nkululeko 0.86.7__tar.gz → 0.87.0__tar.gz

{nkululeko-0.86.7 → nkululeko-0.87.0}/CHANGELOG.md

@@ -1,6 +1,14 @@
 Changelog
 =========
 
+Version 0.87.0
+--------------
+* added class probability output and uncertainty analysis
+
+Version 0.86.8
+--------------
+* handle single feature sets as strings in the config
+
 Version 0.86.7
 --------------
 * handles now audformat tables where the target is in a file index

{nkululeko-0.86.7/nkululeko.egg-info → nkululeko-0.87.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: nkululeko
-Version: 0.86.7
+Version: 0.87.0
 Summary: Machine learning audio prediction experiments based on templates
 Home-page: https://github.com/felixbur/nkululeko
 Author: Felix Burkhardt
@@ -51,6 +51,7 @@ Requires-Dist: pylatex
   - [t-SNE plots](#t-sne-plots)
   - [Data distribution](#data-distribution)
   - [Bias checking](#bias-checking)
+  - [Uncertainty](#uncertainty)
 - [Documentation](#documentation)
 - [Installation](#installation)
 - [Usage](#usage)
@@ -113,6 +114,13 @@ In cases you might wonder if there's bias in your data. You can try to detect th
 
 <img src="meta/images/emotion-pesq.png" width="500px"/>
 
+### Uncertainty
+Nkululeko estimates uncertainty of model decision (only for classifiers) with entropy over the class-probabilities or logits per sample.
+
+<img src="meta/images/uncertainty.png" width="500px"/>
+
+
+
 ## Documentation
 The documentation, along with extensions of installation, usage, INI file format, and examples, can be found [nkululeko.readthedocs.io](https://nkululeko.readthedocs.io).
 
@@ -343,6 +351,14 @@ F. Burkhardt, Johannes Wagner, Hagen Wierstorf, Florian Eyben and Björn Schulle
 Changelog
 =========
 
+Version 0.87.0
+--------------
+* added class probability output and uncertainty analysis
+
+Version 0.86.8
+--------------
+* handle single feature sets as strings in the config
+
 Version 0.86.7
 --------------
 * handles now audformat tables where the target is in a file index

{nkululeko-0.86.7 → nkululeko-0.87.0}/README.md

@@ -7,6 +7,7 @@
   - [t-SNE plots](#t-sne-plots)
   - [Data distribution](#data-distribution)
   - [Bias checking](#bias-checking)
+  - [Uncertainty](#uncertainty)
 - [Documentation](#documentation)
 - [Installation](#installation)
 - [Usage](#usage)
@@ -69,6 +70,13 @@ In cases you might wonder if there's bias in your data. You can try to detect th
 
 <img src="meta/images/emotion-pesq.png" width="500px"/>
 
+### Uncertainty
+Nkululeko estimates uncertainty of model decision (only for classifiers) with entropy over the class-probabilities or logits per sample.
+
+<img src="meta/images/uncertainty.png" width="500px"/>
+
+
+
 ## Documentation
 The documentation, along with extensions of installation, usage, INI file format, and examples, can be found [nkululeko.readthedocs.io](https://nkululeko.readthedocs.io).
 

