Recherche avancée

Sur d’autres sites (17648)

• libavfi/dnn : add LibTorch as one of DNN backend

  15 mars 2024, par Wenbin Chen

  libavfi/dnn : add LibTorch as one of DNN backend
  PyTorch is an open source machine learning framework that accelerates
  
  the path from research prototyping to production deployment. Official
  
  website : https://pytorch.org/. We call the C++ library of PyTorch as
  
  LibTorch, the same below.
  To build FFmpeg with LibTorch, please take following steps as
  
  reference :
  
  1. download LibTorch C++ library in
  
   https://pytorch.org/get-started/locally/,
  
  please select C++/Java for language, and other options as your need.
  
  Please download cxx11 ABI version :
  
   (libtorch-cxx11-abi-shared-with-deps-*.zip).
  
  2. unzip the file to your own dir, with command
  
  unzip libtorch-shared-with-deps-latest.zip -d your_dir
  
  3. export libtorch_root/libtorch/include and
  
  libtorch_root/libtorch/include/torch/csrc/api/include to $PATH
  
  export libtorch_root/libtorch/lib/ to $LD_LIBRARY_PATH
  
  4. config FFmpeg with ../configure —enable-libtorch \
  
   —extra-cflag=-I/libtorch_root/libtorch/include \
  
   —extra-cflag=-I/libtorch_root/libtorch/include/torch/csrc/api/include \
  
   —extra-ldflags=-L/libtorch_root/libtorch/lib/
  
  5. make
  To run FFmpeg DNN inference with LibTorch backend :
  
  ./ffmpeg -i input.jpg -vf \
  
  dnn_processing=dnn_backend=torch:model=LibTorch_model.pt -y output.jpg
  The LibTorch_model.pt can be generated by Python with torch.jit.script()
  
  api. https://pytorch.org/tutorials/advanced/cpp_export.html. This is
  
  pytorch official guide about how to convert and load torchscript model.
  
  Please note, torch.jit.trace() is not recommanded, since it does
  
  not support ambiguous input size.
  Signed-off-by : Ting Fu <ting.fu@intel.com>
  
  Signed-off-by : Wenbin Chen <wenbin.chen@intel.com>
  
  Reviewed-by : Guo Yejun <yejun.guo@intel.com>
  

  • [DH] configure
  • [DH] libavfilter/dnn/Makefile
  • [DH] libavfilter/dnn/dnn_backend_torch.cpp
  • [DH] libavfilter/dnn/dnn_interface.c
  • [DH] libavfilter/dnn_filter_common.c
  • [DH] libavfilter/dnn_interface.h
  • [DH] libavfilter/vf_dnn_processing.c

• Process GIF using FFmpeg libraries - can't find parser at av_parser_init

  24 novembre 2018, par natario

  I am playing with ffmpeg libs, namely libswscale and libavcodec for now. My goal is resize GIF files. From my ridiculous understanding, I think I need to

  • decode the GIF and get an AVFrame
  • process the frame with libswscale
  • encode again into GIF

  But I am stuck at step 1. Based on official sample at https://github.com/FFmpeg/FFmpeg/blob/master/doc/examples/decode_video.c , I need to get a parser :

  codec = avcodec_find_decoder(AV_CODEC_ID_GIF);
  
  parser = av_parser_init(codec->id);

  But no parser is found. I am not touching parser in my configure call, so I take all :

  Enabled parsers:
  
  aac                       cavsvideo                 dvbsub                    h263                      mpegvideo                 sipr                      xma
  
  aac_latm                  cook                      dvd_nav                   h264                      opus                      tak
  
  ac3                       dca                       dvdsub                    hevc                      png                       vc1
  
  adx                       dirac                     flac                      mjpeg                     pnm                       vorbis
  
  av1                       dnxhd                     g729                      mlp                       rv30                      vp3
  
  avs2                      dpx                       gsm                       mpeg4video                rv40                      vp8
  
  bmp                       dvaudio                   h261                      mpegaudio                 sbc                       vp9

  What am I doing wrong ? If this is the wrong approach, what is the correct one ?

• three ways to achieve an H264 file

  18 mars 2016, par Kindermann

  here I have three ways to get an H264 file, like all forensic scientists, I am very curious about the differences between them :

  1.

  ffmpeg -i video.mp4 video.h264

  2.

  ffmpeg -i video.mp4 -vcodec copy -an -f h264 video.h264

  3. Using the example "demuxing_decoding.c" provided on the ffmpeg official website :
  http://ffmpeg.org/doxygen/trunk/demuxing_decoding_8c-example.html

  Obviously, the first one does the transformation, and the second one does the demuxing. They render different H264 files which however have similar file sizes(in my case, it’s about say 24 MB). Suprisingly, the third one, which is also supposed to do the demuxing job, renders an H264 file with 8.4 GB ! Why ?

  What I wondered is really, how the interiors of these three methods work ?(The third one is already in source code, therefore it’s quite easy to have an insight) What about the first two commands ? What APIs are called when executing these two commands and how those APIs are called(namely, in what kind of sequences they are called) and things like that.
  One thing that is also important to me is, i have no idea how I can trace the execution routines of ffmpeg command lines. I want to see what’s going on behind ffmpeg commands in source code version. Is it possible ?

  I appreciate any comment.