nkululeko 0.86.1__py3-none-any.whl → 0.86.2__py3-none-any.whl

nkululeko/constants.py

@@ -1,2 +1,2 @@
-VERSION="0.86.1"
+VERSION="0.86.2"
 SAMPLING_RATE = 16000

nkululeko/models/model_mlp.py

@@ -14,12 +14,12 @@ from nkululeko.losses.loss_softf1loss import SoftF1Loss
 
 
 class MLP_model(Model):
-    """MLP = multi layer perceptron"""
+    """MLP = multi layer perceptron."""
 
     is_classifier = True
 
     def __init__(self, df_train, df_test, feats_train, feats_test):
-        """Constructor taking the configuration and all dataframes"""
+        """Constructor taking the configuration and all dataframes."""
         super().__init__(df_train, df_test, feats_train, feats_test)
         super().set_model_type("ann")
         self.name = "mlp"

nkululeko/models/model_tuned.py

@@ -1,5 +1,6 @@
 """Code based on @jwagner."""
 
+import ast
 import dataclasses
 import json
 import os
@@ -51,8 +52,9 @@ class TunedModel(BaseModel):
         self.batch_size = int(self.util.config_val("MODEL", "batch_size", "8"))
         self.util.debug(f"batch size: {self.batch_size}")
         self.learning_rate = float(
-            self.util.config_val("MODEL", "learning_rate", 0.0001)
+            self.util.config_val("MODEL", "learning_rate", "0.0001")
         )
+        self.max_duration = float(self.util.config_val("MODEL", "max_duration", "8.0"))
         self.df_train, self.df_test = df_train, df_test
         self.epoch_num = int(self.util.config_val("EXP", "epochs", 1))
         drop = self.util.config_val("MODEL", "drop", False)
@@ -64,11 +66,10 @@ class TunedModel(BaseModel):
 
     def _init_model(self):
         model_path = "facebook/wav2vec2-large-robust-ft-swbd-300h"
-        pretrained_model = self.util.config_val(
-            "MODEL", "pretrained_model", model_path)
+        pretrained_model = self.util.config_val("MODEL", "pretrained_model", model_path)
         self.num_layers = None
         self.sampling_rate = 16000
-        self.max_duration_sec = 8.0
+        self.max_duration_sec = self.max_duration
         self.accumulation_steps = 4
 
         # print finetuning information via debug
@@ -235,8 +236,8 @@ class TunedModel(BaseModel):
     def train(self):
         """Train the model."""
         model_root = self.util.get_path("model_dir")
-        log_root = os.path.join(self.util.get_exp_dir(), "log")
-        audeer.mkdir(log_root)
+        self.log_root = os.path.join(self.util.get_exp_dir(), "log")
+        audeer.mkdir(self.log_root)
         self.torch_root = audeer.path(model_root, "torch")
         conf_file = os.path.join(self.torch_root, "config.json")
         if os.path.isfile(conf_file):
@@ -244,20 +245,35 @@ class TunedModel(BaseModel):
             self.load(self.run, self.epoch_num)
             return
         targets = pd.DataFrame(self.dataset["train"]["targets"])
-        counts = targets[0].value_counts().sort_index()
 
         if self.is_classifier:
-            train_weights = 1 / counts
-            train_weights /= train_weights.sum()
-            self.util.debug(f"train weights: {train_weights}")
-            criterion = torch.nn.CrossEntropyLoss(
-                weight=torch.Tensor(train_weights).to("cuda"),
-            )
+            criterion = self.util.config_val("MODEL", "loss", "cross")
+            if criterion == "cross":
+                if self.util.config_val("MODEL", "class_weight", False):
+                    counts = targets[0].value_counts().sort_index()
+                    train_weights = 1 / counts
+                    train_weights /= train_weights.sum()
+                    self.util.debug(f"train weights: {train_weights}")
+                    criterion = torch.nn.CrossEntropyLoss(
+                        weight=torch.Tensor(train_weights).to("cuda"),
+                    )
+                else:
+                    criterion = torch.nn.CrossEntropyLoss()
+            else:
+                self.util.error(f"criterion {criterion} not supported for classifier")
         else:
-            criterion = ConcordanceCorCoeff()
+            self.criterion = self.util.config_val("MODEL", "loss", "ccc")
+            if criterion == "1-ccc":
+                criterion = ConcordanceCorCoeff()
+            elif criterion == "mse":
+                criterion = torch.nn.MSELoss()
+            elif criterion == "mae":
+                criterion = torch.nn.L1Loss()
+            else:
+                self.util.error(f"criterion {criterion} not supported for regressor")
 
