_cont)
+        is_continue_lists.append(is_continue)
+
+    print("Top {} Paths:".format(len(optimal_paths)))
+    for i, path in enumerate(optimal_paths):
+        path_indices = [node.index for node in path]
+        print("Path {}: Cost: {}, Nodes: {}".format(i + 1, top_k_paths_info[i][0], path_indices))
+
+    return optimal_paths, is_continue_lists
+
+
+def test_fn(model, device, iteration, candidate_json_path, test_path, cfg, audio_path, **kwargs):
+    create_graph = kwargs["create_graph"]
+    torch.set_grad_enabled(False)
+    pool_path = candidate_json_path.replace("data_json", "cached_graph").replace(".json", ".pkl")
+    graph = igraph.Graph.Read_Pickle(fname=pool_path)
+    # print(len(graph.vs))
+
+    save_dir = os.path.join(test_path, f"retrieved_motions_{iteration}")
+    os.makedirs(save_dir, exist_ok=True)
+
+    actual_model = model.module if isinstance(model, torch.nn.parallel.DistributedDataParallel) else model
+    actual_model.eval()
+
+    # with open(candidate_json_path, 'r') as f:
+    #     candidate_data = json.load(f)
+    all_motions = {}
+    for i, node in enumerate(graph.vs):
+        if all_motions.get(node["name"]) is None:
+            all_motions[node["name"]] = [node["axis_angle"].reshape(-1)]
+        else:
+            all_motions[node["name"]].append(node["axis_angle"].reshape(-1))
+    for k, v in all_motions.items():
+        all_motions[k] = np.stack(v)  # T, J*3
+        # print(k, all_motions[k].shape)
+
+    window_size = cfg.data.pose_length
+    motion_high_all = []
+    motion_low_all = []
+    for k, v in all_motions.items():
+        motion_tensor = torch.from_numpy(v).float().to(device).unsqueeze(0)
+        _, t, _ = motion_tensor.shape
+
+        if t >= window_size:
+            num_chunks = t // window_size
+            motion_high_list = []
+            motion_low_list = []
+
+            for i in range(num_chunks):
+                start_idx = i * window_size
+                end_idx = start_idx + window_size
+                motion_slice = motion_tensor[:, start_idx:end_idx, :]
+
+                motion_features = actual_model.get_motion_features(motion_slice)
+
+                motion_low = motion_features["motion_low"].cpu().numpy()
+                motion_high = motion_features["motion_cls"].unsqueeze(0).repeat(1, motion_low.shape[1], 1).cpu().numpy()
+
+                motion_high_list.append(motion_high[0])
+                motion_low_list.append(motion_low[0])
+
+            remain_length = t % window_size
+            if remain_length > 0:
+                start_idx = t - window_size
+                motion_slice = motion_tensor[:, start_idx:, :]
+
+                motion_features = actual_model.get_motion_features(motion_slice)
+                # motion_high = motion_features["motion_high_weight"].cpu().numpy()
+                motion_low = motion_features["motion_low"].cpu().numpy()
+                motion_high = motion_features["motion_cls"].unsqueeze(0).repeat(1, motion_low.shape[1], 1).cpu().numpy()
+
+                motion_high_list.append(motion_high[0][-remain_length:])
+                motion_low_list.append(motion_low[0][-remain_length:])
+
+            motion_high_all.append(np.concatenate(motion_high_list, axis=0))
+            motion_low_all.append(np.concatenate(motion_low_list, axis=0))
+
+        else:  # t < window_size:
+            gap = window_size - t
+            motion_slice = torch.cat(
+                [motion_tensor, torch.zeros((motion_tensor.shape[0], gap, motion_tensor.shape[2])).to(motion_tensor.device)], 1
+            )
+            motion_features = actual_model.get_motion_features(motion_slice)
+            # motion_high = motion_features["motion_high_weight"].cpu().numpy()
+            motion_low = motion_features["motion_low"].cpu().numpy()
+            motion_high = motion_features["motion_cls"].unsqueeze(0).repeat(1, motion_low.shape[1], 1).cpu().numpy()
+
+            motion_high_all.append(motion_high[0][:t])
+            motion_low_all.append(motion_low[0][:t])
+
+    motion_high_all = np.concatenate(motion_high_all, axis=0)
+    motion_low_all = np.concatenate(motion_low_all, axis=0)
+    # print(motion_high_all.shape, motion_low_all.shape, len(graph.vs))
+    motion_low_all = motion_low_all / np.linalg.norm(motion_low_all, axis=1, keepdims=True)
+    motion_high_all = motion_high_all / np.linalg.norm(motion_high_all, axis=1, keepdims=True)
+    assert motion_high_all.shape[0] == len(graph.vs)
+    assert motion_low_all.shape[0] == len(graph.vs)
+
+    for i, node in enumerate(graph.vs):
+        node["motion_high"] = motion_high_all[i]
+        node["motion_low"] = motion_low_all[i]
+
+    graph = graph_pruning(graph)
+    # for gradio, use a subgraph
+    if len(graph.vs) > 1800:
+        gap = len(graph.vs) - 1800
+        start_d = random.randint(0, 1800)
+        graph.delete_vertices(range(start_d, start_d + gap))
+    ascc_2 = graph.clusters(mode="STRONG")
+    graph = ascc_2.giant()
+
+    # drop the id of gt
+    idx = 0
+    audio_waveform, sr = librosa.load(audio_path)
+    audio_waveform = librosa.resample(audio_waveform, orig_sr=sr, target_sr=cfg.data.audio_sr)
+    audio_tensor = torch.from_numpy(audio_waveform).float().to(device).unsqueeze(0)
+
+    target_length = audio_tensor.shape[1] // cfg.data.audio_sr * 30
+    window_size = int(cfg.data.audio_sr * (cfg.data.pose_length / 30))
+    _, t = audio_tensor.shape
+    audio_low_list = []
+    audio_high_list = []
+
+    if t >= window_size:
+        num_chunks = t // window_size
+        # print(num_chunks, t % window_size)
+        for i in range(num_chunks):
+            start_idx = i * window_size
+            end_idx = start_idx + window_size
+            # print(start_idx, end_idx, window_size)
+            audio_slice = audio_tensor[:, start_idx:end_idx]
+
+            model_out_candidates = actual_model.get_audio_features(audio_slice)
+            audio_low = model_out_candidates["audio_low"]
+            # audio_high = model_out_candidates["audio_high_weight"]
+            audio_high = model_out_candidates["audio_cls"].unsqueeze(0).repeat(1, audio_low.shape[1], 1)
+            # print(audio_low.shape, audio_high.shape)
+
+            audio_low = F.normalize(audio_low, dim=2)[0].cpu().numpy()
+            audio_high = F.normalize(audio_high, dim=2)[0].cpu().numpy()
+
+            audio_low_list.append(audio_low)
+            audio_high_list.append(audio_high)
+            # print(audio_low.shape, audio_high.shape)
+
+        remain_length = t % window_size
+        if remain_length > 1:
+            start_idx = t - window_size
+            audio_slice = audio_tensor[:, start_idx:]
+
+            model_out_candidates = actual_model.get_audio_features(audio_slice)
+            audio_low = model_out_candidates["audio_low"]
+            # audio_high = model_out_candidates["audio_high_weight"]
+            audio_high = model_out_candidates["audio_cls"].unsqueeze(0).repeat(1, audio_low.shape[1], 1)
+
+            gap = target_length - np.concatenate(audio_low_list, axis=0).shape[1]
+            audio_low = F.normalize(audio_low, dim=2)[0][-gap:].cpu().numpy()
+            audio_high = F.normalize(audio_high, dim=2)[0][-gap:].cpu().numpy()
+
+            # print(audio_low.shape, audio_high.shape)
+            audio_low_list.append(audio_low)
+            audio_high_list.append(audio_high)
+    else:
+        gap = window_size - t
+        audio_slice = audio_tensor
+        model_out_candidates = actual_model.get_audio_features(audio_slice)
+        audio_low = model_out_candidates["audio_low"]
+        # audio_high = model_out_candidates["audio_high_weight"]
+        audio_high = model_out_candidates["audio_cls"].unsqueeze(0).repeat(1, audio_low.shape[1], 1)
+        audio_low = F.normalize(audio_low, dim=2)[0].cpu().numpy()
+        audio_high = F.normalize(audio_high, dim=2)[0].cpu().numpy()
+        audio_low_list.append(audio_low)
+        audio_high_list.append(audio_high)
+
+    audio_low_all = np.concatenate(audio_low_list, axis=0)
+    audio_high_all = np.concatenate(audio_high_list, axis=0)
+    path_list, is_continue_list = search_path_dp(graph, audio_low_all, audio_high_all, top_k=1, search_mode="both")
+
+    res_motion = []
+    counter = 0
+    wav2lip_checkpoint_path = os.path.join(SCRIPT_PATH, "Wav2Lip/checkpoints/wav2lip_gan.pth")  # Update this path to your Wav2Lip checkpoint
+    wav2lip_script_path = os.path.join(SCRIPT_PATH, "Wav2Lip/inference.py")
+    for path, is_continue in zip(path_list, is_continue_list):
+        if False:
+            # time is limited if we create graph on hugging face, lets skip blending.
+            res_motion_current = path_visualization(
+                graph,
+                path,
+                is_continue,
+                os.path.join(save_dir, f"audio_{idx}_retri_{counter}.mp4"),
+                audio_path=audio_path,
+                return_motion=True,
+                verbose_continue=True,
+            )
+            video_temp_path = os.path.join(save_dir, f"audio_{idx}_retri_{counter}.mp4")
+        else:
+            res_motion_current = path_visualization_v2(
+                graph,
+                path,
+                is_continue,
+                os.path.join(save_dir, f"audio_{idx}_retri_{counter}.mp4"),
+                audio_path=None,
+                return_motion=True,
+                verbose_continue=True,
+            )
+            video_temp_path = os.path.join(save_dir, f"audio_{idx}_retri_{counter}.mp4")
+            video_reader = VideoReader(video_temp_path)
+            video_np = []
+            for i in range(len(video_reader)):
+                if i == 0:
+                    continue
+                video_frame = video_reader[i].asnumpy()
+                video_np.append(Image.fromarray(video_frame))
+            adjusted_video_pil = adjust_statistics_to_match_reference([video_np])
+            save_videos_from_pil(
+                adjusted_video_pil[0], os.path.join(save_dir, f"audio_{idx}_retri_{counter}.mp4"), fps=graph.vs[0]["fps"], bitrate=2000000
+            )
+
+        audio_temp_path = audio_path
+        lipsync_output_path = os.path.join(save_dir, f"audio_{idx}_retri_{counter}.mp4")
+        cmd_wav2lip_1 = f"cd Wav2Lip; python {wav2lip_script_path} --checkpoint_path {wav2lip_checkpoint_path} --face {video_temp_path} --audio {audio_temp_path} --outfile {lipsync_output_path} --nosmooth --out_height 720"
+        subprocess.run(cmd_wav2lip_1, shell=True)
+
+        res_motion.append(res_motion_current)
+        np.savez(os.path.join(save_dir, f"audio_{idx}_retri_{counter}.npz"), motion=res_motion_current)
+
+        start_node = path[1].index
+        end_node = start_node + 100
+
+    if create_graph:
+        # time is limited if create graph, let us skip the second video
+        result = [
+            os.path.join(save_dir, f"audio_{idx}_retri_0.mp4"),
+            os.path.join(save_dir, f"audio_{idx}_retri_0.mp4"),
+            os.path.join(save_dir, f"audio_{idx}_retri_0.npz"),
+            os.path.join(save_dir, f"audio_{idx}_retri_0.npz"),
+        ]
+        return result
+
+    print(f"delete gt-nodes {start_node}, {end_node}")
+    nodes_to_delete = list(range(start_node, end_node))
+    graph.delete_vertices(nodes_to_delete)
+    graph = graph_pruning(graph)
+    path_list, is_continue_list = search_path_dp(graph, audio_low_all, audio_high_all, top_k=1, search_mode="both")
+    res_motion = []
+    counter = 1
+    for path, is_continue in zip(path_list, is_continue_list):
+        res_motion_current = path_visualization_v2(
+            graph,
+            path,
+            is_continue,
+            os.path.join(save_dir, f"audio_{idx}_retri_{counter}.mp4"),
+            audio_path=None,
+            return_motion=True,
+            verbose_continue=True,
+        )
+        video_temp_path = os.path.join(save_dir, f"audio_{idx}_retri_{counter}.mp4")
+        video_reader = VideoReader(video_temp_path)
+        video_np = []
+        for i in range(len(video_reader)):
+            if i == 0:
+                continue
+            video_frame = video_reader[i].asnumpy()
+            video_np.append(Image.fromarray(video_frame))
+        adjusted_video_pil = adjust_statistics_to_match_reference([video_np])
+        save_videos_from_pil(
+            adjusted_video_pil[0], os.path.join(save_dir, f"audio_{idx}_retri_{counter}.mp4"), fps=graph.vs[0]["fps"], bitrate=2000000
+        )
+
+        audio_temp_path = audio_path
+        lipsync_output_path = os.path.join(save_dir, f"audio_{idx}_retri_{counter}.mp4")
+        cmd_wav2lip_2 = f"cd Wav2Lip; python {wav2lip_script_path} --checkpoint_path {wav2lip_checkpoint_path} --face {video_temp_path} --audio {audio_temp_path} --outfile {lipsync_output_path} --nosmooth --out_height 720"
+        subprocess.run(cmd_wav2lip_2, shell=True)
+        res_motion.append(res_motion_current)
+        np.savez(os.path.join(save_dir, f"audio_{idx}_retri_{counter}.npz"), motion=res_motion_current)
+
+    result = [
+        os.path.join(save_dir, f"audio_{idx}_retri_0.mp4"),
+        os.path.join(save_dir, f"audio_{idx}_retri_1.mp4"),
+        os.path.join(save_dir, f"audio_{idx}_retri_0.npz"),
+        os.path.join(save_dir, f"audio_{idx}_retri_1.npz"),
+    ]
+    return result
+
+
+def init_class(module_name, class_name, config, **kwargs):
+    module = importlib.import_module(module_name)
+    model_class = getattr(module, class_name)
+    instance = model_class(config, **kwargs)
+    return instance
+
+
+def seed_everything(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+
+
+def prepare_all(yaml_name):
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--config", type=str, default=yaml_name)
+    parser.add_argument("--debug", action="store_true", help="Enable debugging mode")
+    parser.add_argument("overrides", nargs=argparse.REMAINDER)
+    args = parser.parse_args()
+    if args.config.endswith(".yaml"):
+        config = OmegaConf.load(args.config)
+        config.exp_name = os.path.basename(args.config)[:-5]
+    else:
+        raise ValueError("Unsupported config file format. Only .yaml files are allowed.")
+    save_dir = os.path.join(OUTPUT_DIR, config.exp_name)
+    os.makedirs(save_dir, exist_ok=True)
+    return config
+
+
+def save_first_10_seconds(video_path, output_path="./save_video.mp4", max_length=512):
+    if os.path.exists(output_path):
+        os.remove(output_path)
+
+    cap = cv2.VideoCapture(video_path)
+
+    if not cap.isOpened():
+        return
+
+    fps = int(cap.get(cv2.CAP_PROP_FPS))
+    original_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    original_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+
+    # Calculate the aspect ratio and resize dimensions
+    if original_width >= original_height:
+        new_width = max_length
+        new_height = int(original_height * (max_length / original_width))
+    else:
+        new_height = max_length
+        new_width = int(original_width * (max_length / original_height))
+
+    fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+    out = cv2.VideoWriter(output_path.replace(".mp4", "_fps.mp4"), fourcc, fps, (new_width, new_height))
+
+    frames_to_save = fps * 20
+    frame_count = 0
+
+    while cap.isOpened() and frame_count < frames_to_save:
+        ret, frame = cap.read()
+        if not ret:
+            break
+        # Resize the frame while keeping the aspect ratio
+        resized_frame = cv2.resize(frame, (new_width, new_height))
+        # resized_frame = frame
+        out.write(resized_frame)
+        frame_count += 1
+
+    cap.release()
+    out.release()
+    command = [
+        "ffmpeg",
+        "-i",
+        output_path.replace(".mp4", "_fps.mp4"),
+        "-vf",
+        "minterpolate=fps=30:mi_mode=mci:mc_mode=aobmc:vsbmc=1",
+        output_path,
+    ]
+    subprocess.run(command)
+    os.remove(output_path.replace(".mp4", "_fps.mp4"))
+
+
+character_name_to_yaml = {
+    "speaker8_jjRWaMCWs44_00-00-30.16_00-00-33.32.mp4": "./datasets/data_json/youtube_test/speaker8.json",
+    "speaker7_iuYlGRnC7J8_00-00-0.00_00-00-3.25.mp4": "./datasets/data_json/youtube_test/speaker7.json",
+    "speaker9_o7Ik1OB4TaE_00-00-38.15_00-00-42.33.mp4": "./datasets/data_json/youtube_test/speaker9.json",
+    "1wrQ6Msp7wM_00-00-39.69_00-00-45.68.mp4": "./datasets/data_json/youtube_test/speaker1.json",
+    "101099-00_18_09-00_18_19.mp4": "./datasets/data_json/show_oliver_test/Stupid_Watergate_-_Last_Week_Tonight_with_John_Oliver_HBO-FVFdsl29s_Q.mkv.json",
+}
+
+
+TARGET_SR = 16000
+OUTPUT_DIR = os.path.join(SCRIPT_PATH, "outputs/")
+
+
+# @spaces.GPU(duration=200)
+def tango(audio_path, character_name, seed, create_graph=False, video_folder_path=None):
+    shutil.rmtree(OUTPUT_DIR, ignore_errors=True)
+    os.makedirs(OUTPUT_DIR, exist_ok=True)
+    cfg_file = os.path.join(SCRIPT_PATH, "configs/gradio.yaml")
+    cfg = prepare_all(cfg_file)
+    cfg.seed = seed
+    seed_everything(cfg.seed)
+    experiment_ckpt_dir = os.path.join(OUTPUT_DIR, cfg.exp_name)
+    saved_audio_path = os.path.join(OUTPUT_DIR, "saved_audio.wav")
+    sample_rate, audio_waveform = audio_path
+    sf.write(saved_audio_path, audio_waveform, sample_rate)
+
+    audio_waveform, sample_rate = librosa.load(saved_audio_path)
+    # print(audio_waveform.shape)
+    resampled_audio = librosa.resample(audio_waveform, orig_sr=sample_rate, target_sr=TARGET_SR)
+    required_length = int(TARGET_SR * (128 / 30)) * 2
+    resampled_audio = resampled_audio[:required_length]
+    sf.write(saved_audio_path, resampled_audio, TARGET_SR)
+    audio_path = saved_audio_path
+
+    yaml_name = os.path.join(SCRIPT_PATH, "datasets/data_json/youtube_test/speaker1.json")
+    cfg.data.test_meta_paths = yaml_name
+    print(yaml_name)
+
+    video_folder_path = os.path.join(OUTPUT_DIR, "tmpvideo")
+    if os.path.basename(character_name) not in character_name_to_yaml.keys():
+        create_graph = True
+        # load video, and save it to "./save_video.mp4 for the first 20s of the video."
+        os.makedirs(video_folder_path, exist_ok=True)
+        save_first_10_seconds(character_name, os.path.join(video_folder_path, "save_video.mp4"))
+
+    if create_graph:
+        data_save_path = os.path.join(OUTPUT_DIR, "tmpdata")
+        json_save_path = os.path.join(OUTPUT_DIR, "save_video.json")
+        graph_save_path = os.path.join(OUTPUT_DIR, "save_video.pkl")
+        cmd_smplx = f"cd ./SMPLer-X/ && python app.py --video_folder_path {video_folder_path} --data_save_path {data_save_path} --json_save_path {json_save_path} && cd .."
+        subprocess.run(cmd_smplx, shell=True)
+        print("cmd_smplx: ", cmd_smplx)
+        cmd_graph = f"python ./create_graph.py --json_save_path {json_save_path} --graph_save_path {graph_save_path}"
+        subprocess.run(cmd_graph, shell=True)
+        print("cmd_graph: ", cmd_graph)
+        cfg.data.test_meta_paths = json_save_path
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    smplx_model = smplx.create(
+        "./emage/smplx_models/",
+        model_type="smplx",
+        gender="NEUTRAL_2020",
+        use_face_contour=False,
+        num_betas=300,
+        num_expression_coeffs=100,
+        ext="npz",
+        use_pca=False,
+    )
+    model = init_class(cfg.model.name_pyfile, cfg.model.class_name, cfg)
+    for param in model.parameters():
+        param.requires_grad = False
+    model.smplx_model = smplx_model
+    model.get_motion_reps = get_motion_reps_tensor
+    assert torch.cuda.is_available(), "CUDA is not available"
+    device = torch.device("cuda:0")
+    smplx_model = smplx_model.to(device).eval()
+    model = model.to(device)
+    model.smplx_model = model.smplx_model.to(device)
+
+    checkpoint_path = os.path.join(SCRIPT_PATH, "datasets/cached_ckpts/ckpt.pth")
+    checkpoint = torch.load(checkpoint_path)
+    state_dict = checkpoint["model_state_dict"]
+    new_state_dict = {k.replace("module.", ""): v for k, v in state_dict.items()}
+    model.load_state_dict(new_state_dict, strict=False)
+
+    test_path = os.path.join(experiment_ckpt_dir, f"test_{0}")
+    os.makedirs(test_path, exist_ok=True)
+    result = test_fn(model, device, 0, cfg.data.test_meta_paths, test_path, cfg, audio_path, create_graph=create_graph)
+    gc.collect()
+    torch.cuda.empty_cache()
+    return result
+
+
+examples_audio = [
+    ["./datasets/cached_audio/example_male_voice_9_seconds.wav"],
+    ["./datasets/cached_audio/example_female_voice_9_seconds.wav"],
+]
+
+examples_video = [
+    ["./datasets/cached_audio/speaker8_jjRWaMCWs44_00-00-30.16_00-00-33.32.mp4"],
+    ["./datasets/cached_audio/speaker7_iuYlGRnC7J8_00-00-0.00_00-00-3.25.mp4"],
+    ["./datasets/cached_audio/speaker9_o7Ik1OB4TaE_00-00-38.15_00-00-42.33.mp4"],
+    ["./datasets/cached_audio/1wrQ6Msp7wM_00-00-39.69_00-00-45.68.mp4"],
+    ["./datasets/cached_audio/101099-00_18_09-00_18_19.mp4"],
+]
+
+combined_examples = [
+    ["./datasets/cached_audio/example_female_voice_9_seconds.wav", "./datasets/cached_audio/female_test_V1.mp4", 2024],
+]
+
+
+def make_demo():
+    with gr.Blocks(analytics_enabled=False) as Interface:
+        gr.Markdown(
+            """
+        <div style="display: flex; justify-content: center; align-items: center; text-align: center;">
+          <div>
+            <h1>TANGO</h1>
+            <span>Generating full-body talking videos from audio and reference video</span>
+            <h2 style='font-weight: 450; font-size: 1rem; margin: 0rem'>\
+              <a href='https://h-liu1997.github.io/'>Haiyang Liu</a>, \
+              <a href='https://yangxingchao.github.io/'>Xingchao Yang</a>, \
+              <a href=''>Tomoya Akiyama</a>, \
+              <a href='https://sky24h.github.io/'> Yuantian Huang</a>, \
+              <a href=''>Qiaoge Li</a>, \
+              <a href='https://www.tut.ac.jp/english/university/faculty/cs/164.html'>Shigeru Kuriyama</a>, \
+              <a href='https://taketomitakafumi.sakura.ne.jp/web/en/'>Takafumi Taketomi</a>\
+            </h2>
+            <br>
+            <div style="display: flex; justify-content: center; align-items: center; text-align: center;">
+              <a href="https://arxiv.org/abs/2410.04221"><img src="https://img.shields.io/badge/arXiv-2410.04221-blue"></a>
+              &nbsp;
+              <a href="https://pantomatrix.github.io/TANGO/"><img src="https://img.shields.io/badge/Project_Page-TANGO-orange" alt="Project Page"></a>
+              &nbsp;
+              <a href="https://github.com/CyberAgentAILab/TANGO"><img src="https://img.shields.io/badge/Github-Code-green"></a>
+              &nbsp;
+              <a href="https://github.com/CyberAgentAILab/TANGO"><img src="https://img.shields.io/github/stars/CyberAgentAILab/TANGO
+              "></a>
+            </div>
+          </div>
+        </div>
+        """
+        )
+
+        # Create a gallery with 5 videos
+        with gr.Row():
+            gr.Video(value="./datasets/cached_audio/demo1.mp4", label="Demo 0", watermark="./datasets/watermark.png")
+            gr.Video(value="./datasets/cached_audio/demo2.mp4", label="Demo 1", watermark="./datasets/watermark.png")
+            gr.Video(value="./datasets/cached_audio/demo3.mp4", label="Demo 2", watermark="./datasets/watermark.png")
+            gr.Video(value="./datasets/cached_audio/demo4.mp4", label="Demo 3", watermark="./datasets/watermark.png")
+            gr.Video(value="./datasets/cached_audio/demo5.mp4", label="Demo 4", watermark="./datasets/watermark.png")
+        with gr.Row():
+            gr.Video(value="./datasets/cached_audio/demo6.mp4", label="Demo 5", watermark="./datasets/watermark.png")
+            gr.Video(value="./datasets/cached_audio/demo0.mp4", label="Demo 6", watermark="./datasets/watermark.png")
+            gr.Video(value="./datasets/cached_audio/demo7.mp4", label="Demo 7", watermark="./datasets/watermark.png")
+            gr.Video(value="./datasets/cached_audio/demo8.mp4", label="Demo 8", watermark="./datasets/watermark.png")
+            gr.Video(value="./datasets/cached_audio/demo9.mp4", label="Demo 9", watermark="./datasets/watermark.png")
+
+        with gr.Row():
+            gr.Markdown(
+                """
+              <div style="display: flex; justify-content: center; align-items: center; text-align: center;">
+              This is an open-source project running locally, operates in low-quality mode. Some generated results from high-quality mode are shown above.
+              <br>
+              News:
+              <br>
+              [10/15]: Add watermark, fix bugs on custom character by downgrades to py3.9, fix bugs to support audio less than 4s.
+              </div>
+              """
+            )
+
+        with gr.Row():
+            with gr.Column(scale=4):
+                video_output_1 = gr.Video(
+                    label="Generated video - 1",
+                    interactive=False,
+                    autoplay=False,
+                    loop=False,
+                    show_share_button=True,
+                    watermark="./datasets/watermark.png",
+                )
+            with gr.Column(scale=4):
+                video_output_2 = gr.Video(
+                    label="Generated video - 2",
+                    interactive=False,
+                    autoplay=False,
+                    loop=False,
+                    show_share_button=True,
+                    watermark="./datasets/watermark.png",
+                )
+            with gr.Column(scale=1):
+                file_output_1 = gr.File(label="Download 3D Motion and Visualize in Blender")
+                file_output_2 = gr.File(label="Download 3D Motion and Visualize in Blender")
+                gr.Markdown("""
+                <div style="display: flex; justify-content: center; align-items: center; text-align: left;">
+                Details of the low-quality mode:
+                <br>
+                1. lower resolution, video resized as long-side 512 and keep aspect ratio.
+                <br>
+                2. subgraph instead of full-graph, causing noticeable "frame jumps". 
+                <br>
+                3. only use the first 8s of your input audio.
+                <br>
+                4. only use the first 20s of your input video for custom character. if you custom character, it will only generate one video result without "smoothing" for saving time.
+                <br>
+                5. use open-source tools like SMPLerX-s-model, Wav2Lip, and FiLM for faster processing. 
+                <br>
+                <br>
+                Feel free to open an issue on GitHub or contact the authors if this does not meet your needs.
+                </div>
+                """)
+
+        with gr.Row():
+            with gr.Column(scale=1):
+                audio_input = gr.Audio(label="Upload your audio")
+                seed_input = gr.Number(label="Seed", value=2024, interactive=True)
+            with gr.Column(scale=2):
+                gr.Examples(
+                    examples=examples_audio,
+                    inputs=[audio_input],
+                    outputs=[video_output_1, video_output_2, file_output_1, file_output_2],
+                    label="Select existing Audio examples",
+                    cache_examples=False,
+                )
+            with gr.Column(scale=1):
+                video_input = gr.Video(label="Your Character", elem_classes="video")
+            with gr.Column(scale=2):
+                gr.Examples(
+                    examples=examples_video,
+                    inputs=[video_input],  # Correctly refer to video input
+                    outputs=[video_output_1, video_output_2, file_output_1, file_output_2],
+                    label="Character Examples",
+                    cache_examples=False,
+                )
+
+        # Fourth row: Generate video button
+        with gr.Row():
+            run_button = gr.Button("Generate Video")
+
+        # Define button click behavior
+        run_button.click(
+            fn=tango,
+            inputs=[audio_input, video_input, seed_input],
+            outputs=[video_output_1, video_output_2, file_output_1, file_output_2],
+        )
+
+        with gr.Row():
+            with gr.Column(scale=4):
+                gr.Examples(
+                    examples=combined_examples,
+                    inputs=[audio_input, video_input, seed_input],  # Both audio and video as inputs
+                    outputs=[video_output_1, video_output_2, file_output_1, file_output_2],
+                    fn=tango,  # Function that processes both audio and video inputs
+                    label="Select Combined Audio and Video Examples (Cached)",
+                    cache_examples=True,
+                )
+
+    return Interface
+
+
+if __name__ == "__main__":
+    os.environ["MASTER_ADDR"] = "127.0.0.1"
+    os.environ["MASTER_PORT"] = "8675"
+
+    demo = make_demo()
+    demo.launch(share=True)

