RetinaNet

An implementation of RetinaNet in PyTorch.

• Installation

• Training

• COCO 2017

• Pascal VOC

• Custom Dataset

• Evaluation

• Todo

• Credits

Installation

1. Install PyTorch and torchvision.

2. For faster data augmentation, install pillow-simd:

pip uninstall -y pillow
pip install pillow-simd

Training

COCO 2017

1. First, install pycocotools:

git clone https://github.com/pdollar/coco/cd coco/PythonAPI
make
python setup.py installcd ../..
rm -r coco

1. Then download COCO 2017 into ./datasets/COCO/:

cd datasets
mkdir COCOcd COCO

If your using wget:

wget http://images.cocodataset.org/zips/train2017.zip &&wget http://images.cocodataset.org/zips/val2017.zip &&wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip

If your using aria2c (recommended on for higher bandwidth connections and for allowing resumption of the download. Tune the number of max concurrent downloads (-j) and max connections per server (-x) as needed:

aria2c -x 10 -j 10 http://images.cocodataset.org/zips/train2017.zip &&aria2c -x 10 -j 10 http://images.cocodataset.org/zips/val2017.zip &&aria2c -x 10 -j 10 http://images.cocodataset.org/annotations/annotations_trainval2017.zip

unzip *.zip
rm *.zip

Then just run:

python train_coco.py

Pascal VOC

cd datasets
mkdir VOCcd VOC

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar &&wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar &&wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar

If your using aria2c (recommended on for higher bandwidth connections and for allowing resumption of the download. Tune the number of max concurrent downloads (-j) and max connections per server (-x) as needed:

aria2c -x 10 -j 10 http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar &&aria2c -x 10 -j 10 http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar &&aria2c -x 10 -j 10 http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar

tar xf *.tar
rm *.tar

Then just run:

python train_voc.py

Custom Dataset

Lots to write here. :wink:

Evaluation

To evaluate an image on a trained model:

python eval.py [checkpoint_path] [image_path]

This will create an image (output.jpg) with bounding box annotations.

Todo

1. Finish converting the COCO dataset class to work with batches.

2. Train COCO 2017 for 90,000 iterations and save a reusable checkpoint.

3. Try training on Pascal VOC and add download instructions.

4. Produce bounding box outputs for a few sanity check images.

5. Upload trained weights to Github releases.

6. Train on the magic proprietary dataset .

Credits