Support SHOW MODEL FEATURE DETAILS

Signed-off-by: Misiu Godfrey <misiu.godfrey@kraken.mapd.com>
SINHASantos · Sep 1, 2023 · 5d05a5f · 5d05a5f
1 parent 226faca
commit 5d05a5f
Show file tree

Hide file tree

Showing 8 changed files with 444 additions and 53 deletions.
diff --git a/Catalog/DdlCommandExecutor.cpp b/Catalog/DdlCommandExecutor.cpp
@@ -661,6 +661,9 @@ ExecutionResult DdlCommandExecutor::execute(bool read_only_mode) {
     result = ShowModelsCommand{*ddl_data_, session_ptr_}.execute(read_only_mode);
   } else if (ddl_command_ == "SHOW_MODEL_DETAILS") {
     result = ShowModelDetailsCommand{*ddl_data_, session_ptr_}.execute(read_only_mode);
+  } else if (ddl_command_ == "SHOW_MODEL_FEATURE_DETAILS") {
+    result =
+        ShowModelFeatureDetailsCommand{*ddl_data_, session_ptr_}.execute(read_only_mode);
   } else if (ddl_command_ == "ALTER_SERVER") {
     result = AlterForeignServerCommand{*ddl_data_, session_ptr_}.execute(read_only_mode);
   } else if (ddl_command_ == "ALTER_DATABASE") {
@@ -2091,7 +2094,6 @@ ShowModelDetailsCommand::ShowModelDetailsCommand(
 ExecutionResult ShowModelDetailsCommand::execute(bool read_only_mode) {
   auto execute_read_lock = legacylockmgr::getExecuteReadLock();
 
-  std::vector<std::string> labels{"model_name", "model_type", "training_query"};
   std::vector<TargetMetaInfo> label_infos;
   label_infos.emplace_back("model_name", SQLTypeInfo(kTEXT, true));
   label_infos.emplace_back("model_type", SQLTypeInfo(kTEXT, true));
@@ -2171,6 +2173,134 @@ std::vector<std::string> ShowModelDetailsCommand::getFilteredModelNames() {
   }
 }
 
+ShowModelFeatureDetailsCommand::ShowModelFeatureDetailsCommand(
+    const DdlCommandData& ddl_data,
+    std::shared_ptr<Catalog_Namespace::SessionInfo const> session_ptr)
+    : DdlCommand(ddl_data, session_ptr) {
+  if (!g_enable_ml_functions) {
+    throw std::runtime_error(
+        "Cannot show model feature details. ML functions are disabled.");
+  }
+  if (g_restrict_ml_model_metadata_to_superusers) {
+    // Check if user is super user
+    const auto& current_user = session_ptr->get_currentUser();
+    if (!current_user.isSuper) {
+      throw std::runtime_error(
+          "Cannot show model feature details. Showing model information to "
+          "non-superusers is "
+          "disabled.");
+    }
+  }
+}
+
+ExecutionResult ShowModelFeatureDetailsCommand::execute(bool read_only_mode) {
+  auto execute_read_lock = legacylockmgr::getExecuteReadLock();
+  auto& ddl_payload = extractPayload(ddl_data_);
+  CHECK(ddl_payload.HasMember("modelName")) << "Model name missing.";
+  const auto model_name = ddl_payload["modelName"].GetString();
+  const auto model_metadata = g_ml_models.getModelMetadata(model_name);
+  const auto model_type = model_metadata.getModelType();
+  std::vector<TargetMetaInfo> label_infos;
+  label_infos.emplace_back("feature_id", SQLTypeInfo(kBIGINT, true));
+  label_infos.emplace_back("feature", SQLTypeInfo(kTEXT, true));
+  label_infos.emplace_back("sub_feature_id", SQLTypeInfo(kBIGINT, true));
+  label_infos.emplace_back("sub_feature", SQLTypeInfo(kTEXT, true));
+  std::vector<double> extra_metadata;
+  std::vector<std::vector<double>> eigenvectors;
+  auto features = model_metadata.getFeatures();
+  const auto model = g_ml_models.getModel(model_name);
+  auto cat_sub_features = model->getCatFeatureKeys();
+  switch (model_type) {
+    case MLModelType::LINEAR_REG: {
+      label_infos.emplace_back("coefficient", SQLTypeInfo(kDOUBLE, true));
+      const auto linear_reg_model =
+          std::dynamic_pointer_cast<LinearRegressionModel>(model);
+      extra_metadata = linear_reg_model->getCoefs();
+      break;
+    }
+#ifdef HAVE_ONEDAL
+    case MLModelType::RANDOM_FOREST_REG: {
+      const auto random_forest_reg_model =
+          std::dynamic_pointer_cast<RandomForestRegressionModel>(model);
+      extra_metadata = random_forest_reg_model->getVariableImportanceScores();
+      if (!extra_metadata.empty()) {
+        label_infos.emplace_back("feature_importance", SQLTypeInfo(kDOUBLE, true));
+      }
+      break;
+    }
+    case MLModelType::PCA: {
+      label_infos.emplace_back("eigenvalue", SQLTypeInfo(kDOUBLE, true));
+      label_infos.emplace_back("eigenvector", SQLTypeInfo(kTEXT, true));
+      const auto pca_model = std::dynamic_pointer_cast<PcaModel>(model);
+      extra_metadata = pca_model->getEigenvalues();
+      eigenvectors = pca_model->getEigenvectors();
+      CHECK_EQ(eigenvectors.size(), extra_metadata.size());
+
+      break;
+    }
+#endif  // HAVE_ONEDAL
+    default: {
+      break;
+    }
+  }
+  const int64_t num_features = static_cast<int64_t>(features.size());
+  std::vector<RelLogicalValues::RowValues> logical_values;
+  if (model_type == MLModelType::LINEAR_REG) {
+    logical_values.emplace_back(RelLogicalValues::RowValues{});
+    logical_values.back().emplace_back(genLiteralBigInt(0));
+    logical_values.back().emplace_back(genLiteralStr("intercept"));
+    logical_values.back().emplace_back(genLiteralBigInt(1));
+    logical_values.back().emplace_back(genLiteralStr(""));
+    logical_values.back().emplace_back(genLiteralDouble(extra_metadata[0]));
+    extra_metadata.erase(extra_metadata.begin());
+  }
+  int64_t physical_feature_idx = 0;
+  for (int64_t feature_idx = 0; feature_idx < num_features; ++feature_idx) {
+    int64_t num_sub_features =
+        feature_idx >= static_cast<int64_t>(cat_sub_features.size())
+            ? 0
+            : static_cast<int64_t>(cat_sub_features[feature_idx].size());
+    const bool has_sub_features = num_sub_features > 0;
+    num_sub_features = num_sub_features == 0 ? 1 : num_sub_features;
+    for (int64_t sub_feature_idx = 0; sub_feature_idx < num_sub_features;
+         ++sub_feature_idx) {
+      logical_values.emplace_back(RelLogicalValues::RowValues{});
+      // Make feature id one-based
+      logical_values.back().emplace_back(genLiteralBigInt(feature_idx + 1));
+      logical_values.back().emplace_back(genLiteralStr(features[feature_idx]));
+      logical_values.back().emplace_back(genLiteralBigInt(sub_feature_idx + 1));
+      if (has_sub_features) {
+        logical_values.back().emplace_back(
+            genLiteralStr(cat_sub_features[feature_idx][sub_feature_idx]));
+      } else {
+        logical_values.back().emplace_back(genLiteralStr(""));
+      }
+      if (!extra_metadata.empty()) {
+        logical_values.back().emplace_back(
+            genLiteralDouble(extra_metadata[physical_feature_idx]));
+      }
+      if (!eigenvectors.empty()) {
+        std::ostringstream eigenvector_oss;
+        eigenvector_oss << "[";
+        for (size_t i = 0; i < eigenvectors[physical_feature_idx].size(); ++i) {
+          if (i > 0) {
+            eigenvector_oss << ", ";
+          }
+          eigenvector_oss << eigenvectors[physical_feature_idx][i];
+        }
+        eigenvector_oss << "]";
+        logical_values.back().emplace_back(genLiteralStr(eigenvector_oss.str()));
+      }
+      physical_feature_idx++;
+    }
+  }
+  // Create ResultSet
+  std::shared_ptr<ResultSet> rSet = std::shared_ptr<ResultSet>(
+      ResultSetLogicalValuesBuilder::create(label_infos, logical_values));
+
+  return ExecutionResult(rSet, label_infos);
+}
+
 EvaluateModelCommand::EvaluateModelCommand(
     const DdlCommandData& ddl_data,
     std::shared_ptr<Catalog_Namespace::SessionInfo const> session_ptr)

