Need help with decord?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

764 Stars 64 Forks Apache License 2.0 348 Commits 65 Opened issues


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

Services available


Need anything else?

Contributors list


CI Build Release Build PyPI Downloads


is a reverse procedure of
. 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

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

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.

Table of contents

Preliminary benchmark

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

Speed up


Install via pip

Simply use

pip install decord

Supported platforms:

  • [x] Linux
  • [x] Mac OS >= 10.12, python>=3.5
  • [x] Windows

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

Install from source


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
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

Build the shared library in source root directory:

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

you can specify

to enable NVDEC hardware accelerated decoding:

Note that if you encountered the issue(e.g., the offical nvidia docker, see #102) with
, it's probably due to the missing link for
, you can manually find(
ldconfig -p | grep libnvcuvid
) and link the library to
to allow
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 $PYTHONPATH, you will need to install numpy separately
echo "PYTHONPATH=$PYTHONPATH:$pwd" >> ~/.bashrc
source ~/.bashrc
# option 2: install with setuptools
python3 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

Then go to root directory build shared library:

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

Install python bindings:

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


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

When dependencies are ready, open command line prompt:

cd your-workspace
git clone --recursive
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


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


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


seek to start batch = print('frame shape:', batch.shape) print('numpy frames:', batch.asnumpy())


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 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 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('', 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

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

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.