Skip to content

Latest commit

 

History

History
159 lines (128 loc) · 5.62 KB

INSTALL.md

File metadata and controls

159 lines (128 loc) · 5.62 KB
Raw

Installation

git clone --recursive https://github.com/open-mmlab/mmaction.git

Requirements

  • Linux
  • Python 3.5+
  • PyTorch 1.0+
  • CUDA 9.0+
  • NVCC 2+
  • GCC 4.9+
  • ffmpeg 4.0+
  • mmcv. Note that you are strongly recommended to clone the master branch and build from scratch since some of the features have not been added in the latest release.
  • Decord
  • dense_flow

Note that the last two will be contained in this codebase as a submodule.

Install Pre-requisites

Install Decord (Optional)

Decord is an efficient video loader with smart shuffling. This is required when you want to use videos as the input format for training and is more efficient than OpenCV, useds by mmcv. If you just want to have a quick experience with MMAction, you can simply skip this step. The installation steps follow decord's documentation.

(a) install the required packages by running:

# official PPA comes with ffmpeg 2.8, which lacks tons of features, we use ffmpeg 4.0 here
sudo apt-get install -y software-properties-common
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 
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
sudo apt-get install ffmpeg

(b) Build the library from source

cd third_party/decord
mkdir build && cd build
cmake .. -DUSE_CUDA=0
make

(c) Install python binding

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

Install dense_flow (Optional)

Dense_flow is used to calculate the optical flow of videos. If you just want to have a quick experience with MMAction without taking pain of installing opencv, you can skip this step.

Note that Dense_flow now supports OpenCV 4.1.0, 3.1.0 and 2.4.13. The master branch is for 4.1.0. For those with 2.4.13, please refer to the lines with strikethrough.

(a) Install OpenCV=2.4.13 
cd third_party/
sudo apt-get -qq install libopencv-dev build-essential checkinstall cmake pkg-config yasm libjpeg-dev libjasper-dev libavcodec-dev libavformat-dev libswscale-dev libdc1394-22-dev libgstreamer1.0-dev libgstreamer-plugins-base1.0-dev libv4l-dev python-dev python-numpy libtbb-dev libqt4-dev libgtk2.0-dev libfaac-dev libmp3lame-dev libopencore-amrnb-dev libopencore-amrwb-dev libtheora-dev libvorbis-dev libxvidcore-dev x264 v4l-utils

wget -O OpenCV-2.4.13.zip https://github.com/Itseez/opencv/archive/2.4.13.zip
unzip OpenCV-2.4.13.zip

cd opencv-2.4.13
mkdir build && cd build
cmake -D CMAKE_BUILD_TYPE=RELEASE -D WITH_TBB=ON -D WITH_V4L=ON ..
make -j32
cd ../../../

(a) Install OpenCV=4.1.0

  1. (For CUDA 10.0 only) CUDA 9.x should have no problem. Video decoder is deprecated in CUDA 10.0. To handle this, download NVIDIA VIDEO CODEC SDK and copy the header files to your cuda path (/usr/local/cuda-10.0/include/ for example). Note that you may have to do as root.
unzip Video_Codec_SDK_9.0.20.zip
cp Video_Codec_SDK_9.0.20/include/nvcuvid.h /usr/local/cuda-10.0/include/
cp Video_Codec_SDK_9.0.20/include/cuviddec.h /usr/local/cuda-10.0/include/
cp Video_Codec_SDK_9.0.20/Lib/linux/stubs/x86_64/libnvcuvid.so /usr/local/cuda-10.0/lib64/libnvcuvid.so.1
  1. Obtain required packages for building OpenCV 4.1.0 (duplicated with requirements for Decord in part)
sudo apt-get install -y liblapack-dev libatlas-base-dev
sudo apt-get install -y libgstreamer1.0-dev libgstreamer-plugins-base1.0-dev libavcodec-dev libavformat-dev libswscale-dev libdc1394-22-dev
sudo apt-get install -y software-properties-common
sudo add-apt-repository ppa:jonathonf/ffmpeg-4
sudo apt update
sudo apt install -y ffmpeg
  1. Obtain OpenCV 4.1.0 and its extra modules (optflow, etc.) by
cd third_party
wget -O OpenCV-4.1.0.zip wget https://github.com/opencv/opencv/archive/4.1.0.zip
unzip OpenCV-4.1.0.zip
wget -O OpenCV_contrib-4.1.0.zip https://github.com/opencv/opencv_contrib/archive/4.1.0.zip
unzip OpenCV_contrib-4.1.0.zip
  1. Build OpenCV 4.1.0 from scratch (due to some custom settings)
cd opencv-4.1.0
mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Release -DWITH_CUDA=ON -DOPENCV_EXTRA_MODULES_PATH=../../opencv_contrib-4.1.0/modules/ -DWITH_TBB=ON -DBUILD_opencv_cnn_3dobj=OFF -DBUILD_opencv_dnn=OFF -DBUILD_opencv_dnn_modern=OFF -DBUILD_opencv_dnns_easily_fooled=OFF -DOPENCV_ENABLE_NONFREE=ON ..
make -j

Note that -DOPENCV_ENABLE_NONFREE=ON is explicitly set to enable warped flow proposed in TSN. You can skip this argument to speed up the compilation if you do not intend to use it.

(b) Build dense_flow

cd third_party/dense_flow
# dense_flow dependencies
sudo apt-get -qq install libzip-dev libboost-all-dev
mkdir build && cd build
# deprecated:
# OpenCV_DIR=../../opencv-2.4.13/build cmake ..
OpenCV_DIR=../../opencv-4.1.0/build cmake ..
make -j

Install MMAction

(a) Install Cython

pip install cython

(b) Compile CUDA extensions

./compile.sh

(c) Install mmaction

python setup.py develop

Please refer to DATASET.md to get familar with the data preparation and to GETTING_STARTED.md to use MMAction.