{nkululeko-0.86.7 → nkululeko-0.87.0}/nkululeko/constants.py

@@ -1,2 +1,2 @@
-VERSION="0.86.7"
+VERSION="0.87.0"
 SAMPLING_RATE = 16000

{nkululeko-0.86.7 → nkululeko-0.87.0}/nkululeko/data/dataset_csv.py

@@ -23,6 +23,9 @@ class Dataset_CSV(Dataset):
         root = os.path.dirname(data_file)
         audio_path = self.util.config_val_data(self.name, "audio_path", "./")
         df = pd.read_csv(data_file)
+        # trim all string values
+        df_obj = df.select_dtypes("object")
+        df[df_obj.columns] = df_obj.apply(lambda x: x.str.strip())
         # special treatment for segmented dataframes with only one column:
         if "start" in df.columns and len(df.columns) == 4:
             index = audformat.segmented_index(
@@ -49,8 +52,7 @@ class Dataset_CSV(Dataset):
                     .map(lambda x: root + "/" + audio_path + "/" + x)
                     .values
                 )
-                df = df.set_index(df.index.set_levels(
-                    file_index, level="file"))
+                df = df.set_index(df.index.set_levels(file_index, level="file"))
             else:
                 if not isinstance(df, pd.DataFrame):
                     df = pd.DataFrame(df)
@@ -59,27 +61,24 @@ class Dataset_CSV(Dataset):
                         lambda x: root + "/" + audio_path + "/" + x
                     )
                 )
-        else: # absolute path is True
+        else:  # absolute path is True
             if audformat.index_type(df.index) == "segmented":
                 file_index = (
-                    df.index.levels[0]
-                    .map(lambda x: audio_path + "/" + x)
-                    .values
+                    df.index.levels[0].map(lambda x: audio_path + "/" + x).values
                 )
-                df = df.set_index(df.index.set_levels(
-                    file_index, level="file"))
+                df = df.set_index(df.index.set_levels(file_index, level="file"))
             else:
                 if not isinstance(df, pd.DataFrame):
                     df = pd.DataFrame(df)
-                df = df.set_index(df.index.to_series().apply(
-                    lambda x: audio_path + "/" + x ))
+                df = df.set_index(
+                    df.index.to_series().apply(lambda x: audio_path + "/" + x)
+                )
 
         self.df = df
         self.db = None
         self.got_target = True
         self.is_labeled = self.got_target
-        self.start_fresh = eval(
-            self.util.config_val("DATA", "no_reuse", "False"))
+        self.start_fresh = eval(self.util.config_val("DATA", "no_reuse", "False"))
         is_index = False
         try:
             if self.is_labeled and not "class_label" in self.df.columns:
@@ -106,8 +105,7 @@ class Dataset_CSV(Dataset):
                 f" {self.got_gender}, got age: {self.got_age}"
             )
         self.util.debug(r_string)
-        glob_conf.report.add_item(ReportItem(
-            "Data", "Loaded report", r_string))
+        glob_conf.report.add_item(ReportItem("Data", "Loaded report", r_string))
 
     def prepare(self):
         super().prepare()

{nkululeko-0.86.7 → nkululeko-0.87.0}/nkululeko/demo.py

@@ -30,10 +30,8 @@ from transformers import pipeline
 
 
 def main(src_dir):
-    parser = argparse.ArgumentParser(
-        description="Call the nkululeko DEMO framework.")
-    parser.add_argument("--config", default="exp.ini",
-                        help="The base configuration")
+    parser = argparse.ArgumentParser(description="Call the nkululeko DEMO framework.")
+    parser.add_argument("--config", default="exp.ini", help="The base configuration")
     parser.add_argument(
         "--file", help="A file that should be processed (16kHz mono wav)"
     )
@@ -84,8 +82,7 @@ def main(src_dir):
     )
 
     def print_pipe(files, outfile):
-        """
-        Prints the pipeline output for a list of files, and optionally writes the results to an output file.
+        """Prints the pipeline output for a list of files, and optionally writes the results to an output file.
 
         Args:
             files (list): A list of file paths to process through the pipeline.
@@ -108,8 +105,7 @@ def main(src_dir):
                 f.write("\n".join(results))
 
     if util.get_model_type() == "finetune":
-        model_path = os.path.join(
-            util.get_exp_dir(), "models", "run_0", "torch")
+        model_path = os.path.join(util.get_exp_dir(), "models", "run_0", "torch")
         pipe = pipeline("audio-classification", model=model_path)
         if args.file is not None:
             print_pipe([args.file], args.outfile)

{nkululeko-0.86.7 → nkululeko-0.87.0}/nkululeko/experiment.py

@@ -5,13 +5,13 @@ import pickle
 import random
 import time
 
+import audeer
+import audformat
 import numpy as np
 import pandas as pd
 from sklearn.preprocessing import LabelEncoder
 
