PyTorch-RetinaNet
Train RetinaNet with Focal Loss in PyTorch.
Reference: [1] Focal Loss for Dense Object Detection
Installation
Note that load_checkpoint.sh requires you to be on the UW CSE network (a VPN suffices).
cd YOUR_CATKIN_WS/src
git clone https://github.com/personalrobotics/pytorch_retinanet.git
cd ./pytorch_retinanet
./load_checkpoint.sh
catkin build pytorch_retinanet
source $(catkin locate)/devel/setup.bash
Preparing a dataset
Make a symlink or put your dataset for training in data directory. RetinaNet currently supports annotations in xml format.
For example, if your dataset name is sample_data, all the images should be in data/sample_data/images, and annotations should be saved in data/sample_data/annotations/xmls. Then, define a label map (id - name pairs) in data/sample_data/sample_label_map.pbtxt.
Please check the data/sample_data for actual exmaples. When the images, xmls, and label_map are ready, run
cd ./data/sample_data
generate_listdata.py sample
in the sample directory to generate sample_ann_train.txt and sample_ann_test.txt. These two files will be used ListDataset (from data/listdataset.py).
Train the model
Change the src/pytorch_retinanet/config/config.py according to your machine environment and the dataset location.
Build and run the training script:
catkin build pytorch_retinanet
source $(catkin locate)/devel/setup.bash
cd ./script
./train.py