We are using H2O version 3.14.0.3 with Flow (v. 0.7.7) on a Hadoop Cluster with three Nodes. In the flow web interface there are some model defaults (especially for Deep Learning) we adjust every time (e.g. lowering Epochs to 0.1). If we do not adjust the parameter the training takes hours due to the volume of our datasets.
Would I be able to configure the default parameter settings?
Those are the default hyperparameter values for H2O algorithms which are hardcoded in the Java code and propagated through all the APIs (R, Python, Scala, and Flow). The only way to change the default values is to set them manually in the client interface (like you are already doing), or you can fork the H2O source code, change them in Java, re-compile H2O and use your forked version with the modified settings.
You might be able to save a flow file with those settings and re-use that for your experiments. This way you'd only need to change the value for the data file (single change) vs many changes to various hyperparameter values.
Lastly, if you are comfortable with R or Python, you could very easily write a script that has your preferred settings and use that instead. If you want to use some existing R or Python Deep Learning code, there are some examples here.
Collected from the Internet
Please contact [email protected] to delete if infringement.
Comments