LeoXing1996/decord

An efficient video loader for deep learning with smart shuffling that's super easy to digest

Decord

Decord is a reverse procedure of Record. It provides convenient video slicing methods based on a thin wrapper on top of hardware accelerated video decoders, e.g.

• FFMPEG/LibAV(Done)
• Nvidia Codecs(Done)
• Intel Codecs

Decord was designed to handle awkward video shuffling experience in order to provide smooth experiences similar to random image loader for deep learning.

Decord is also able to decode audio from both video and audio files. One can slice video and audio together to get a synchronized result; hence providing a one-stop solution for both video and audio decoding.

What's Difference with Original Repo

We change the some cpp code according to this PR and this issue comment.

We test the modification with following system environment, and cannot ensure it works on other environments.

# Operation System
cat /etc/os-release

# Output
NAME="CentOS Linux"
VERSION="7 (Core)"
ID="centos"
ID_LIKE="rhel fedora"
VERSION_ID="7"
PRETTY_NAME="CentOS Linux 7 (Core)"
ANSI_COLOR="0;31"
CPE_NAME="cpe:/o:centos:centos:7"
HOME_URL="https://www.centos.org/"
BUG_REPORT_URL="https://bugs.centos.org/"

CENTOS_MANTISBT_PROJECT="CentOS-7"
CENTOS_MANTISBT_PROJECT_VERSION="7"
REDHAT_SUPPORT_PRODUCT="centos"
REDHAT_SUPPORT_PRODUCT_VERSION="7"

# GCC
gcc --version

# Output
gcc (GCC) 10.2.0
Copyright © 2020 Free Software Foundation, Inc.
本程序是自由软件；请参看源代码的版权声明。本软件没有任何担保；
包括没有适销性和某一专用目的下的适用性担保。

# Cmake
cmake --version

# Output
cmake version 3.25.1

CMake suite maintained and supported by Kitware (kitware.com/cmake).

# NVCC
nvcc -V

# Output
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Wed_Sep_21_10:33:58_PDT_2022
Cuda compilation tools, release 11.8, V11.8.89
Build cuda_11.8.r11.8/compiler.31833905_0

# Python
python --version

# Output
Python 3.10.14

How to Compile GPU version without sudo?

1. Install FFMPEG via conda-forge

conda install -c conda-forge ffmpeg

# check installation
ffmpeg -hwaccels

# Outputs
ffmpeg version 5.1.2 Copyright (c) 2000-2022 the FFmpeg developers
  built with gcc 11.3.0 (conda-forge gcc 11.3.0-19)
  configuration: --prefix=/home/conda/feedstock_root/build_artifacts/ffmpeg_1674566204550/ ....
  libavutil      57. 28.100 / 57. 28.100
  libavcodec     59. 37.100 / 59. 37.100
  libavformat    59. 27.100 / 59. 27.100
  libavdevice    59.  7.100 / 59.  7.100
  libavfilter     8. 44.100 /  8. 44.100
  libswscale      6.  7.100 /  6.  7.100
  libswresample   4.  7.100 /  4.  7.100
  libpostproc    56.  6.100 / 56.  6.100
Hardware acceleration methods:
cuda
vaapi

2. Clone this Repo with submodules

git clone --recursive https://github.com/LeoXing1996/decord

3. Start to compile!

cd decord
mkdir build && cd build
cmake .. -DUSE_CUDA=ON -DCMAKE_BUILD_TYPE=Release -DFFMPEG_DIR=${CONDA_PREFIX} -DUSE_CUDA=YOU_CUDA_HOME
make

4. Install Python Bindings

cd ../python
python setup.py install --user

5. Test!

import time
from decord import VideoReader, gpu

frame_list = []
with open("YOU_VIDEO.mp4", "rb") as f:
    vr = VideoReader(f, ctx=gpu(0))
    s_ = time.time()

    for s in range(0, len(vr), 100):
        e = min(s + 100, len(vr))
        frame = vr.get_batch([i for i in range(s, e)]).asnumpy()
        frame_list.append(frame)
        print(s, e, frame.shape)

    e_ = time.time()

print(f'Total time: {e_ - s_}')

Table of contents

• Decord
  • What's Difference with Original Repo
  • How to Compile GPU version without sudo?
• Table of contents
  • Preliminary benchmark
  • Installation
    • Install via pip
    • Install from source
      • Linux
      • Mac OS
      • Windows
  • Usage
    • VideoReader
    • VideoLoader
    • AudioReader
    • AVReader
  • Bridges for deep learning frameworks:

Preliminary benchmark

Decord is good at handling random access patterns, which is rather common during neural network training.