assets/app.py

@@ -0,0 +1,149 @@
+import os
+import shutil
+import argparse
+import re
+import json
+import numpy as np
+import cv2
+import torch
+from tqdm import tqdm
+
+try:
+    import mmpose  # noqa: F401
+except Exception as e:
+    print(e)
+    print("mmpose error, installing transformer_utils")
+    os.system("pip install ./main/transformer_utils")
+
+
+def extract_frame_number(file_name):
+    match = re.search(r"(\d{5})", file_name)
+    if match:
+        return int(match.group(1))
+    return None
+
+
+def merge_npz_files(npz_files, output_file):
+    npz_files = sorted(npz_files, key=lambda x: extract_frame_number(os.path.basename(x)))
+    merged_data = {}
+    for file in npz_files:
+        data = np.load(file)
+        for key in data.files:
+            if key not in merged_data:
+                merged_data[key] = []
+            merged_data[key].append(data[key])
+    for key in merged_data:
+        merged_data[key] = np.stack(merged_data[key], axis=0)
+    np.savez(output_file, **merged_data)
+
+
+def npz_to_npz(pkl_path, npz_path):
+    # Load the pickle file
+    pkl_example = np.load(pkl_path, allow_pickle=True)
+    n = pkl_example["expression"].shape[0]  # Assuming this is the batch size
+    full_pose = np.concatenate(
+        [
+            pkl_example["global_orient"],
+            pkl_example["body_pose"],
+            pkl_example["jaw_pose"],
+            pkl_example["leye_pose"],
+            pkl_example["reye_pose"],
+            pkl_example["left_hand_pose"],
+            pkl_example["right_hand_pose"],
+        ],
+        axis=1,
+    )
+    # print(full_pose.shape)
+    np.savez(
+        npz_path,
+        betas=np.zeros(300),
+        poses=full_pose.reshape(n, -1),
+        expressions=np.zeros((n, 100)),
+        trans=pkl_example["transl"].reshape(n, -1),
+        model="smplx2020",
+        gender="neutral",
+        mocap_frame_rate=30,
+    )
+
+
+def get_json(root_dir, output_dir):
+    clips = []
+    dirs = os.listdir(root_dir)
+    all_length = 0
+    for dir in dirs:
+        if not dir.endswith(".mp4"):
+            continue
+        video_id = dir[:-4]
+        root = root_dir
+        try:
+            length = np.load(os.path.join(root, video_id + ".npz"), allow_pickle=True)["poses"].shape[0]
+            all_length += length
+        except Exception as e:
+            print("cant open ", dir, e)
+            continue
+        clip = {
+            "video_id": video_id,
+            "video_path": root,
+            # "audio_path": root,
+            "motion_path": root,
+            "mode": "test",
+            "start_idx": 0,
+            "end_idx": length,
+        }
+        clips.append(clip)
+    if all_length < 1:
+        print(f"skip due to total frames is less than 1500 for {root_dir}")
+        return 0
+    else:
+        with open(output_dir, "w") as f:
+            json.dump(clips, f, indent=4)
+        return all_length
+
+
+def infer(video_input, in_threshold, num_people, render_mesh, inferer, OUT_FOLDER):
+    shutil.rmtree(f"{OUT_FOLDER}/smplx", ignore_errors=True)
+    os.makedirs(f"{OUT_FOLDER}/smplx", exist_ok=True)
+    multi_person = num_people
+    cap = cv2.VideoCapture(video_input)
+    video_name = os.path.basename(video_input)
+    success = 1
+    frame = 0
+    while success:
+        success, original_img = cap.read()
+        if not success:
+            break
+        frame += 1
+        _, _, _ = inferer.infer(original_img, in_threshold, frame, multi_person, not (render_mesh))
+    cap.release()
+    npz_files = [os.path.join(OUT_FOLDER, "smplx", x) for x in os.listdir(os.path.join(OUT_FOLDER, "smplx"))]
+
+    merge_npz_files(npz_files, os.path.join(OUT_FOLDER, video_name.replace(".mp4", ".npz")))
+    shutil.rmtree(f"{OUT_FOLDER}/smplx", ignore_errors=True)
+    npz_to_npz(os.path.join(OUT_FOLDER, video_name.replace(".mp4", ".npz")), os.path.join(OUT_FOLDER, video_name.replace(".mp4", ".npz")))
+    source = video_input
+    destination = os.path.join(OUT_FOLDER, video_name.replace(".mp4", ".npz")).replace(".npz", ".mp4")
+    shutil.copy(source, destination)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--video_folder_path", type=str, default="")
+    parser.add_argument("--data_save_path", type=str, default="")
+    parser.add_argument("--json_save_path", type=str, default="")
+    args = parser.parse_args()
+    video_folder = args.video_folder_path
+
+    DEFAULT_MODEL = "smpler_x_s32"
+    OUT_FOLDER = args.data_save_path
+    os.makedirs(OUT_FOLDER, exist_ok=True)
+    num_gpus = 1 if torch.cuda.is_available() else -1
+    index = torch.cuda.current_device()
+    from main.inference import Inferer
+
+    inferer = Inferer(DEFAULT_MODEL, num_gpus, OUT_FOLDER)
+
+    for video_input in tqdm(os.listdir(video_folder)):
+        if not video_input.endswith(".mp4"):
+            continue
+        infer(os.path.join(video_folder, video_input), 0.5, False, False, inferer, OUT_FOLDER)
+    get_json(OUT_FOLDER, args.json_save_path)

