link system ffmpeg to torchvision. Don't install ffmpeg that comes with torchvision

[nikky4D]

🐛 Describe the bug

I want to install (via conda) torchvision without installing the ffmpeg that comes with it. I already have a hardware accelerated ffmpeg build on my computer and would like to keep using it. can you give me some idea on how I can do this with torchvision?

What I tried:
I tried to build from source but that requires pytorch nightly which clashes with other libraries that depend on pytorch 0.12 so I can't use pytorch nightly.

Versions

yTorch version: 1.12.1
Is debug build: False
CUDA used to build PyTorch: 11.3
ROCM used to build PyTorch: N/A

OS: Ubuntu 18.04.5 LTS (x86_64)
GCC version: (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
Clang version: Could not collect
CMake version: version 3.14.0
Libc version: glibc-2.27

Python version: 3.9.13 (main, Aug 25 2022, 23:26:10) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-48-generic-x86_64-with-glibc2.27
Is CUDA available: True
CUDA runtime version: 11.3.122
GPU models and configuration:
GPU 0: NVIDIA RTX A5000
GPU 1: NVIDIA RTX A5000
GPU 2: NVIDIA RTX A5000
GPU 3: NVIDIA RTX A5000

Nvidia driver version: 470.141.03
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.0.4
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.0.4
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.0.4
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.0.4
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.0.4
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.0.4
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.0.4
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

Versions of relevant libraries:
[pip3] numpy==1.23.3
[pip3] pytorch-lightning==1.6.0
[pip3] pytorchvideo==0.1.5
[pip3] torch==1.13.0.dev20220914+cu113
[pip3] torchaudio==0.12.1
[pip3] torchmetrics==0.7.3
[pip3] torchvision==0.13.1
[conda] blas 1.0 mkl
[conda] cudatoolkit 11.3.1 h2bc3f7f_2
[conda] mkl 2021.4.0 h06a4308_640
[conda] mkl-service 2.4.0 py39h7f8727e_0
[conda] mkl_fft 1.3.1 py39hd3c417c_0
[conda] mkl_random 1.2.2 py39h51133e4_0
[conda] numpy 1.23.3 pypi_0 pypi
[conda] numpy-base 1.23.1 py39ha15fc14_0
[conda] pytorch 1.12.1 py3.9_cuda11.3_cudnn8.3.2_0 pytorch
[conda] pytorch-lightning 1.6.0 pypi_0 pypi
[conda] pytorch-mutex 1.0 cuda pytorch
[conda] pytorchvideo 0.1.5 pypi_0 pypi
[conda] torch 1.13.0.dev20220914+cu113 pypi_0 pypi
[conda] torchaudio 0.12.1 py39_cu113 pytorch
[conda] torchmetrics 0.7.3 pypi_0 pypi
[conda] torchvision 0.13.1 pypi_0 pypi

[pmeier]

I'll leave the FFMPEG issue up to our video folks, but you could simply

git checkout v0.13.1

and build that from source. That is the version compatible with torch==1.12.1.

[bjuncek]

Hi @nikky4D and thanks for your question.
First and foremost, I'm not sure why there would be FFMPEG built when you build your torchvision from conda (on official channels; conda-forge version may include FFMPEG).

If you want to have the code build against your own FFMPEG, you'd have to compile TV from source, and make sure your FFMPEG installation is discoverable by TV (see setup.py, usually requires that it's included in path and library path env variables).

Let me know if you have any issues and I'd be happy to help you further.