def test_output_shape_is_correct(self):
loss_obj = losses.PretrainedCREPEEmbeddingLoss()
input_audio = tf.random.uniform((3, 16000), dtype=tf.float32)
target_audio = tf.random.uniform((3, 16000), dtype=tf.float32)
loss = loss_obj(input_audio, target_audio)
self.assertListEqual([], loss.shape.as_list())
def test_output_shape_is_correct(self):
loss_obj = losses.PretrainedCREPEEmbeddingLoss()
input_audio = tf.ones((3, 16000), dtype=tf.float32)
target_audio = tf.ones((3, 16000), dtype=tf.float32)
loss = loss_obj(input_audio, target_audio)
self.assertListEqual([], loss.shape.as_list())
self.assertTrue(np.isfinite(loss))
def setUp(self):
"""Create some dummy input data for the chain."""
super().setUp()
# Create a network output dictionary.
self.nn_outputs = {
'audio': tf.ones((3, 8000), dtype=tf.float32),
'audio_synth': tf.ones((3, 8000), dtype=tf.float32),
'magnitudes': tf.ones((3, 200, 2), dtype=tf.float32),
'f0_hz': 200 + tf.ones((3, 200, 1), dtype=tf.float32),
}
# Create Processors.
spectral_loss = losses.SpectralLoss()
crepe_loss = losses.PretrainedCREPEEmbeddingLoss(name='crepe_loss')
# Create DAG for testing.
self.dag = [
(spectral_loss, ['audio', 'audio_synth']),
(crepe_loss, ['audio', 'audio_synth']),
]
self.expected_outputs = ['spectral_loss', 'crepe_loss']