assets/inference.py

@@ -0,0 +1,125 @@
+# This script is modified from https://github.com/caizhongang/SMPLer-X/blob/main/main/inference.py
+# Licensed under:
+"""
+S-Lab License 1.0
+
+Copyright 2022 S-Lab
+Redistribution and use for non-commercial purpose in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
+1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
+3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+4. In the event that redistribution and/or use for commercial purpose in source or binary forms, with or without modification is required, please contact the contributor(s) of the work.
+"""
+
+import os
+import sys
+import os.path as osp
+import numpy as np
+import torchvision.transforms as transforms
+import torch.backends.cudnn as cudnn
+import torch
+
+CUR_DIR = osp.dirname(os.path.abspath(__file__))
+sys.path.insert(0, osp.join(CUR_DIR, "..", "main"))
+sys.path.insert(0, osp.join(CUR_DIR, "..", "common"))
+from config import cfg
+from mmdet.apis import init_detector, inference_detector
+from utils.inference_utils import process_mmdet_results
+
+
+class Inferer:
+    def __init__(self, pretrained_model, num_gpus, output_folder):
+        self.output_folder = output_folder
+        self.device = torch.device("cuda") if (num_gpus > 0) else torch.device("cpu")
+        config_path = osp.join(CUR_DIR, "./config", f"config_{pretrained_model}.py")
+        ckpt_path = osp.join(CUR_DIR, "../pretrained_models", f"{pretrained_model}.pth.tar")
+        cfg.get_config_fromfile(config_path)
+        cfg.update_config(num_gpus, ckpt_path, output_folder, self.device)
+        self.cfg = cfg
+        cudnn.benchmark = True
+
+        # load model
+        from base import Demoer
+
+        demoer = Demoer()
+        demoer._make_model()
+        demoer.model.eval()
+        self.demoer = demoer
+        checkpoint_file = osp.join(CUR_DIR, "../pretrained_models/mmdet/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth")
+        config_file = osp.join(CUR_DIR, "../pretrained_models/mmdet/mmdet_faster_rcnn_r50_fpn_coco.py")
+        model = init_detector(config_file, checkpoint_file, device=self.device)  # or device='cuda:0'
+        self.model = model
+
+    def infer(self, original_img, iou_thr, frame, multi_person=False, mesh_as_vertices=False):
+        from utils.preprocessing import process_bbox, generate_patch_image
+
+        mesh_paths = []
+        smplx_paths = []
+        # prepare input image
+        transform = transforms.ToTensor()
+        vis_img = original_img.copy()
+        original_img_height, original_img_width = original_img.shape[:2]
+
+        ## mmdet inference
+        mmdet_results = inference_detector(self.model, original_img)
+
+        pred_instance = mmdet_results.pred_instances.cpu().numpy()
+        bboxes = np.concatenate((pred_instance.bboxes, pred_instance.scores[:, None]), axis=1)
+        bboxes = bboxes[pred_instance.labels == 0]
+        bboxes = np.expand_dims(bboxes, axis=0)
+        mmdet_box = process_mmdet_results(bboxes, cat_id=0, multi_person=True)
+
+        # save original image if no bbox
+        if len(mmdet_box[0]) < 1:
+            return original_img, [], []
+
+        num_bbox = 1
+        mmdet_box = mmdet_box[0]
+
+        ## loop all detected bboxes
+        for bbox_id in range(num_bbox):
+            mmdet_box_xywh = np.zeros((4))
+            mmdet_box_xywh[0] = mmdet_box[bbox_id][0]
+            mmdet_box_xywh[1] = mmdet_box[bbox_id][1]
+            mmdet_box_xywh[2] = abs(mmdet_box[bbox_id][2] - mmdet_box[bbox_id][0])
+            mmdet_box_xywh[3] = abs(mmdet_box[bbox_id][3] - mmdet_box[bbox_id][1])
+
+            # skip small bboxes by bbox_thr in pixel
+            if mmdet_box_xywh[2] < 50 or mmdet_box_xywh[3] < 150:
+                continue
+
+            bbox = process_bbox(mmdet_box_xywh, original_img_width, original_img_height)
+            img, img2bb_trans, bb2img_trans = generate_patch_image(original_img, bbox, 1.0, 0.0, False, self.cfg.input_img_shape)
+            img = transform(img.astype(np.float32)) / 255
+            img = img.to(cfg.device)[None, :, :, :]
+            inputs = {"img": img}
+            targets = {}
+            meta_info = {}
+
+            # mesh recovery
+            with torch.no_grad():
+                out = self.demoer.model(inputs, targets, meta_info, "test")
+
+            ## save single person param
+            smplx_pred = {}
+            smplx_pred["global_orient"] = out["smplx_root_pose"].reshape(-1, 3).cpu().numpy()
+            smplx_pred["body_pose"] = out["smplx_body_pose"].reshape(-1, 3).cpu().numpy()
+            smplx_pred["left_hand_pose"] = out["smplx_lhand_pose"].reshape(-1, 3).cpu().numpy()
+            smplx_pred["right_hand_pose"] = out["smplx_rhand_pose"].reshape(-1, 3).cpu().numpy()
+            smplx_pred["jaw_pose"] = out["smplx_jaw_pose"].reshape(-1, 3).cpu().numpy()
+            smplx_pred["leye_pose"] = np.zeros((1, 3))
+            smplx_pred["reye_pose"] = np.zeros((1, 3))
+            smplx_pred["betas"] = out["smplx_shape"].reshape(-1, 10).cpu().numpy()
+            smplx_pred["expression"] = out["smplx_expr"].reshape(-1, 10).cpu().numpy()
+            smplx_pred["transl"] = out["cam_trans"].reshape(-1, 3).cpu().numpy()
+            save_path_smplx = os.path.join(self.output_folder, "smplx")
+            os.makedirs(save_path_smplx, exist_ok=True)
+
+            npz_path = os.path.join(save_path_smplx, f"{frame:05}_{bbox_id}.npz")
+            np.savez(npz_path, **smplx_pred)
+            smplx_paths.append(npz_path)
+
+            vis_img = None
+            mesh_paths = None
+        return vis_img, mesh_paths, smplx_paths

