def testBasicOneHotEncoderOctaveSquash(self):
steps_per_beat = 4
transpose_to_key = 0
min_note = 48
max_note = 84
num_classes = max_note - min_note + 2
melody = melodies_lib.Melody()
melody.from_event_list(
[NO_EVENT, 84, 86, 52, NO_EVENT, NOTE_OFF, 65, 67, NOTE_OFF, 69, 71, 72,
NO_EVENT, NOTE_OFF, 38, 40, NOTE_OFF])
transformed_melody = [NO_EVENT, 24, 26, 4, NO_EVENT, NOTE_OFF, 17, 19,
NOTE_OFF, 21, 23, 24, NO_EVENT, NOTE_OFF, 2, 4]
expected_inputs = [one_hot(note + 2, num_classes)
for note in transformed_melody]
expected_labels = ([note + 2 for note in transformed_melody[1:]] +
[NOTE_OFF + 2])
expected_sequence_example = make_sequence_example(expected_inputs,
expected_labels)
sequence_example, _ = sequence_to_melodies.basic_one_hot_encoder(
melody, steps_per_beat, min_note, max_note, transpose_to_key)
self.assertEqual(expected_sequence_example, sequence_example)
def testBasicOneHotEncoderTransposeKey(self):
steps_per_beat = 4
transpose_to_key = 0
min_note = 48
max_note = 84
num_classes = max_note - min_note + 2
melody = melodies_lib.Melody()
melody.from_event_list(
[NO_EVENT, 61, 63, 65, NO_EVENT, NOTE_OFF, 66, 68, NOTE_OFF, 70, 72, 73,
NO_EVENT, NOTE_OFF, 75, 77, NOTE_OFF])
transformed_melody = [NO_EVENT, 12, 14, 16, NO_EVENT, NOTE_OFF, 17, 19,
NOTE_OFF, 21, 23, 24, NO_EVENT, NOTE_OFF, 26, 28]
expected_inputs = [one_hot(note + 2, num_classes)
for note in transformed_melody]
expected_labels = ([note + 2 for note in transformed_melody[1:]] +
[NOTE_OFF + 2])
expected_sequence_example = make_sequence_example(expected_inputs,
expected_labels)
sequence_example, _ = sequence_to_melodies.basic_one_hot_encoder(
melody, steps_per_beat, min_note, max_note, transpose_to_key)
self.assertEqual(expected_sequence_example, sequence_example)
def testBasicOneHotEncoder(self):
steps_per_beat = 4
transpose_to_key = 0
min_note = 48
max_note = 84
num_classes = max_note - min_note + 2
melody = melodies_lib.Melody()
melody.from_event_list(
[NO_EVENT, 60, 62, 64, NO_EVENT, NOTE_OFF, 65, 67, NOTE_OFF, 69, 71, 72,
NO_EVENT, NOTE_OFF, 74, 76, 77, 79, NO_EVENT, NOTE_OFF])
transformed_melody = [NO_EVENT, 12, 14, 16, NO_EVENT, NOTE_OFF, 17, 19,
NOTE_OFF, 21, 23, 24, NO_EVENT, NOTE_OFF, 26, 28,
29, 31, NO_EVENT, NOTE_OFF]
expected_inputs = ([one_hot(note + 2, num_classes)
for note in transformed_melody] +
[one_hot(0, num_classes)] * 12)
expected_labels = [note + 2 for note in transformed_melody[1:]] + [0] * 13
expected_sequence_example = make_sequence_example(expected_inputs,
expected_labels)
sequence_example, _ = sequence_to_melodies.basic_one_hot_encoder(
melody, steps_per_beat, min_note, max_note, transpose_to_key)
self.assertEqual(expected_sequence_example, sequence_example)
def sampler_loop(graph, decoder, checkpoint_dir, primer, num_gen_steps):
"""Generate many melodies simulatneously given a primer.
Generate melodies by sampling from model output and feeding it back into
the model as input at every step.
Args:
graph: A tf.Graph instance containing the graph to sample from.
decoder: A function that converts model output and reconstruction data into
a Melody object. The method takes two inputs: A list of integers which
are the softmax classes chosen at each step, and reconstruction data
returned by the encoder. It returns a melodies_lib.Melody.
checkpoint_dir: Directory to look for most recent model checkpoint in.
primer: A Melody object.
num_gen_steps: How many time steps to generate.
Returns:
List of generated melodies, each as a Melody object.
"""
softmax = graph.get_collection('softmax')[0]
initial_state = graph.get_collection('initial_state')[0]
final_state = graph.get_collection('final_state')[0]
melody_sequence = graph.get_collection('melody_sequence')[0]
lengths = graph.get_collection('lengths')[0]
with graph.as_default():
saver = tf.train.Saver()
session = tf.Session(graph=graph)
logging.info('Checkpoint dir: %s', checkpoint_dir)
checkpoint_file = tf.train.latest_checkpoint(checkpoint_dir)
saver.restore(session, checkpoint_file)
batch_size = softmax.get_shape()[0].value
# Convert primer Melody to model inputs.
sequence_example, encoder_information = sequence_to_melodies.basic_one_hot_encoder(primer)
primer_input = [
list(i.float_list.value)
for i in sequence_example.feature_lists.feature_list['inputs'].feature]
# Run model over primer sequence.
primer_input_batch = np.tile([primer_input], (batch_size, 1, 1))
state = session.run(
final_state,
feed_dict={initial_state: np.zeros(initial_state.get_shape().as_list()),
melody_sequence: primer_input_batch,
lengths: np.full(batch_size, len(primer),
dtype=int)})
# Sample from model repeatedly to generate melodies.
generated_sequences = [list() for i in xrange(batch_size)]
last_outputs = [melody_sequence] * batch_size
singleton_lengths = np.full(batch_size, 1, dtype=int)
for i in xrange(num_gen_steps):
input_batch = np.transpose(
[make_onehot(last_outputs, basic_rnn_ops.NUM_CLASSES)], (1, 0, 2))
state, batch_softmax = session.run(
[final_state, softmax],
feed_dict={initial_state: state,
melody_sequence: input_batch,
lengths: singleton_lengths})
last_outputs = [
np.random.choice(basic_rnn_ops.NUM_CLASSES, p=p_dist.flatten())
for p_dist in batch_softmax]
for generated_seq, next_output in zip(generated_sequences, last_outputs):
generated_seq.append(next_output)
primer_event_list = list(primer)
generated_melodies = []
for seq in generated_sequences:
melody = melodies_lib.Melody(steps_per_bar=primer.steps_per_bar)
melody.from_event_list(
primer_event_list + list(decoder(seq, encoder_information)))
generated_melodies.append(melody)
return generated_melodies
