Adding CMake, FFmpeg 5+ and VS2019/2022 support

[wo80]

I've been working on Visual Studio compatibility lately (well, picked up work from a year ago), see https://github.com/wo80/essentia/

Here's what's in the fork:

• CMake build script
• RougeVector compatibility with Visual Studio 2019/2022
• Compatibility with FFmpeg 4/5/6/7

There's a GitHub workflow and a short guide available showing how to compile the project.

If you are interested in merging this, I think the changes in the cmake branch should be split up into smaller pull requests. I'd suggest the following changes to go into separate PRs and that detailed discussion being done there:

1. Remove using namespace std from headers, see CppCoreGuidelines#Rs-using-directive (wo80@b510967)
2. Update C++ code (wo80@38bd16b, wo80@4c27b35, wo80@57c2e6f, wo80@877df4d)
3. Update googletest
4. Update AudioContext (wo80@7733e28, wo80@087af17, wo80@57918d0, wo80@246eff3)
5. Update RougeVector (wo80@5d761dd, wo80@07c48f2)
6. Remove outdated Visual Studio project files (and win32 3rd party code)
7. Add CMake support
8. Add Github workflow to test Visual Studio build

Regarding point 6: having the Visual Studio project files doesn't make sense with the CMake script in place. The win32 3rd party code is completely out of date, so I'd suggest, if you want to keep the option to cross-compile for Windows, all dependencies should be pulled from the online repos (like done in my GitHub workflow) and the versions being updated on a regular basis. I might update the shell scripts, if I find the time.

CMake build tested on

• Windows 10, Visual Studio 2022
• Windows 10, msys2, MinGW64, GCC 13
• Debian 12, GCC 12
• Fedora 39, GCC 13
• Manjaro, GCC 13 and clang

WAF build ./waf --with-examples --with-cpptests --std c++17 tested on

• Debian 12, GCC 12
• Fedora 39, GCC 13

C++ unit tests are passing on all platforms.

Python wheel

I have tested the Python wheel only on Debian. The essentia_python_tutorial.ipynb notebook runs without problems. I've also been able to build the wheel for Windows (see GitHub workflow of the cmake-python branch), but haven't done much testing. The following code runs fine in the Python prompt:

import os
os.add_dll_directory("C:/usr/bin")
import essentia
import essentia.standard
print(dir(essentia.standard))
help(essentia.standard.MFCC)

I'm compiling against shared library dependencies (FFmpeg etc.), so the location where those are residing has to be added with os.add_dll_directory on Windows.

---

I've tried to find the affected issues and pull requests. No guarantee for completeness, though.

• Build on Windows #204
• Python bindings in Windows #243
• Compiling using native windows Visual Studio compiler #330
• Review outdate files in packaging/win32 #338
• Win32 static build #730
• Compiling windows MSVC #853
• Include Essentia in Visual Studio 2017 #854
• Getting Essentia to work on Windows #1130
• Error when trying to cross-compile for Windows #1134
• pip install essentia fails on windows 10 #1157
• std::binary_function removed on c++17 #1206
• Starting migrate to CMake & Cpp 20 stardand #1207
• Support ffmpeg version 5 #1248
• Static builds with GAIA #1265
• Support for x86_64 with mingw #1311
• Windows compilation fails with RogueVector #1360
• Please provide updated binaries for windows #1340
• MSVC Vamp plugin builds for Windows #1402
• Build fails with "error: ‘av_register_all’ was not declared in this scope" #1411
• Compilation instructions are broken #1439
• Changes for Visual Studio compilation #396
• Python win changes to build dll #820
• add small github actions script for a windows build #1113
• Fixed C++17 compatibility issues #1370

[Lovecraft-github]

could you please explain to me how to crosscompile a windows static binary of essentia with gaia?
I tried to follow the instructions in this repo but only manage to produce the binaries without gaia. the 3 party doesn't seem to offer an option for gaia like the linux one.

[wo80]

@Lovecraft-github To me, Gaia and the related machine learning models look abandoned. That's why I haven't done any tests besides checking that the Gaia cmake project at https://github.com/wo80/gaia/tree/cmake compiles and links ok with the Essentia cmake project. If you haven't tried yet, I suggest you use the code from that branch. Please also read the guide - after compiling and installing Gaia, you will have to configure Essentia with

cmake -B build -DUSE_GAIA2=ON

I haven't tried to cross-compile the cmake project. The whole point of this, was to get it working with Visual Studio. So if you are on Linux trying to compile for Windows, I'm probably not much of a help.

