This repository contains our 3 networks we worked with:
- The CSLNet according to Laina et al.
- A MaskRCNN network using the Detectron2 framework
- A Combined network using the Detectron2 framework which combines the MaskRCNN network and the CSLNet by implementing a custom CSL head
Run the CSL network with python csl/run.py --argument2 --argument2 .... The --command argument has to be
init, train or call and is always required.
Initializes the CSL model for further training. It basically loads the model with pretrained encoder weights and outputs
a csl.pth file containing the model. With the --workspace argument you can specify the output directory for that file.
Starts training the CSL model. Possible arguments are:
--workspace Workspace directory containing the .pth model.
--dataset Directory containing the annotated data (.json files)
--segloss The segmentation loss. Either ce or dice
--normalize Flag. If set, normalize gaussian kernels on ground truth heatmaps to [0-1]
--batch The batch size
--lambdah The lambda value according to the paper. Weighing parameter between segmentation and localisation
--sigma The sigma value for the gaussian kernels applied to the ground truth heatmaps
--learningrate The learning rate
Does inference on a trained model. Specify the dataset with --dataset and the directory containing the .pth file with
--workspace
Before you can use the MaskRCNN network you have to register the dataset properly. The dataset directory has to contain
the images in .png format and their corresponding annotations in .json files. To register it, run:
python detectron2_commons/register_dataset.py --dataset [path/to/dataset/] --output [output/directory]. This generates
.json file describing the dataset.
Now you can run the network with python maskrcnn/detectron_run.py --config [path/to/config] --dataset[path/to/dataset] [--train]. The config file is a .yaml file according to the Detectron2 standard (like in maskrcnn/configs). The
dataset directory has to contain the previously generated .json file. The --train specifies whether to train or not.
Just as with the MaskRCNN network you have to register the dataset with the register_dataset.py script. Then start
the network with python combined/combined_run.py .... The arguments are identical to the MaskRCNN network.