You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In conjunction with #3891 and #3890, a new problem appears:
If you use the concrete-class variant of specifying Updater (to enable you to set the momentum), you get the method setLearningRate(..). If using the Builder, you get the fluent variant of it as such:
However, this won't work, as it is the learningRate set directly on the NeuralNetConfiguration.Builder that is used. This defaults to 1e-1. So evidently the build employs the updater.setLearningRate(..), overwriting whatever the user has set on it.
Maybe a solution could be to not overwrite if the user has explicitly set it, i.e. use the idiom where it is default double.NaN, only overwrite if it still is NaN when building. Or something like that.
Also - as I've understood, the Updater can be layer-specific. This would then need to be handled for this case too - i.e. setting a specific learning rate on an Updater for a specific layer should not be overriden by the base config.
The text was updated successfully, but these errors were encountered:
In conjunction with #3891 and #3890, a new problem appears:
If you use the concrete-class variant of specifying Updater (to enable you to set the momentum), you get the method setLearningRate(..). If using the Builder, you get the fluent variant of it as such:
However, this won't work, as it is the learningRate set directly on the NeuralNetConfiguration.Builder that is used. This defaults to 1e-1. So evidently the build employs the updater.setLearningRate(..), overwriting whatever the user has set on it.
Maybe a solution could be to not overwrite if the user has explicitly set it, i.e. use the idiom where it is default double.NaN, only overwrite if it still is NaN when building. Or something like that.
Also - as I've understood, the Updater can be layer-specific. This would then need to be handled for this case too - i.e. setting a specific learning rate on an Updater for a specific layer should not be overriden by the base config.
The text was updated successfully, but these errors were encountered: