A lightweight yet powerful audio-to-MIDI converter with pitch bend detection
Python audiolightweightmachine-learningmidimusicpitch-detectionpolyphonicpythontranscriptiontypescript
Add TFLite serialized model
David Rubinstein5ceb6450f265
diff --git a/basic_pitch/__init__.py b/basic_pitch/__init__.pyindex d8b78b2..64761f2 100644--- a/basic_pitch/__init__.py+++ b/basic_pitch/__init__.py@@ -38,6 +38,19 @@ except ImportError:"reinstall basic-pitch with `pip install 'basic-pitch[tf]'`")+try:+ import tflite_runtime.interpreter as tflite++ TFLITE_PRESENT = True+except ImportError:+ TFLITE_PRESENT = False+ logging.warning(+ "tflite-runtime is not installed. "+ "If you plan to use a TFLite Model, "+ "reinstall basic-pitch with `pip install 'basic-pitch tflite-runtime'` or "+ "`pip install 'basic-pitch[tf]'"+ )+try:import onnxruntime as ort@@ -52,6 +65,8 @@ except ImportError:if TF_PRESENT:_filename = "nmp"+elif TFLITE_PRESENT:+ _filename = "nmp.tflite"elif ONNX_PRESENT:_filename = "nmp.onnx"diff --git a/basic_pitch/inference.py b/basic_pitch/inference.pyindex a116059..36c130e 100644--- a/basic_pitch/inference.py+++ b/basic_pitch/inference.py@@ -54,7 +54,7 @@ from basic_pitch.constants import (ANNOTATIONS_FPS,FFT_HOP,)-from basic_pitch import ONNX_PRESENT, TF_PRESENT+from basic_pitch import ONNX_PRESENT, TF_PRESENT, TFLITE_PRESENTfrom basic_pitch.commandline_printing import (generating_file_message,no_tf_warnings,@@ -67,6 +67,7 @@ import basic_pitch.note_creation as inferclass Model:class MODEL_TYPES(enum.Enum):TENSORFLOW = enum.auto()+ TFLITE = enum.auto()ONNX = enum.auto()def __init__(self, model_path: Union[pathlib.Path, str]):@@ -78,6 +79,17 @@ class Model:except Exception:pass++ if TFLITE_PRESENT:+ try:+ self.model_type = Model.MODEL_TYPES.TFLITE+ self.interpreter = tflite.Interpreter(model_path=model_path)+ self.interpreter.allocate_tensors()+ self.model = self.interpreter.get_signature_runner()+ return+ except Exception:+ pass+if ONNX_PRESENT:try:self.model_type = Model.MODEL_TYPES.ONNX@@ -95,6 +107,8 @@ class Model:def predict(self, x: npt.NDArray[np.float32]) -> Dict[str, npt.NDArray[np.float32]]:if self.model_type == Model.MODEL_TYPES.TENSORFLOW:return {k: v.numpy() for k, v in cast(tf.keras.Model, self.model(x)).items()}+ elif self.model_type == Model.MODEL_TYPES.TFLITE:+ return cast(tflite.Interpreter, self.model)(x)elif self.model_type == Model.MODEL_TYPES.ONNX:return cast(ort.InferenceSession, self.model).run(["note", "onset", "contour"], {"input": x})diff --git a/basic_pitch/predict.py b/basic_pitch/predict.pyindex ecdbf5d..dfec692 100644--- a/basic_pitch/predict.py+++ b/basic_pitch/predict.py@@ -38,7 +38,7 @@ def main() -> None:default=ICASSP_2022_MODEL_PATH,help="path to the saved model directory. Defaults to a ICASSP 2022 model. ""The preferred model is determined by the first library available in "- "[tensorflow, onnx]",+ "[tensorflow, tensorflow-lite, onnx]",)parser.add_argument("--save-midi",diff --git a/basic_pitch/saved_models/icassp_2022/nmp.tflite b/basic_pitch/saved_models/icassp_2022/nmp.tflitenew file mode 100644index 0000000..85a41beBinary files /dev/null and b/basic_pitch/saved_models/icassp_2022/nmp.tflite differdiff --git a/setup.cfg b/setup.cfgindex 756c6d3..a06c8e4 100644--- a/setup.cfg+++ b/setup.cfg@@ -40,6 +40,7 @@ install_requires =pretty_midi>=0.2.9resampy>=0.2.2scipy>=1.4.1+ tflite-runtime>=2.4.1; platform_system == "Linux"typing_extensions[options.entry_points]