assets/transforms.py

@@ -0,0 +1,344 @@
+# This script is modified from https://github.com/caizhongang/SMPLer-X/blob/main/common/utils/transforms.py
+# Licensed under:
+"""
+S-Lab License 1.0
+
+Copyright 2022 S-Lab
+Redistribution and use for non-commercial purpose in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
+1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
+3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+4. In the event that redistribution and/or use for commercial purpose in source or binary forms, with or without modification is required, please contact the contributor(s) of the work.
+"""
+
+"""
+Function rotation_matrix_to_angle_axis, rotation_matrix_to_quaternion, and quaternion_to_angle_axis are
+modified from https://github.com/eglxiang/torchgeometry/blob/master/torchgeometry/core/conversions.py
+The original code is licensed under the License: https://github.com/eglxiang/torchgeometry/blob/master/LICENSE
+We modified the code to make it compatible with the torch>=1.9.0.
+"""
+
+import torch
+import numpy as np
+from config import cfg
+from torch.nn import functional as F
+
+
+def cam2pixel(cam_coord, f, c):
+    x = cam_coord[:, 0] / cam_coord[:, 2] * f[0] + c[0]
+    y = cam_coord[:, 1] / cam_coord[:, 2] * f[1] + c[1]
+    z = cam_coord[:, 2]
+    return np.stack((x, y, z), 1)
+
+
+def pixel2cam(pixel_coord, f, c):
+    x = (pixel_coord[:, 0] - c[0]) / f[0] * pixel_coord[:, 2]
+    y = (pixel_coord[:, 1] - c[1]) / f[1] * pixel_coord[:, 2]
+    z = pixel_coord[:, 2]
+    return np.stack((x, y, z), 1)
+
+
+def world2cam(world_coord, R, t):
+    cam_coord = np.dot(R, world_coord.transpose(1, 0)).transpose(1, 0) + t.reshape(1, 3)
+    return cam_coord
+
+
+def cam2world(cam_coord, R, t):
+    world_coord = np.dot(np.linalg.inv(R), (cam_coord - t.reshape(1, 3)).transpose(1, 0)).transpose(1, 0)
+    return world_coord
+
+
+def rigid_transform_3D(A, B):
+    n, dim = A.shape
+    centroid_A = np.mean(A, axis=0)
+    centroid_B = np.mean(B, axis=0)
+    H = np.dot(np.transpose(A - centroid_A), B - centroid_B) / n
+    U, s, V = np.linalg.svd(H)
+    R = np.dot(np.transpose(V), np.transpose(U))
+    if np.linalg.det(R) < 0:
+        s[-1] = -s[-1]
+        V[2] = -V[2]
+        R = np.dot(np.transpose(V), np.transpose(U))
+
+    varP = np.var(A, axis=0).sum()
+    c = 1 / varP * np.sum(s)
+
+    t = -np.dot(c * R, np.transpose(centroid_A)) + np.transpose(centroid_B)
+    return c, R, t
+
+
+def rigid_align(A, B):
+    c, R, t = rigid_transform_3D(A, B)
+    A2 = np.transpose(np.dot(c * R, np.transpose(A))) + t
+    return A2
+
+
+def transform_joint_to_other_db(src_joint, src_name, dst_name):
+    src_joint_num = len(src_name)
+    dst_joint_num = len(dst_name)
+
+    new_joint = np.zeros(((dst_joint_num,) + src_joint.shape[1:]), dtype=np.float32)
+    for src_idx in range(len(src_name)):
+        name = src_name[src_idx]
+        if name in dst_name:
+            dst_idx = dst_name.index(name)
+            new_joint[dst_idx] = src_joint[src_idx]
+
+    return new_joint
+
+
+def rotation_matrix_to_angle_axis(rotation_matrix):
+    """Convert 3x4 rotation matrix to Rodrigues vector
+
+    Args:
+        rotation_matrix (Tensor): rotation matrix.
+
+    Returns:
+        Tensor: Rodrigues vector transformation.
+
+    Shape:
+        - Input: :math:`(N, 3, 4)`
+        - Output: :math:`(N, 3)`
+
+    Example:
+        >>> input = torch.rand(2, 3, 4)  # Nx4x4
+        >>> output = tgm.rotation_matrix_to_angle_axis(input)  # Nx3
+    """
+    # todo add check that matrix is a valid rotation matrix
+    quaternion = rotation_matrix_to_quaternion(rotation_matrix)
+    return quaternion_to_angle_axis(quaternion)
+
+
+def rotation_matrix_to_quaternion(rotation_matrix, eps=1e-6):
+    """Convert 3x4 rotation matrix to 4d quaternion vector
+
+    This algorithm is based on algorithm described in
+    https://github.com/KieranWynn/pyquaternion/blob/master/pyquaternion/quaternion.py#L201
+
+    Args:
+        rotation_matrix (Tensor): the rotation matrix to convert.
+
+    Return:
+        Tensor: the rotation in quaternion
+
+    Shape:
+        - Input: :math:`(N, 3, 4)`
+        - Output: :math:`(N, 4)`
+
+    Example:
+        >>> input = torch.rand(4, 3, 4)  # Nx3x4
+        >>> output = tgm.rotation_matrix_to_quaternion(input)  # Nx4
+    """
+    if not torch.is_tensor(rotation_matrix):
+        raise TypeError("Input type is not a torch.Tensor. Got {}".format(type(rotation_matrix)))
+
+    if len(rotation_matrix.shape) > 3:
+        raise ValueError("Input size must be a three dimensional tensor. Got {}".format(rotation_matrix.shape))
+    if not rotation_matrix.shape[-2:] == (3, 4):
+        raise ValueError("Input size must be a N x 3 x 4  tensor. Got {}".format(rotation_matrix.shape))
+
+    rmat_t = torch.transpose(rotation_matrix, 1, 2)
+
+    mask_d2 = rmat_t[:, 2, 2] < eps
+
+    mask_d0_d1 = rmat_t[:, 0, 0] > rmat_t[:, 1, 1]
+    mask_d0_nd1 = rmat_t[:, 0, 0] < -rmat_t[:, 1, 1]
+
+    t0 = 1 + rmat_t[:, 0, 0] - rmat_t[:, 1, 1] - rmat_t[:, 2, 2]
+    q0 = torch.stack([rmat_t[:, 1, 2] - rmat_t[:, 2, 1], t0, rmat_t[:, 0, 1] + rmat_t[:, 1, 0], rmat_t[:, 2, 0] + rmat_t[:, 0, 2]], -1)
+    t0_rep = t0.repeat(4, 1).t()
+
+    t1 = 1 - rmat_t[:, 0, 0] + rmat_t[:, 1, 1] - rmat_t[:, 2, 2]
+    q1 = torch.stack([rmat_t[:, 2, 0] - rmat_t[:, 0, 2], rmat_t[:, 0, 1] + rmat_t[:, 1, 0], t1, rmat_t[:, 1, 2] + rmat_t[:, 2, 1]], -1)
+    t1_rep = t1.repeat(4, 1).t()
+
+    t2 = 1 - rmat_t[:, 0, 0] - rmat_t[:, 1, 1] + rmat_t[:, 2, 2]
+    q2 = torch.stack([rmat_t[:, 0, 1] - rmat_t[:, 1, 0], rmat_t[:, 2, 0] + rmat_t[:, 0, 2], rmat_t[:, 1, 2] + rmat_t[:, 2, 1], t2], -1)
+    t2_rep = t2.repeat(4, 1).t()
+
+    t3 = 1 + rmat_t[:, 0, 0] + rmat_t[:, 1, 1] + rmat_t[:, 2, 2]
+    q3 = torch.stack([t3, rmat_t[:, 1, 2] - rmat_t[:, 2, 1], rmat_t[:, 2, 0] - rmat_t[:, 0, 2], rmat_t[:, 0, 1] - rmat_t[:, 1, 0]], -1)
+    t3_rep = t3.repeat(4, 1).t()
+
+    mask_c0 = mask_d2 * mask_d0_d1
+
+    """
+    Modified the code from the original source to make it compatible with the torch>=1.9.0
+    Original code:
+    mask_c1 = mask_d2 * (1 - mask_d0_d1)
+    mask_c2 = (1 - mask_d2) * mask_d0_nd1
+    mask_c3 = (1 - mask_d2) * (1 - mask_d0_nd1)
+    """
+    # From here
+    inv_mask_d0_d1 = ~mask_d0_d1
+    inv_mask_d0_nd1 = ~mask_d0_nd1
+    inv_mask_d2 = ~mask_d2
+    mask_c1 = mask_d2 * inv_mask_d0_d1
+    mask_c2 = inv_mask_d2 * mask_d0_nd1
+    mask_c3 = inv_mask_d2 * inv_mask_d0_nd1
+    # Until here
+
+    mask_c0 = mask_c0.view(-1, 1).type_as(q0)
+    mask_c1 = mask_c1.view(-1, 1).type_as(q1)
+    mask_c2 = mask_c2.view(-1, 1).type_as(q2)
+    mask_c3 = mask_c3.view(-1, 1).type_as(q3)
+
+    q = q0 * mask_c0 + q1 * mask_c1 + q2 * mask_c2 + q3 * mask_c3
+    q /= torch.sqrt(
+        t0_rep * mask_c0
+        + t1_rep * mask_c1  # noqa
+        + t2_rep * mask_c2
+        + t3_rep * mask_c3
+    )  # noqa
+    q *= 0.5
+    return q
+
+
+def quaternion_to_angle_axis(quaternion: torch.Tensor) -> torch.Tensor:
+    """Convert quaternion vector to angle axis of rotation.
+
+    Adapted from ceres C++ library: ceres-solver/include/ceres/rotation.h
+
+    Args:
+        quaternion (torch.Tensor): tensor with quaternions.
+
+    Return:
+        torch.Tensor: tensor with angle axis of rotation.
+
+    Shape:
+        - Input: :math:`(*, 4)` where `*` means, any number of dimensions
+        - Output: :math:`(*, 3)`
+
+    Example:
+        >>> quaternion = torch.rand(2, 4)  # Nx4
+        >>> angle_axis = tgm.quaternion_to_angle_axis(quaternion)  # Nx3
+    """
+    if not torch.is_tensor(quaternion):
+        raise TypeError("Input type is not a torch.Tensor. Got {}".format(type(quaternion)))
+
+    if not quaternion.shape[-1] == 4:
+        raise ValueError("Input must be a tensor of shape Nx4 or 4. Got {}".format(quaternion.shape))
+    # unpack input and compute conversion
+    q1: torch.Tensor = quaternion[..., 1]
+    q2: torch.Tensor = quaternion[..., 2]
+    q3: torch.Tensor = quaternion[..., 3]
+    sin_squared_theta: torch.Tensor = q1 * q1 + q2 * q2 + q3 * q3
+
+    sin_theta: torch.Tensor = torch.sqrt(sin_squared_theta)
+    cos_theta: torch.Tensor = quaternion[..., 0]
+    two_theta: torch.Tensor = 2.0 * torch.where(cos_theta < 0.0, torch.atan2(-sin_theta, -cos_theta), torch.atan2(sin_theta, cos_theta))
+
+    k_pos: torch.Tensor = two_theta / sin_theta
+    k_neg: torch.Tensor = 2.0 * torch.ones_like(sin_theta)
+    k: torch.Tensor = torch.where(sin_squared_theta > 0.0, k_pos, k_neg)
+
+    angle_axis: torch.Tensor = torch.zeros_like(quaternion)[..., :3]
+    angle_axis[..., 0] += q1 * k
+    angle_axis[..., 1] += q2 * k
+    angle_axis[..., 2] += q3 * k
+    return angle_axis
+
+
+def rot6d_to_axis_angle(x):
+    batch_size = x.shape[0]
+
+    x = x.view(-1, 3, 2)
+    a1 = x[:, :, 0]
+    a2 = x[:, :, 1]
+    b1 = F.normalize(a1)
+    b2 = F.normalize(a2 - torch.einsum("bi,bi->b", b1, a2).unsqueeze(-1) * b1)
+    b3 = torch.cross(b1, b2)
+    rot_mat = torch.stack((b1, b2, b3), dim=-1)  # 3x3 rotation matrix
+
+    rot_mat = torch.cat([rot_mat, torch.zeros((batch_size, 3, 1)).to(cfg.device).float()], 2)  # 3x4 rotation matrix
+    axis_angle = rotation_matrix_to_angle_axis(rot_mat).reshape(-1, 3)  # axis-angle
+    axis_angle[torch.isnan(axis_angle)] = 0.0
+    return axis_angle
+
+
+def sample_joint_features(img_feat, joint_xy):
+    height, width = img_feat.shape[2:]
+    x = joint_xy[:, :, 0] / (width - 1) * 2 - 1
+    y = joint_xy[:, :, 1] / (height - 1) * 2 - 1
+    grid = torch.stack((x, y), 2)[:, :, None, :]
+    img_feat = F.grid_sample(img_feat, grid, align_corners=True)[:, :, :, 0]  # batch_size, channel_dim, joint_num
+    img_feat = img_feat.permute(0, 2, 1).contiguous()  # batch_size, joint_num, channel_dim
+    return img_feat
+
+
+def soft_argmax_2d(heatmap2d):
+    batch_size = heatmap2d.shape[0]
+    height, width = heatmap2d.shape[2:]
+    heatmap2d = heatmap2d.reshape((batch_size, -1, height * width))
+    heatmap2d = F.softmax(heatmap2d, 2)
+    heatmap2d = heatmap2d.reshape((batch_size, -1, height, width))
+
+    accu_x = heatmap2d.sum(dim=(2))
+    accu_y = heatmap2d.sum(dim=(3))
+
+    accu_x = accu_x * torch.arange(width).float().to(cfg.device)[None, None, :]
+    accu_y = accu_y * torch.arange(height).float().to(cfg.device)[None, None, :]
+
+    accu_x = accu_x.sum(dim=2, keepdim=True)
+    accu_y = accu_y.sum(dim=2, keepdim=True)
+
+    coord_out = torch.cat((accu_x, accu_y), dim=2)
+    return coord_out
+
+
+def soft_argmax_3d(heatmap3d):
+    batch_size = heatmap3d.shape[0]
+    depth, height, width = heatmap3d.shape[2:]
+    heatmap3d = heatmap3d.reshape((batch_size, -1, depth * height * width))
+    heatmap3d = F.softmax(heatmap3d, 2)
+    heatmap3d = heatmap3d.reshape((batch_size, -1, depth, height, width))
+
+    accu_x = heatmap3d.sum(dim=(2, 3))
+    accu_y = heatmap3d.sum(dim=(2, 4))
+    accu_z = heatmap3d.sum(dim=(3, 4))
+
+    accu_x = accu_x * torch.arange(width).float().to(cfg.device)[None, None, :]
+    accu_y = accu_y * torch.arange(height).float().to(cfg.device)[None, None, :]
+    accu_z = accu_z * torch.arange(depth).float().to(cfg.device)[None, None, :]
+
+    accu_x = accu_x.sum(dim=2, keepdim=True)
+    accu_y = accu_y.sum(dim=2, keepdim=True)
+    accu_z = accu_z.sum(dim=2, keepdim=True)
+
+    coord_out = torch.cat((accu_x, accu_y, accu_z), dim=2)
+    return coord_out
+
+
+def restore_bbox(bbox_center, bbox_size, aspect_ratio, extension_ratio):
+    bbox = bbox_center.view(-1, 1, 2) + torch.cat(
+        (-bbox_size.view(-1, 1, 2) / 2.0, bbox_size.view(-1, 1, 2) / 2.0), 1
+    )  # xyxy in (cfg.output_hm_shape[2], cfg.output_hm_shape[1]) space
+    bbox[:, :, 0] = bbox[:, :, 0] / cfg.output_hm_shape[2] * cfg.input_body_shape[1]
+    bbox[:, :, 1] = bbox[:, :, 1] / cfg.output_hm_shape[1] * cfg.input_body_shape[0]
+    bbox = bbox.view(-1, 4)
+
+    # xyxy -> xywh
+    bbox[:, 2] = bbox[:, 2] - bbox[:, 0]
+    bbox[:, 3] = bbox[:, 3] - bbox[:, 1]
+
+    # aspect ratio preserving bbox
+    w = bbox[:, 2]
+    h = bbox[:, 3]
+    c_x = bbox[:, 0] + w / 2.0
+    c_y = bbox[:, 1] + h / 2.0
+
+    mask1 = w > (aspect_ratio * h)
+    mask2 = w < (aspect_ratio * h)
+    h[mask1] = w[mask1] / aspect_ratio
+    w[mask2] = h[mask2] * aspect_ratio
+
+    bbox[:, 2] = w * extension_ratio
+    bbox[:, 3] = h * extension_ratio
+    bbox[:, 0] = c_x - bbox[:, 2] / 2.0
+    bbox[:, 1] = c_y - bbox[:, 3] / 2.0
+
+    # xywh -> xyxy
+    bbox[:, 2] = bbox[:, 2] + bbox[:, 0]
+    bbox[:, 3] = bbox[:, 3] + bbox[:, 1]
+    return bbox

audio_0_retri_0_watermarked.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7af0a48abf6efb3195378081c8018b809e25ee6e24e5d731cfa989b3d671fe43
+size 1022865

configs/gradio.yaml

@@ -0,0 +1,77 @@
+wandb_project: 'TANGO'
+exp_name: 'debug'
+
+wandb_entity: ''
+wandb_key: ""
+wandb_log_dir: '/content/outputs/wandb'
+output_dir: ./outputs/
+log_period: 1
+seed: 42
+
+data:
+  name_pyfile: "datasets.beat2_v5"
+  class_name: "BEAT2Dataset"
+  train_bs: 2
+  meta_paths:
+    - "./datasets/data_json/show-oliver-s40_w128.json"
+  # test_meta_paths: "./datasets/data_json/show_oliver_test/Abortion_Laws_-_Last_Week_Tonight_with_John_Oliver_HBO-DRauXXz6t0Y.webm.json"
+  "test_meta_paths": "./datasets/data_json/show_oliver_test/Stupid_Watergate_-_Last_Week_Tonight_with_John_Oliver_HBO-FVFdsl29s_Q.mkv.json"
+  pose_norm: False
+  pose_fps: 30
+  rot6d: True
+  pose_dims: 825
+  pose_length: 128
+  stride: 20
+  test_length: 128
+  audio_sr: 16000
+  audio_fps: 16000
+
+model:
+  name_pyfile: "models.jointembedding_high_env0"
+  class_name: "JointEmbedding"
+  motion_f: 256
+  audio_rep: wave16k
+  audio_sr: 16000
+  audio_fps: 16000
+  audio_norm: False
+  audio_f: 256
+  word_rep: textgrid
+  word_index_num: 11195
+  word_dims: 300
+  facial_rep: smplxflame_30
+  facial_dims: 100
+  facial_norm: False
+  facial_f: 0
+  f_pre_encoder: null
+  f_encoder: null
+  f_fix_pre: False
+  id_rep: onehot
+  speaker_f: 0
+  hidden_size: 512
+  n_layer: 1
+  motion_dim: 825
+ 
+validation:
+  val_loss_steps: 1
+  validation_steps: 1000
+  # guidance_scale: 3.5
+  # denoising_steps: 20 
+
+solver:
+  gradient_accumulation_steps: 1
+  # mixed_precision: 'fp16'
+  # enable_xformers_memory_efficient_attention: True 
+  gradient_checkpointing: False 
+  max_train_steps: 5000000
+  max_grad_norm: 1.0
+  # lr
+  learning_rate: 2e-5
+  scale_lr: False 
+  lr_warmup_steps: 50
+  lr_scheduler: 'constant'
+  # optimizer
+  use_8bit_adam: False 
+  adam_beta1: 0.9
+  adam_beta2: 0.999
+  adam_weight_decay:  1.0e-2
+  adam_epsilon: 1.0e-8

