Merge pull request #233 from sunya-ch/exporter

add MAPE to estimator result and update exported figure

Author: rootfs

cmd/cmd_plot.py

@@ -134,7 +134,7 @@ def metadata_plot(args, energy_source, metadata_df, output_folder, name):
             ax = axes
         else:
             ax = axes[i]
-        sns.boxplot(data=metadata_df, x="feature_group", y="mae", hue="trainer", ax=ax, hue_order=sorted(metadata_df['trainer'].unique()), showfliers=False, palette="Set3")
+        sns.barplot(data=metadata_df, x="feature_group", y="mae", hue="trainer", ax=ax, hue_order=sorted(metadata_df['trainer'].unique()), errorbar=None, palette="Set3")
         ax.set_title(component)
         ax.set_ylabel("MAE (Watt)")
         ax.set_xlabel("Feature Group")
@@ -144,6 +144,7 @@ def metadata_plot(args, energy_source, metadata_df, output_folder, name):
        #  ax.legend(bbox_to_anchor=(1.05, 1.05))
     plt.suptitle("Pipieline metadata of {} {}".format(energy_source.upper(), args.output_type))
     plt.tight_layout()
+    plt.legend(frameon=False)
     filename = os.path.join(output_folder, name + ".png")
     fig.savefig(filename)
     plt.close()

cmd/main.py

@@ -547,10 +547,11 @@ def estimate(args):
                         label_power_columns = [col for col in power_labels if energy_component in col and col != predicted_power_colname]
                         sum_power_label = data.groupby([TIMESTAMP_COL]).mean()[label_power_columns].sum(axis=1).sort_index()
                         sum_predicted_power = data_with_prediction.groupby([TIMESTAMP_COL]).sum().sort_index()[predicted_power_colname]
-                        mae, mse = compute_error(sum_power_label, sum_predicted_power)
+                        mae, mse, mape = compute_error(sum_power_label, sum_predicted_power)
                         summary_item = dict()
                         summary_item["MAE"] = mae
                         summary_item["MSE"] = mse
+                        summary_item["MAPE"] = mape
                         summary_item["n"] = len(sum_predicted_power)
                         summary_item["energy_component"] = energy_component
                         summary_item["energy_source"] = energy_source

src/estimate/__init__.py

@@ -9,4 +9,4 @@
 from estimate_common import compute_error
 
 from model import load_model, get_background_containers
-from model import default_predicted_col_func, get_predicted_power_colname, get_predicted_background_power_colname, get_dynamic_power_colname, get_predicted_dynamic_power_colname, get_predicted_dynamic_background_power_colname, get_label_power_colname, get_reconstructed_power_colname
+from model import default_predicted_col_func, get_predicted_power_colname, get_predicted_background_power_colname, get_dynamic_power_colname, get_predicted_dynamic_power_colname, get_predicted_dynamic_background_power_colname, get_label_power_colname, get_reconstructed_power_colname, default_idle_predicted_col_func

src/estimate/model/estimate_common.py

@@ -1,5 +1,6 @@
 keras_enabled = True
 import cpuinfo
+import numpy as np
 
 cpu_info = cpuinfo.get_cpu_info()
 
@@ -61,7 +62,20 @@ def load_model_by_keras(model_path, model_filename):
 def load_model_by_json(model_path, model_filename):
     return load_json(model_path, model_filename)
 
+# return mae, mse, mape
 def compute_error(predicted_power, actual_powers):
     mse = mean_squared_error(actual_powers, predicted_power)
     mae = mean_absolute_error(actual_powers, predicted_power)
-    return mae, mse
+    actual_power_values = list(actual_powers)
+    predicted_power_values = list(predicted_power)
+    if len(actual_powers) == 0:
+        mape = -1
+    else:
+        non_zero_predicted_powers = np.array([predicted_power_values[i] for i in range(len(predicted_power_values)) if actual_power_values[i] > 0])
+        if len(non_zero_predicted_powers) == 0:
+            mape = -1
+        else:
+            non_zero_y_test = np.array([y for y in actual_powers if y > 0])
+            absolute_percentage_errors = np.abs((non_zero_y_test - non_zero_predicted_powers) / non_zero_y_test) * 100
+            mape = np.mean(absolute_percentage_errors)
+    return mae, mse, mape

src/estimate/model/model.py

@@ -30,6 +30,9 @@
 def default_predicted_col_func(energy_component):
     return "default_{}_power".format(energy_component)
 
+def default_idle_predicted_col_func(energy_component):
+    return "default_idle_{}_power".format(energy_component)
+
 def get_background_containers(idle_data):
     return pd.unique(idle_data[valid_container_query]["container_name"])
 
@@ -122,6 +125,12 @@ def append_prediction(self, data, predicted_col_func=default_predicted_col_func)
     def print_log(self, message):
         print("{} model: {}".format(self.model_name, message))
 
+    def append_idle_prediction(self, data, predicted_col_func=default_idle_predicted_col_func):
+        idle_data = data.copy() 
+        features = self.estimator.features
+        idle_data[features] = 0
+        return self.append_prediction(idle_data, predicted_col_func)
+    
 def load_model(model_path):
     metadata = load_metadata(model_path)
     if metadata is not None:

tests/estimator_model_test.py

@@ -35,7 +35,7 @@ def test_model(group_path, model_name, test_data_with_label, power_columns, powe
         predicted_power_colname = default_predicted_col_func(energy_component)
         sum_power_label = test_data_with_label.groupby([TIMESTAMP_COL]).mean()[label_power_columns].sum(axis=1).sort_index()
         sum_predicted_power = data_with_prediction.groupby([TIMESTAMP_COL]).sum().sort_index()[predicted_power_colname]
-        mae, mse = compute_error(sum_power_label, sum_predicted_power)
+        mae, mse, mape = compute_error(sum_power_label, sum_predicted_power)
         if power_range is None:
             power_range = sum_power_label.max() - sum_power_label.min()
         percent = mae/power_range
@@ -44,6 +44,7 @@ def test_model(group_path, model_name, test_data_with_label, power_columns, powe
         item['mae'] = mae
         item['mse'] = mse
         item['%mae'] = percent * 100
+        item['mape'] = mape
         items += [item]
     return pd.DataFrame(items), data_with_prediction, model