Capture intermediate results during optimization

bhupendra_patilbhupendra_patil Administrator, Employee, Member Posts: 168  RM Data Scientist
edited November 2018 in Knowledge Base

Rapidminer provides various optimization techniques including for Paramter Optimization, Feature Selection & Feature Generation. Rapidminer automatically delivers the best models, best parameter sets as output of these operators in most cases, however sometime the user may want to capture the intermediatary results that were obtained after every iteration. This is easily possible with the Log operator.

 

In case you are looking to actuall save intermediate models and results look at the discussion here http://community.rapidminer.com/t5/RapidMiner-Studio/Backup-and-retrieve-modeling-Training-status/m-p/32263#M22981 

 

 

The following example shows how to capture several key parameters during "Parameter Optimization" 

The overall process is attached and the top level Process looks like below 

optimization logs.png

The optimization operator is trying various combinations of C & gamma values, for the SVM learner as you will see in the settings by clicking on "Edit Parameter Settings"

optimization logs inside.png

As you notice we are spliting the data sets into two parts (50%) each and then building model on one and testing on the other.

 

The key operator where we capture results at end of every combination is in the log operator. By clicking on the "Edit List" the user can specify what metrics to capture.

log settings.png

Once the process is completed, you should see an output tab called as Log with results like below. These are the intermediate results after every iterations.

log output.png

 

As you will notice on the log operator you can also write these outputs to a file, by specifying the "filename" parameter. 

Additonally if you need to store intermediate models or performance vectors you can also save them as needed using the store operator and dynamically changing paths using macros.

 

Sign In or Register to comment.