lsrock1 / maskrcnn_benchmark.cpp

Implementation maskrcnn-benchmark, pytorch c++ frontend

maskrcnn_benchmark.cpp

faster rcnn cpp implementation based on maskrcnn-benchmark

Codes

All code architecture from maskrcnn-benchmark

Installation

export INSTALL_DIR=$PWD
git clone --recursive https://github.com/lsrock1/maskrcnn_benchmark.cpp.git
cd $INSTALL_DIR

// if opencv is not installed
// opencv install
cd maskrcnn_benchmark.cpp/lib/opencv
mkdir build
cd build
cmake ..
make install

// if yaml-cpp is not installed
// yaml-cpp install
cd $INSTALL_DIR
cd maskrcnn_benchmark.cpp/lib/yaml-cpp
mkdir build
cd build
cmake ..
make install

// Download pytorch cpp (cxx11 ABI)
// place libtorch folder in maskrcnn_benchmark.cpp/lib directory
// if you are using without cuda and get dyld: Library not loaded: @rpath/libmklml.dylib error
// https://github.com/pytorch/pytorch/issues/14165

cd $INSTALL_DIR
cd maskrcnn_benchmark.cpp
mkdir build
cd build
cmake ..
make

//run inference r50-fpn
run.out ../configs/e2e_faster_rcnn_R_50_FPN_1x.yaml inference

Datasets

Download coco datasets
make directory under maskrcnn_benchmark.cpp

datasets
  - coco
    - train2017
    - val2017
  - annotations
    - instances_val2017.json
    - instances_train2017.json

Results

Device: RTX2080ti 1 GPU, cuda 10, cudnn 7, Ubuntu 16.04

ResNet

VoVNet

ResNets are 1x and VoVNets are 2x

backbone	type	lr sched	inference total batch	inference time(s/im)	box AP	Speed Improvement
R-50-FPN(python)	Fast	1x	8	0.05989	0.368	0
R-50-FPN(cpp)	Fast	1x	8	0.05520	0.368	7.8%
R-101-FPN(python)	Fast	1x	8	0.07627	0.391	0
R-101-FPN(cpp)	Fast	1x	8	0.07176	0.391	5.9%
VoV-39(python)	Fast	2x	8	0.06479	0.398	0
VoV-39(cpp)	Fast	2x	8	0.05949	0.398	8.1%
VoV-57(python)	Fast	2x	8	0.07224	0.409	0
VoV-57(cpp)	Fast	2x	8	0.06713	0.409	7%

Warning

In Progress.

• It doesn't support training yet.(Testing!)

TODO

• concat dataset
• python jit -> cpp model code
• multi GPU training(code complete but bug exists in libtorch)
• cmake install
• clean up code

Requirements

• Yaml-cpp
• gtest
• libtorch >= 1.2
• rapidjson
• opencv

MODELS

Weight python to cpp Download into /models

name	from	link(dropbox)	link(google drive)
R-50(backbone only)	python-pretrained	link	link
R-101(backbone only)	python-pretrained	link	link
R-50-FPN	python-trained	link	link
R-101-FPN	python-trained	link	link
R-50-C4	python-trained	link	link
V-39-FPN	python-trained	link	link
V-57-FPN	python-trained	link	link