Python 3.12 / TensorFlow 2.16+ (Keras 3) Support

.github/workflows/tox.yml

@@ -14,15 +14,16 @@ jobs:
           - ubuntu-latest
           - windows-latest
         py:
+          - "3.12"
           - "3.11"
           - "3.10"
-          - "3.9"
-          - "3.8"
         include:
           - os: macos-latest-xlarge
             py: "3.10.11"
           - os: macos-latest-xlarge
             py: "3.11.8"
+          - os: macos-latest-xlarge
+            py: "3.12"
     steps:
       - name: Setup python for test ${{ matrix.py }}
         uses: actions/setup-python@v4

README.md

@@ -44,7 +44,7 @@ To update Basic Pitch to the latest version, add `--upgrade` to the above comman
 
 #### Compatible Environments:
 - MacOS, Windows and Ubuntu operating systems
-- Python versions 3.7, 3.8, 3.9, 3.10, 3.11
+- Python versions 3.10, 3.11, 3.12
 - **For Mac M1 hardware, we currently only support python version 3.10. Otherwise, we suggest using a virtual machine.**
 
 

basic_pitch/__init__.py

@@ -19,7 +19,6 @@
 import logging
 import pathlib
 
-
 try:
     import coremltools
 
@@ -79,7 +78,8 @@ class FilenameSuffix(enum.Enum):
 
 
 if TF_PRESENT:
-    _default_model_type = FilenameSuffix.tf
+    _tf_major_minor = tuple(int(part) for part in tensorflow.__version__.split(".")[:2])
+    _default_model_type = FilenameSuffix.tf if _tf_major_minor < (2, 16) else FilenameSuffix.tflite
 elif CT_PRESENT:
     _default_model_type = FilenameSuffix.coreml
 elif TFLITE_PRESENT:

basic_pitch/constants.py

@@ -19,7 +19,6 @@
 
 from enum import Enum
 
-
 SEMITONES_PER_OCTAVE = 12  # for frequency bin calculations
 
 FFT_HOP = 256

basic_pitch/inference.py

@@ -39,8 +39,11 @@
 try:
     import tflite_runtime.interpreter as tflite
 except ImportError:
-    if TF_PRESENT:
-        import tensorflow.lite as tflite
+    try:
+        from ai_edge_litert import interpreter as tflite
+    except ImportError:
+        if TF_PRESENT:
+            import tensorflow.lite as tflite
 
 try:
     import onnxruntime as ort

basic_pitch/layers/signal.py

@@ -80,7 +80,7 @@ def padded_window(window_length: int, dtype: tf.dtypes.DType = tf.float32) -> tf
             self.spec = tf.keras.layers.Lambda(
                 lambda x: tf.pad(
                     x,
-                    [[0, 0] for _ in range(input_shape.rank - 1)] + [[self.fft_length // 2, self.fft_length // 2]],
+                    [[0, 0] for _ in range(len(input_shape) - 1)] + [[self.fft_length // 2, self.fft_length // 2]],
                     mode=self.pad_mode,
                 )
             )
@@ -161,7 +161,7 @@ class NormalizedLog(tf.keras.layers.Layer):
 
     def build(self, input_shape: tf.Tensor) -> None:
         self.squeeze_batch = lambda batch: batch
-        rank = input_shape.rank
+        rank = len(input_shape)
         if rank == 4:
             assert input_shape[1] == 1, "If the rank is 4, the second dimension must be length 1"
             self.squeeze_batch = lambda batch: tf.squeeze(batch, axis=1)

basic_pitch/models.py

@@ -32,6 +32,7 @@
 from basic_pitch.layers import signal, nnaudio
 
 tfkl = tf.keras.layers
+keras_ops = tf.keras.ops if hasattr(tf.keras, "ops") else tf
 
 MAX_N_SEMITONES = int(np.floor(12.0 * np.log2(0.5 * AUDIO_SAMPLE_RATE / ANNOTATIONS_BASE_FREQUENCY)))
 
@@ -184,7 +185,7 @@ def get_cqt(inputs: tf.Tensor, n_harmonics: int, use_batchnorm: bool) -> tf.Tens
         bins_per_octave=12 * CONTOURS_BINS_PER_SEMITONE,
     )(x)
     x = signal.NormalizedLog()(x)
-    x = tf.expand_dims(x, -1)
+    x = keras_ops.expand_dims(x, -1)
     if use_batchnorm:
         x = tfkl.BatchNormalization()(x)
     return x
@@ -263,7 +264,7 @@ def model(
         x_contours = nn.FlattenFreqCh(name=contour_name)(x_contours)  # contour output
 
         # reduced contour output as input to notes
-        x_contours_reduced = tf.expand_dims(x_contours, -1)
+        x_contours_reduced = keras_ops.expand_dims(x_contours, -1)
     else:
         x_contours_reduced = x_contours
 

basic_pitch/nn.py

@@ -18,7 +18,6 @@
 from typing import Any, List
 
 import tensorflow as tf
-import tensorflow.keras.backend as K
 
 from basic_pitch.layers.math import log_base_b
 
@@ -97,7 +96,7 @@ class FlattenAudioCh(tf.keras.layers.Layer):
 
     def call(self, x: tf.Tensor) -> tf.Tensor:
         """x: (batch, time, ch)"""
-        shapes = K.int_shape(x)
+        shapes = x.shape
         tf.assert_equal(shapes[2], 1)
         return tf.squeeze(x, axis=2)
 