Installation

Install via pip

Simply use

pip install decord

Supported platforms:

• Linux
• Mac OS >= 10.12, python>=3.5
• Windows

Note that only CPU versions are provided with PYPI now. Please build from source to enable GPU acclerator.

Install from source

Linux

Install the system packages for building the shared library, for Debian/Ubuntu users, run:

# official PPA comes with ffmpeg 2.8, which lacks tons of features, we use ffmpeg 4.0 here
sudo add-apt-repository ppa:jonathonf/ffmpeg-4 # for ubuntu20.04 official PPA is already version 4.2, you may skip this step
sudo apt-get update
sudo apt-get install -y build-essential python3-dev python3-setuptools make cmake
sudo apt-get install -y ffmpeg libavcodec-dev libavfilter-dev libavformat-dev libavutil-dev
# note: make sure you have cmake 3.8 or later, you can install from cmake official website if it's too old

Clone the repo recursively(important)

git clone --recursive https://github.com/dmlc/decord

Build the shared library in source root directory:

cd decord
mkdir build && cd build
cmake .. -DUSE_CUDA=0 -DCMAKE_BUILD_TYPE=Release
make

you can specify -DUSE_CUDA=ON or -DUSE_CUDA=/path/to/cuda or -DUSE_CUDA=ON -DCMAKE_CUDA_COMPILER=/path/to/cuda/nvcc to enable NVDEC hardware accelerated decoding:

cmake .. -DUSE_CUDA=ON -DCMAKE_BUILD_TYPE=Release

Note that if you encountered the an issue with libnvcuvid.so (e.g., see #102), it's probably due to the missing link for libnvcuvid.so, you can manually find it (ldconfig -p | grep libnvcuvid) and link the library to CUDA_TOOLKIT_ROOT_DIR\lib64 to allow decord smoothly detect and link the correct library.

To specify a customized FFMPEG library path, use `-DFFMPEG_DIR=/path/to/ffmpeg".

Install python bindings:

cd ../python
# option 1: add python path to $PYTHONPATH, you will need to install numpy separately
pwd=$PWD
echo "PYTHONPATH=$PYTHONPATH:$pwd" >> ~/.bashrc
source ~/.bashrc
# option 2: install with setuptools
python3 setup.py install --user

Mac OS

Installation on macOS is similar to Linux. But macOS users need to install building tools like clang, GNU Make, cmake first.

Tools like clang and GNU Make are packaged in Command Line Tools for macOS. To install:

xcode-select --install

To install other needed packages like cmake, we recommend first installing Homebrew, which is a popular package manager for macOS. Detailed instructions can be found on its homepage.

After installation of Homebrew, install cmake and ffmpeg by:

brew install cmake ffmpeg
# note: make sure you have cmake 3.8 or later, you can install from cmake official website if it's too old

Clone the repo recursively(important)

git clone --recursive https://github.com/dmlc/decord

Then go to root directory build shared library:

cd decord
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make

Install python bindings:

cd ../python
# option 1: add python path to $PYTHONPATH, you will need to install numpy separately
pwd=$PWD
echo "PYTHONPATH=$PYTHONPATH:$pwd" >> ~/.bash_profile
source ~/.bash_profile
# option 2: install with setuptools
python3 setup.py install --user

Windows

For windows, you will need CMake and Visual Studio for C++ compilation.

• First, install git, cmake, ffmpeg and python. You can use Chocolatey to manage packages similar to Linux/Mac OS.
• Second, install Visual Studio 2017 Community, this my take some time.

When dependencies are ready, open command line prompt:

cd your-workspace
git clone --recursive https://github.com/dmlc/decord
cd decord
mkdir build
cd build
cmake -DCMAKE_CXX_FLAGS="/DDECORD_EXPORTS" -DCMAKE_CONFIGURATION_TYPES="Release" -G "Visual Studio 15 2017 Win64" ..
# open `decord.sln` and build project

Usage

Decord provides minimal API set for bootstraping. You can also check out jupyter notebook examples.

