Output Documentation

DANNCE training and predictions produces several different output files, which are explained in more detail here.

com-train. Produces outputs in the com_train_dir indicated in your io.yaml file (Default: ./COM/train_results.

• logs (directory). Directory of tensorboard log files. To visualize your loss curves, launch tensorboard --logdir=./logs/ from the terminal when inside ./COM/train_results.
• fullmodel_weights (directory). Directory containing a full model file at the end of training (i.e. architecture + weights + optimizer state).
• copy_params.mat. File containing the values of all configuration parameters.
• training.csv. File containing the loss and metric values for each epoch.
• train_samples.pickle. File containing references to the training samples used.
• val_samples.pickle. File containing references to the validation samples used. If this file is referenced in the load_valid argument in a subsequent training session, com-train will use these validation samples for validation. Alternatively, the data_split_seed can be specified as a configuration option to ensure validation reproducibility.
• weights.{epoch}-{loss}.hdf5. Network weights for the indicated epochs.

com-predict. Produces outputs in the com_predict_dir indicated in your io.yaml file (Default: ./COM/predict_results.

• com3d.mat. File containing 3D COM coordinates, sampleIDs, and associated metadata.
• com3d.pickle. An expanded file containing individual 2D COM predictions in addition to 3D triangulation. If using this file in dannce, an expanded set of COM operations are enabled (e.g. filtering COM based on network confidence).