nkululeko 0.77.12__tar.gz → 0.77.14__tar.gz

{nkululeko-0.77.12 → nkululeko-0.77.14}/CHANGELOG.md

@@ -1,6 +1,14 @@
 Changelog
 =========
 
+Version 0.77.14
+--------------
+* fixed bug preventing bin scaling to work
+
+Version 0.77.13
+--------------
+* added bins scaler
+
 Version 0.77.12
 --------------
 * fixed bug with scatter plots for numeric targets

{nkululeko-0.77.12/nkululeko.egg-info → nkululeko-0.77.14}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: nkululeko
-Version: 0.77.12
+Version: 0.77.14
 Summary: Machine learning audio prediction experiments based on templates
 Home-page: https://github.com/felixbur/nkululeko
 Author: Felix Burkhardt
@@ -314,6 +314,14 @@ F. Burkhardt, Johannes Wagner, Hagen Wierstorf, Florian Eyben and Björn Schulle
 Changelog
 =========
 
+Version 0.77.14
+--------------
+* fixed bug preventing bin scaling to work
+
+Version 0.77.13
+--------------
+* added bins scaler
+
 Version 0.77.12
 --------------
 * fixed bug with scatter plots for numeric targets

{nkululeko-0.77.12 → nkululeko-0.77.14}/nkululeko/constants.py

@@ -1,2 +1,2 @@
-VERSION="0.77.12"
+VERSION="0.77.14"
 SAMPLING_RATE = 16000

{nkululeko-0.77.12 → nkululeko-0.77.14}/nkululeko/experiment.py

@@ -577,6 +577,9 @@ class Experiment:
                 scale_feats,
             )
             self.feats_train, self.feats_test = self.scaler_feats.scale()
+            # store versions
+            self.util.save_to_store(self.feats_train, "feats_train_scaled")
+            self.util.save_to_store(self.feats_test, "feats_test_scaled")
 
     def init_runmanager(self):
         """Initialize the manager object for the runs."""

{nkululeko-0.77.12 → nkululeko-0.77.14}/nkululeko/runmanager.py

@@ -70,7 +70,8 @@ class Runmanager:
             plot_epoch_progression = self.util.config_val(
                 "PLOT", "epoch_progression", 0
             )
-            if plot_epoch_progression:
+            epoch_num = int(self.util.config_val("EXP", "epochs", 1))
+            if epoch_num > 1 and plot_epoch_progression:
                 plot_name_suggest = self.util.get_exp_name()
                 plot_name = (
                     self.util.config_val("PLOT", "name", plot_name_suggest)
@@ -81,7 +82,8 @@ class Runmanager:
             # remember the best run
             best_report = self.get_best_result(self.reports)
             plot_best_model = self.util.config_val("PLOT", "best_model", False)
-            if plot_best_model:
+
+            if epoch_num > 1 and plot_best_model:
                 plot_name_suggest = self.util.get_exp_name()
                 plot_name = (
                     self.util.config_val("PLOT", "name", plot_name_suggest)

{nkululeko-0.77.12 → nkululeko-0.77.14}/nkululeko/scaler.py

@@ -1,8 +1,9 @@
 # scaler.py
 
-
+import numpy as np
 from sklearn.preprocessing import StandardScaler, RobustScaler
 import pandas as pd
+
 from nkululeko.utils.util import Util
 
 
@@ -32,6 +33,8 @@ class Scaler:
             self.scaler = RobustScaler()
         elif scaler_type == "speaker":
             self.scaler = StandardScaler()
+        elif scaler_type == "bins":
+            pass
         else:
             self.util.error("unknown scaler: " + scaler_type)
         self.scaler_type = scaler_type
@@ -56,10 +59,13 @@ class Scaler:
             return self.speaker_scale()
 
     def scale_all(self):
-        self.scaler.fit(self.feats_train.values)
-        self.feats_train = self.scale_df(self.feats_train)
-        if self.feats_test is not None:
-            self.feats_test = self.scale_df(self.feats_test)
+        if self.scaler_type != "bins":
+            self.scaler.fit(self.feats_train.values)
+            self.feats_train = self.scale_df(self.feats_train)
+            if self.feats_test is not None:
+                self.feats_test = self.scale_df(self.feats_test)
+        else:
+            self.bin_to_three()
         return self.feats_train, self.feats_test
 
     def scale_df(self, df):
@@ -80,3 +86,23 @@ class Scaler:
                 feats_df.loc[indices, :]
             )
         return feats_df
+
+    def bin_to_three(self):
+        feats_bin_train = pd.DataFrame(index=self.feats_train.index)
+        feats_bin_test = pd.DataFrame(index=self.feats_test.index)
+        for c in self.feats_train.columns:
+            b1 = np.quantile(self.feats_train[c], 0.33)
+            b2 = np.quantile(self.feats_train[c], 0.66)
+            feats_bin_train[c] = self._bin(self.feats_train[c].values, b1, b2).values
+            feats_bin_test[c] = self._bin(self.feats_test[c].values, b1, b2).values
+        self.feats_train = feats_bin_train
+        self.feats_test = feats_bin_test
+
+    def _bin(self, series, b1, b2):
+        bins = [-1000000, b1, b2, 1000000]
+        labels = [0, 0.5, 1]
+        result = np.digitize(series, bins) - 1
+        result = pd.Series(result)
+        for i, l in enumerate(labels):
+            result = result.replace(i, str(l))
+        return result

{nkululeko-0.77.12 → nkululeko-0.77.14}/nkululeko/utils/util.py

@@ -195,6 +195,7 @@ class Util:
             ["MODEL", "learning_rate"],
             ["MODEL", "k_fold_cross"],
             ["FEATS", "balancing"],
+            ["FEATS", "scale"],
         ]
         for option in options:
             return_string += self._get_value_descript(option[0], option[1])
@@ -306,7 +307,13 @@ class Util:
         elif format == "csv":
             return audformat.utils.read_csv(name)
         else:
-            self.error(f"unkown store format: {format}")
+            self.error(f"unknown store format: {format}")
+
+    def save_to_store(self, df, name):
+        store = self.get_path("store")
+        store_format = self.config_val("FEATS", "store_format", "pkl")
+        storage = f"{store}{name}.{store_format}"
+        self.write_store(df, storage, store_format)
 
     def copy_flags(self, df_source, df_target):
         if hasattr(df_source, "is_labeled"):

{nkululeko-0.77.12 → nkululeko-0.77.14/nkululeko.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: nkululeko
-Version: 0.77.12
+Version: 0.77.14
 Summary: Machine learning audio prediction experiments based on templates
 Home-page: https://github.com/felixbur/nkululeko
 Author: Felix Burkhardt
@@ -314,6 +314,14 @@ F. Burkhardt, Johannes Wagner, Hagen Wierstorf, Florian Eyben and Björn Schulle
 Changelog
 =========
 
+Version 0.77.14
+--------------
+* fixed bug preventing bin scaling to work
+
+Version 0.77.13
+--------------
+* added bins scaler
+
 Version 0.77.12
 --------------
 * fixed bug with scatter plots for numeric targets