update

Author: typhoonzero

go/codegen/experimental/codegen.go

@@ -65,7 +65,8 @@ func parseToIR(sqlProgram string, session *pb.Session) ([]ir.SQLFlowStmt, error)
 	for _, sql := range sqls {
 		if sql.IsExtendedSyntax() {
 			if sql.Train {
-				r, err = ir.GenerateTrainStmt(sql.SQLFlowSelectStmt)
+				// TODO(typhoonzero): use feature derivation at runtime, call GenerateTrainStmt only.
+				r, err = ir.GenerateTrainStmtWithInferredColumns(sql.SQLFlowSelectStmt, session.DbConnStr, "", "", false, false)
 			} else if sql.ShowTrain {
 				r, err = ir.GenerateShowTrainStmt(sql.SQLFlowSelectStmt)
 			} else if sql.Explain {

go/codegen/experimental/codegen_test.go

@@ -21,9 +21,7 @@ import (
 )
 
 func TestXGBCodegen(t *testing.T) {
-	t.Skip("enable this test when finished generate code using runtime.feature.*")
-
-	sql := "SELECT * FROM iris TO TRAIN xgboost.gbtree WITH objective=\"binary:logistic\",num_class=3 LABEL class INTO sqlflow_models.xgb_classification;"
+	sql := "SELECT * FROM iris.train TO TRAIN xgboost.gbtree WITH objective=\"binary:logistic\",num_class=3 LABEL class INTO sqlflow_models.xgb_classification;"
 	s := &pb.Session{DbConnStr: database.GetTestingMySQLURL()}
 	_, err := GenerateCodeCouler(sql, s)
 	if err != nil {

go/codegen/experimental/xgboost.go

@@ -68,8 +68,9 @@ func XGBoostGenerateTrain(trainStmt *ir.TrainStmt, session *pb.Session) (string,
 		delete(params["train."], "num_workers")
 	}
 
+	// TODO(typhoonzero): use feature derivation at runtime.
 	if len(trainStmt.Features) != 1 {
-		return "", fmt.Errorf("xgboost only support 1 feature column set, received %d", len(trainStmt.Features))
+		return "", fmt.Errorf("xgboost only support 0 or 1 feature column set, received %d", len(trainStmt.Features))
 	}
 
 	featureColumnCode, featureFieldDesc, labelFieldDesc, err := deriveFeatureColumnCodeAndFieldDescs(trainStmt.Features["feature_columns"], trainStmt.Label)
@@ -128,16 +129,19 @@ def step_entry():
     feature_metas = json.loads('''{{.FieldDescJSON}}''')
     label_meta = json.loads('''{{.LabelJSON}}''')
 
-    feature_column_names = [{{range .FeatureColumnNames}}
-    "{{.}}",
-    {{end}}]
-
     ds = "{{.DataSource}}"
     is_pai = False
     pai_train_table = ""
     select = "{{.Select}}"
     val_select = "{{.ValidationSelect}}"
 
+    # Derive feature columns at runtime like:
+    # fcmap, fc_label = infer_feature_columns(conn, select, features, label, n=1000)
+
+    feature_column_names = [{{range .FeatureColumnNames}}
+    "{{.}}",
+    {{end}}]
+
     # NOTE: in the current implementation, we are generating a transform_fn from COLUMN clause. 
     # The transform_fn is executed during the process of dumping the original data into DMatrix SVM file.
     feature_column_list = [{{.FeatureColumnCode}}]