@@ -111,7 +110,7 @@ class FlattenFreqCh(tf.keras.layers.Layer):
     """
 
     def call(self, x: tf.Tensor) -> tf.Tensor:
-        shapes = K.int_shape(x)
+        shapes = x.shape
         batch_size = tf.shape(x)[0]
         time_dim = shapes[1]
         freq_dim = shapes[2]

basic_pitch/predict.py

@@ -27,7 +27,6 @@
 )
 from basic_pitch.inference import Model
 
-
 os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
 
 

basic_pitch/train.py

@@ -150,7 +150,6 @@ def main(
     model.compile(
         loss=loss,
         optimizer=tf.keras.optimizers.Adam(learning_rate),
-        sample_weight_mode={"contour": None, "note": None, "onset": None},
     )
 
     logging.info("--- Model Training specs ---")

pyproject.toml

@@ -4,32 +4,35 @@ version = "0.4.0"
 description = "Basic Pitch, a lightweight yet powerful audio-to-MIDI converter with pitch bend detection."
 readme = "README.md"
 keywords = []
+requires-python = ">=3.10"
 classifiers = [
 	"Development Status :: 5 - Production/Stable",
 	"Natural Language :: English",
 	"Operating System :: POSIX :: Linux",
 	"Operating System :: MacOS :: MacOS X",
 	"Operating System :: Microsoft :: Windows",
 	"Programming Language :: Python",
-	"Programming Language :: Python :: 3.8",
-	"Programming Language :: Python :: 3.9",
 	"Programming Language :: Python :: 3.10",
 	"Programming Language :: Python :: 3.11",
+	"Programming Language :: Python :: 3.12",
 	"Programming Language :: Python :: Implementation :: CPython",
 ]
 dependencies = [
 	"coremltools; platform_system == 'Darwin'",
-	"librosa>=0.8.0",
+	"librosa>=0.10.0",
 	"mir_eval>=0.6.0",
-	"numpy>=1.18",
+	"numpy>=1.21",
 	"onnxruntime; platform_system == 'Windows' and python_version < '3.11'",
+	"onnxruntime; platform_system == 'Windows' and python_version >= '3.12'",
 	"pretty_midi>=0.2.9",
-	"resampy>=0.2.2,<0.4.3",
+	"resampy>=0.2.2",
 	"scikit-learn",
-	"scipy>=1.4.1",
-	"tensorflow>=2.4.1,<2.15.1; platform_system != 'Darwin' and python_version >= '3.11'",
-	"tensorflow-macos>=2.4.1,<2.15.1; platform_system == 'Darwin' and python_version > '3.11'",
+	"scipy>=1.7.0",
+	"tensorflow>=2.4.1,<2.16.0; platform_system != 'Darwin' and python_version >= '3.11' and python_version < '3.12'",
+	"tensorflow>=2.16.0; platform_system != 'Darwin' and python_version >= '3.12'",
+	"tensorflow>=2.13.0; platform_system == 'Darwin' and python_version >= '3.12'",
 	"tflite-runtime; platform_system == 'Linux' and python_version < '3.11'",
+	"ai-edge-litert; platform_system != 'Windows' and python_version >= '3.12'",
 	"typing_extensions",
 ]
 
@@ -68,8 +71,10 @@ test = [
 	"mido"
 ]
 tf = [
-	"tensorflow>=2.4.1,<2.15.1; platform_system != 'Darwin'",
-	"tensorflow-macos>=2.4.1,<2.15.1; platform_system == 'Darwin' and python_version > '3.7'",
+	"tensorflow>=2.4.1,<2.16.0; platform_system != 'Darwin' and python_version < '3.12'",
+	"tensorflow>=2.16.0; platform_system != 'Darwin' and python_version >= '3.12'",
+	"tensorflow>=2.13.0; platform_system == 'Darwin'",
+	"tensorboard>=2.4.1",
 ]
 coreml = ["coremltools"] 
 onnx = ["onnxruntime"]

tests/data/conftest.py

@@ -12,21 +12,21 @@
 SLAKH_TEST_INDEX = json.load(open(RESOURCES_PATH / "data" / "slakh" / "dummy_index.json"))
 
 
-@pytest.fixture  # type: ignore[misc]
+@pytest.fixture  # type: ignore[untyped-decorator]
 def mock_slakh_index() -> None:  # type: ignore[misc]
     with mock.patch("mirdata.datasets.slakh.Dataset.download"):
         with mock.patch("mirdata.datasets.slakh.Dataset._index", new=SLAKH_TEST_INDEX):
             yield
 
