fixing tests

.github/environment-ci.yml

@@ -27,8 +27,8 @@ dependencies:
   - pip:
     - "keras<3.0.0"
     - "tensorflow>=2.12.0,<2.16"
-    - "torch==2.0.0"
-    - "torchaudio==2.0.1"
+    - "torch==1.13"
+    - "torchaudio"
     - "essentia"
     - "soundfile>=0.12.1"
     - "opencv-python~=4.6.0"

compiam/separation/singing_voice_extraction/cold_diff_sep/__init__.py

@@ -94,6 +94,7 @@ def separate(
         clusters=5,
         scheduler=4,
         chunk_size=3,
+        normalize_input=True,
         gpu="-1",
     ):
         """Separate singing voice from mixture.
@@ -103,6 +104,8 @@ def separate(
             relevant if the input is an array of data instead of a filepath.
         :param clusters: Number of clusters to use to build the separation masks.
         :param scheduler: Scheduler factor to weight the clusters to be more or less restirctive with the interferences.
+        :param chunk_size: Size of the chunks to process the audio signal.
+        :param normalize_input: Normalize the input audio signal.
         :param gpu: Id of the available GPU to use (-1 by default, to run on CPU), use string: '0', '1', etc.
         :return: Singing voice signal.
         """
@@ -153,6 +156,12 @@ def separate(
                 f"Downsampling to mono... your audio is stereo, \
                     and the model is trained on mono audio."
             )
+
+        if normalize_input:
+            # Normalizing audio for better performance overall
+            mean = tf.reduce_mean(mixture, keepdims=True)
+            std = tf.math.reduce_std(mixture, keepdims=True)
+            mixture = (mixture - mean) / (1e-6 + std)
 
         output_voc = np.zeros(mixture.shape)
         hopsized_chunk = int((chunk_size * self.sample_rate) / 2)

tests/melody/test_deepsrgm.py

@@ -43,8 +43,11 @@ def _get_features():
         feat = deepsrgm.get_features(
             os.path.join(TESTDIR, "resources", "melody", "pitch_test.wav")
         )
-    feat_1 = deepsrgm.get_features(np.zeros(44100))
-    feat_2 = deepsrgm.get_features(
+    feat_1 = deepsrgm.get_features(np.zeros([44100]))
+    feat_2 = deepsrgm.get_features(np.zeros([1, 44100]))
+    feat_3 = deepsrgm.get_features(np.zeros([2, 44100]))
+    feat_3 = deepsrgm.get_features(np.zeros([44100, 2]))
+    feat_4 = deepsrgm.get_features(
         os.path.join(TESTDIR, "resources", "melody", "pitch_test.wav")
     )
 

tests/melody/test_essentia_extractors.py

@@ -1,5 +1,6 @@
 import os
 import pytest
+import librosa
 
 import numpy as np
 
@@ -15,9 +16,9 @@ def _predict_normalized_pitch():
     pitch = melodia.extract(
         os.path.join(TESTDIR, "resources", "melody", "pitch_test.wav")
     )
-    pitch_2 = melodia.extract(np.zeros(44100))
-    pitch_3 = melodia.extract(np.zeros(2, 44100))  # Testing input array
-    pitch_4 = melodia.extract(np.zeros(44100, 2))  # Testing input array
+    pitch_2 = melodia.extract(np.zeros([44100]))
+    pitch_3 = melodia.extract(np.zeros([2, 44100]))  # Testing input array
+    pitch_4 = melodia.extract(np.zeros([44100, 2]))  # Testing input array
 
     assert isinstance(pitch, np.ndarray)
     assert np.shape(pitch) == (699, 2)
@@ -70,11 +71,10 @@ def _predict_normalized_pitch():
     tonic = tonic_multipitch.extract(
         os.path.join(TESTDIR, "resources", "melody", "pitch_test.wav")
     )
-    tonic_2 = tonic_multipitch.extract(np.zeros(44100))  # Testing input array
-    tonic_3 = tonic_multipitch.extract(np.zeros(2, 44100))  # Testing input array
-    tonic_4 = tonic_multipitch.extract(np.zeros(44100, 2))  # Testing input array
-
-
+    audio = librosa.load(os.path.join(TESTDIR, "resources", "melody", "pitch_test.wav"), sr=44100)[0]
+    tonic_2 = tonic_multipitch.extract(audio)  # Testing input array
+    tonic_3 = tonic_multipitch.extract(np.stack([audio, audio]))  # Testing input array
+    tonic_4 = tonic_multipitch.extract(np.stack([audio, audio]).T)  # Testing input array
 
     assert isinstance(tonic, float)
     assert tonic == 157.64892578125