-import audeer
-import audformat
-
+import nkululeko.glob_conf as glob_conf
 from nkululeko.data.dataset import Dataset
 from nkululeko.data.dataset_csv import Dataset_CSV
 from nkululeko.demo_predictor import Demo_predictor
@@ -19,8 +19,6 @@ from nkululeko.feat_extract.feats_analyser import FeatureAnalyser
 from nkululeko.feature_extractor import FeatureExtractor
 from nkululeko.file_checker import FileChecker
 from nkululeko.filter_data import DataFilter
-from nkululeko.filter_data import filter_min_dur
-import nkululeko.glob_conf as glob_conf
 from nkululeko.plots import Plots
 from nkululeko.reporting.report import Report
 from nkululeko.runmanager import Runmanager
@@ -109,7 +107,8 @@ class Experiment:
         # print keys/column
         dbs = ",".join(list(self.datasets.keys()))
         labels = self.util.config_val("DATA", "labels", False)
-        auto_labels = list(next(iter(self.datasets.values())).df[self.target].unique())
+        auto_labels = list(
+            next(iter(self.datasets.values())).df[self.target].unique())
         if labels:
             self.labels = ast.literal_eval(labels)
             self.util.debug(f"Using target labels (from config): {labels}")
@@ -159,7 +158,8 @@ class Experiment:
                 data.split()
                 data.prepare_labels()
                 self.df_test = pd.concat(
-                    [self.df_test, self.util.make_segmented_index(data.df_test)]
+                    [self.df_test, self.util.make_segmented_index(
+                        data.df_test)]
                 )
                 self.df_test.is_labeled = data.is_labeled
             self.df_test.got_gender = self.got_gender
@@ -260,7 +260,8 @@ class Experiment:
                     test_cats = self.df_test[self.target].unique()
                 else:
                     # if there is no target, copy a dummy label
-                    self.df_test = self._add_random_target(self.df_test).astype("str")
+                    self.df_test = self._add_random_target(
+                        self.df_test).astype("str")
                 train_cats = self.df_train[self.target].unique()
                 # print(f"df_train: {pd.DataFrame(self.df_train[self.target])}")
                 # print(f"train_cats with target {self.target}: {train_cats}")
@@ -268,7 +269,8 @@ class Experiment:
                 if type(test_cats) == np.ndarray:
                     self.util.debug(f"Categories test (nd.array): {test_cats}")
                 else:
-                    self.util.debug(f"Categories test (list): {list(test_cats)}")
+                    self.util.debug(
+                        f"Categories test (list): {list(test_cats)}")
             if type(train_cats) == np.ndarray:
                 self.util.debug(f"Categories train (nd.array): {train_cats}")
             else:
@@ -291,7 +293,8 @@ class Experiment:
 
         target_factor = self.util.config_val("DATA", "target_divide_by", False)
         if target_factor:
-            self.df_test[self.target] = self.df_test[self.target] / float(target_factor)
+            self.df_test[self.target] = self.df_test[self.target] / \
+                float(target_factor)
             self.df_train[self.target] = self.df_train[self.target] / float(
                 target_factor
             )
@@ -314,14 +317,16 @@ class Experiment:
     def plot_distribution(self, df_labels):
         """Plot the distribution of samples and speaker per target class and biological sex"""
         plot = Plots()
-        sample_selection = self.util.config_val("EXPL", "sample_selection", "all")
+        sample_selection = self.util.config_val(
+            "EXPL", "sample_selection", "all")
         plot.plot_distributions(df_labels)
         if self.got_speaker:
             plot.plot_distributions_speaker(df_labels)
 
     def extract_test_feats(self):
         self.feats_test = pd.DataFrame()
-        feats_name = "_".join(ast.literal_eval(glob_conf.config["DATA"]["tests"]))
+        feats_name = "_".join(ast.literal_eval(
+            glob_conf.config["DATA"]["tests"]))
         feats_types = self.util.config_val_list("FEATS", "type", ["os"])
         self.feature_extractor = FeatureExtractor(
             self.df_test, feats_types, feats_name, "test"
@@ -338,9 +343,17 @@ class Experiment:
 