VideoReader

VideoReader is used to access frames directly from video files.

from decord import VideoReader
from decord import cpu, gpu

vr = VideoReader('examples/flipping_a_pancake.mkv', ctx=cpu(0))
# a file like object works as well, for in-memory decoding
with open('examples/flipping_a_pancake.mkv', 'rb') as f:
  vr = VideoReader(f, ctx=cpu(0))
print('video frames:', len(vr))
# 1. the simplest way is to directly access frames
for i in range(len(vr)):
    # the video reader will handle seeking and skipping in the most efficient manner
    frame = vr[i]
    print(frame.shape)

# To get multiple frames at once, use get_batch
# this is the efficient way to obtain a long list of frames
frames = vr.get_batch([1, 3, 5, 7, 9])
print(frames.shape)
# (5, 240, 320, 3)
# duplicate frame indices will be accepted and handled internally to avoid duplicate decoding
frames2 = vr.get_batch([1, 2, 3, 2, 3, 4, 3, 4, 5]).asnumpy()
print(frames2.shape)
# (9, 240, 320, 3)

# 2. you can do cv2 style reading as well
# skip 100 frames
vr.skip_frames(100)
# seek to start
vr.seek(0)
batch = vr.next()
print('frame shape:', batch.shape)
print('numpy frames:', batch.asnumpy())

VideoLoader

VideoLoader is designed for training deep learning models with tons of video files. It provides smart video shuffle techniques in order to provide high random access performance (We know that seeking in video is super slow and redundant). The optimizations are underlying in the C++ code, which are invisible to user.

from decord import VideoLoader
from decord import cpu, gpu

vl = VideoLoader(['1.mp4', '2.avi', '3.mpeg'], ctx=[cpu(0)], shape=(2, 320, 240, 3), interval=1, skip=5, shuffle=1)
print('Total batches:', len(vl))

for batch in vl:
    print(batch[0].shape)

Shuffling video can be tricky, thus we provide various modes:

shuffle = -1  # smart shuffle mode, based on video properties, (not implemented yet)
shuffle = 0  # all sequential, no seeking, following initial filename order
shuffle = 1  # random filename order, no random access for each video, very efficient
shuffle = 2  # random order
shuffle = 3  # random frame access in each video only

AudioReader

AudioReader is used to access samples directly from both video(if there's an audio track) and audio files.

from decord import AudioReader
from decord import cpu, gpu

# You can specify the desired sample rate and channel layout
# For channels there are two options: default to the original layout or mono
ar = AudioReader('example.mp3', ctx=cpu(0), sample_rate=44100, mono=False)
print('Shape of audio samples: ', ar.shape())
# To access the audio samples
print('The first sample: ', ar[0])
print('The first five samples: ', ar[0:5])
print('Get a batch of samples: ', ar.get_batch([1,3,5]))

AVReader

AVReader is a wraper for both AudioReader and VideoReader. It enables you to slice the video and audio simultaneously.

from decord import AVReader
from decord import cpu, gpu

av = AVReader('example.mov', ctx=cpu(0))
# To access both the video frames and corresponding audio samples
audio, video = av[0:20]
# Each element in audio will be a batch of samples corresponding to a frame of video
print('Frame #: ', len(audio))
print('Shape of the audio samples of the first frame: ', audio[0].shape)
print('Shape of the first frame: ', video.asnumpy()[0].shape)
# Similarly, to get a batch
audio2, video2 = av.get_batch([1,3,5])

Bridges for deep learning frameworks:

It's important to have a bridge from decord to popular deep learning frameworks for training/inference

• Apache MXNet (Done)
• Pytorch (Done)
• TensorFlow (Done)

Using bridges for deep learning frameworks are simple, for example, one can set the default tensor output to mxnet.ndarray:

import decord
vr = decord.VideoReader('examples/flipping_a_pancake.mkv')
print('native output:', type(vr[0]), vr[0].shape)
# native output: <class 'decord.ndarray.NDArray'>, (240, 426, 3)
# you only need to set the output type once
decord.bridge.set_bridge('mxnet')
print(type(vr[0], vr[0].shape))
# <class 'mxnet.ndarray.ndarray.NDArray'> (240, 426, 3)
# or pytorch and tensorflow(>=2.2.0)
decord.bridge.set_bridge('torch')
decord.bridge.set_bridge('tensorflow')
# or back to decord native format
decord.bridge.set_bridge('native')