         # set push_to_hub value, default false
-        push = self.util.config_val("MODEL", "push_to_hub", False)
+        push = eval(self.util.config_val("MODEL", "push_to_hub", "False"))
 
         class Trainer(transformers.Trainer):
             def compute_loss(
@@ -284,10 +300,20 @@ class TunedModel(BaseModel):
         num_steps = max(1, num_steps)
 
         metrics_for_best_model = self.measure.upper()
+        if metrics_for_best_model == "UAR":
+            greater_is_better = True
+        elif metrics_for_best_model == "CCC":
+            greater_is_better = True
+        elif metrics_for_best_model == "MSE":
+            greater_is_better = False
+        elif metrics_for_best_model == "MAE":
+            greater_is_better = False
+        else:
+            self.util.error(f"unknown metric/measure: {metrics_for_best_model}")
 
         training_args = transformers.TrainingArguments(
             output_dir=model_root,
-            logging_dir=log_root,
+            logging_dir=self.log_root,
             per_device_train_batch_size=self.batch_size,
             per_device_eval_batch_size=self.batch_size,
             gradient_accumulation_steps=self.accumulation_steps,
@@ -301,7 +327,7 @@ class TunedModel(BaseModel):
             learning_rate=self.learning_rate,
             save_total_limit=2,
             metric_for_best_model=metrics_for_best_model,
-            greater_is_better=True,
+            greater_is_better=greater_is_better,
             load_best_model_at_end=True,
             remove_unused_columns=False,
             report_to="none",
@@ -319,8 +345,15 @@ class TunedModel(BaseModel):
             tokenizer=self.processor.feature_extractor,
             callbacks=[transformers.integrations.TensorBoardCallback()],
         )
+
         trainer.train()
         trainer.save_model(self.torch_root)
+        log_file = os.path.join(
+            self.log_root,
+            "log.txt",
+        )
+        with open(log_file, "w") as text_file:
+            print(trainer.state.log_history, file=text_file)
         self.util.debug(f"saved best model to {self.torch_root}")
         self.load(self.run, self.epoch)
 
@@ -351,8 +384,30 @@ class TunedModel(BaseModel):
             self.run,
             self.epoch_num,
         )
+        self._plot_epoch_progression(report)
         return report
 
+    def _plot_epoch_progression(self, report):
+        log_file = os.path.join(
+            self.log_root,
+            "log.txt",
+        )
+        with open(log_file, "r") as file:
+            data = file.read()
+        list = ast.literal_eval(data)
+        epochs, vals, loss = [], [], []
+        for index, tp in enumerate(list):
+            try:
+                epochs.append(tp["epoch"])
+                measure = self.measure.upper()
+                vals.append(tp[f"eval_{measure}"])
+                loss.append(tp["eval_loss"])
+            except KeyError:
+                del epochs[-1]
+                # print(f'no value at {index}')
+        df = pd.DataFrame({"results": vals, "losses": loss}, index=epochs)
+        report.plot_epoch_progression_finetuned(df)
+
     def predict_sample(self, signal):
         """Predict one sample"""
         prediction = {}