[dbogdanov]

Firstly, thank you very much @wo80 for the PR!!! We'll be looking at it soon.

To clarify the situation, we are planning to get rid of dependency on Gaia. It is there for the SVM classification models support (GaiaTransform algorithm), however those are outdated by now. Our current TensorFlow models have a significantly better performance in terms of accuracy and generalization, so we need to remove Gaia and add support for TensorFlow on Windows instead.

[Lovecraft-github]

@wo80 I managed to build it for linux, but I didn't find a way to do it for windows. It was a long shot to ask you. I'm with you that seems abandoned. Thanks anyway.

@dbogdanov So we could say that gaia2 is officially gone. Well, that at least solves the problem of continuing trying :-P. It is planned to release static binaries or installer for windows in a near future? or make it possible to cross-compiling it?. And last, any estimation when it will happen?

[wo80]

@Lovecraft-github I've updated some of the CMake modules which should also fix problems finding the Gaia package. The latest code is in the msvc-shared-lib branch - in case you still want to use the old models.

@dbogdanov I've also fixed issues with the Tensorflow support. I have linked with recent Tensorflow v2.15.0 and everything builds fine. I've successfully run the streaming_musicnn_predict example with the feature-extractors/musicnn/msd-musicnn-1.pb model. I tried streaming_vggish_predict with classification-heads/genre_dortmund/genre_dortmund-audioset-vggish-1.pb, but I'm getting

[   INFO   ] TensorflowPredict: Successfully loaded graph file: D:\projects\tensorflow\models\genre_dortmund-audioset-vggish-1.pb
TensorflowPredict: 'model/Sigmoid' is not a valid node name of this graph.
TensorflowPredict: Available node names are:
model/Placeholder, dense/kernel, dense/kernel/read, dense/bias, dense/bias/read, model/dense/MatMul,
model/dense/BiasAdd, model/dense/Relu, dense_1/kernel, dense_1/kernel/read, dense_1/bias, dense_1/bias/read,
model/dense_1/MatMul, model/dense_1/BiasAdd, model/Softmax.

I have to say, it's a bit unclear which models are supposed to work with which examples...

[gotev]

Hi everyone, first of all thank you for the joint efforts and time spent creating and maintaining this wonder!

I started using the lib successfully on macOS and Linux with python bindings, and now I've been curious to make it work also on Windows. It's not working out of the box with classic pip install essentia as pointed out by many users, so by researching further, I've come across this issue and followed the guide, but still no success.

Not using Windows for 12 years straight, so bear with me if I'm missing something obvious 😄

Going to describe what I've done in reproducible detail. Every shell command has been issued in a PowerShell run as administrator.

A. Build environment setup from scratch

1. Create a new VM with: Win10_22H2_English_x64 ISO
2. Run a PowerShell as Administrator. Install chocolatey package manager, python, git and vscode:

   Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned
   # Install chocolatey package manager copying command from: https://chocolatey.org/install
   # and then:
   choco install -y python git vscode vscode-python

3. Install CMake and Build Tools

   choco install -y cmake --installargs 'ADD_CMAKE_TO_PATH=System'
   choco install -y visualstudio2022buildtools

4. Open Visual Studio Installer
   • Select Workloads
   • Select Desktop Development with C++
   • Click on Modify
   • Wait for the installation to be completed
5. Restart
6. Make a VM snapshot (just in case)

B. Get essentia sources and build dependencies

Navigate to your favorite working directory, then Run Developer PowerShell for VS 2022 as Administrator and:

# clone essentia
git clone --recurse-submodules https://github.com/wo80/essentia.git -b cmake
cd essentia
# install vcpkg locally, to resolve C++ lib dependencies
git clone https://github.com/microsoft/vcpkg
.\vcpkg\bootstrap-vcpkg.bat
# get dependencies with vcpkg
.\vcpkg\vcpkg install eigen3
.\vcpkg\vcpkg install fftw3
.\vcpkg\vcpkg install libyaml
# this one is veery long, as it also compiles ffmpeg (may take up to 40 mins)
.\vcpkg\vcpkg install chromaprint
.\vcpkg\vcpkg install libsamplerate
.\vcpkg\vcpkg install taglib
pip install numpy

C. Build essentia with python bindings

From the same Developer PowerShell for VS 2022 run as Administrator of the previous paragraph (or a new one in the essentia sources folder):

# configure essentia build
cmake -B build -D BUILD_PYTHON_BINDINGS=ON -S . "-DCMAKE_TOOLCHAIN_FILE=vcpkg/scripts/buildsystems/vcpkg.cmake"
# build and test essentia
cmake --build build --config Release
ctest --test-dir build --output-on-failure -C Release

D. Outcome

The build is successful and the tests are passing. I see also the compiled wheel inside the build folder: essentia-2.1b6.dev0-cp312-cp312-win32.whl but it's for Win32, even if I've done everything on Win64, so I can't install it in my virtual env with the usual:

.\venv\Scripts\activate
pip install ..\essentia\build\wheel\essentia-2.1b6.dev0-cp312-cp312-win32.whl

I'm getting:

ERROR: essentia-2.1b6.dev0-cp312-cp312-win32.whl is not a supported wheel on this platform.

the output from cmake -B build -D BUILD_PYTHON_BINDINGS=ON -S . "-DCMAKE_TOOLCHAIN_FILE=vcpkg/scripts/buildsystems/vcpkg.cmake" is the following:

-- Selecting Windows SDK version 10.0.22621.0 to target Windows 10.0.19045.
-- Found Python3: C:/Python312/libs/python312.lib (found version "3.12.2") found components: Development NumPy Interpreter Development.Module Development.Embed
-- Found Python3: C:/Python312/python.exe (found version "3.12.2") found components: Interpreter Development Development.Module Development.Embed
-- Found Python3: C:/Python312/python.exe (found version "3.12.2") found components: Interpreter
-- Configuring done (1.4s)
-- Generating done (1.9s)
-- Build files have been written to: C:/Users/Admin/Desktop/essentia/build

Where am I doing it wrong? How to make it compile for Win64 instead? Any hint is appreciated.

[wo80]

@gotev You could try specifying the target platform explictily, like cmake -B build -A x64 ...

And can you post the complete output of the CMake configure step?

[gotev]

By keeping everything else as my previous comment, except for the commands run at paragraph C, substituted by:

if (Test-Path .\build) { Remove-Item -Path .\build -Recurse -Force }
cmake -G "Visual Studio 17 2022" -A x64 -B build -D BUILD_PYTHON_BINDINGS=ON -S . "-DCMAKE_TOOLCHAIN_FILE=vcpkg/scripts/buildsystems/vcpkg.cmake"
# build and test essentia
cmake --build build --config Release
ctest --test-dir build --output-on-failure -C Release

I'm still getting essentia-2.1b6.dev0-cp312-cp312-win32.whl for win32 and not x64. By trying to install the wheel no matter the name of the file, I'm still getting the above mentioned error.

During the final steps of the build, I intercepted this from the console output:

Creating library C:/Users/Admin/Desktop/essentia/build/src/python/Release/_essentia.cp312-win_amd64.lib and object
   C:/Users/Admin/Desktop/essentia/build/src/python/Release/_essentia.cp312-win_amd64.exp
  _essentia.vcxproj -> C:\Users\Admin\Desktop\essentia\build\src\python\Release\_essentia.cp312-win_amd64.pyd
  Building Custom Rule C:/Users/Admin/Desktop/essentia/src/python/CMakeLists.txt
  Processing c:\users\admin\desktop\essentia\build\src\python
    Installing build dependencies: started
    Installing build dependencies: finished with status 'done'
    Getting requirements to build wheel: started
    Getting requirements to build wheel: finished with status 'done'
    Installing backend dependencies: started
    Installing backend dependencies: finished with status 'done'
    Preparing metadata (pyproject.toml): started
    Preparing metadata (pyproject.toml): finished with status 'done'
  Building wheels for collected packages: essentia
    Building wheel for essentia (pyproject.toml): started
    Building wheel for essentia (pyproject.toml): finished with status 'done'
    Created wheel for essentia: filename=essentia-2.1b6.dev0-cp312-cp312-win32.whl size=1978664 sha256=430f279d9b4d2be2
  ebe9c974d0525a7e46d2a6130a5fa0875236ac25143ba19e
    Stored in directory: C:\Users\Admin\AppData\Local\Temp\pip-ephem-wheel-cache-odjj24xg\wheels\2b\11\7b\0b358554c85e0
  36a6917648cc3dceefc2379e67ac2793b5169
  Successfully built essentia