 
-@pytest.fixture  # type: ignore[misc]
+@pytest.fixture  # type: ignore[untyped-decorator]
 def mock_medleydb_pitch_index() -> None:  # type: ignore[misc]
     with mock.patch("mirdata.datasets.medleydb_pitch.Dataset.download"):
         with mock.patch("mirdata.datasets.medleydb_pitch.Dataset._index", new=MEDLEYDB_PITCH_TEST_INDEX):
             yield
 
 
-@pytest.fixture  # type: ignore[misc]
+@pytest.fixture  # type: ignore[untyped-decorator]
 def mock_maestro_index() -> None:  # type: ignore[misc]
     index_with_metadata = MAESTRO_TEST_INDEX
     metadata = {mdata["midi_filename"].split(".")[0]: mdata for mdata in METADATA_TEST_INDEX}
@@ -36,14 +36,14 @@ def mock_maestro_index() -> None:  # type: ignore[misc]
                 yield
 
 
-@pytest.fixture  # type: ignore[misc]
+@pytest.fixture  # type: ignore[untyped-decorator]
 def mock_guitarset_index() -> None:  # type: ignore[misc]
     with mock.patch("mirdata.datasets.guitarset.Dataset.download"):
         with mock.patch("mirdata.datasets.guitarset.Dataset._index", new=GUITAR_SET_TEST_INDEX):
             yield
 
 
-@pytest.fixture  # type: ignore[misc]
+@pytest.fixture  # type: ignore[untyped-decorator]
 def mock_ikala_index() -> None:  # type: ignore[misc]
     with mock.patch("mirdata.datasets.ikala.Dataset.download"):
         with mock.patch("mirdata.datasets.ikala.Dataset._index", new=IKALA_TEST_INDEX):

tests/data/test_medleydb_pitch.py

@@ -25,7 +25,6 @@
     create_input_data,
 )
 
-
 # TODO: Create test_medleydb_pitch_to_tf_example
 
 

tests/test_callbacks.py

@@ -56,6 +56,6 @@ def test_visualize_callback_on_epoch_end(tmpdir: str) -> None:
         contours=True,
     )
 
-    vc.model = MockModel()
+    vc.set_model(MockModel())
 
     vc.on_epoch_end(1, {"loss": np.random.random(), "val_loss": np.random.random()})

tests/test_inference.py

@@ -58,7 +58,7 @@ def test_predict() -> None:
 
     # Check that model output has the expected length according to the last frame second computed downstream
     # (via model_frames_to_time) with to a few frames of tolerance
-    audio_length_s = librosa.get_duration(filename=test_audio_path)
+    audio_length_s = librosa.get_duration(path=test_audio_path)
     n_model_output_frames = model_output["note"].shape[0]
     last_frame_s = note_creation.model_frames_to_time(n_model_output_frames)[-1]
     np.testing.assert_allclose(last_frame_s, audio_length_s, atol=2 * ANNOTATION_HOP)

tests/test_nn.py

@@ -67,7 +67,6 @@ def test_defaults(self) -> None:
         model.compile(
             loss="binary_crossentropy",
             optimizer=tf.keras.optimizers.Adam(0.1),
-            sample_weight_mode=None,
         )
 
         model.fit(
@@ -102,7 +101,6 @@ def test_fractions(self) -> None:
         model.compile(
             loss="binary_crossentropy",
             optimizer=tf.keras.optimizers.Adam(0.1),
-            sample_weight_mode=None,
         )
 
         model.fit(

tox.ini

@@ -1,5 +1,5 @@
 [tox]
-envlist = py38,py39,py310,py311,full,data,manifest,check-formatting,lint,mypy
+envlist = py310,py311,py312,full,data,manifest,check-formatting,lint,mypy
 skipsdist = True
 usedevelop = True
 requires =
@@ -19,6 +19,7 @@ setenv =
 
 # Tests with TF
 [testenv:full]
+basepython = python3.12
 deps = -e .[dev]
 commands =
     pytest tests {posargs}
@@ -28,8 +29,9 @@ setenv =
     LINE_LENGTH = "120"
 
 [testenv:data]
+basepython = python3.12
 deps = -e .[data]
-commands = 
+commands =
     pytest tests/data {posargs}
 
 [testenv:manifest]
@@ -38,27 +40,27 @@ skip_install = true
 commands = check-manifest --ignore 'tests'
 
 [testenv:check-formatting]
-basepython = python3.10
+basepython = python3.12
 deps = black
 skip_install = true
 commands =
-    black {env:SOURCE} tests --line-length {env:LINE_LENGTH} --diff --check
+    black {env:SOURCE} tests --line-length {env:LINE_LENGTH} --target-version py312 --diff --check
 
 [testenv:format]
-basepython = python3.10
+basepython = python3.12
 deps = black
 skip_install = true
 commands =
-    black {env:SOURCE} tests --line-length {env:LINE_LENGTH}
+    black {env:SOURCE} tests --line-length {env:LINE_LENGTH} --target-version py312
 
 [testenv:lint]
-basepython = python3.10
+basepython = python3.12
 deps = flake8
 skip_install = true
 commands = flake8
 
 [testenv:mypy]
-basepython = python3.10
+basepython = python3.12
 deps =
     mypy
 commands = mypy basic_pitch tests --strict --ignore-missing-imports --allow-subclassing-any