Contains hyper-parameter configurations for models.
Hyper-parameter search-space is specificed in /config/config.py
. Default values tuned during paper experiments are defined in /config/<dataset>/<outcome>/<domain>/config_<model>_<strategy>.json
.
Individual experiments can be specified with a combination of --domain_shift
and --outcome
parameters. A subset of models and Continual learning strategies can be evaluated with --models
and --strategies
respectively. To re-run hyperparameter tuning pass the --validate
flag.
Example:
python main.py --domain_shift hospital --outcome mortality_48h --models CNN --strategies EWC Replay
Flag | Arg(s) | Meaning |
---|---|---|
--domain_shift |
region hospital age ethnicity |
Domain shift exhibited between tasks |
--outcome |
mortality_48h Shock_4h Shock_12h ARF_4h ARF_12h |
Outcome to predict |
--models |
MLP CNN RNN LSTM GRU Transformer |
Model(s) to evaluate |
--strategies |
Naive Cumulative EWC OnlineEWC LwF SI GEM AGEM Replay GDumb |
Continual learning strategy(s) to evaluate |
--validate |
Re-tune hyper-parameters | |
--num_samples |
<int> |
Budget for hyper-parameter search |