nkululeko/plots.py

@@ -1,21 +1,23 @@
 # plots.py
-import pandas as pd
+import ast
+
 import matplotlib.pyplot as plt
-from sklearn.manifold import TSNE
-import seaborn as sns
 import numpy as np
-import ast
+import pandas as pd
 from scipy import stats
-from nkululeko.utils.util import Util
-import nkululeko.utils.stats as su
+import seaborn as sns
+from sklearn.manifold import TSNE
+
 import nkululeko.glob_conf as glob_conf
-from nkululeko.reporting.report_item import ReportItem
 from nkululeko.reporting.defines import Header
+from nkululeko.reporting.report_item import ReportItem
+import nkululeko.utils.stats as su
+from nkululeko.utils.util import Util
 
 
 class Plots:
     def __init__(self):
-        """Initializing the util system"""
+        """Initializing the util system."""
         self.util = Util("plots")
         self.format = self.util.config_val("PLOT", "format", "png")
         self.target = self.util.config_val("DATA", "target", "emotion")
@@ -138,8 +140,7 @@ class Plots:
                             df, att1, class_label, att1, type_s
                         )
                     else:
-                        ax, caption = self._plot2cont(
-                            df, class_label, att1, type_s)
+                        ax, caption = self._plot2cont(df, class_label, att1, type_s)
                 self._save_plot(
                     ax,
                     caption,
@@ -152,8 +153,7 @@ class Plots:
                 att1 = att[0]
                 att2 = att[1]
                 if att1 == self.target or att2 == self.target:
-                    self.util.debug(
-                        f"no need to correlate {self.target} with itself")
+                    self.util.debug(f"no need to correlate {self.target} with itself")
                     return
                 if att1 not in df:
                     self.util.error(f"unknown feature: {att1}")
@@ -168,8 +168,7 @@ class Plots:
                     if self.util.is_categorical(df[att1]):
                         if self.util.is_categorical(df[att2]):
                             # class_label = cat, att1 = cat, att2 = cat
-                            ax, caption = self._plot2cat(
-                                df, att1, att2, att1, type_s)
+                            ax, caption = self._plot2cat(df, att1, att2, att1, type_s)
                         else:
                             # class_label = cat, att1 = cat, att2 = cont
                             ax, caption = self._plotcatcont(
@@ -190,8 +189,7 @@ class Plots:
                     if self.util.is_categorical(df[att1]):
                         if self.util.is_categorical(df[att2]):
                             # class_label = cont, att1 = cat, att2 = cat
-                            ax, caption = self._plot2cat(
-                                df, att1, att2, att1, type_s)
+                            ax, caption = self._plot2cat(df, att1, att2, att1, type_s)
                         else:
                             # class_label = cont, att1 = cat, att2 = cont
                             ax, caption = self._plot2cont_cat(
@@ -205,8 +203,7 @@ class Plots:
                             )
                         else:
                             # class_label = cont, att1 = cont, att2 = cont
-                            ax, caption = self._plot2cont(
-                                df, att1, att2, type_s)
+                            ax, caption = self._plot2cont(df, att1, att2, type_s)
 
                 self._save_plot(
                     ax, caption, f"Correlation of {att1} and {att2}", filename, type_s
@@ -238,8 +235,7 @@ class Plots:
         )
 
     def _check_binning(self, att, df):
-        bin_reals_att = eval(self.util.config_val(
-            "EXPL", f"{att}.bin_reals", "False"))
+        bin_reals_att = eval(self.util.config_val("EXPL", f"{att}.bin_reals", "False"))
         if bin_reals_att:
             self.util.debug(f"binning continuous variable {att} to categories")
             att_new = f"{att}_binned"
@@ -342,8 +338,7 @@ class Plots:
 
