Requirements

References

Installation instructions

git clone https://github.com/delmalih/MIAS-mammography-obj-detection

1. Faster R-CNN instructions

conda create --name faster-r-cnnconda activate faster-r-cnnconda install ipython pipcd MIAS-mammography-obj-detectionpip install -r requirements.txtcd ..

# install pytorchconda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0 -c pytorchexport INSTALL_DIR=$PWD# install pycocotoolscd $INSTALL_DIRgit clone https://github.com/cocodataset/cocoapi.gitcd cocoapi/PythonAPIpython setup.py build_ext install# install cityscapesScriptscd $INSTALL_DIRgit clone https://github.com/mcordts/cityscapesScripts.gitcd cityscapesScripts/python setup.py build_ext install# install apexcd $INSTALL_DIRgit clone https://github.com/NVIDIA/apex.gitcd apexpython setup.py install --cuda_ext --cpp_ext# install PyTorch Detectioncd $INSTALL_DIRgit clone https://github.com/facebookresearch/maskrcnn-benchmark.gitcd maskrcnn-benchmarkpython setup.py build developcd $INSTALL_DIRunset INSTALL_DIR

2. RetinaNet instructions

conda create --name retinanet python=3.6conda activate retinanetconda install ipython pipcd MIAS-mammography-obj-detectionpip install -r requirements.txtcd ..pip install tensorflow-gpu==1.9pip install keras==2.2.5

# clone keras-retinanet repogit clone https://github.com/fizyr/keras-retinanetcd keras-retinanetpip install .python setup.py build_ext --inplace

3. FCOS instructions

conda create --name fcosconda activate fcosconda install ipython pipcd MIAS-mammography-obj-detectionpip install -r requirements.txtcd ..

How it works

1. Download the MIAS Database

mkdir mias-db && cd mias-dbwget http://peipa.essex.ac.uk/pix/mias/all-mias.tar.gztar -zxvf all-mias.tar.gzrm all-mias.tar.gz && cd ..

2. Generate COCO or VOC augmented data

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)>

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

3. How to run a training

3.1 Faster R-CNN

  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"    },  }

    # 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)

python train.py --config-file mias_config.yml

3.2 RetinaNet

cd retinanetconda deactivate && conda activate retinanetpython 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>

tensorboard --logdir <Path to the tensorboard directory>

3.3 FCOS

cd fcosconda deactivate && conda activate fcospython train.py --config-file <Path to the config file> \                OUTPUT_DIR <Path to the output dir for the logs>

4. How to run an inference

4.1 Faster R-CNN

cd faster-r-cnnconda deactivate && conda activate faster-r-cnnpython 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>

4.2 RetinaNet

cd retinanetconda deactivate && conda activate retinanetpython 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>

4.3 FCOS

cd fcosconda deactivate && conda activate fcospython 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>

Results

Poster : Breast Cancer Detection Contest