Torchaudio is not detecting FFmpeg

[ruliworst]

🐛 Describe the bug

Hi everyone,

I am trying to use torchaudio with FFmpeg but for some reason it is not being detected.

import torchaudio
print(torchaudio.list_audio_backends())

I added the path 'C:\Program Files\ffmpeg\bin' to the PATH environment variable. As you can see I can execute ffmpeg from cmd. However, when I execute torchaudio.list_audio_backends() only 'soundfile' is detected. I tested this in a Python PIP (I can't use conda) virtual environment with version 3.10 and in another one with version 3.12 but, it didn't work in both. Torchaudio and torch versions are 2.2.1. I don't know if it is a problem with version compatibility but in the official documentation I saw ffmpeg 6 is supported and I checked the code which tries to load the libraries.

I checked a lot of related issues but I couldn't find the right solution.

Any suggestions are welcome! Thanks.

Versions

Collecting environment information...

PyTorch version: 2.2.1+cpu
Is debug build: False
CUDA used to build PyTorch: Could not collect
ROCM used to build PyTorch: N/A
OS: Microsoft Windows Server 2022 Standard
GCC version: Could not collect
Clang version: Could not collect
CMake version: Could not collect
Libc version: N/A

Python version: 3.10.7 (tags/v3.10.7:6cc6b13, Sep  5 2022, 14:08:36) [MSC v.1933 64 bit (AMD64)] (64-bit runtime)
Python platform: Windows-10-10.0.20348-SP0
Is CUDA available: False
CUDA runtime version: 12.3.103
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: GPU 0: NVIDIA A100-PCIE-40GB
Nvidia driver version: 546.12
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
'wmic' is not recognized as an internal or external command,
operable program or batch file.

Versions of relevant libraries:

[pip3] numpy==1.26.4
[pip3] pytorch-lightning==2.2.4
[pip3] pytorch-metric-learning==2.5.0
[pip3] torch==2.2.1
[pip3] torch-audiomentations==0.11.1
[pip3] torch-pitch-shift==1.2.4
[pip3] torchaudio==2.2.1
[pip3] torchmetrics==1.4.0
[conda] Could not collect

[ruliworst]

Loading manullay the binaries is working but, I would like it was automatic.

ctypes.CDLL('C:/Program Files/ffmpeg/bin/avcodec-60.dll')
ctypes.CDLL('C:/Program Files/ffmpeg/bin/avdevice-60.dll')
ctypes.CDLL('C:/Program Files/ffmpeg/bin/avfilter-9.dll')
ctypes.CDLL('C:/Program Files/ffmpeg/bin/avformat-60.dll')
ctypes.CDLL('C:/Program Files/ffmpeg/bin/avutil-58.dll')

I still don't know why it is not working without preloading libraries.

[ruliworst]

If I locate the libraries in the same folder of libtorio_ffmpeg6.pyd , they are detected while importing. That is, in Lib/site-packages/torio/lib

EDIT: I have been checking the internal code and in file torchaudio\_extension\__init__.py line 24 there is a code which is not being executed:

if os.name == "nt" and (3, 8) <= sys.version_info < (3, 9):
    _init_dll_path()

This code adds the paths from PATH environment variable to the DLL search path. However, I would like to know why it is only executed in Python 3.8 version.

Thank you.

[chislon]

Ran into this same unexpected behavior as well.

If _init_dll_path() is allowed to run it works as expected

[HarikalarKutusu]

Same problem here, struggling for hours. Win 11, Python 3.12.4, using pip, ffmpeg 6.1.2 with shared libraries, path OK, can run on CMD, restarted multiple times, etc... Using torchaudio==2.4.0+cu124 thou...

ctypes.CDLL workaround also did not work in my case. Only "soundfile" can be used. Adding this to the code did not help either:

os.add_dll_directory("C:\\Programs\\ffmpeg-6.1\\bin\\")

PS: I'm trying to convert my code to use torch/torchaudio. Before torchaudio I've been using PyAV without any problems - using ffmpeg for a decade. torchaudio's ffmpeg documentation is also very cryptic. As far as I can see, torchaudio moved to ffmpeg as the main cross-platform library, but still too many problems.

[chislonchow]

Hi HarikalarKutusu, do you have all the required ffmpeg DLL files?

In my case I need all of these for ffmpeg to be detected, in addition to executing _init_dll_path from the file modified in torchaudio:

avcodec-60.dll avdevice-60.dll avfilter-9.dll avformat-60.dll avutil-58.dll postproc-57.dll swresample-4.dll swscale-7.dll

Using the dlls from this ffmpeg shared build: https://github.com/GyanD/codexffmpeg/releases/tag/6.1.1

[HarikalarKutusu]

I used [ffmpeg-builds/win64-gpl-shared-6.1](https://github.com/users/BtbN/packages/container/package/ffmpeg-builds%2Fwin64-gpl-shared-6.1), the naming of the .dll files are the same (v6.1.2 but IMO should not effect this).

I did not update the _init_dll_path directly in tourchaudio distribution, but used os.add_dll_directory() in my code before calling any torch code. This is what it does if ran:

def _init_dll_path():
    # On Windows Python-3.8+ has `os.add_dll_directory` call,
    # which is called to configure dll search path.
    # To find cuda related dlls we need to make sure the
    # conda environment/bin path is configured Please take a look:
    # https://stackoverflow.com/questions/59330863/cant-import-dll-module-in-python
    # Please note: if some path can't be added using add_dll_directory we simply ignore this path
    for path in os.environ.get("PATH", "").split(";"):
        if os.path.exists(path):
            try:
                os.add_dll_directory(path)
            except Exception:
                pass

Let me try the one from your link... EDIT: No joy

[HarikalarKutusu]

Joy! I was wrong. I imported the function and called it.

from torchaudio._extension.utils import _init_dll_path
...
def my_transcode_function():
    _init_dll_path()
   ...

Also changed the v3.8 limitation code to be safe.

# if os.name == "nt" and (3, 8) <= sys.version_info < (3, 9):
#     _init_dll_path()
if os.name == "nt" and (3, 8) <= sys.version_info < (3, 99):
    _init_dll_path()

It now works. Thank you for the guidance @chislonchow !