         """
         df_train, df_test = self.df_train, self.df_test
-        feats_name = "_".join(ast.literal_eval(glob_conf.config["DATA"]["databases"]))
+        feats_name = "_".join(ast.literal_eval(
+            glob_conf.config["DATA"]["databases"]))
         self.feats_test, self.feats_train = pd.DataFrame(), pd.DataFrame()
-        feats_types = self.util.config_val_list("FEATS", "type", [])
+        feats_types = self.util.config_val("FEATS", "type", "os")
+        # Ensure feats_types is always a list of strings
+        if isinstance(feats_types, str):
+            if feats_types.startswith("[") and feats_types.endswith("]"):
+                feats_types = ast.literal_eval(feats_types)
+            else:
+                feats_types = [feats_types]
+        # print(f"feats_types: {feats_types}")
         # for some models no features are needed
         if len(feats_types) == 0:
             self.util.debug("no feature extractor specified.")
@@ -372,7 +385,8 @@ class Experiment:
                 f"test feats ({self.feats_test.shape[0]}) != test labels"
                 f" ({self.df_test.shape[0]})"
             )
-            self.df_test = self.df_test[self.df_test.index.isin(self.feats_test.index)]
+            self.df_test = self.df_test[self.df_test.index.isin(
+                self.feats_test.index)]
             self.util.warn(f"new test labels shape: {self.df_test.shape[0]}")
 
         self._check_scale()
@@ -387,7 +401,8 @@ class Experiment:
         """Augment the selected samples."""
         from nkululeko.augmenting.augmenter import Augmenter
 
-        sample_selection = self.util.config_val("AUGMENT", "sample_selection", "all")
+        sample_selection = self.util.config_val(
+            "AUGMENT", "sample_selection", "all")
         if sample_selection == "all":
             df = pd.concat([self.df_train, self.df_test])
         elif sample_selection == "train":
@@ -482,7 +497,8 @@ class Experiment:
         """
         from nkululeko.augmenting.randomsplicer import Randomsplicer
 
-        sample_selection = self.util.config_val("AUGMENT", "sample_selection", "all")
+        sample_selection = self.util.config_val(
+            "AUGMENT", "sample_selection", "all")
         if sample_selection == "all":
             df = pd.concat([self.df_train, self.df_test])
         elif sample_selection == "train":
@@ -503,7 +519,8 @@ class Experiment:
         plot_feats = eval(
             self.util.config_val("EXPL", "feature_distributions", "False")
         )
-        sample_selection = self.util.config_val("EXPL", "sample_selection", "all")
+        sample_selection = self.util.config_val(
+            "EXPL", "sample_selection", "all")
         # get the data labels
         if sample_selection == "all":
             df_labels = pd.concat([self.df_train, self.df_test])
@@ -566,7 +583,8 @@ class Experiment:
             for scat_target in scat_targets:
                 if self.util.is_categorical(df_labels[scat_target]):
                     for scatter in scatters:
-                        plots.scatter_plot(df_feats, df_labels, scat_target, scatter)
+                        plots.scatter_plot(
+                            df_feats, df_labels, scat_target, scatter)
                 else:
                     self.util.debug(
                         f"{self.name}: binning continuous variable to categories"
@@ -657,7 +675,8 @@ class Experiment:
         preds = best.preds
         speakers = self.df_test.speaker.values
         print(f"{len(truths)} {len(preds)} {len(speakers) }")
-        df = pd.DataFrame(data={"truth": truths, "pred": preds, "speaker": speakers})
+        df = pd.DataFrame(
+            data={"truth": truths, "pred": preds, "speaker": speakers})
         plot_name = "result_combined_per_speaker"
         self.util.debug(
             f"plotting speaker combination ({function}) confusion matrix to"
@@ -733,7 +752,6 @@ class Experiment:
         if model.is_ann():
             print("converting to onnx from torch")
         else:
-            from skl2onnx import to_onnx
 
             print("converting to onnx from sklearn")
         # save the rest

{nkululeko-0.86.7 → nkululeko-0.87.0}/nkululeko/feature_extractor.py

@@ -39,16 +39,20 @@ class FeatureExtractor:
         self.feats = pd.DataFrame()
         for feats_type in self.feats_types:
             store_name = f"{self.data_name}_{feats_type}"
-            self.feat_extractor = self._get_feat_extractor(store_name, feats_type)
+            self.feat_extractor = self._get_feat_extractor(
+                store_name, feats_type)
             self.feat_extractor.extract()
             self.feat_extractor.filter()
-            self.feats = pd.concat([self.feats, self.feat_extractor.df], axis=1)
+            self.feats = pd.concat(
+                [self.feats, self.feat_extractor.df], axis=1)
         return self.feats
 