     def describe_df(self, name, df, target, filename):
         """Make a stacked barplot of samples and speakers per sex and target values. speaker, gender and target columns must be present"""
-        fig_dir = self.util.get_path(
-            "fig_dir") + "../"  # one up because of the runs
+        fig_dir = self.util.get_path("fig_dir") + "../"  # one up because of the runs
         sampl_num = df.shape[0]
         sex_col = "gender"
         if target == "gender":
@@ -392,8 +387,7 @@ class Plots:
         dim_num = int(self.util.config_val("EXPL", "scatter.dim", 2))
         # one up because of the runs
         fig_dir = self.util.get_path("fig_dir") + "../"
-        sample_selection = self.util.config_val(
-            "EXPL", "sample_selection", "all")
+        sample_selection = self.util.config_val("EXPL", "sample_selection", "all")
         filename = f"{label}_{self.util.get_feattype_name()}_{sample_selection}_{dimred_type}_{str(dim_num)}d"
         filename = f"{fig_dir}{filename}.{self.format}"
         self.util.debug(f"computing {dimred_type}, this might take a while...")
@@ -435,8 +429,7 @@ class Plots:
 
         if dim_num == 2:
             plot_data = np.vstack((data.T, labels)).T
-            plot_df = pd.DataFrame(
-                data=plot_data, columns=("Dim_1", "Dim_2", "label"))
+            plot_df = pd.DataFrame(data=plot_data, columns=("Dim_1", "Dim_2", "label"))
             # plt.tight_layout()
             ax = (
                 sns.FacetGrid(plot_df, hue="label", height=6)

nkululeko/reporting/reporter.py

@@ -300,6 +300,23 @@ class Reporter:
     def get_result(self):
         return self.result
 
+    def plot_epoch_progression_finetuned(self, df):
+        plot_name_suggest = self.util.get_exp_name()
+        fig_dir = self.util.get_path("fig_dir")
+        plot_name = (
+            self.util.config_val("PLOT", "name", plot_name_suggest)
+            + "_epoch_progression"
+        )
+        ax = df.plot()
+        fig = ax.figure
+        plt.xlabel("epochs")
+        plt.ylabel(f"{self.MEASURE}")
+        plot_path = f"{fig_dir}{plot_name}.{self.format}"
+        plt.savefig(plot_path)
+        self.util.debug(f"plotted epoch progression to {plot_path}")
+        plt.close(fig)
+        fig.clear()
+
     def plot_epoch_progression(self, reports, out_name):
         fig_dir = self.util.get_path("fig_dir")
         results, losses, train_results, losses_eval = [], [], [], []

{nkululeko-0.86.1.dist-info → nkululeko-0.86.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: nkululeko
-Version: 0.86.1
+Version: 0.86.2
 Summary: Machine learning audio prediction experiments based on templates
 Home-page: https://github.com/felixbur/nkululeko
 Author: Felix Burkhardt
@@ -334,6 +334,10 @@ F. Burkhardt, Johannes Wagner, Hagen Wierstorf, Florian Eyben and Björn Schulle
 Changelog
 =========
 
+Version 0.86.2
+--------------
+* plots epoch progression for finetuned models now
+
 Version 0.86.1
 --------------
 * functionality to push to hub

