Fix explain into when shap values is not list (#2853)

* fix explain into when shap values is not list

* update

* add e2e test

* update

* update

* clean

Author: typhoonzero

go/cmd/sqlflowserver/e2e_pai_maxcompute_test.go

@@ -378,16 +378,16 @@ func CasePAIMaxComputeTrainXGBoost(t *testing.T) {
 	a := assert.New(t)
 
 	trainSQL := fmt.Sprintf(`SELECT * FROM %s
-	TO TRAIN xgboost.gbtree
-	WITH
-		objective="multi:softprob",
-		train.num_boost_round = 30,
-		eta = 0.4,
-		num_class = 3,
-		train.batch_size=10,
-		validation.select="select * from %s"
-	LABEL class
-	INTO e2etest_xgb_classi_model;`, caseTrainTable, caseTrainTable)
+TO TRAIN xgboost.gbtree
+WITH
+	objective="multi:softprob",
+	train.num_boost_round = 30,
+	eta = 0.4,
+	num_class = 3,
+	train.batch_size=10,
+	validation.select="select * from %s"
+LABEL class
+INTO e2etest_xgb_classi_model;`, caseTrainTable, caseTrainTable)
 	_, _, _, err := connectAndRunSQL(trainSQL)
 	a.NoError(err, "Run trainSQL error.")
 
@@ -405,13 +405,20 @@ INTO %s.e2etest_xgb_evaluate_result;`, caseTestTable, caseDB)
 	_, _, _, err = connectAndRunSQL(evalSQL)
 	a.NoError(err, "Run evalSQL error.")
 
-	explainSQL := fmt.Sprintf(`SELECT * FROM %s
-TO EXPLAIN e2etest_xgb_classi_model
-WITH label_col=class
-USING TreeExplainer
-INTO %s.e2etest_xgb_explain_result;`, caseTrainTable, caseDB)
-	_, _, _, err = connectAndRunSQL(explainSQL)
-	a.NoError(err, "Run explainSQL error.")
+	titanicTrain := fmt.Sprintf(`SELECT * FROM %s.sqlflow_titanic_train
+TO TRAIN xgboost.gbtree
+WITH objective="binary:logistic"
+LABEL survived
+INTO e2etest_xgb_titanic;`, caseDB)
+	_, _, _, err = connectAndRunSQL(titanicTrain)
+	a.NoError(err, "Run titanicTrain error.")
+
+	titanicExplain := fmt.Sprintf(`SELECT * FROM %s.sqlflow_titanic_train
+TO EXPLAIN e2etest_xgb_titanic
+WITH label_col=survived
+INTO %s.e2etest_titanic_explain_result;`, caseDB, caseDB)
+	_, _, _, err = connectAndRunSQL(titanicExplain)
+	a.NoError(err, "Run titanicExplain error.")
 }
 
 func CasePAIMaxComputeTrainCustomModel(t *testing.T) {

go/codegen/xgboost/codegen_explain.go

@@ -57,6 +57,7 @@ func Explain(explainStmt *ir.ExplainStmt, session *pb.Session) (string, error) {
 		FeatureColumnNames:   fs,
 		FeatureColumnCode:    featureColumnCode,
 		LabelJSON:            string(l),
+		ResultTable:          explainStmt.Into,
 		IsPAI:                tf.IsPAI(),
 		PAIExplainTable:      explainStmt.TmpExplainTable,
 	}

go/codegen/xgboost/template_explain.go

@@ -25,6 +25,7 @@ type explainFiller struct {
 	FeatureColumnNames   []string
 	FeatureColumnCode    string
 	LabelJSON            string
+	ResultTable          string
 	IsPAI                bool
 	PAIExplainTable      string
 }
@@ -48,10 +49,11 @@ transform_fn = xgboost_extended.feature_column.ComposedColumnTransformer(feature
 explain(
     datasource='''{{.DataSource}}''',
     select='''{{.DatasetSQL}}''',
-	feature_field_meta=feature_field_meta,
-	feature_column_names=feature_column_names,
+    feature_field_meta=feature_field_meta,
+    feature_column_names=feature_column_names,
     label_meta=label_meta,
     summary_params=summary_params,
+    result_table="{{.ResultTable}}",
     is_pai="{{.IsPAI}}" == "true",
     pai_explain_table="{{.PAIExplainTable}}",
     transform_fn=transform_fn,

