chainer-mask-rcnn

Chainer Implementation of Mask R-CNN.

Features

• ResNet50, ResNet101 backbone.
• VOC and COCO training examples.
• Reproduced result of original work (ResNet50, COCO).
• Weight copy from pretrained model at facebookresearch/Detectron.
• Training with batch size >= 2.
• Support FPN backbones.
• Keypoint detection.

Fig 1. Mask R-CNN, ResNet50, 8GPU, Ours, COCO 31.4 mAP@50:95

COCO Results

Model	Implementation	N gpu training	mAP@50:95	Log
Mask R-CNN, ResNet50	Ours	8	31.5 - 31.8	Log
Mask R-CNN, ResNet50	Detectron	8	31.4 (30.8 after copied)	Log
FCIS, ResNet50	FCIS	8	27.1	-

Inference

# you can use your trained model
./demo.py logs/<YOUR_TRAINING_LOG> --img <IMAGE_PATH_OR_URL>

# COCO Example: Mask R-CNN, ResNet50, 31.4 mAP@50:95
cd examples/coco
LOG_DIR=logs/20180730_081433
mkdir -p $LOG_DIR
pip install gdown
gdown https://drive.google.com/uc?id=1XC-Mx4HX0YBIy0Fbp59EjJFOF7a3XK0R -O $LOG_DIR/snapshot_model.npz
gdown https://drive.google.com/uc?id=1fXHanL2pBakbkv83wn69QhI6nM6KjrzL -O $LOG_DIR/params.yaml
./demo.py $LOG_DIR

# copy weight from caffe2 to chainer
cd examples/coco
./convert_caffe2_to_chainer.py  # or download from https://drive.google.com/open?id=1WOEtVnxqYdHl35pAyIcp-H0HtTjI-l3V
./demo.py logs/R-50-C4_x1_caffe2_to_chainer --img https://raw.githubusercontent.com/facebookresearch/Detectron/master/demo/33823288584_1d21cf0a26_k.jpg
./demo.py logs/R-50-C4_x1_caffe2_to_chainer --img https://raw.githubusercontent.com/facebookresearch/Detectron/master/demo/17790319373_bd19b24cfc_k.jpg

Fig 2. Mask R-CNN, ResNet50, 8GPU, Copied from Detectron, COCO 31.4 mAP@50:95

Installation & Training

Single GPU Training

# Install Chainer Mask R-CNN.
pip install opencv-python
pip install .

# Run training!
cd examples/coco && ./train.py --gpu 0

Multi GPU Training

# Install OpenMPI
wget https://www.open-mpi.org/software/ompi/v3.0/downloads/openmpi-3.0.0.tar.gz
tar zxvf openmpi-3.0.0.tar.gz
cd openmpi-3.0.0
./configure --with-cuda
make -j4
sudo make install
sudo ldconfig

# Install NCCL
# dpkg -i nccl-repo-ubuntu1404-2.1.4-ga-cuda8.0_1-1_amd64.deb
dpkg -i nccl-repo-ubuntu1604-2.1.15-ga-cuda9.1_1-1_amd64.deb
sudo apt update
sudo apt install libnccl2 libnccl-dev

# Install ChainerMN
pip install chainermn

# Finally, install Chainer Mask R-CNN.
pip install opencv-python
pip install .

# Run training!
cd examples/coco && mpirun -n 4 ./train.py --multi-node

Testing

pip install flake8 pytest
flake8 .
pytest -v tests