Added makeModel call to R/H2O to allow for hand made glm models with …

…efficient scoring.

R/h2o-package/R/Algorithms.R

@@ -1396,6 +1396,14 @@ h2o.predict <- function(object, newdata, ...) {
     stop(paste("Prediction has not yet been implemented for", class(object)))
 }
 
+
+h2o.makeGLMModel <- function(model, beta) {
+  if( missing(model) || class(model) != "H2OGLMModel") 
+    stop('Must specify source glm model')      
+   res = .h2o.__remoteSend(model@data@h2o, .h2o.__GLMMakeModel, model=model@key, names = paste(names(beta),sep=","), beta = as.vector(beta))   
+   .h2o.get.glm(model@data@h2o, as.character(res$destination_key), FALSE)
+}
+
 h2o.confusionMatrix <- function(data, reference) {
   if(class(data) != "H2OParsedData") stop("data must be an H2O parsed dataset")
   if(class(reference) != "H2OParsedData") stop("reference must be an H2O parsed dataset")

R/h2o-package/R/Internal.R

@@ -179,6 +179,8 @@ h2o.setLogPath <- function(path, type) {
 .h2o.__PAGE_RemoveVec = "2/RemoveVec.json"
 .h2o.__PAGE_Order = "2/Order.json"
 
+.h2o.__GLMMakeModel = "2/GLMMakeModel.json"
+
 # client -- Connection object returned from h2o.init().
 # page   -- URL to access within the H2O server.
 # parms  -- List of parameters to send to the server.

R/h2o-package/R/models.R

@@ -195,9 +195,9 @@ function(h2o, key, lambda_idx = -1, return_all_lambda = TRUE, pre = "", data = N
   names(result$coefficients) <- pre$json$glm_model$coefficients_names[idxes]
   if (!is.null(result$normalized_coefficients))
     names(result$normalized_coefficients) <- pre$json$glm_model$coefficients_names[idxes]
-  if(.isBinomial(pre)) {  # build and set the confusion matrix
+  if(.isBinomial(pre) && !is.null(valid) && !is.null(valid$cms)) {  # build and set the confusion matrix
     cm_ind <- trunc(100*result$best_threshold) + 1
-    if ( trunc(100 * result$best_threshold) + 1 > length(valid$cms)) {
+    if (trunc(100 * result$best_threshold) + 1 > length(valid$cms)) {
       threshs <- pre$json$glm_model$submodels[[lambda_idx]]$validation$thresholds
       cm_ind <- which(threshs == result$best_threshold)
     }

src/main/java/hex/glm/GLMModel.java

@@ -52,6 +52,7 @@ public class GLMModel extends Model implements Comparable<GLMModel> {
   @API(help="index of lambda_value giving best results")
   int best_lambda_idx;
 
+  public DataInfo dinfo(){return data_info;}
   public Key [] xvalModels() {
     if(submodels == null)return null;
     for(Submodel sm:submodels)
@@ -151,9 +152,7 @@ public int compareTo(GLMModel m){
   // fully expanded beta used for scoring
   private double [] global_beta;
 
-  public void setBestSubmodel(double lambda){
 
-  }
 
   @API(help="Validation of the null model")
   public GLMValidation null_validation;
@@ -220,6 +219,25 @@ public Submodel(double lambda , double [] beta, double [] norm_beta, long run_ti
   @API(help = "Variable importances", json=true)
   VarImp variable_importances;
 
+  public GLMModel(GLM2 parameters, Key selfKey, Key dataKey, GLMParams glmp, String [] coefficients_names, double [] beta, DataInfo dinfo, double threshold) {
+    super(selfKey,dataKey,dinfo._adaptedFrame,null);
+    this.parameters = (GLM2)parameters.clone();
+    submodels = new Submodel[]{new Submodel(0,beta,null,-1,-1,false)};
+    this.coefficients_names = coefficients_names;
+    alpha = 0;
+    lambda_max = Double.NaN;
+    this.threshold = threshold;
+    useAllFactorLevels = dinfo._useAllFactorLevels;
+    global_beta = submodels[0].beta.clone();
+    this.glm = glmp;
+    this.ymu = Double.NaN;
+    this.prior = Double.NaN;
+    this.warnings = new String[]{"Hand made model."};
+    this.data_info = dinfo;
+    this.beta_eps = Double.NaN;
+    this.job_key = null;
+    best_lambda_idx = 0;
+  }
   public GLMModel(GLM2 job, Key selfKey, DataInfo dinfo, GLMParams glm, GLMValidation nullVal, double beta_eps, double alpha, double lambda_max, double ymu, double prior) {
     super(selfKey,job.source._key == null ? dinfo._frameKey : job.source._key,dinfo._adaptedFrame, /* priorClassDistribution */ null);
     parameters = Job.hygiene((GLM2) job.clone());

src/main/java/water/api/RequestServer.java

@@ -105,7 +105,7 @@ public enum API_VERSION {
     // Register Inspect2 just for viewing frames
     registerRequest(new Inspect2());
     registerRequest(new MMStats());
-
+    registerRequest(new GLMMakeModel());
     // FVec models
     Request.addToNavbar(registerRequest(new DeepLearning()),"Deep Learning",                   "Model");
     Request.addToNavbar(registerRequest(new GLM2()),        "Generalized Linear Model",        "Model");