refactor: move audio package from utils to top

setup.py

@@ -51,13 +51,13 @@ setup(
     license='MIT License',
     packages=[
         'spleeter',
+        'spleeter.audio',
         'spleeter.commands',
         'spleeter.model',
         'spleeter.model.functions',
         'spleeter.model.provider',
         'spleeter.resources',
         'spleeter.utils',
-        'spleeter.utils.audio',
     ],
     package_data={'spleeter.resources': ['*.json']},
     python_requires='>=3.6, <3.8',

spleeter/commands/separate.py

@@ -20,8 +20,8 @@ import tensorflow as tf
 import numpy as np
 # pylint: enable=import-error
 
-from ..utils.audio.adapter import get_audio_adapter
-from ..utils.audio.convertor import to_n_channels
+from ..audio.adapter import get_audio_adapter
+from ..audio.convertor import to_n_channels
 from ..utils.estimator import create_estimator
 from ..utils.tensor import set_tensor_shape
 

spleeter/commands/train.py

@@ -13,9 +13,9 @@ from functools import partial
 import tensorflow as tf
 # pylint: enable=import-error
 
+from ..audio.adapter import get_audio_adapter
 from ..dataset import get_training_dataset, get_validation_dataset
 from ..model import model_fn
-from ..utils.audio.adapter import get_audio_adapter
 from ..utils.logging import get_logger
 
 __email__ = 'research@deezer.com'

spleeter/dataset.py

@@ -2,15 +2,16 @@
 # coding: utf8
 
 """
-    Module for building data preprocessing pipeline using the tensorflow data
-    API.
-    Data preprocessing such as audio loading, spectrogram computation, cropping,
-    feature caching or data augmentation is done using a tensorflow dataset object
-    that output a tuple (input_, output) where:
-    - input_ is a dictionary with a single key that contains the (batched) mix
-    spectrogram of audio samples
-    - output is a dictionary of spectrogram of the isolated tracks (ground truth)
+    Module for building data preprocessing pipeline using the tensorflow
+    data API. Data preprocessing such as audio loading, spectrogram
+    computation, cropping, feature caching or data augmentation is done
+    using a tensorflow dataset object that output a tuple (input_, output)
+    where:
 
+    -   input is a dictionary with a single key that contains the (batched)
+        mix spectrogram of audio samples
+    -   output is a dictionary of spectrogram of the isolated tracks
+        (ground truth)
 """
 
 import time
@@ -23,10 +24,10 @@ import numpy as np
 import tensorflow as tf
 # pylint: enable=import-error
 
-from .utils.audio.convertor import (
+from .audio.convertor import (
     db_uint_spectrogram_to_gain,
     spectrogram_to_db_uint)
-from .utils.audio.spectrogram import (
+from .audio.spectrogram import (
     compute_spectrogram_tf,
     random_pitch_shift,
     random_time_stretch)

spleeter/separator.py

@@ -20,9 +20,9 @@ from multiprocessing import Pool
 from pathlib import Path
 from os.path import join
 
+from .audio.adapter import get_default_audio_adapter
+from .audio.convertor import to_stereo
 from .model import model_fn
-from .utils.audio.adapter import get_default_audio_adapter
-from .utils.audio.convertor import to_stereo
 from .utils.configuration import load_configuration
 from .utils.estimator import create_estimator, to_predictor