A lightweight yet powerful audio-to-MIDI converter with pitch bend detection
Python audiolightweightmachine-learningmidimusicpitch-detectionpolyphonicpythontranscriptiontypescript
convert tf.debugging.assert_none_equal to standard assert for single value checks, remove model train callback test bc basically duplicate with more work.
Benjie Genchele2f68fc27627
diff --git a/basic_pitch/data/tf_example_deserialization.py b/basic_pitch/data/tf_example_deserialization.pyindex acb2f99..ffaeb73 100644--- a/basic_pitch/data/tf_example_deserialization.py+++ b/basic_pitch/data/tf_example_deserialization.py@@ -73,10 +73,7 @@ def prepare_datasets(# check that the base dataset returned by ds_function is FINITEfor ds in [ds_train, ds_validation]:- tf.debugging.assert_none_equal(- tf.cast(tf.data.experimental.cardinality(ds), tf.int32),- tf.data.experimental.INFINITE_CARDINALITY,- )+ assert tf.cast(tf.data.experimental.cardinality(ds), tf.int32) != tf.data.experimental.INFINITE_CARDINALITY# training datasetif training_shuffle_buffer_size > 0:@@ -137,10 +134,7 @@ def prepare_visualization_datasets(# check that the base dataset returned by ds_function is FINITEfor ds in [ds_train, ds_validation]:- tf.debugging.assert_none_equal(- tf.cast(tf.data.experimental.cardinality(ds), tf.int32),- tf.data.experimental.INFINITE_CARDINALITY,- )+ assert tf.cast(tf.data.experimental.cardinality(ds), tf.int32) != tf.data.experimental.INFINITE_CARDINALITY# training datasetds_train = ds_train.repeat().batch(batch_size).prefetch(tf.data.AUTOTUNE)diff --git a/tests/test_callbacks.py b/tests/test_callbacks.pyindex bf3929e..e5f5788 100644--- a/tests/test_callbacks.py+++ b/tests/test_callbacks.py@@ -24,8 +24,6 @@ from typing import Dictfrom basic_pitch.callbacks import VisualizeCallbackfrom basic_pitch.constants import AUDIO_N_SAMPLES, ANNOTATIONS_N_SEMITONES, ANNOT_N_FRAMES-os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"-class MockModel(tf.keras.Model):def __init__(self) -> None:@@ -62,27 +60,3 @@ def test_visualize_callback_on_epoch_end(tmpdir: str) -> None:vc.model = MockModel()vc.on_epoch_end(1, {"loss": np.random.random(), "val_loss": np.random.random()})---def test_visualize_callback_on_epoch_end_with_model(tmpdir: str) -> None:- model = MockModel()- model.compile(optimizer="adam", loss="mse")-- batch_size = 2 # needs to be at least 2 bc validation_split required-- x_train = np.random.random((batch_size, AUDIO_N_SAMPLES, 1))- y_train = {- key: np.random.random((batch_size, ANNOTATIONS_N_SEMITONES, ANNOT_N_FRAMES))- for key in ["onset", "contour", "note"]- }-- vc = VisualizeCallback(- train_ds=create_mock_dataset(),- validation_ds=create_mock_dataset(),- tensorboard_dir=str(tmpdir),- original_validation_ds=create_mock_dataset(),- contours=True,- )-- history = model.fit(x_train, y_train, epochs=1, validation_split=0.5, callbacks=[vc], verbose=0)- assert history