configs/gradio_speaker1.yaml

@@ -0,0 +1,77 @@
+wandb_project: 'TANGO'
+exp_name: 'debug'
+
+wandb_entity: ''
+wandb_key: ""
+wandb_log_dir: '/content/outputs/wandb'
+output_dir: ./outputs/
+log_period: 1
+seed: 42
+
+data:
+  name_pyfile: "datasets.beat2_v5"
+  class_name: "BEAT2Dataset"
+  train_bs: 2
+  meta_paths:
+    - "./datasets/data_json/show-oliver-s40_w128.json"
+  # test_meta_paths: "./datasets/data_json/show_oliver_test/Abortion_Laws_-_Last_Week_Tonight_with_John_Oliver_HBO-DRauXXz6t0Y.webm.json"
+  "test_meta_paths": "./datasets/data_json/youtube_test/speaker1.json"
+  pose_norm: False
+  pose_fps: 30
+  rot6d: True
+  pose_dims: 825
+  pose_length: 128
+  stride: 20
+  test_length: 128
+  audio_sr: 16000
+  audio_fps: 16000
+
+model:
+  name_pyfile: "models.jointembedding_high_env0"
+  class_name: "JointEmbedding"
+  motion_f: 256
+  audio_rep: wave16k
+  audio_sr: 16000
+  audio_fps: 16000
+  audio_norm: False
+  audio_f: 256
+  word_rep: textgrid
+  word_index_num: 11195
+  word_dims: 300
+  facial_rep: smplxflame_30
+  facial_dims: 100
+  facial_norm: False
+  facial_f: 0
+  f_pre_encoder: null
+  f_encoder: null
+  f_fix_pre: False
+  id_rep: onehot
+  speaker_f: 0
+  hidden_size: 512
+  n_layer: 1
+  motion_dim: 825
+ 
+validation:
+  val_loss_steps: 1
+  validation_steps: 1000
+  # guidance_scale: 3.5
+  # denoising_steps: 20 
+
+solver:
+  gradient_accumulation_steps: 1
+  # mixed_precision: 'fp16'
+  # enable_xformers_memory_efficient_attention: True 
+  gradient_checkpointing: False 
+  max_train_steps: 5000000
+  max_grad_norm: 1.0
+  # lr
+  learning_rate: 2e-5
+  scale_lr: False 
+  lr_warmup_steps: 50
+  lr_scheduler: 'constant'
+  # optimizer
+  use_8bit_adam: False 
+  adam_beta1: 0.9
+  adam_beta2: 0.999
+  adam_weight_decay:  1.0e-2
+  adam_epsilon: 1.0e-8

configs/gradio_speaker7.yaml

@@ -0,0 +1,77 @@
+wandb_project: 'TANGO'
+exp_name: 'debug'
+
+wandb_entity: ''
+wandb_key: ""
+wandb_log_dir: '/content/outputs/wandb'
+output_dir: ./outputs/
+log_period: 1
+seed: 42
+
+data:
+  name_pyfile: "datasets.beat2_v5"
+  class_name: "BEAT2Dataset"
+  train_bs: 2
+  meta_paths:
+    - "./datasets/data_json/show-oliver-s40_w128.json"
+  # test_meta_paths: "./datasets/data_json/show_oliver_test/Abortion_Laws_-_Last_Week_Tonight_with_John_Oliver_HBO-DRauXXz6t0Y.webm.json"
+  "test_meta_paths": "./datasets/data_json/youtube_test/speaker7.json"
+  pose_norm: False
+  pose_fps: 30
+  rot6d: True
+  pose_dims: 825
+  pose_length: 128
+  stride: 20
+  test_length: 128
+  audio_sr: 16000
+  audio_fps: 16000
+
+model:
+  name_pyfile: "models.jointembedding_high_env0"
+  class_name: "JointEmbedding"
+  motion_f: 256
+  audio_rep: wave16k
+  audio_sr: 16000
+  audio_fps: 16000
+  audio_norm: False
+  audio_f: 256
+  word_rep: textgrid
+  word_index_num: 11195
+  word_dims: 300
+  facial_rep: smplxflame_30
+  facial_dims: 100
+  facial_norm: False
+  facial_f: 0
+  f_pre_encoder: null
+  f_encoder: null
+  f_fix_pre: False
+  id_rep: onehot
+  speaker_f: 0
+  hidden_size: 512
+  n_layer: 1
+  motion_dim: 825
+ 
+validation:
+  val_loss_steps: 1
+  validation_steps: 1000
+  # guidance_scale: 3.5
+  # denoising_steps: 20 
+
+solver:
+  gradient_accumulation_steps: 1
+  # mixed_precision: 'fp16'
+  # enable_xformers_memory_efficient_attention: True 
+  gradient_checkpointing: False 
+  max_train_steps: 5000000
+  max_grad_norm: 1.0
+  # lr
+  learning_rate: 2e-5
+  scale_lr: False 
+  lr_warmup_steps: 50
+  lr_scheduler: 'constant'
+  # optimizer
+  use_8bit_adam: False 
+  adam_beta1: 0.9
+  adam_beta2: 0.999
+  adam_weight_decay:  1.0e-2
+  adam_epsilon: 1.0e-8

configs/gradio_speaker8.yaml

@@ -0,0 +1,77 @@
+wandb_project: 'TANGO'
+exp_name: 'debug'
+
+wandb_entity: ''
+wandb_key: ""
+wandb_log_dir: '/content/outputs/wandb'
+output_dir: ./outputs/
+log_period: 1
+seed: 42
+
+data:
+  name_pyfile: "datasets.beat2_v5"
+  class_name: "BEAT2Dataset"
+  train_bs: 2
+  meta_paths:
+    - "./datasets/data_json/show-oliver-s40_w128.json"
+  # test_meta_paths: "./datasets/data_json/show_oliver_test/Abortion_Laws_-_Last_Week_Tonight_with_John_Oliver_HBO-DRauXXz6t0Y.webm.json"
+  "test_meta_paths": "./datasets/data_json/youtube_test/speaker8.json"
+  pose_norm: False
+  pose_fps: 30
+  rot6d: True
+  pose_dims: 825
+  pose_length: 128
+  stride: 20
+  test_length: 128
+  audio_sr: 16000
+  audio_fps: 16000
+
+model:
+  name_pyfile: "models.jointembedding_high_env0"
+  class_name: "JointEmbedding"
+  motion_f: 256
+  audio_rep: wave16k
+  audio_sr: 16000
+  audio_fps: 16000
+  audio_norm: False
+  audio_f: 256
+  word_rep: textgrid
+  word_index_num: 11195
+  word_dims: 300
+  facial_rep: smplxflame_30
+  facial_dims: 100
+  facial_norm: False
+  facial_f: 0
+  f_pre_encoder: null
+  f_encoder: null
+  f_fix_pre: False
+  id_rep: onehot
+  speaker_f: 0
+  hidden_size: 512
+  n_layer: 1
+  motion_dim: 825
+ 
+validation:
+  val_loss_steps: 1
+  validation_steps: 1000
+  # guidance_scale: 3.5
+  # denoising_steps: 20 
+
+solver:
+  gradient_accumulation_steps: 1
+  # mixed_precision: 'fp16'
+  # enable_xformers_memory_efficient_attention: True 
+  gradient_checkpointing: False 
+  max_train_steps: 5000000
+  max_grad_norm: 1.0
+  # lr
+  learning_rate: 2e-5
+  scale_lr: False 
+  lr_warmup_steps: 50
+  lr_scheduler: 'constant'
+  # optimizer
+  use_8bit_adam: False 
+  adam_beta1: 0.9
+  adam_beta2: 0.999
+  adam_weight_decay:  1.0e-2
+  adam_epsilon: 1.0e-8

configs/gradio_speaker9.yaml

@@ -0,0 +1,77 @@
+wandb_project: 'TANGO'
+exp_name: 'debug'
+
+wandb_entity: ''
+wandb_key: ""
+wandb_log_dir: '/content/outputs/wandb'
+output_dir: ./outputs/
+log_period: 1
+seed: 42
+
+data:
+  name_pyfile: "datasets.beat2_v5"
+  class_name: "BEAT2Dataset"
+  train_bs: 2
+  meta_paths:
+    - "./datasets/data_json/show-oliver-s40_w128.json"
+  # test_meta_paths: "./datasets/data_json/show_oliver_test/Abortion_Laws_-_Last_Week_Tonight_with_John_Oliver_HBO-DRauXXz6t0Y.webm.json"
+  "test_meta_paths": "./datasets/data_json/youtube_test/speaker9.json"
+  pose_norm: False
+  pose_fps: 30
+  rot6d: True
+  pose_dims: 825
+  pose_length: 128
+  stride: 20
+  test_length: 128
+  audio_sr: 16000
+  audio_fps: 16000
+
+model:
+  name_pyfile: "models.jointembedding_high_env0"
+  class_name: "JointEmbedding"
+  motion_f: 256
+  audio_rep: wave16k
+  audio_sr: 16000
+  audio_fps: 16000
+  audio_norm: False
+  audio_f: 256
+  word_rep: textgrid
+  word_index_num: 11195
+  word_dims: 300
+  facial_rep: smplxflame_30
+  facial_dims: 100
+  facial_norm: False
+  facial_f: 0
+  f_pre_encoder: null
+  f_encoder: null
+  f_fix_pre: False
+  id_rep: onehot
+  speaker_f: 0
+  hidden_size: 512
+  n_layer: 1
+  motion_dim: 825
+ 
+validation:
+  val_loss_steps: 1
+  validation_steps: 1000
+  # guidance_scale: 3.5
+  # denoising_steps: 20 
+
+solver:
+  gradient_accumulation_steps: 1
+  # mixed_precision: 'fp16'
+  # enable_xformers_memory_efficient_attention: True 
+  gradient_checkpointing: False 
+  max_train_steps: 5000000
+  max_grad_norm: 1.0
+  # lr
+  learning_rate: 2e-5
+  scale_lr: False 
+  lr_warmup_steps: 50
+  lr_scheduler: 'constant'
+  # optimizer
+  use_8bit_adam: False 
+  adam_beta1: 0.9
+  adam_beta2: 0.999
+  adam_weight_decay:  1.0e-2
+  adam_epsilon: 1.0e-8

