readme.md

graph-extraction-networks

Extraction of graphs from skeletonised images — a TensorFlow implementation.

---

Related: data generation, sample GUI implementation

The graph extraction consists of a node extraction part, an edge extraction part and a combination scheme for the edge predictions.

Node Extraction

The nodes as well as their attributes are extracted using a model based on the U-Net.

Positions	Degrees	Types
	0, 1, 2, 3, 4	end, crossing, border

Current best model: b_16_d2 [id: pqphq89g] (wandb) (yaml) (h5)

Training

1. Define configuration:
   • Dataset config -- change data_path and validation_fraction as needed.
   • Model config in configs, e.g.sample_nodes_nn.yaml.
2. Set config filepath in nodes_nn.py (L7)
3. Run nodes_nn.py.

   nohup python nodes_nn.py > nodes_nn_log.txt &
   

Testing

1. Initialise variables:
   • Set network type to NetworkType.NODES_NN in evaluate.py (L5).
   • Populate model_ids list with the IDs of the model(s) to be tested in evaluate.py (L6).
   • Ensure that eval_nodes_nn.yaml is set in the second argument of run.get_configs in evaluate.py (L11).
2. Model configuration and weights — ensure that these files exist:
   • Model config yaml file in configs, e.g.sample_nodes_nn.yaml.
   • Model weights in the path used in evaluate.py (L19).
     Default path: wandb/{model_id}.h5.
3. Choose the type of evaluation to be done, modify/comment out what's not needed:
   • Visualise batch predictions (L31 - L33)
   • Evaluate metrics per test image (L35 - L41)
4. Run evaluate.py.

Edge Extraction

Prediction of the existence of an edge between a pair of nodes using a VGG-based model.

Current best model: baseline (wandb) (yaml) (h5)

Combination Scheme

Parallel batched edge predictions.