Full output from:

cmake -G "Visual Studio 17 2022" -A x64 -B build -D BUILD_PYTHON_BINDINGS=ON -S . "-DCMAKE_TOOLCHAIN_FILE=vcpkg/scripts/buildsystems/vcpkg.cmake"

is the following:

-- Selecting Windows SDK version 10.0.22621.0 to target Windows 10.0.19045.
-- The C compiler identification is MSVC 19.39.33521.0
-- The CXX compiler identification is MSVC 19.39.33521.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: C:/Program Files (x86)/Microsoft Visual Studio/2022/BuildTools/VC/Tools/MSVC/14.39.33519/bin/Hostx64/x64/cl.exe - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: C:/Program Files (x86)/Microsoft Visual Studio/2022/BuildTools/VC/Tools/MSVC/14.39.33519/bin/Hostx64/x64/cl.exe - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found FFmpeg: C:/Users/Admin/Desktop/essentia/vcpkg/installed/x64-windows/debug/lib/avcodec.lib;C:/Users/Admin/Desktop/essentia/vcpkg/installed/x64-windows/debug/lib/avformat.lib;C:/Users/Admin/Desktop/essentia/vcpkg/installed/x64-windows/debug/lib/avutil.lib;C:/Users/Admin/Desktop/essentia/vcpkg/installed/x64-windows/debug/lib/swresample.lib
-- Found SampleRate: C:/Users/Admin/Desktop/essentia/vcpkg/installed/x64-windows/debug/lib/samplerate.lib (found version "0.2.2")
-- Found Taglib: C:/Users/Admin/Desktop/essentia/vcpkg/installed/x64-windows/debug/lib/tag.lib (found version "2.0")
-- Found Chromaprint: C:/Users/Admin/Desktop/essentia/vcpkg/installed/x64-windows/debug/lib/chromaprint.lib (found version "1.5.1")
-- Found YAML: C:/Users/Admin/Desktop/essentia/vcpkg/installed/x64-windows/debug/lib/yaml.lib (found version "0.1")
-- Found Python3: C:/Python312/libs/python312.lib (found version "3.12.2") found components: Development NumPy Interpreter Development.Module Development.Embed
-- Found FFTW3f: C:/Users/Admin/Desktop/essentia/vcpkg/installed/x64-windows/debug/lib/fftw3f.lib (found version "3.3.9")
-- Found Python3: C:/Python312/python.exe (found version "3.12.2") found components: Interpreter Development Development.Module Development.Embed
-- Found Python3: C:/Python312/python.exe (found version "3.12.2") found components: Interpreter
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Failed
-- Looking for pthread_create in pthreads
-- Looking for pthread_create in pthreads - not found
-- Looking for pthread_create in pthread
-- Looking for pthread_create in pthread - not found
-- Found Threads: TRUE
-- Configuring done (31.0s)
-- Generating done (0.5s)
-- Build files have been written to: C:/Users/Admin/Desktop/essentia/build

[gotev]

Got it one step further. Executed steps A and B from my first comment and then:

if (Test-Path .\build) { Remove-Item -Path .\build -Recurse -Force }
cmake -G "Visual Studio 17 2022" -A x64 -B build -D BUILD_PYTHON_BINDINGS=ON -S . "-DCMAKE_TOOLCHAIN_FILE=vcpkg/scripts/buildsystems/vcpkg.cmake"

Then I manually edited:

code .\build\src\python\setup.py

by adding this line:

options={'bdist_wheel': {'plat_name': 'win_amd64'}},

after line 86, so the setup in the file looks like this:

setup(
    ..
    cmdclass = {
        'bdist': bdist,
        'sdist': sdist,
        'install_lib': install_lib,
        'build_ext': cmake_build_ext,
    },
    options={'bdist_wheel': {'plat_name': 'win_amd64'}},
)

Then, by running:

cmake --build build --config Release
ctest --test-dir build --output-on-failure -C Release

the lib compiles, tests passes and the 64 bit wheel gets generated: build\wheel\essentia-2.1b6.dev0-cp312-cp312-win_amd64.whl

I'm able to install it successfully in the virtualenv:

.\venv\Scripts\activate
pip install ..\essentia\build\wheel\essentia-2.1b6.dev0-cp312-cp312-win_amd64.whl

