Stop deepLeaning learning stage on conditions
Most of our deepLearning curves looks like :
loss | metrics |
they are raising until a few epochs then become more stable. As every epochs are time consuming, the purpose of this feature is to allow the user to stop the learning stage regarding some conditions which might be :
-
Stop if during an interval of epochs (which can be a configuration file input), the parameter (validation loss, train loss, kappa, fscore or oa) the evolution is under a threshold.
-
Keep the learning stage going as long as performance are increasing. Stop if the derivative become < 0 (across an epoch interval).
-
...
feel free to add some condition to stop the learning step.