     def extract_sample(self, signal, sr):
         return self.feat_extractor.extract_sample(signal, sr)
 
     def _get_feat_extractor(self, store_name, feats_type):
+        if isinstance(feats_type, list) and len(feats_type) == 1:
+            feats_type = feats_type[0]
         feat_extractor_class = self._get_feat_extractor_class(feats_type)
         if feat_extractor_class is None:
             self.util.error(f"unknown feats_type: {feats_type}")
@@ -103,13 +107,15 @@ class FeatureExtractor:
         prefix, _, ext = feats_type.partition("-")
         from importlib import import_module
 
-        module = import_module(f"nkululeko.feat_extract.feats_{prefix.lower()}")
+        module = import_module(
+            f"nkululeko.feat_extract.feats_{prefix.lower()}")
         class_name = f"{prefix.capitalize()}"
         return getattr(module, class_name)
 
     def _get_feat_extractor_by_name(self, feats_type):
         from importlib import import_module
 
-        module = import_module(f"nkululeko.feat_extract.feats_{feats_type.lower()}")
+        module = import_module(
+            f"nkululeko.feat_extract.feats_{feats_type.lower()}")
         class_name = f"{feats_type.capitalize()}Set"
         return getattr(module, class_name)

{nkululeko-0.86.7 → nkululeko-0.87.0}/nkululeko/modelrunner.py

@@ -85,7 +85,7 @@ class Modelrunner:
                     f"run: {self.run} epoch: {epoch}: result: {test_score_metric}"
                 )
                 # print(f"performance: {performance.split(' ')[1]}")
-                performance = float(test_score_metric.split(' ')[1])
+                performance = float(test_score_metric.split(" ")[1])
                 if performance > self.best_performance:
                     self.best_performance = performance
                     self.best_epoch = epoch
@@ -204,15 +204,15 @@ class Modelrunner:
                 self.df_train, self.df_test, self.feats_train, self.feats_test
             )
         elif model_type == "cnn":
-            from nkululeko.models.model_cnn import CNN_model
+            from nkululeko.models.model_cnn import CNNModel
 
-            self.model = CNN_model(
+            self.model = CNNModel(
                 self.df_train, self.df_test, self.feats_train, self.feats_test
             )
         elif model_type == "mlp":
-            from nkululeko.models.model_mlp import MLP_model
+            from nkululeko.models.model_mlp import MLPModel
 
-            self.model = MLP_model(
+            self.model = MLPModel(
                 self.df_train, self.df_test, self.feats_train, self.feats_test
             )
         elif model_type == "mlp_reg":

{nkululeko-0.86.7 → nkululeko-0.87.0}/nkululeko/models/model.py

@@ -247,8 +247,25 @@ class Model:
                 self.clf.fit(feats, labels)
 
     def get_predictions(self):
-        predictions = self.clf.predict(self.feats_test.to_numpy())
-        return predictions
+        #        predictions = self.clf.predict(self.feats_test.to_numpy())
+        if self.util.exp_is_classification():
+            # make a dataframe for the class probabilities
+            proba_d = {}
+            for c in self.clf.classes_:
+                proba_d[c] = []
+            # get the class probabilities
+            predictions = self.clf.predict_proba(self.feats_test.to_numpy())
+            # pred = self.clf.predict(features)
+            for i, c in enumerate(self.clf.classes_):
+                proba_d[c] = list(predictions.T[i])
+            probas = pd.DataFrame(proba_d)
+            probas = probas.set_index(self.feats_test.index)
+            predictions = probas.idxmax(axis=1).values
+        else:
+            predictions = self.clf.predict(self.feats_test.to_numpy())
+            probas = None
+
+        return predictions, probas
 
     def predict(self):
         if self.feats_test.isna().to_numpy().any():
@@ -263,13 +280,16 @@ class Model:
             )
             return report
         """Predict the whole eval feature set"""
-        predictions = self.get_predictions()
+        predictions, probas = self.get_predictions()
+
         report = Reporter(
             self.df_test[self.target].to_numpy().astype(float),
             predictions,
             self.run,
             self.epoch,
+            probas=probas,
         )
+        report.print_probabilities()
         return report
 
     def get_type(self):