Processing c:\users\admin\desktop\essentia\build\wheel\essentia-2.1b6.dev0-cp312-cp312-win_amd64.whl
Collecting numpy>=1.8.2 (from essentia==2.1b6.dev0)
  Using cached numpy-1.26.4-cp312-cp312-win_amd64.whl.metadata (61 kB)
Collecting pyyaml (from essentia==2.1b6.dev0)
  Using cached PyYAML-6.0.1-cp312-cp312-win_amd64.whl.metadata (2.1 kB)
Collecting six (from essentia==2.1b6.dev0)
  Using cached six-1.16.0-py2.py3-none-any.whl.metadata (1.8 kB)
Using cached numpy-1.26.4-cp312-cp312-win_amd64.whl (15.5 MB)
Using cached PyYAML-6.0.1-cp312-cp312-win_amd64.whl (138 kB)
Using cached six-1.16.0-py2.py3-none-any.whl (11 kB)
Installing collected packages: six, pyyaml, numpy, essentia
Successfully installed essentia-2.1b6.dev0 numpy-1.26.4 pyyaml-6.0.1 six-1.16.0

But then when I try to import it:

python

Python 3.12.2 (tags/v3.12.2:6abddd9, Feb  6 2024, 21:26:36) [MSC v.1937 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.
>>> import essentia
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Users\Admin\Desktop\myscript\venv\Lib\site-packages\essentia\__init__.py", line 1, in <module>
    from . import _essentia
ImportError: DLL load failed while importing _essentia: The specified module could not be found.

It fails with this error 🤔

[wo80]

That manual editing of setup.py should not be necessary. When I compile, I get essentia-2.1b6.dev0-cp312-cp312-win_amd64.whl. Ultimately, python/pip is responsible for the final name of the wheel, and the command is setup here:
https://github.com/wo80/essentia/cmake/modules/python-wheel.cmake

The error you are getting is probably due to vcpkg building shared libs, and python does not know where to find them. You need to add that path manually (the path used below is just a guess from the output you provided):

import os
os.add_dll_directory("C:/Users/Admin/Desktop/essentia/vcpkg/installed/x64-windows/debug/bin")
import essentia
...

[gotev]

@wo80 great! Yeah the path is correct. I can now import the library:

python

Python 3.12.2 (tags/v3.12.2:6abddd9, Feb  6 2024, 21:26:36) [MSC v.1937 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.
>>> import os
>>> os.add_dll_directory("C:/Users/Admin/Desktop/essentia/vcpkg/installed/x64-windows/debug/bin")
<AddedDllDirectory('C:/Users/Admin/Desktop/essentia/vcpkg/installed/x64-windows/debug/bin')>
>>> import essentia
>>>

but I wonder if there's a way to install those libs in standard dirs wth vcpkg (inside the virtualenv), so I don't have to write this platform specific python code for Windows.

This line caught my attention:

Python 3.12.2 (tags/v3.12.2:6abddd9, Feb  6 2024, 21:26:36) [MSC v.1937 64 bit (AMD64)] on win32

that on win32 at the end may be causing win32 wheel build by default. I agree setup.py shouldn't need manual patching. Installed python through chocolatey. How did you install python and what does it display to you when you enter the REPL?

[wo80]

How did you install python and what does it display to you when you enter the REPL?

I used the installer provided on python.org, and I'm quite confident, that chocolatey does the same.

Python 3.12.1 (tags/v3.12.1:2305ca5, Dec  7 2023, 22:03:25) [MSC v.1937 64 bit (AMD64)] on win32

but I wonder if there's a way to install those libs in standard dirs with vcpkg, so I don't have to write this platform specific python code for Windows.

First, take a look at https://docs.python.org/3/whatsnew/3.8.html#bpo-36085-whatsnew . That means, shared dependencies have to be installed in a system path or the wheel directory. I'm not a python guy, but my guess is, even in case this is done, some manual loading might be necessary in __init__.py on the Windows platform. I'll test copying the dependencies to the wheel directory tomorrow (it's getting late here 🥱).

I have made improvements linking with static dependencies in wo80@61d8947 , but there are still problems - for example linking the FFmpeg libraries statically, you also need to add Windows' bcrypt.lib, which the CMake build script doesn't do right now.

