Storage of best models

Location

The best trained models, for the main task and the different sub-tasks, will be stored in /data/lisa/exp/best_models/emotiw/, in a sub-directory indicating which task it was trained on. For instance, a model trained for emotion recognition on static faces will be stored in /data/lisa/exp/best_models/emotiw/static_faces.

Keep only the best trained model for an architecture, of a run of experiments here. Do not save all the models trained during hyper-parameter optimization there.

Naming scheme

The saved model should follow this naming convention: <date (YYYY.mm.dd)>_<username>_<description>.<extension>.

• The date when the model was trained, for instance 2013.04.01, so we have an idea of the chronology;
• The (DIRO) username of the person having run the experiment (not necessarily the developer of the model);
• Some description of the model;
• (optionally) an extension indicating the file type.

If your model is contained in several files, these files should be in a directory (or compressed archive) that follows that naming convention.

Model description

Each of the saved models should come with a text file with the same name (<date (YYYY.mm.dd)>_<username>_<description>.txt) containing the values for the hyper-parameters used in that model, as well as all the other useful information.

For instance, it could contain the path to a YAML file that was used to train the model, as well as the command line used to start the training.