{nkululeko-0.86.1.dist-info → nkululeko-0.86.2.dist-info}/RECORD

@@ -2,7 +2,7 @@ nkululeko/__init__.py,sha256=62f8HiEzJ8rG2QlTFJXUCMpvuH3fKI33DoJSj33mscc,63
 nkululeko/aug_train.py,sha256=YhuZnS_WVWnun9G-M6g5n6rbRxoVREz6Zh7k6qprFNQ,3194
 nkululeko/augment.py,sha256=4MG0apTAG5RgkuJrYEjGgDdbodZWi_HweSPNI1JJ5QA,3051
 nkululeko/cacheddataset.py,sha256=lIJ6hUo5LoxSrzXtWV8mzwO7wRtUETWnOQ4ws2XfL1E,969
-nkululeko/constants.py,sha256=pZ3DZYgXdEpxfaj-mnI6q21TyYMa2QQG_sKa6CBxCCA,39
+nkululeko/constants.py,sha256=E9mXpvAI5IDamRRXgBlBH8XGTw1xEjEBzNibjhFPEFc,39
 nkululeko/demo.py,sha256=8bl15Kitoesnz8oa8yrs52T6YCSOhWbbq9PnZ8Hj6D0,3232
 nkululeko/demo_feats.py,sha256=sAeGFojhEj9WEDFtG3SzPBmyYJWLF2rkbpp65m8Ujo4,2025
 nkululeko/demo_predictor.py,sha256=es56xbT8ifkS_vnrlb5NTZT54gNmeUtNlA4zVA_gnN8,4757
@@ -17,7 +17,7 @@ nkululeko/modelrunner.py,sha256=iCmfJxsS2UafcikjRdUqPQuqQMOYA-Ctr3et3HeNR3c,1045
 nkululeko/multidb.py,sha256=fG3VukEWP1vreVN4gB1IRXxwwg4jLftsSEYtu0o1f78,5634
 nkululeko/nkuluflag.py,sha256=PGWSmZz-PiiHLgcZJAoGOI_Y-sZDVI1ksB8p5r7riWM,3725
 nkululeko/nkululeko.py,sha256=Kn3s2E3yyH8cJ7z6lkMxrnqtCxTu7-qfe9Zr_ONTD5g,1968
-nkululeko/plots.py,sha256=nd9tF_61DyAx7oGZF8gTrHXazkgFjFe4eClxu1nQ_XU,23276
+nkululeko/plots.py,sha256=C2mwQFK0Vxfl5ZM7CO87tULDoEf7G16ek0nU77bhOc4,23070
 nkululeko/predict.py,sha256=sF091sSSLnEWcISx9ZcULLie3tY5XeFsQJd6b3vrxFg,2409
 nkululeko/resample.py,sha256=2d9eao_0sLrGZ_KSl8OVKsPor3BkFrlmMhrpB9WelIs,4267
 nkululeko/runmanager.py,sha256=eTM1DNQKt1lxYhzt4vZyZluPXW9sWlIJHNQzex4lkJU,7624
@@ -82,13 +82,13 @@ nkululeko/models/model_gmm.py,sha256=hZ9UO36KNf48qa3J-xkWIicIj9-TApmt21zNES2vEOs
 nkululeko/models/model_knn.py,sha256=KlnrJfwiVnmXZrAaYGFrKA2f5sznvTzSJQ8-5etOP0k,599
 nkululeko/models/model_knn_reg.py,sha256=j7YFfVm6xOR2d9yBYdQiwwqYfqkX0JynX_qLCvkr1fk,610
 nkululeko/models/model_lin_reg.py,sha256=0D7mSnSwK82lNWDMwHYRyq3FmGa6y-DHDGg4qUe85q4,422
-nkululeko/models/model_mlp.py,sha256=JtC83GYKtqCTW00rUm_xKSKjAsdMUAsqtnBfEFZBCwA,9854
+nkululeko/models/model_mlp.py,sha256=xMirtYax3bLBz_0kkC0M4Rc6-KQY05NNKHQGw7rbum8,9856
 nkululeko/models/model_mlp_regression.py,sha256=PO5qyfjgAJH8hawhmeXDaUThyXDYdM642dQHkO0NY7c,10204
 nkululeko/models/model_svm.py,sha256=rsME3KvKvNG7bdE5lbvYUu85WZhaASZxxmdNDIVJRZ4,940
 nkululeko/models/model_svr.py,sha256=_YZeksqB3eBENGlg3g9RwYFlk9rQQ-XCeNBKLlGGVoE,725
 nkululeko/models/model_tree.py,sha256=rf16faUm4o2LJgkoYpeY998b8DQIvXZ73_m1IS3TnnE,417