EDIT: you can try building static libraries with vcpkg using vcpkg install taglib:x64-windows-static etc. and then configure CMake with the -DENABLE_STATIC_DEPENDENCIES=ON option. I'm not sure how the FFmpeg dependency is handled by vcpgk, though.

[gotev]

Retried again all the steps, by also recreating the VM from scratch and explicitly specifying x64 architecture when installing python from chocolatey (it's using the installer from python.org) and also dependencies with vcpkg. To sum it up, those are all the updated steps from scratch:

A. Build environment setup from scratch

1. Create a new VM with: Win10_22H2_English_x64 ISO
2. Run a PowerShell as Administrator. Install chocolatey package manager, python, git and vscode:

   Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned
   # Install chocolatey package manager copying command from: https://chocolatey.org/install
   # and then:
   choco install -y python --x64
   choco install -y git vscode vscode-python

3. Install CMake and Build Tools

   choco install -y cmake --installargs 'ADD_CMAKE_TO_PATH=System'
   choco install -y visualstudio2022buildtools

4. Open Visual Studio Installer
   • Select Workloads
   • Select Desktop Development with C++
   • Click on Modify
   • Wait for the installation to be completed
5. Manually check and install windows updates (from Start > Settings > Updates)
6. Restart
7. Make a VM snapshot (just in case)

B. Get essentia sources and build dependencies

Navigate to your favorite working directory, then Run Developer PowerShell for VS 2022 as Administrator and:

# clone essentia
git clone --recurse-submodules https://github.com/wo80/essentia.git -b cmake
cd essentia
# install vcpkg locally, to resolve C++ lib dependencies
git clone https://github.com/microsoft/vcpkg
.\vcpkg\bootstrap-vcpkg.bat
# get dependencies with vcpkg
.\vcpkg\vcpkg install eigen3:x64-windows
.\vcpkg\vcpkg install fftw3:x64-windows
.\vcpkg\vcpkg install libyaml:x64-windows
.\vcpkg\vcpkg install libsamplerate:x64-windows
.\vcpkg\vcpkg install taglib:x64-windows
# this one is veery long, as it also compiles ffmpeg (may take up to 40 mins)
.\vcpkg\vcpkg install chromaprint:x64-windows
pip install numpy

C. Build essentia with python bindings

From the same Developer PowerShell for VS 2022 run as Administrator of the previous paragraph (or a new one in the essentia sources folder):

if (Test-Path .\build) { Remove-Item -Path .\build -Recurse -Force }
cmake -G "Visual Studio 17 2022" -A x64 -B build -D BUILD_PYTHON_BINDINGS=ON -S . "-DCMAKE_TOOLCHAIN_FILE=vcpkg/scripts/buildsystems/vcpkg.cmake"
# build and test essentia
cmake --build build --config Release
ctest --test-dir build --output-on-failure -C Release

After the build completes:

cmake --build build --config Release

I'm still getting only essentia-2.1b6.dev0-cp312-cp312-win32.whl wheel inside the build\wheel folder.

To workaround this (haven't found what's causing the issue yet, but replicated it the same from scratch 2 times) and prepare a 64 bit wheel with the needed DLLs bundled in, without manually editing any file, I found this way:

pip install wheel setuptools delvewheel

# from inside the essentia sources root directory, after the build is completed
cd .\build\src\python
# manually build the 64 bit wheel
python setup.py bdist_wheel --plat-name win_amd64
# get to the wheel output directory
cd ..\..\wheel
# find the exact name of the 64 bit wheel
$wheel_name = (Get-ChildItem -Path . -Filter "*amd64.whl" | Select-Object -First 1).Name
# show missing DLLs
delvewheel show $wheel_name --add-path ..\..\vcpkg\installed\x64-windows\bin
# integrate missing DLLs in the wheel and repackage it
delvewheel repair $wheel_name --add-path ..\..\vcpkg\installed\x64-windows\bin -w .

