kaknikhil commented on a change in pull request #451: DL: Update evaluate and predict for multi model outputs URL: https://github.com/apache/madlib/pull/451#discussion_r338189265
########## File path: src/ports/postgres/modules/deep_learning/madlib_keras_fit_multiple_model.py_in ########## @@ -47,6 +47,27 @@ mb_dep_var_col = MINIBATCH_OUTPUT_DEPENDENT_COLNAME_DL mb_indep_var_col = MINIBATCH_OUTPUT_INDEPENDENT_COLNAME_DL dist_key_col = DISTRIBUTION_KEY_COLNAME +""" +FitMultipleModel: This class implements the Model Hopper technique for +training multiple models in parallel. The goal of this function is to train +multiple different models on the same data with different compile parameters. +The main advantage of this method over running the existing fit function in a +loop is avoiding inaccuracies caused by the model averaging. The basic idea of +model hopper is simple. Let's assume that there are n segments and c*n model +configurations. We begin with distributing these configs to segments. After +that, each segment trains their c models on the data they have locally for one +iteration. Once we have these models, we move them to a different segment +(hopping) as well as receive a different set of models. Once we have the new +models, we use our segments data to refine them (similar to the warm start +functionality). Once every model hops through every segment, we consider an +iteration complete. + +This method ensures that we don't have to average any model and the Review comment: There are a couple of things to add to this comment but I will take care of that in a future PR. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