create_graph.py

@@ -0,0 +1,507 @@
+"""
+input: json file with video, audio, motion paths
+output: igraph object with nodes containing video, audio, motion, position, velocity, axis_angle, previous, next, frame, fps
+
+preprocess:
+1. assume you have a video for one speaker in folder, listed in
+    -- video_a.mp4
+    -- video_b.mp4
+    run process_video.py to extract frames and audio
+"""
+
+import os
+import json
+import smplx
+import torch
+import igraph
+import numpy as np
+import subprocess
+import utils.rotation_conversions as rc
+from moviepy.editor import VideoClip, AudioFileClip
+from tqdm import tqdm
+import imageio
+import tempfile
+import argparse
+import time
+
+SCRIPT_PATH = os.path.dirname(os.path.realpath(__file__))
+
+
+def get_motion_reps_tensor(motion_tensor, smplx_model, pose_fps=30, device="cuda"):
+    bs, n, _ = motion_tensor.shape
+    motion_tensor = motion_tensor.float().to(device)
+    motion_tensor_reshaped = motion_tensor.reshape(bs * n, 165)
+
+    output = smplx_model(
+        betas=torch.zeros(bs * n, 300, device=device),
+        transl=torch.zeros(bs * n, 3, device=device),
+        expression=torch.zeros(bs * n, 100, device=device),
+        jaw_pose=torch.zeros(bs * n, 3, device=device),
+        global_orient=torch.zeros(bs * n, 3, device=device),
+        body_pose=motion_tensor_reshaped[:, 3 : 21 * 3 + 3],
+        left_hand_pose=motion_tensor_reshaped[:, 25 * 3 : 40 * 3],
+        right_hand_pose=motion_tensor_reshaped[:, 40 * 3 : 55 * 3],
+        return_joints=True,
+        leye_pose=torch.zeros(bs * n, 3, device=device),
+        reye_pose=torch.zeros(bs * n, 3, device=device),
+    )
+
+    joints = output["joints"].reshape(bs, n, 127, 3)[:, :, :55, :]
+    dt = 1 / pose_fps
+    init_vel = (joints[:, 1:2] - joints[:, 0:1]) / dt
+    middle_vel = (joints[:, 2:] - joints[:, :-2]) / (2 * dt)
+    final_vel = (joints[:, -1:] - joints[:, -2:-1]) / dt
+    vel = torch.cat([init_vel, middle_vel, final_vel], dim=1)
+
+    position = joints
+    rot_matrices = rc.axis_angle_to_matrix(motion_tensor.reshape(bs, n, 55, 3))
+    rot6d = rc.matrix_to_rotation_6d(rot_matrices).reshape(bs, n, 55, 6)
+
+    init_vel_ang = (motion_tensor[:, 1:2] - motion_tensor[:, 0:1]) / dt
+    middle_vel_ang = (motion_tensor[:, 2:] - motion_tensor[:, :-2]) / (2 * dt)
+    final_vel_ang = (motion_tensor[:, -1:] - motion_tensor[:, -2:-1]) / dt
+    angular_velocity = torch.cat([init_vel_ang, middle_vel_ang, final_vel_ang], dim=1).reshape(bs, n, 55, 3)
+
+    rep15d = torch.cat([position, vel, rot6d, angular_velocity], dim=3).reshape(bs, n, 55 * 15)
+
+    return {
+        "position": position,
+        "velocity": vel,
+        "rotation": rot6d,
+        "axis_angle": motion_tensor,
+        "angular_velocity": angular_velocity,
+        "rep15d": rep15d,
+    }
+
+
+def get_motion_reps(motion, smplx_model, pose_fps=30):
+    gt_motion_tensor = motion["poses"]
+    n = gt_motion_tensor.shape[0]
+    bs = 1
+    gt_motion_tensor = torch.from_numpy(gt_motion_tensor).float().to(device).unsqueeze(0)
+    gt_motion_tensor_reshaped = gt_motion_tensor.reshape(bs * n, 165)
+    output = smplx_model(
+        betas=torch.zeros(bs * n, 300).to(device),
+        transl=torch.zeros(bs * n, 3).to(device),
+        expression=torch.zeros(bs * n, 100).to(device),
+        jaw_pose=torch.zeros(bs * n, 3).to(device),
+        global_orient=torch.zeros(bs * n, 3).to(device),
+        body_pose=gt_motion_tensor_reshaped[:, 3 : 21 * 3 + 3],
+        left_hand_pose=gt_motion_tensor_reshaped[:, 25 * 3 : 40 * 3],
+        right_hand_pose=gt_motion_tensor_reshaped[:, 40 * 3 : 55 * 3],
+        return_joints=True,
+        leye_pose=torch.zeros(bs * n, 3).to(device),
+        reye_pose=torch.zeros(bs * n, 3).to(device),
+    )
+    joints = output["joints"].detach().cpu().numpy().reshape(n, 127, 3)[:, :55, :]
+    dt = 1 / pose_fps
+    init_vel = (joints[1:2] - joints[0:1]) / dt
+    middle_vel = (joints[2:] - joints[:-2]) / (2 * dt)
+    final_vel = (joints[-1:] - joints[-2:-1]) / dt
+    vel = np.concatenate([init_vel, middle_vel, final_vel], axis=0)
+    position = joints
+    rot_matrices = rc.axis_angle_to_matrix(gt_motion_tensor.reshape(1, n, 55, 3))[0]
+    rot6d = rc.matrix_to_rotation_6d(rot_matrices).reshape(n, 55, 6).cpu().numpy()
+
+    init_vel = (motion["poses"][1:2] - motion["poses"][0:1]) / dt
+    middle_vel = (motion["poses"][2:] - motion["poses"][:-2]) / (2 * dt)
+    final_vel = (motion["poses"][-1:] - motion["poses"][-2:-1]) / dt
+    angular_velocity = np.concatenate([init_vel, middle_vel, final_vel], axis=0).reshape(n, 55, 3)
+
+    rep15d = np.concatenate([position, vel, rot6d, angular_velocity], axis=2).reshape(n, 55 * 15)
+    return {
+        "position": position,
+        "velocity": vel,
+        "rotation": rot6d,
+        "axis_angle": motion["poses"],
+        "angular_velocity": angular_velocity,
+        "rep15d": rep15d,
+        "trans": motion["trans"],
+    }
+
+
+def create_graph(json_path, smplx_model):
+    fps = 30
+    data_meta = json.load(open(json_path, "r"))
+    graph = igraph.Graph(directed=True)
+    global_i = 0
+    for data_item in data_meta:
+        video_path = os.path.join(data_item["video_path"], data_item["video_id"] + ".mp4")
+        # audio_path = os.path.join(data_item['audio_path'], data_item['video_id'] +  ".wav")
+        motion_path = os.path.join(data_item["motion_path"], data_item["video_id"] + ".npz")
+        video_id = data_item.get("video_id", "")
+        motion = np.load(motion_path, allow_pickle=True)
+        motion_reps = get_motion_reps(motion, smplx_model)
+        position = motion_reps["position"]
+        velocity = motion_reps["velocity"]
+        trans = motion_reps["trans"]
+        axis_angle = motion_reps["axis_angle"]
+        # audio, sr = librosa.load(audio_path, sr=None)
+        # audio = librosa.resample(audio, orig_sr=sr, target_sr=16000)
+        all_frames = []
+        reader = imageio.get_reader(video_path)
+        all_frames = []
+        for frame in reader:
+            all_frames.append(frame)
+        video_frames = np.array(all_frames)
+        min_frames = min(len(video_frames), position.shape[0])
+        position = position[:min_frames]
+        velocity = velocity[:min_frames]
+        video_frames = video_frames[:min_frames]
+        # print(min_frames)
+        for i in tqdm(range(min_frames)):
+            if i == 0:
+                previous = -1
+                next_node = global_i + 1
+            elif i == min_frames - 1:
+                previous = global_i - 1
+                next_node = -1
+            else:
+                previous = global_i - 1
+                next_node = global_i + 1
+            graph.add_vertex(
+                idx=global_i,
+                name=video_id,
+                motion=motion_reps,
+                position=position[i],
+                velocity=velocity[i],
+                axis_angle=axis_angle[i],
+                trans=trans[i],
+                # audio=audio[],
+                video=video_frames[i],
+                previous=previous,
+                next=next_node,
+                frame=i,
+                fps=fps,
+            )
+            global_i += 1
+    return graph
+
+
+def create_edges(graph):
+    adaptive_length = [-4, -3, -2, -1, 1, 2, 3, 4]
+    # print()
+    for i, node in enumerate(graph.vs):
+        current_position = node["position"]
+        current_velocity = node["velocity"]
+        current_trans = node["trans"]
+        # print(current_position.shape, current_velocity.shape)
+        avg_position = np.zeros(current_position.shape[0])
+        avg_velocity = np.zeros(current_position.shape[0])
+        avg_trans = 0
+        count = 0
+        for node_offset in adaptive_length:
+            idx = i + node_offset
+            if idx < 0 or idx >= len(graph.vs):
+                continue
+            if node_offset < 0:
+                if graph.vs[idx]["next"] == -1:
+                    continue
+            else:
+                if graph.vs[idx]["previous"] == -1:
+                    continue
+            # add check
+            other_node = graph.vs[idx]
+            other_position = other_node["position"]
+            other_velocity = other_node["velocity"]
+            other_trans = other_node["trans"]
+            # print(other_position.shape, other_velocity.shape)
+            avg_position += np.linalg.norm(current_position - other_position, axis=1)
+            avg_velocity += np.linalg.norm(current_velocity - other_velocity, axis=1)
+            avg_trans += np.linalg.norm(current_trans - other_trans, axis=0)
+            count += 1
+
+        if count == 0:
+            continue
+        threshold_position = avg_position / count
+        threshold_velocity = avg_velocity / count
+        threshold_trans = avg_trans / count
+        # print(threshold_position, threshold_velocity, threshold_trans)
+        for j, other_node in enumerate(graph.vs):
+            if i == j:
+                continue
+            if j == node["previous"] or j == node["next"]:
+                graph.add_edge(i, j, is_continue=1)
+                continue
+            other_position = other_node["position"]
+            other_velocity = other_node["velocity"]
+            other_trans = other_node["trans"]
+            position_similarity = np.linalg.norm(current_position - other_position, axis=1)
+            velocity_similarity = np.linalg.norm(current_velocity - other_velocity, axis=1)
+            trans_similarity = np.linalg.norm(current_trans - other_trans, axis=0)
+            if trans_similarity < threshold_trans:
+                if np.sum(position_similarity < threshold_position) >= 45 and np.sum(velocity_similarity < threshold_velocity) >= 45:
+                    graph.add_edge(i, j, is_continue=0)
+
+    print(f"nodes: {len(graph.vs)}, edges: {len(graph.es)}")
+    in_degrees = graph.indegree()
+    out_degrees = graph.outdegree()
+    avg_in_degree = sum(in_degrees) / len(in_degrees)
+    avg_out_degree = sum(out_degrees) / len(out_degrees)
+    print(f"Average In-degree: {avg_in_degree}")
+    print(f"Average Out-degree: {avg_out_degree}")
+    print(f"max in degree: {max(in_degrees)}, max out degree: {max(out_degrees)}")
+    print(f"min in degree: {min(in_degrees)}, min out degree: {min(out_degrees)}")
+    # igraph.plot(graph, target="/content/test.png", bbox=(1000, 1000), vertex_size=10)
+    return graph
+
+
+def random_walk(graph, walk_length, start_node=None):
+    if start_node is None:
+        start_node = np.random.choice(graph.vs)
+    walk = [start_node]
+    is_continue = [1]
+    for _ in range(walk_length):
+        current_node = walk[-1]
+        neighbor_indices = graph.neighbors(current_node.index, mode="OUT")
+        if not neighbor_indices:
+            break
+        next_idx = np.random.choice(neighbor_indices)
+        edge_id = graph.get_eid(current_node.index, next_idx)
+        is_cont = graph.es[edge_id]["is_continue"]
+        walk.append(graph.vs[next_idx])
+        is_continue.append(is_cont)
+    return walk, is_continue
+
+
+def path_visualization(graph, path, is_continue, save_path, verbose_continue=False, audio_path=None, return_motion=False):
+    all_frames = [node["video"] for node in path]
+    average_dis_continue = 1 - sum(is_continue) / len(is_continue)
+    if verbose_continue:
+        print("average_dis_continue:", average_dis_continue)
+
+    fps = graph.vs[0]["fps"]
+    duration = len(all_frames) / fps
+
+    def make_frame(t):
+        idx = min(int(t * fps), len(all_frames) - 1)
+        return all_frames[idx]
+
+    video_only_path = f"/tmp/video_only_{time.time()}.mp4"  # Temporary file
+    video_clip = VideoClip(make_frame, duration=duration)
+    video_clip.write_videofile(video_only_path, codec="libx264", fps=fps, audio=False)
+
+    # Optionally, ensure audio and video durations match
+    if audio_path is not None:
+        audio_clip = AudioFileClip(audio_path)
+        video_duration = video_clip.duration
+        audio_duration = audio_clip.duration
+
+        if audio_duration > video_duration:
+            # Trim the audio
+            trimmed_audio_path = "trimmed_audio.aac"
+            audio_clip = audio_clip.subclip(0, video_duration)
+            audio_clip.write_audiofile(trimmed_audio_path)
+            audio_input = trimmed_audio_path
+        else:
+            audio_input = audio_path
+
+        # Use FFmpeg to combine video and audio
+        ffmpeg_command = [
+            "ffmpeg",
+            "-y",
+            "-i",
+            video_only_path,
+            "-i",
+            audio_input,
+            "-c:v",
+            "copy",
+            "-c:a",
+            "aac",
+            "-strict",
+            "experimental",
+            save_path,
+        ]
+        subprocess.check_call(ffmpeg_command)
+
+        # Clean up temporary files if necessary
+        os.remove(video_only_path)
+        if audio_input != audio_path:
+            os.remove(audio_input)
+
+    if return_motion:
+        all_motion = [node["axis_angle"] for node in path]
+        all_motion = np.stack(all_motion, 0)
+        return all_motion
+
+
+def generate_transition_video(frame_start_path, frame_end_path, output_video_path):
+    import subprocess
+    import os
+
+    # Define the path to your model and inference script
+    model_path = os.path.join(SCRIPT_PATH, "frame-interpolation-pytorch/film_net_fp32.pt")
+    inference_script = os.path.join(SCRIPT_PATH, "frame-interpolation-pytorch/inference.py")
+
+    # Build the command to run the inference script
+    command = [
+        "python",
+        inference_script,
+        model_path,
+        frame_start_path,
+        frame_end_path,
+        "--save_path",
+        output_video_path,
+        "--gpu",
+        "--frames",
+        "3",
+        "--fps",
+        "30",
+    ]
+
+    # Run the command
+    try:
+        subprocess.run(command, check=True)
+        print(f"Generated transition video saved at {output_video_path}")
+    except subprocess.CalledProcessError as e:
+        print(f"Error occurred while generating transition video: {e}")
+
+
+def path_visualization_v2(graph, path, is_continue, save_path, verbose_continue=False, audio_path=None, return_motion=False):
+    """
+    this is for hugging face demo for fast interpolation. our paper use a diffusion based interpolation method
+    """
+    all_frames = [node["video"] for node in path]
+    average_dis_continue = 1 - sum(is_continue) / len(is_continue)
+    if verbose_continue:
+        print("average_dis_continue:", average_dis_continue)
+    duration = len(all_frames) / graph.vs[0]["fps"]
+
+    # First loop: Confirm where blending is needed
+    discontinuity_indices = []
+    for i, cont in enumerate(is_continue):
+        if cont == 0:
+            discontinuity_indices.append(i)
+
+    # Identify blending positions without overlapping
+    blend_positions = []
+    processed_frames = set()
+    for i in discontinuity_indices:
+        # Define the frames for blending: i-2 to i+2
+        start_idx = i - 2
+        end_idx = i + 2
+        # Check index boundaries
+        if start_idx < 0 or end_idx >= len(all_frames):
+            continue  # Skip if indices are out of bounds
+        # Check for overlapping frames
+        overlap = any(idx in processed_frames for idx in range(i - 1, i + 2))
+        if overlap:
+            continue  # Skip if frames have been processed
+        # Mark frames as processed
+        processed_frames.update(range(i - 1, i + 2))
+        blend_positions.append(i)
+
+    # Second loop: Perform blending
+    temp_dir = tempfile.mkdtemp(prefix="blending_frames_")
+    for i in tqdm(blend_positions):
+        start_frame_idx = i - 2
+        end_frame_idx = i + 2
+        frame_start = all_frames[start_frame_idx]
+        frame_end = all_frames[end_frame_idx]
+        frame_start_path = os.path.join(temp_dir, f"frame_{start_frame_idx}.png")
+        frame_end_path = os.path.join(temp_dir, f"frame_{end_frame_idx}.png")
+        # Save the start and end frames as images
+        imageio.imwrite(frame_start_path, frame_start)
+        imageio.imwrite(frame_end_path, frame_end)
+
+        # Call FiLM API to generate video
+        generated_video_path = os.path.join(temp_dir, f"generated_{start_frame_idx}_{end_frame_idx}.mp4")
+        generate_transition_video(frame_start_path, frame_end_path, generated_video_path)
+
+        # Read the generated video frames
+        reader = imageio.get_reader(generated_video_path)
+        generated_frames = [frame for frame in reader]
+        reader.close()
+
+        # Replace the middle three frames (i-1, i, i+1) in all_frames
+        total_generated_frames = len(generated_frames)
+        if total_generated_frames < 5:
+            print(f"Generated video has insufficient frames ({total_generated_frames}). Skipping blending at position {i}.")
+            continue
+        middle_start = 1  # Start index for middle 3 frames
+        middle_frames = generated_frames[middle_start : middle_start + 3]
+        for idx, frame_idx in enumerate(range(i - 1, i + 2)):
+            all_frames[frame_idx] = middle_frames[idx]
+
+    # Create the video clip
+    def make_frame(t):
+        idx = min(int(t * graph.vs[0]["fps"]), len(all_frames) - 1)
+        return all_frames[idx]
+
+    video_clip = VideoClip(make_frame, duration=duration)
+    if audio_path is not None:
+        audio_clip = AudioFileClip(audio_path)
+        video_clip = video_clip.set_audio(audio_clip)
+    video_clip.write_videofile(save_path, codec="libx264", fps=graph.vs[0]["fps"], audio_codec="aac")
+
+    if return_motion:
+        all_motion = [node["axis_angle"] for node in path]
+        all_motion = np.stack(all_motion, 0)
+        return all_motion
+
+
+def graph_pruning(graph):
+    ascc = graph.clusters(mode="STRONG")
+    lascc = ascc.giant()
+    print(f"before nodes: {len(graph.vs)}, edges: {len(graph.es)}")
+    print(f"after nodes: {len(lascc.vs)}, edges: {len(lascc.es)}")
+    in_degrees = lascc.indegree()
+    out_degrees = lascc.outdegree()
+    avg_in_degree = sum(in_degrees) / len(in_degrees)
+    avg_out_degree = sum(out_degrees) / len(out_degrees)
+    print(f"Average In-degree: {avg_in_degree}")
+    print(f"Average Out-degree: {avg_out_degree}")
+    print(f"max in degree: {max(in_degrees)}, max out degree: {max(out_degrees)}")
+    print(f"min in degree: {min(in_degrees)}, min out degree: {min(out_degrees)}")
+    return lascc
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--json_save_path", type=str, default="")
+    parser.add_argument("--graph_save_path", type=str, default="")
+    args = parser.parse_args()
+    json_path = args.json_save_path
+    print("json_path", json_path)
+    graph_path = args.graph_save_path
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    smplx_model = (
+        smplx.create(
+            os.path.join(SCRIPT_PATH, "emage/smplx_models/"),
+            model_type="smplx",
+            gender="NEUTRAL_2020",
+            use_face_contour=False,
+            num_betas=300,
+            num_expression_coeffs=100,
+            ext="npz",
+            use_pca=False,
+        )
+        .to(device)
+        .eval()
+    )
+
+    # single_test
+    # graph = create_graph('/content/drive/MyDrive/003_Codes/TANGO/datasets/data_json/show_oliver_test/Abortion_Laws_-_Last_Week_Tonight_with_John_Oliver_HBO-DRauXXz6t0Y.webm.json')
+    graph = create_graph(json_path, smplx_model)
+    graph = create_edges(graph)
+    # pool_path = "/content/drive/MyDrive/003_Codes/TANGO-JointEmbedding/datasets/oliver_test/show-oliver-test.pkl"
+    # graph = igraph.Graph.Read_Pickle(fname=pool_path)
+    # graph = igraph.Graph.Read_Pickle(fname="/content/drive/MyDrive/003_Codes/TANGO-JointEmbedding/datasets/oliver_test/test.pkl")
+
+    walk, is_continue = random_walk(graph, 100)
+    motion = path_visualization(graph, walk, is_continue, "./test.mp4", audio_path=None, verbose_continue=True, return_motion=True)
+    # print(motion.shape)
+    save_graph = graph.write_pickle(fname=graph_path)
+    graph = graph_pruning(graph)
+
+    # show-oliver
+    # json_path = "/content/drive/MyDrive/003_Codes/TANGO/datasets/data_json/show_oliver_test/"
+    # pre_node_path = "/content/drive/MyDrive/003_Codes/TANGO/datasets/cached_graph/show_oliver_test/"
+    # for json_file in tqdm(os.listdir(json_path)):
+    #     graph = create_graph(os.path.join(json_path, json_file))
+    #     graph = create_edges(graph)
+    #     if not len(graph.vs) >= 1500:
+    #         print(f"skip: {len(graph.vs)}", json_file)
+    #     graph.write_pickle(fname=os.path.join(pre_node_path, json_file.split(".")[0] + ".pkl"))
+    #     print(f"Graph saved at {json_file.split('.')[0]}.pkl")

datasets/beat2_v5.py

@@ -0,0 +1,80 @@
+import json
+import torch
+from torch.utils import data
+import numpy as np
+import librosa
+import textgrid as tg
+import os
+import math
+
+
+class BEAT2Dataset(data.Dataset):
+    def __init__(self, cfg, split):
+        data_meta_paths = cfg.data.meta_paths
+        vid_meta = []
+        for data_meta_path in data_meta_paths:
+            vid_meta.extend(json.load(open(data_meta_path, "r")))
+        self.vid_meta = [item for item in vid_meta if item.get("mode") == split]
+        self.mean = 0  # np.load(cfg.data.mean_path) if cfg.data.mean_path is not None else 0
+        self.std = 1  # np.load(cfg.data.std_path) if cfg.data.std_path is not None else 1
+        self.joint_mask = None  # cfg.data.joint_mask if cfg.data.joint_mask is not None else None
+        self.data_list = self.vid_meta
+        self.fps = cfg.data.pose_fps
+        self.audio_sr = cfg.data.audio_sr
+        self.use_text = False  # cfg.data.use_text
+
+    def __len__(self):
+        return len(self.data_list)
+
+    @staticmethod
+    def normalize(motion, mean, std):
+        return (motion - mean) / (std + 1e-7)
+
+    @staticmethod
+    def inverse_normalize(motion, mean, std):
+        return motion * std + mean
+
+    @staticmethod
+    def select_joints(motion, joint_mask):
+        return motion[:, joint_mask]
+
+    @staticmethod
+    def unselect_joints(motion, joint_mask):
+        # for visualization
+        full_motion = np.zeros((motion.shape[0], joint_mask.shape[0]))
+        full_motion[:, joint_mask] = motion
+
+    def __getitem__(self, item):
+        data = self.data_list[item]
+        motion = np.load(os.path.join(data["video_path"], data["video_id"] + ".npy"))
+        sdx = data["start_idx"]
+        edx = data["end_idx"]
+
+        SMPLX_FPS = 30
+        motion = motion[sdx:edx]
+        audio = np.load(os.path.join(data["audio_path"], data["video_id"] + "_text.npz"), allow_pickle=True)
+        sdx_audio = math.floor(sdx * (1 / SMPLX_FPS * 50))
+        edx_audio = sdx_audio + int((edx - sdx) * 50 / SMPLX_FPS) + 1
+        cached_audio_low = audio["wav2vec2_low"][sdx_audio:edx_audio]
+        cached_audio_high = audio["wav2vec2_high"][sdx_audio:edx_audio]
+        bert_time_aligned = audio["bert_time_aligned"][sdx_audio:edx_audio]
+
+        motion_tensor = torch.from_numpy(motion).float()  # T x D
+        cached_audio_low = torch.from_numpy(cached_audio_low).float()
+        cached_audio_high = torch.from_numpy(cached_audio_high).float()
+        bert_time_aligned = torch.from_numpy(bert_time_aligned).float()
+
+        audio_wave, sr = librosa.load(os.path.join(data["audio_path"], data["video_id"] + ".wav"))
+        audio_wave = librosa.resample(audio_wave, orig_sr=sr, target_sr=self.audio_sr)
+        sdx_audio = sdx * int(1 / SMPLX_FPS * self.audio_sr)
+        edx_audio = edx * int(1 / SMPLX_FPS * self.audio_sr)
+        audio_wave = audio_wave[sdx_audio:edx_audio]
+        audio_tensor = torch.from_numpy(audio_wave).float()
+
+        return dict(
+            cached_rep15d=motion_tensor,
+            cached_audio_low=cached_audio_low,
+            cached_audio_high=cached_audio_high,
+            bert_time_aligned=bert_time_aligned,
+            audio_tensor=audio_tensor,
+        )

datasets/cached_audio/101099-00_18_09-00_18_19.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:addd2c332242bf4e234adee59d8220f85a3ba4e587e145ed8aece0c9f4b8c358
+size 1393776

datasets/cached_audio/1wrQ6Msp7wM_00-00-39.69_00-00-45.68.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0eb3ec8a6ded1a3e378b6b8745695beff96cdc4976570c1d070d688ab1dbeba
+size 2569514

datasets/cached_audio/demo1.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e01aee2e94689d95514749887d5c5ab77455b7e1adf4c8bded9f72e9c69b2db0
+size 1142106

datasets/cached_audio/demo3.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5914ea9b0001ad9f1ee7e9595f73a36d75ea17aae36294f32bee70ca3439a956
+size 1378144

datasets/cached_audio/demo4.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2929191b3089a22503538a29d38d5444b2d245715e0767e7af29a2341cfe9a8f
+size 1054816

datasets/cached_audio/demo5.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87d40d0740a740934e74a376e3125cd0ee01332c563faf91613787f99ab9110a
+size 1348398

datasets/cached_audio/demo7.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06deda4ba3eed683ff774d32a5999a2c624f956a7f95d77d2d0c3bd943f069c8
+size 1120862

datasets/cached_audio/demo9.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b172832933ffb13fffaa8dd649e5aa7130aad1c5e25d53e331f87ed1a4815b63
+size 1284539

datasets/cached_audio/female_test_V1.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b89e47ea2febb05fef6a321822c9d14cd4fd752f7fcb2e27f28abec3104f5823
+size 2727168

datasets/cached_audio/speaker12_10_BVHw8aCPATM_00-01-05.0_00-01-10.0.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:926444bb200639713b1d0d48ea1ff544685c1dc24b9f1d42e8133724563e18bd
+size 1577443

datasets/cached_audio/speaker7_iuYlGRnC7J8_00-00-0.00_00-00-3.25.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffb58134d03dd7dabe2bfc587ea615c540cf0c161b20c754f95b74de07379bb9
+size 1679489