Installation

Overview

This package is distributed as part of the ACCV-Lab package, and can be installed as described in the Installation Guide.

Note

In most cases, it is recommended to follow the the general ACCV-Lab installation guide (see above) to install the package.

Independent Installation

If needed, the package can be installed independently of the ACCV-Lab package and the default environment using the steps outlined in this section.

Prerequisites

CUDA: >= 11.0
Operating System: Linux
Python: 3.8+
cmake: >= 3.21
scikit-build

Step 1: Pull Base Image

Choose an appropriate NVIDIA PyTorch container from NVIDIA Deep Learning Framework Documentation. Images released after 22.11 are required for successful compilation.

Example using PyTorch 23.02:

docker pull nvcr.io/nvidia/pytorch:23.02-py3

Step 2: Install Dependencies

Install `scikit-build`

pip install scikit-build

Install `accvlab_build_config`

cd build_config  # ← Assuming you are in the root directory of the repository
pip install . --no-build-isolation

Build and Install `FFmpeg`

FFMpeg version==4.4 has been tested, and we use this version here.

apt install yasm nasm -y
wget https://ffmpeg.org/releases/ffmpeg-4.4.6.tar.xz
tar -xf ffmpeg-4.4.6.tar.xz

cd ffmpeg-4.4.6
# FFmpeg Libraries used in this package can be built using the following configure options. The configuration 
# ensures that the FFmpeg libraries are built with the minimum required functionality to support 
# the On-demand Video Decoder Package, while also ensuring that only the components of FFmpeg that are 
# licensed under the LGPL license is used.
LDSOFLAGS=-Wl,-rpath,\''$$$$ORIGIN'\'  ./configure --enable-shared --disable-encoders --disable-decoders --enable-decoder=vp9 --arch=x86_64 --prefix=/PATH/TO/YOUR/FFMPEG

make -j
make install
# The FFMPEG_DIR path is used by the On-demand Video Decoder Package to find the FFmpeg libraries during 
# build time. 
export FFMPEG_DIR=/PATH/TO/YOUR/FFMPEG

Step 3: Build and Install

cd packages/on_demand_video_decoder
pip install . --no-build-isolation

Dataset Preparation

Please refer to the Dataset Preparation Guide for instructions on how to prepare the NuScenes (mini) dataset for testing and profiling the video decoder, as well as for potential use of the video dataset for training purposes.

Development

Debugging

For development and debugging purposes, you can trigger a debug build via the environment variables described in the main Installation Guide. The pip install command will internally run the CMake/scikit-build build, so you do not need to call cmake or make manually:

cd packages/on_demand_video_decoder
# Debug build with verbose output
DEBUG_BUILD=1 VERBOSE_BUILD=1 pip install . --no-build-isolation

Performance Profiling

Use NVIDIA Nsight Systems for comprehensive performance analysis:

nsys profile \
    --trace-fork-before-exec true \
    -w true \
    -f true \
    -t cuda,nvtx,osrt,cudnn,cublas,nvvideo \
    --gpu-video-device all \
    -x true \
    -o report \
    python samples/ProfileNuscenesGopDecoder.py

Profile Options Explained:

--trace-fork-before-exec true: Enable fork tracing
-w true: Overwrite existing output files
-f true: Force profiling
-t: Specify trace categories
--gpu-video-device all: Profile all GPU video devices
-x true: Export results
-o: Output file prefix