Recherche avancée

Sur d’autres sites (11794)

• Docker Image for Torchvsion and Torchaudio with FFmpeg with nvidia headers

  22 juillet 2024, par Felipe Marra

  I'm trying to perform GPU video encoding/decoding using PyTorch. This means compiling FFmpeg from source along with Nvidia codec headers.

  


  Currently, my docker image looks like this :

  


  FROM nvcr.io/nvidia/pytorch:24.06-py3

RUN apt-get -yqq update && \
    apt-get install -yq --no-install-recommends ca-certificates expat libgomp1 && \
    apt-get autoremove -y && \
    apt-get clean -y

RUN apt-get install -y bash \
    autoconf \
    automake \
    build-essential \
    cmake \
    git-core \
    libass-dev \
    libfreetype6-dev \
    libgnutls28-dev \
    libmp3lame-dev \
    libnuma1 \
    libnuma-dev \
    libsdl2-dev \
    libtool \
    libva-dev \
    libvdpau-dev \
    libvorbis-dev \
    libxcb1-dev \
    libxcb-shm0-dev \
    libxcb-xfixes0-dev \
    libc6 \
    libc6-dev \
    meson \
    ninja-build \
    pkg-config \
    texinfo \
    unzip \
    wget \
    yasm \
    zlib1g-dev && \
    apt-get -yqq update

# Miniconda
COPY .devcontainer/env.yml .

RUN mkdir -p ~/miniconda3 && \
    wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh && \
    bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3 && \
    rm -rf ~/miniconda3/miniconda.sh && \
    ~/miniconda3/bin/conda init bash && \
    ~/miniconda3/bin/conda init zsh && \
    ~/miniconda3/bin/conda env update --file env.yml

# FFMPEG
RUN git clone https://git.videolan.org/git/ffmpeg/nv-codec-headers.git && \
    cd nv-codec-headers && \
    make install

RUN git clone https://git.ffmpeg.org/ffmpeg.git ffmpeg/ && \
    cd ffmpeg && \
    ./configure \
    --prefix="$CONDA_PREFIX"\
    --enable-nonfree \
    --enable-cuda-nvcc \
    --enable-libnpp \
    --extra-cflags=-I/usr/local/cuda/include \
    --extra-ldflags=-L/usr/local/cuda/lib64 \
    --disable-static \
    --enable-shared && \
    make -k -j 8 && \
    make install && \
    ldconfig

ENV LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH


  


  The env.yml contains the following :

  


  name: base
channels:
  - pytorch
  - nvidia
  - conda-forge
  - defaults
dependencies:
  - pytorch=2.3.1=py3.12_cuda12.1_cudnn8.9.2_0
  - torchvision=0.18.1=py312_cu121
  - torchaudio=2.3.1=py312_cu121
  - pytorch-cuda=12.1=ha16c6d3_5


  


  To test the image, I'm running :

  


  import torchvision
from torchaudio.utils import ffmpeg_utils

print("Library versions:")
print(ffmpeg_utils.get_versions())
print("\nBuild config:")
print(ffmpeg_utils.get_build_config())
print("\nDecoders:")
print([k for k in ffmpeg_utils.get_video_decoders().keys() if "cuvid" in k])
print("\nEncoders:")
print([k for k in ffmpeg_utils.get_video_encoders().keys() if "nvenc" in k])

torchvision.set_video_backend('cuda')


  


  I've also created this repo so that other people will be able to just run this image once its problems get solved.

  


  What is going on is that outside the Conda environment, FFmpeg is configured as expected. But inside it, I'm getting the folowing error :

  


  libopenh264.so.5: cannot open shared object file: No such file or directory


  


  By following this comment on a torchvision issue, I was able to make ffprobe -hide_banner -decoders | grep h264 and ffmpeg -hide_banner -encoders | grep 264 yield the expected outputs, as shown in torchaudio's doc, inside the Conda environment. But then torchaudio wasn't able to find FFmpeg.

  


  I'm new to the whole ecosystem (linux, docker and torch), and would appreciate it if someone could help me build this image, as I think that this should actually be officially provided by torch.

  


• Revision 30244 : On mets à jour le script d’encodage pour tester l’encodage multiple (plus ...

  26 juillet 2009, par kent1@… — Log

  On mets à jour le script d’encodage pour tester l’encodage multiple (plus uniquement en flv)

• stream_decoder : Two read_metadata() fixes from 1.2.1 maintenance branch.

  28 juin 2014, par Erik de Castro Lopo

  stream_decoder : Two read_metadata() fixes from 1.2.1 maintenance branch.
  * Fix leaks in read_metadata_() that could occur because of read errors or
  
    malformed streams.
  
      http://flac.cvs.sourceforge.net/viewvc/flac/flac/src/libFLAC/
  
      stream_decoder.c ?
  
      r1=1.147&r2=1.147.2.1&pathrev=FLAC_RELEASE_1_2_1_MAINTENANCE_BRANCH
  * Fix metadata block initialization bug in read_metadata_().
  
      http://flac.cvs.sourceforge.net/viewvc/flac/flac/src/libFLAC/
  
      stream_decoder.c ?
  
      r1=1.147.2.1&r2=1.147.2.2&pathrev=FLAC_RELEASE_1_2_1_MAINTENANCE_BRANCH
  Patch-from : lvqcl <lvqcl.mail@gmail.com>
  

  • [DH] src/libFLAC/stream_decoder.c