{nkululeko-0.86.7 → nkululeko-0.87.0}/nkululeko/models/model_cnn.py

@@ -5,33 +5,40 @@ Inspired by code from Su Lei
 
 """
 
+import ast
+from collections import OrderedDict
+
+import numpy as np
+import pandas as pd
+from PIL import Image
+from sklearn.metrics import recall_score
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-import torchvision
-import torchvision.transforms as transforms
 from torch.utils.data import Dataset
-import ast
-import numpy as np
-from sklearn.metrics import recall_score
-from collections import OrderedDict
-from PIL import Image
-from traitlets import default
+import torchvision.transforms as transforms
 
-from nkululeko.utils.util import Util
 import nkululeko.glob_conf as glob_conf
+from nkululeko.losses.loss_softf1loss import SoftF1Loss
 from nkululeko.models.model import Model
 from nkululeko.reporting.reporter import Reporter
-from nkululeko.losses.loss_softf1loss import SoftF1Loss
+from nkululeko.utils.util import Util
 
 
-class CNN_model(Model):
-    """CNN = convolutional neural net"""
+class CNNModel(Model):
+    """CNN = convolutional neural net."""
 
     is_classifier = True
 
     def __init__(self, df_train, df_test, feats_train, feats_test):
-        """Constructor taking the configuration and all dataframes"""
+        """Constructor, taking all dataframes.
+
+        Args:
+            df_train (pd.DataFrame): The train labels.
+            df_test (pd.DataFrame): The test labels.
+            feats_train (pd.DataFrame): The train features.
+            feats_test (pd.DataFrame): The test features.
+        """
         super().__init__(df_train, df_test, feats_train, feats_test)
         super().set_model_type("ann")
         self.name = "cnn"
@@ -147,7 +154,20 @@ class CNN_model(Model):
             self.optimizer.step()
         self.loss = (np.asarray(losses)).mean()
 
-    def evaluate_model(self, model, loader, device):
+    def get_probas(self, logits):
+        # make a dataframe for probabilites (logits)
+        proba_d = {}
+        classes = self.df_test[self.target].unique()
+        classes.sort()
+        for c in classes:
+            proba_d[c] = []
+        for i, c in enumerate(classes):
+            proba_d[c] = list(logits.numpy().T[i])
+        probas = pd.DataFrame(proba_d)
+        probas = probas.set_index(self.df_test.index)
+        return probas
+
+    def evaluate(self, model, loader, device):
         logits = torch.zeros(len(loader.dataset), self.class_num)
         targets = torch.zeros(len(loader.dataset))
         model.eval()
@@ -169,14 +189,15 @@ class CNN_model(Model):
         self.loss_eval = (np.asarray(losses)).mean()
         predictions = logits.argmax(dim=1)
         uar = recall_score(targets.numpy(), predictions.numpy(), average="macro")
-        return uar, targets, predictions
+        return uar, targets, predictions, logits
 
     def predict(self):
-        _, truths, predictions = self.evaluate_model(
+        _, truths, predictions, logits = self.evaluate(
             self.model, self.testloader, self.device
         )
-        uar, _, _ = self.evaluate_model(self.model, self.trainloader, self.device)
-        report = Reporter(truths, predictions, self.run, self.epoch)
+        uar, _, _, _ = self.evaluate(self.model, self.trainloader, self.device)
+        probas = self.get_probas(logits)
+        report = Reporter(truths, predictions, self.run, self.epoch, probas=probas)
         try:
             report.result.loss = self.loss
         except AttributeError:  # if the model was loaded from disk the loss is unknown
@@ -189,13 +210,11 @@ class CNN_model(Model):
         return report
 
     def get_predictions(self):
-        _, truths, predictions = self.evaluate_model(
-            self.model, self.testloader, self.device
-        )
+        _, _, predictions, _ = self.evaluate(self.model, self.testloader, self.device)
         return predictions.numpy()
 
     def predict_sample(self, features):
-        """Predict one sample"""
+        """Predict one sample."""
         with torch.no_grad():
             logits = self.model(torch.from_numpy(features).to(self.device))
         a = logits.numpy()