 nkululeko/models/model_tree_reg.py,sha256=IgQcPTE-304HQLYSKPF8Z4ot_Ur9dH01fZjS0nXke_M,428
-nkululeko/models/model_tuned.py,sha256=eiSKFmObn9_VNTqF1lZvWbyyWxvhy1PVjOiIcs3YiGA,18379
+nkululeko/models/model_tuned.py,sha256=xOoY5TROzzTVu3sDtlmEle3V1MAgpf8S3WxO9o4MzV4,20777
 nkululeko/models/model_xgb.py,sha256=Thgx5ESdIok4v72mKh4plxpo4smGcKALWNCJTDScY0M,447
 nkululeko/models/model_xgr.py,sha256=aGBtNGLWjOE_2rICGYGFxmT8DtnHYsIl1lIpMtghHsY,418
 nkululeko/reporting/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -96,7 +96,7 @@ nkululeko/reporting/defines.py,sha256=IsY1YgKRMaABpylVKjBJgJ5bNCEbGCVA_E6pivraqS
 nkululeko/reporting/latex_writer.py,sha256=qiCRSmB4KOD_za4oHu5x-PhwjZohzfo8wecMOwlXZwc,1886
 nkululeko/reporting/report.py,sha256=W0rcigDdjBvxZQ3pZja_gvToILYvaZ1BFtnN2qFRfYI,1060
 nkululeko/reporting/report_item.py,sha256=siWeGNgo4bAE46YBMNcsdf3jTMTy76BO9Fi6DTvDig4,533
-nkululeko/reporting/reporter.py,sha256=eLqwKEUTQ7v5CedzhZP2617qmXGcvi0rjyyFLOBdxtQ,12841
+nkululeko/reporting/reporter.py,sha256=NugmGmS3iwuBJ59jqyuTCKPRpiPLGhnz12z_nlVh69Y,13445
 nkululeko/reporting/result.py,sha256=nSN5or-Py2GPRWHkWpGRh7UCi1W0er7WLEHz8fYLk-A,742
 nkululeko/segmenting/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nkululeko/segmenting/seg_inaspeechsegmenter.py,sha256=pmLHuXsaqvcdYxB4PSW9l1mbQWZZBJFhi_CGabqydas,1947
@@ -105,8 +105,8 @@ nkululeko/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nkululeko/utils/files.py,sha256=UiGAtZRWYjHSvlmPaTMtzyNNGE6qaLaxQkybctS7iRM,4021
 nkululeko/utils/stats.py,sha256=1yUq0FTOyqkU8TwUocJRYdJaqMU5SlOBBRUun9STo2M,2829
 nkululeko/utils/util.py,sha256=mK1MgO14NinrPhavJw72eR_2WN_kBKjVKiEJnzvdO1Q,13946
-nkululeko-0.86.1.dist-info/LICENSE,sha256=0zGP5B_W35yAcGfHPS18Q2B8UhvLRY3dQq1MhpsJU_U,1076
-nkululeko-0.86.1.dist-info/METADATA,sha256=LXoMlzo5QBzABv0fpIDvf4nYDjCJkRCZL1XmffikrRc,37088
-nkululeko-0.86.1.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-nkululeko-0.86.1.dist-info/top_level.txt,sha256=DPFNNSHPjUeVKj44dVANAjuVGRCC3MusJ08lc2a8xFA,10
-nkululeko-0.86.1.dist-info/RECORD,,
+nkululeko-0.86.2.dist-info/LICENSE,sha256=0zGP5B_W35yAcGfHPS18Q2B8UhvLRY3dQq1MhpsJU_U,1076
+nkululeko-0.86.2.dist-info/METADATA,sha256=DmmpMrftBptpWqx7h9US7_4mvMIQbZ5ugzv_4kyBjkM,37170
+nkululeko-0.86.2.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+nkululeko-0.86.2.dist-info/top_level.txt,sha256=DPFNNSHPjUeVKj44dVANAjuVGRCC3MusJ08lc2a8xFA,10
+nkululeko-0.86.2.dist-info/RECORD,,