RLForSeg

Reinforcement learning for Instance Segmentation implementing the concepts discussed here

Installation via conda:

conda env create -f environment.yml

Initiating the training:

python run.py -d </path/to/log/dir> -p <wandb_project_name> -e <wandb_user> -c <conf/file.yaml>

Input data

Train- as well as validation-data is expected to be in separate directories, each containing a set of h5 files. Each file needs the following entries:

- input data (N, H, W) for the feature extractor. 
  Typically the raw image + additional channles 
- A superpixel segmentation, where the n superpixel labels have to be 
  consecutive integers, starting at 0.
- a set of edges (2, E) defining the superpixel graph where 
  each entry corresponds to a superpixel label
optional:
- a ground truth segmentation
- a set of ground truth edge weights (E)
- a set of edge features (E, f) as additional input to the GNN
- a set of node features (n, f) as additional input to the GNN

Custom reward function:

A custom reward function can be implemented in /rewards. It has to be subclassed from RewardFunctionAbc in /rewards/reward_abc.py and implement its two functions. An example for the circles dataset can be found in /rewards/circles_reward.py. The reward class needs to be imported in /rewards/__init__.py. The class name can the be referenced in the confg file. A config file example can be found in /conf/example.yaml.