Then, to test it out:

python -m venv venv
.\venv\Scripts\activate
pip install .\essentia-2.1b6.dev0-cp312-cp312-win_amd64.whl

install output

Processing c:\users\admin\desktop\essentia\build\wheel\essentia-2.1b6.dev0-cp312-cp312-win_amd64.whl
Collecting numpy>=1.8.2 (from essentia==2.1b6.dev0)
  Using cached numpy-1.26.4-cp312-cp312-win_amd64.whl.metadata (61 kB)
Collecting pyyaml (from essentia==2.1b6.dev0)
  Using cached PyYAML-6.0.1-cp312-cp312-win_amd64.whl.metadata (2.1 kB)
Collecting six (from essentia==2.1b6.dev0)
  Using cached six-1.16.0-py2.py3-none-any.whl.metadata (1.8 kB)
Using cached numpy-1.26.4-cp312-cp312-win_amd64.whl (15.5 MB)
Using cached PyYAML-6.0.1-cp312-cp312-win_amd64.whl (138 kB)
Using cached six-1.16.0-py2.py3-none-any.whl (11 kB)
Installing collected packages: six, pyyaml, numpy, essentia
Successfully installed essentia-2.1b6.dev0 numpy-1.26.4 pyyaml-6.0.1 six-1.16.0

python

Python 3.12.2 (tags/v3.12.2:6abddd9, Feb  6 2024, 21:26:36) [MSC v.1937 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.
>>> import essentia
>>>

and boom, it's now imported without any additional statements! Now it's very late, so I'll leave further refinements and tests for another time! Further tunings will certainly be needed to avoid workarounds, enable all the features and be able to use the full fledged library as on macOS and Linux.

[gotev]

@wo80 @dbogdanov managed to get a completely working 64 bit wheel for Windows (with DLLs bundled in the wheel, so no static libs needed) with TensorFlow support. There are some warnings here and there, but it's working 🎉

This is the working recipe from scratch to finish:

A. Build environment setup from scratch

1. Create a new VM with: Win10_22H2_English_x64 ISO
2. Run a PowerShell as Administrator. Install chocolatey package manager, python, git and vscode:

   Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned
   # Install chocolatey package manager copying command from: https://chocolatey.org/install
   # and then:
   choco install -y python --x64
   choco install -y git vscode vscode-python

3. Install CMake and Build Tools

   choco install -y cmake --installargs 'ADD_CMAKE_TO_PATH=System'
   choco install -y visualstudio2022buildtools

4. Open Visual Studio Installer
   • Select Workloads
   • Select Desktop Development with C++
   • Click on Modify
   • Wait for the installation to be completed
5. Manually check and install windows updates (from Start > Settings > Updates)
6. Restart
7. Make a VM snapshot (just in case)

B. Get essentia sources and build dependencies

Navigate to your favorite working directory, then Run Developer PowerShell for VS 2022 as Administrator and:

# clone essentia
git clone --recurse-submodules https://github.com/wo80/essentia.git -b cmake
cd essentia
# install vcpkg locally, to resolve C++ lib dependencies
git clone https://github.com/microsoft/vcpkg
.\vcpkg\bootstrap-vcpkg.bat
# get dependencies with vcpkg
.\vcpkg\vcpkg install eigen3:x64-windows
.\vcpkg\vcpkg install fftw3:x64-windows
.\vcpkg\vcpkg install libyaml:x64-windows
.\vcpkg\vcpkg install libsamplerate:x64-windows
.\vcpkg\vcpkg install taglib:x64-windows
# this one is veery long, as it also compiles ffmpeg (may take up to 40 mins)
.\vcpkg\vcpkg install chromaprint:x64-windows
pip install numpy

Get TensorFlow. Save this in get-tf.ps1 and execute it with .\get-tf.ps1 (credits to @wo80 guide here):

$tfRemoteUrl = "https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-cpu-windows-x86_64-2.15.0.zip"
$tfPatchHeadersUrl = "https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-cpu-linux-x86_64-2.15.0.tar.gz"
$tfTempPath = ".\tf-temp"
$zipFile = ".\tf.zip"
$zipPatchHeadersFile = ".\tf-patch-headers.tar.gz"

if (Test-Path $tfTempPath) {
  Remove-Item -Path $tfTempPath -Recurse -Force
}

mkdir $tfTempPath
Set-Location $tfTempPath

$ProgressPreference = 'SilentlyContinue'

Write-Host "Downloading $tfRemoteUrl ..."
Invoke-WebRequest -Uri $tfRemoteUrl -OutFile $zipFile

Write-Host "Downloading $tfPatchHeadersUrl ..."
Invoke-WebRequest -Uri $tfPatchHeadersUrl -OutFile $zipPatchHeadersFile

$ProgressPreference = 'Continue'

Write-Host "Extracting $tfRemoteUrl ..."
Expand-Archive -Path $zipFile -DestinationPath .