go/executor/executor.go

@@ -272,16 +272,20 @@ func (s *pythonExecutor) ExecuteExplain(cl *ir.ExplainStmt) error {
 		return err
 	}
 	defer db.Close()
+
+	var modelType int
 	if cl.TrainStmt.GetModelKind() == ir.XGBoost {
 		code, err = xgboost.Explain(cl, s.Session)
-		// TODO(typhoonzero): deal with XGBoost model explain result table creation.
+		modelType = pai.ModelTypeXGBoost
 	} else {
 		code, err = tensorflow.Explain(cl, s.Session)
-		if cl.Into != "" {
-			err := createExplainResultTable(db, cl, cl.Into, pai.ModelTypeTF, cl.TrainStmt.Estimator)
-			if err != nil {
-				return err
-			}
+		modelType = pai.ModelTypeTF
+	}
+
+	if cl.Into != "" {
+		err := createExplainResultTable(db, cl, cl.Into, modelType, cl.TrainStmt.Estimator)
+		if err != nil {
+			return err
 		}
 	}
 

go/executor/pai.go

@@ -23,9 +23,10 @@ import (
 	"path"
 	"path/filepath"
 	"regexp"
-	"sqlflow.org/sqlflow/go/verifier"
 	"strings"
 
+	"sqlflow.org/sqlflow/go/verifier"
+
 	"sqlflow.org/sqlflow/go/codegen/optimize"
 
 	"github.com/aliyun/aliyun-oss-go-sdk/oss"
@@ -636,9 +637,13 @@ func getCreateShapResultSQL(db *database.DB, tableName string, selectStmt string
 		return "", err
 	}
 	columnDefList := []string{}
+	columnType := "STRING"
+	if db.DriverName == "mysql" {
+		columnType = "VARCHAR(255)"
+	}
 	for _, fieldName := range flds {
 		if fieldName != labelCol {
-			columnDefList = append(columnDefList, fmt.Sprintf("%s STRING", fieldName))
+			columnDefList = append(columnDefList, fmt.Sprintf("%s %s", fieldName, columnType))
 		}
 	}
 	createStmt := fmt.Sprintf(`CREATE TABLE IF NOT EXISTS %s (%s);`, tableName, strings.Join(columnDefList, ","))

python/runtime/xgboost/dataset.py

@@ -252,20 +252,15 @@ def pai_dataset(filename,
     from subprocess import Popen, PIPE
     from multiprocessing.dummy import Pool  # ThreadPool
     import queue
-
     dname = filename
     if single_file:
         dname = filename + '.dir'
-
     if os.path.exists(dname):
         shutil.rmtree(dname, ignore_errors=True)
 
     os.mkdir(dname)
-
     slice_count = get_pai_table_slice_count(pai_table, nworkers, batch_size)
-
     thread_num = min(int(slice_count / nworkers), 128)
-
     pool = Pool(thread_num)
     complete_queue = queue.Queue()
 
@@ -337,7 +332,7 @@ def pai_download_table_data_worker(dname, feature_metas, feature_column_names,
         feature_column_names, *feature_column_transformers)
 
     conn = PaiIOConnection.from_table(pai_table, slice_id, slice_count)
-    gen = db.db_generator(conn, None)()
+    gen = db.db_generator(conn, None, label_meta=label_meta)()
     selected_cols = db.selected_cols(conn, None)
     filename = "{}/{}.txt".format(dname, slice_id)
     dump_dmatrix(filename,

python/runtime/xgboost/explain.py

@@ -173,23 +173,23 @@ def explain(datasource,
                          pai_explain_table,
                          transform_fn=transform_fn,
                          feature_column_code=feature_column_code)
-
     shap_values, shap_interaction_values, expected_value = xgb_shap_values(x)
-
     if result_table != "":
         if is_pai:
             from runtime.dbapi.paiio import PaiIOConnection
             conn = PaiIOConnection.from_table(result_table)
-            # TODO(typhoonzero): the shape of shap_values is
-            # (3, num_samples, num_features), use the first
-            # dimension here, should find out how to use
-            # the other two.
         else:
             conn = db.connect_with_data_source(datasource)
-
-        write_shap_values(shap_values[0], conn, result_table,
-                          feature_column_names)
-        return
+        # TODO(typhoonzero): the shap_values is may be a
+        # list of shape [3, num_samples, num_features],
+        # use the first dimension here, should find out
+        # when to use the other two. When shap_values is
+        # not a list it can be directly used.
+        if isinstance(shap_values, list):
+            to_write = shap_values[0]
+        else:
+            to_write = shap_values
+        write_shap_values(to_write, conn, result_table, feature_column_names)
 
     if summary_params.get("plot_type") == "decision":
         explainer.plot_and_save(