def yamnet_frames_model(feature_params):
"""Defines the YAMNet waveform-to-class-scores model.
Args:
feature_params: An object with parameter fields to control the feature
calculation.
Returns:
A model accepting (1, num_samples) waveform input and emitting a
(num_patches, num_classes) matrix of class scores per time frame as
well as a (num_spectrogram_frames, num_mel_bins) spectrogram feature
matrix.
"""
print('building frames model')
waveform = layers.Input(batch_shape=(1, None))
print('waveform input done')
# Store the intermediate spectrogram features to use in visualization.
spectrogram = features_lib.waveform_to_log_mel_spectrogram(
tf.squeeze(waveform, axis=0), feature_params)
print('spectrogram output done')
patches = features_lib.spectrogram_to_patches(spectrogram, feature_params)
predictions = yamnet(patches)
print('prediction output done')
frames_model = Model(name='yamnet_frames',
inputs=waveform,
outputs=[predictions, spectrogram])
print('model done')
return frames_model
def yamnet_frames_model(feature_params, fine_tuning=True):
"""Defines the YAMNet waveform-to-class-scores model.
Args:
feature_params: An object with parameter fields to control the feature
calculation.
Returns:
A model accepting (1, num_samples) waveform input and emitting a
(num_patches, num_classes) matrix of class scores per time frame as
well as a (num_spectrogram_frames, num_mel_bins) spectrogram feature
matrix.
"""
waveform = layers.Input(batch_shape=(None, None))
# Store the intermediate spectrogram features to use in visualization.
spectrogram = features_lib.waveform_to_log_mel_spectrogram(
tf.squeeze(waveform, axis=0), feature_params)
patches = features_lib.spectrogram_to_patches(spectrogram, feature_params)
predictions = yamnet(patches, fine_tuning)
if (not fine_tuning):
frames_model = Model(name='yamnet_frames',
inputs=waveform,
outputs=[predictions, spectrogram])
else:
output = layers.Dense(100, activation='relu')(predictions)
output = layers.Dense(4, activation='softmax')(output)
frames_model = Model(name='yamnet_frames',
inputs=waveform,
outputs=[output, spectrogram])
return frames_model
def yamnet_frames_tflite_model(feature_params):
"""Defines the YAMNet waveform-to-class-scores model,
suitable for tflite conversion.
Args:
feature_params: An object with parameter fields to control the feature
calculation.
Returns:
A model accepting (1, num_samples) waveform input and emitting a
(num_patches, num_classes) matrix of class scores per time frame as
well as a (num_spectrogram_frames, num_mel_bins) spectrogram feature
matrix.
"""
num_samples = int(round(params.SAMPLE_RATE * 0.975))
waveform = layers.Input(batch_shape=(1, num_samples))
# Store the intermediate spectrogram features to use in visualization.
spectrogram = features_tflite_lib.waveform_to_log_mel_spectrogram(
tf.squeeze(waveform, axis=0), feature_params)
patches = features_lib.spectrogram_to_patches(spectrogram, feature_params)
predictions = yamnet(patches)
frames_model = Model(name='yamnet_frames',
inputs=waveform,
outputs=[predictions, spectrogram])
return frames_model
def yamnet_frames_tflite_model(feature_params):
num_samples = int(round(params.SAMPLE_RATE * 0.975))
waveform = layers.Input(batch_shape=(1, num_samples))
# Store the intermediate spectrogram features to use in visualization.
spectrogram = features_lib.waveform_to_log_mel_spectrogram(
tf.squeeze(waveform, axis=0), feature_params)
patches = features_lib.spectrogram_to_patches(spectrogram, feature_params)
predictions = yamnet(patches)
frames_model = Model(name='yamnet_frames',
inputs=waveform,
outputs=[predictions, spectrogram])
return frames_model