Write-Host "Extracting $zipPatchHeadersFile ..."
mkdir .\tf-patch-headers
tar -xvzf $zipPatchHeadersFile -C .\tf-patch-headers

Remove-Item $zipFile
Remove-Item $zipPatchHeadersFile

Write-Host "Copying files ..."
Copy-Item .\lib\tensorflow.dll ..\vcpkg\installed\x64-windows\bin -Force
Copy-Item .\lib\tensorflow.lib ..\vcpkg\installed\x64-windows\lib -Force
Copy-Item .\tf-patch-headers\include\* -Destination ..\vcpkg\installed\x64-windows\include -Recurse -Force
Copy-Item .\include\* -Destination ..\vcpkg\installed\x64-windows\include -Recurse -Force
Set-Location ..

C. Build essentia with TensorFlow and python bindings

From the same Developer PowerShell for VS 2022 run as Administrator of the previous paragraph (or a new one in the essentia sources folder):

if (Test-Path .\build) { Remove-Item -Path .\build -Recurse -Force }
cmake -G "Visual Studio 17 2022" -A x64 -B build -D BUILD_PYTHON_BINDINGS=ON -D USE_TENSORFLOW=ON -S . "-DCMAKE_TOOLCHAIN_FILE=vcpkg/scripts/buildsystems/vcpkg.cmake"
# build and test essentia
cmake --build build --config Release
ctest --test-dir build --output-on-failure -C Release

D. Package 64 bit wheel with all needed DLLs

From the same Developer PowerShell for VS 2022 run as Administrator of the previous paragraph (or a new one in the essentia sources folder):

pip install wheel setuptools delvewheel

# from inside the essentia sources root directory, after the build is completed
cd .\build\src\python
# manually build the 64 bit wheel
python setup.py bdist_wheel --plat-name win_amd64
# get to the wheel output directory
cd ..\..\wheel
# find the exact name of the 64 bit wheel
$wheel_name = (Get-ChildItem -Path . -Filter "*amd64.whl" | Select-Object -First 1).Name
# show missing DLLs
delvewheel show $wheel_name --add-path ..\..\vcpkg\installed\x64-windows\bin
# integrate missing DLLs in the wheel and repackage it
delvewheel repair $wheel_name --add-path ..\..\vcpkg\installed\x64-windows\bin -w .

The resulting wheel will be in essentia\build\wheel named like essentia-2.1b6.dev0-cp312-cp312-win_amd64.whl.

E. Install the wheel and import essentia

python -m venv venv
.\venv\Scripts\activate
pip install path\to\wheel.whl

python

Python 3.12.2 (tags/v3.12.2:6abddd9, Feb  6 2024, 21:26:36) [MSC v.1937 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.

>>> import essentia.standard as estd
[   INFO   ] MusicExtractorSVM: no classifier models were configured by default
2024-03-12 08:13:43.420581: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: SSE SSE2 SSE3 SSE4.1 SSE4.2 AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
>>>
>>> loader = estd.MonoLoader(filename="music.mp3")
>>> audio = loader()
>>> key_extractor = estd.KeyExtractor()
>>> key, scale, _ = key_extractor(audio)
>>> print(key)
A
>>> print(scale)
minor
>>> bpm_extractor = estd.RhythmExtractor2013(method="multifeature")
>>> bpm, _, _, _, _ = bpm_extractor(audio)
>>> print(bpm)
128.13800048828125

[wo80]

@gotev That's great! You're the first one reporting back, so thank you for your efforts! Can you test the following in PowerShell:

cd build/wheel
tar -xf essentia-2.1b6.dev0-cp312-cp312-win32.whl
Get-ChildItem -Path essentia -Recurse -Filter "*.pyd"
Select-String -Path essentia-2.1b6.dev0.dist-info/WHEEL -Pattern "Tag:"

If it prints

_essentia.cp312-win_amd64.pyd
essentia-2.1b6.dev0.dist-info\WHEEL:4:Tag: cp312-cp312-win_amd64

then the weird win32 package is actually a valid win_amd64 wheel, and the only thing you need to do to make pip install it, is to rename it (no need for python setup.py bdist_wheel --plat-name win_amd64 in step D).

EDIT:
Maybe also check the following in a Python shell:

import sysconfig
sysconfig.get_platform()

I still want to know where that win32 comes from :)