diff --git a/Catalog/DdlCommandExecutor.h b/Catalog/DdlCommandExecutor.h
@@ -289,6 +289,15 @@ class ShowModelDetailsCommand : public DdlCommand {
   std::vector<std::string> getFilteredModelNames();
 };
 
+class ShowModelFeatureDetailsCommand : public DdlCommand {
+ public:
+  ShowModelFeatureDetailsCommand(
+      const DdlCommandData& ddl_data,
+      std::shared_ptr<Catalog_Namespace::SessionInfo const> session_ptr);
+
+  ExecutionResult execute(bool read_only_mode) override;
+};
+
 class EvaluateModelCommand : public DdlCommand {
  public:
   EvaluateModelCommand(const DdlCommandData& ddl_data,

diff --git a/QueryEngine/TableFunctions/SystemFunctions/os/ML/MLModel.h b/QueryEngine/TableFunctions/SystemFunctions/os/ML/MLModel.h
@@ -327,6 +327,8 @@ class RandomForestRegressionModel : public virtual AbstractTreeModel {
   double out_of_bag_error_;
 };
 
+#endif  // #ifdef HAVE_ONEDAL
+
 class PcaModel : public AbstractMLModel {
  public:
   PcaModel(const std::vector<double>& col_means,
@@ -374,6 +376,4 @@ class PcaModel : public AbstractMLModel {
   std::vector<double> eigenvalues_;
 };
 
-#endif  // #ifdef HAVE_ONEDAL
-
 #endif  // #ifndef __CUDACC__
diff --git a/Tests/DBHandlerTestHelpers.h b/Tests/DBHandlerTestHelpers.h
@@ -712,6 +712,11 @@ class DBHandlerTestFixture : public TestHelpers::TbbPrivateServerKiller {
   size_t getRowCount(const TQueryResult& result) {
     return getRowCount(result.row_set);
   }
+
+  size_t getColumnCount(const TQueryResult& result) {
+    return getColumnCount(result.row_set);
+  }
+
   std::vector<TDatum> getRow(const TQueryResult& result, const size_t index) {
     return getRow(result.row_set, index);
   }