- GCC >= 4.9
- CUDA 9.0 & cuDNN 7.0 (install. instructions)
- Anaconda 3 (install. instructions)
- Faster-RCNN paper: arxiv.org/pdf/1506.01497.pdf
- Faster-RCNN implem. repo.: github.com/facebookresearch/maskrcnn-benchmark
- RetinaNet paper: arxiv.org/pdf/1708.02002.pdf
- RetinaNet implem. repo.: github.com/fizyr/keras-retinanet
- FCOS paper: arxiv.org/pdf/1904.01355.pdf
- FCOS implem. repo.: github.com/tianzhi0549/FCOS
Start by cloning this repo:
git clone https://github.com/delmalih/MIAS-mammography-obj-detection
- First, create an environment :
conda create --name faster-r-cnn
conda activate faster-r-cnn
conda install ipython pip
cd MIAS-mammography-obj-detection
pip install -r requirements.txt
cd ..
- Then, run these commands (ignore if you have already done the FCOS installation) :
# install pytorch
conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0 -c pytorch
export INSTALL_DIR=$PWD
# install pycocotools
cd $INSTALL_DIR
git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI
python setup.py build_ext install
# install cityscapesScripts
cd $INSTALL_DIR
git clone https://github.com/mcordts/cityscapesScripts.git
cd cityscapesScripts/
python setup.py build_ext install
# install apex
cd $INSTALL_DIR
git clone https://github.com/NVIDIA/apex.git
cd apex
python setup.py install --cuda_ext --cpp_ext
# install PyTorch Detection
cd $INSTALL_DIR
git clone https://github.com/facebookresearch/maskrcnn-benchmark.git
cd maskrcnn-benchmark
python setup.py build develop
cd $INSTALL_DIR
unset INSTALL_DIR
- First, create an environment :
conda create --name retinanet python=3.6
conda activate retinanet
conda install ipython pip
cd MIAS-mammography-obj-detection
pip install -r requirements.txt
cd ..
pip install tensorflow-gpu==1.9
pip install keras==2.2.5
- Then, run these commands :
# clone keras-retinanet repo
git clone https://github.com/fizyr/keras-retinanet
cd keras-retinanet
pip install .
python setup.py build_ext --inplace
- Finally, replace the
keras_retinanet/preprocessing/coco.pyfile by this file
- First, create an environment :
conda create --name fcos
conda activate fcos
conda install ipython pip
cd MIAS-mammography-obj-detection
pip install -r requirements.txt
cd ..
- Then, follow these instructions
Run these commands to download to MIAS database :
mkdir mias-db && cd mias-db
wget http://peipa.essex.ac.uk/pix/mias/all-mias.tar.gz
tar -zxvf all-mias.tar.gz
rm all-mias.tar.gz && cd ..
And replace the mias-db/Info.txt by this one
It is possible to generate COCO or VOC annotations from raw data (all-mias folder + Info.txt annotations file) through 2 scripts: generate_{COCO|VOC}_annotations.py :
python generate_{COCO|VOC}_annotations.py --images (or -i) <Path to the images folder> \
--annotations (or -a) <Path to the .txt annotations file> \
--output (or -o) <Path to output folder> \
--aug_fact <Data augmentation factor> \
--train_val_split <Percetange of the train folder (default 0.9)>
For example, to generate 10x augmented COCO annotations, run this command :
python generate_COCO_annotations.py --images ../mias-db/ \
--annotations ../mias-db/Info.txt \
--output ../mias-db/COCO \
--aug_fact 20 \
--train_val_split 0.9
To run a training with the Faster-RCNN:
- Go to the faster-r-cnn directory:
cd faster-r-cnn - Change conda env:
conda deactivate && conda activate faster-r-cnn - Download the Resnet_101_FPN model
- Trim the model:
python trim_detectron_model.py --pretrained_path e2e_faster_rcnn_R_101_FPN_1x.pth --save_path base_model.pth - Edit the
maskrcnn-benchmark/maskrcnn_benchmark/config/paths_catalog.pyfile and put these lines in theDATASETSdictionary :
DATASETS = {
...,
"mias_train_cocostyle": {
"img_dir": "<PATH_TO_'mias-db'_folder>/<COCO_FOLDER>/images/train",
"ann_file": "<PATH_TO_'mias-db'_folder>/<COCO_FOLDER>/annotations/instances_train.json"
},
"mias_val_cocostyle": {
"img_dir": "<PATH_TO_'mias-db'_folder>/<COCO_FOLDER>/images/val",
"ann_file": "<PATH_TO_'mias-db'_folder>/<COCO_FOLDER>/annotations/instances_val.json"
},
}
- In the
maskrcnn-benchmark/maskrcnn_benchmark/data/datasets/coco.py, comment line 84 to 92 :
# if anno and "segmentation" in anno[0]:
# masks = [obj["segmentation"] for obj in anno]
# masks = SegmentationMask(masks, img.size, mode='poly')
# target.add_field("masks", masks)
# if anno and "keypoints" in anno[0]:
# keypoints = [obj["keypoints"] for obj in anno]
# keypoints = PersonKeypoints(keypoints, img.size)
# target.add_field("keypoints", keypoints)
- Run this command :
python train.py --config-file mias_config.yml
To run a training with the retinanet :
cd retinanet
conda deactivate && conda activate retinanet
python train.py --compute-val-loss \ # Computer val loss or not
--tensorboard-dir <Path to the tensorboard directory> \
--batch-size <Batch size> \
--epochs <Nb of epochs> \
coco <Path to the COCO dataset>
And if you want to see the tensorboard, run on another window :
tensorboard --logdir <Path to the tensorboard directory>
To run a training with the FCOS Object Detector :
- Follow these instructions
- Run this command :
cd fcos
conda deactivate && conda activate fcos
python train.py --config-file <Path to the config file> \
OUTPUT_DIR <Path to the output dir for the logs>
To run an inference, you need a pre-trained model. Run this command:
cd faster-r-cnn
conda deactivate && conda activate faster-r-cnn
python inference.py --config-file <Path to the config file> \
MODEL.WEIGHT <Path to weights of the model to load> \
TEST.IMS_PER_BATCH <Nb of images per batch>
- Put the images you want to run an inference on, in
<Name of COCO dataset>/<Name of folder> - Run this command :
cd retinanet
conda deactivate && conda activate retinanet
python inference.py --snapshot <Path of the model snapshot> \
--set_name <Name of the inference folder in the COCO dataset> \
coco <Path to the COCO dataset>
cd fcos
conda deactivate && conda activate fcos
python inference.py --config-file <Path to the config file> \
MODEL.WEIGHT <Path to weights of the model to load> \
TEST.IMS_PER_BATCH <Nb of images per batch>
| Metric | Faster-RCNN | RetinaNet | FCOS |
|---|---|---|---|
| mAP | 98,70% | 94,97% | 98,20% |
| Precision | 94,12% | 100,00% | 94,44% |
| Recall | 98,65% | 94,72% | 98,20% |
| F1-score | 96,22% | 96,93% | 96,25% |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent