nkululeko 0.95.9__tar.gz → 0.96.0__tar.gz

{nkululeko-0.95.9 → nkululeko-0.96.0}/CHANGELOG.md

@@ -1,6 +1,10 @@
 Changelog
 =========
 
+Version 0.96.0 (25-07-14)
+--------------------------
+* added Bert feature extractor
+
 Version 0.95.9 (25-07-14)
 --------------------------
 * added google translation

{nkululeko-0.95.9/nkululeko.egg-info → nkululeko-0.96.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nkululeko
-Version: 0.95.9
+Version: 0.96.0
 Summary: Machine learning audio prediction experiments based on templates
 Home-page: https://github.com/felixbur/nkululeko
 Author: Felix Burkhardt

{nkululeko-0.95.9 → nkululeko-0.96.0}/nkululeko/constants.py

@@ -1,2 +1,2 @@
-VERSION="0.95.9"
+VERSION="0.96.0"
 SAMPLING_RATE = 16000

nkululeko-0.96.0/nkululeko/feat_extract/feats_bert.py

@@ -0,0 +1,105 @@
+import os
+
+import pandas as pd
+from tqdm import tqdm
+import transformers
+import torch 
+from transformers import BertTokenizer, BertModel
+
+from nkululeko.feat_extract.featureset import Featureset
+import nkululeko.glob_conf as glob_conf
+
+
+class Bert(Featureset):
+    """Class to extract bert embeddings"""
+
+    def __init__(self, name, data_df, feat_type):
+        """Constructor.
+
+        If_train is needed to distinguish from test/dev sets,
+        because they use the codebook from the training
+        """
+        super().__init__(name, data_df, feat_type)
+        cuda = "cuda" if torch.cuda.is_available() else "cpu"
+        self.device = self.util.config_val("MODEL", "device", cuda)
+        self.model_initialized = False
+        if feat_type == "bert":
+            self.feat_type = "bert-base-uncased"
+        else:
+            self.feat_type = feat_type
+
+    def init_model(self):
+        # load model
+        self.util.debug(f"loading {self.feat_type} model...")
+        model_path = self.util.config_val(
+            "FEATS", "bert.model", f"google-bert/{self.feat_type}"
+        )
+        config = transformers.AutoConfig.from_pretrained(model_path)
+        layer_num = config.num_hidden_layers
+        hidden_layer = int(self.util.config_val("FEATS", "bert.layer", "0"))
+        config.num_hidden_layers = layer_num - hidden_layer
+        self.util.debug(f"using hidden layer #{config.num_hidden_layers}")
+
+        self.tokenizer = BertTokenizer.from_pretrained(model_path)
+        self.model = BertModel.from_pretrained(model_path, config=config).to(
+            self.device
+        )
+        print(f"initialized {self.feat_type} model on {self.device}")
+        self.model.eval()
+        self.model_initialized = True
+
+    def extract(self):
+        """Extract the features or load them from disk if present."""
+        store = self.util.get_path("store")
+        storage = os.path.join(store, f"{self.name}.pkl")
+        extract = self.util.config_val("FEATS", "needs_feature_extraction", False)
+        no_reuse = eval(self.util.config_val("FEATS", "no_reuse", "False"))
+        if extract or no_reuse or not os.path.isfile(storage):
+            if not self.model_initialized:
+                self.init_model()
+            self.util.debug(
+                f"extracting {self.feat_type} embeddings, this might take a while..."
+            )
+            emb_series = pd.Series(index=self.data_df.index, dtype=object)
+            for idx, row in tqdm(self.data_df.iterrows(), total=len(self.data_df)):
+                file = idx[0]
+                text = row['text']
+                emb = self.get_embeddings(text, file)
+                emb_series[idx] = emb
+            # print(f"emb_series shape: {emb_series.shape}")
+            self.df = pd.DataFrame(emb_series.values.tolist(), index=self.data_df.index)
+            # print(f"df shape: {self.df.shape}")
+            self.df.to_pickle(storage)
+            try:
+                glob_conf.config["DATA"]["needs_feature_extraction"] = "false"
+            except KeyError:
+                pass
+        else:
+            self.util.debug(f"reusing extracted {self.feat_type} embeddings")
+            self.df = pd.read_pickle(storage)
+            if self.df.isnull().values.any():
+                self.util.error(
+                    f"got nan: {self.df.shape} {self.df.isnull().sum().sum()}"
+                )
+
+    def get_embeddings(self, text, file):
+        r"""Extract embeddings from raw audio signal."""
+        try:
+            with torch.no_grad():
+                inputs = self.tokenizer(text, return_tensors="pt")
+                outputs = self.model(**inputs)
+                # mean pooling 
+                y = torch.mean(outputs[0], dim=1)
+                y = y.ravel()
+        except RuntimeError as re:
+            print(str(re))
+            self.util.error(f"couldn't extract file: {file}")
+            y = None
+        if y is None:
+            return None
+        return y.detach().cpu().numpy()
+
+    def extract_sample(self, text):
+        self.init_model()
+        feats = self.get_embeddings(text, "no file")
+        return feats

{nkululeko-0.95.9 → nkululeko-0.96.0}/nkululeko/feature_extractor.py

@@ -80,7 +80,7 @@ class FeatureExtractor:
             return MLD_set
 
         elif feats_type.startswith(
-            ("wav2vec2", "hubert", "wavlm", "spkrec", "whisper", "ast", "emotion2vec")
+            ("bert", "wav2vec2", "hubert", "wavlm", "spkrec", "whisper", "ast", "emotion2vec")
         ):
             return self._get_feat_extractor_by_prefix(feats_type)
 

{nkululeko-0.95.9 → nkululeko-0.96.0/nkululeko.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nkululeko
-Version: 0.95.9
+Version: 0.96.0
 Summary: Machine learning audio prediction experiments based on templates
 Home-page: https://github.com/felixbur/nkululeko
 Author: Felix Burkhardt

{nkululeko-0.95.9 → nkululeko-0.96.0}/nkululeko.egg-info/SOURCES.txt

@@ -125,6 +125,7 @@ nkululeko/feat_extract/feats_analyser.py
 nkululeko/feat_extract/feats_ast.py
 nkululeko/feat_extract/feats_auddim.py
 nkululeko/feat_extract/feats_audmodel.py
+nkululeko/feat_extract/feats_bert.py
 nkululeko/feat_extract/feats_clap.py
 nkululeko/feat_extract/feats_emotion2vec.py
 nkululeko/feat_extract/feats_hubert.py

{nkululeko-0.95.9 → nkululeko-0.96.0}/pyproject.toml

@@ -7,7 +7,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "nkululeko"
-version = "0.95.9"
+version = "0.96.0"
 description = "Machine learning audio prediction experiments based on templates"
 authors = [
     {name = "Felix Burkhardt", email = "fxburk@gmail.com"},