def __init__(self, path='onsets-frames'):
"""Load the Onset-Frames Model (arXiv:1710.11153 [cs.SD]) and pretrained weights from the path using tensorflow and magenta.
Args:
path (str, optional): The path the model weights. Defaults to 'onsets-frames'.
"""
tf.disable_eager_execution()
tf.disable_v2_behavior()
tf.logging.set_verbosity(tf.logging.ERROR)
self.config = configs.CONFIG_MAP['onsets_frames']
self.hparams = self.config.hparams
self.hparams.use_cudnn = False
self.hparams.batch_size = 1
self.checkpoint_dir = path
self.examples = tf.placeholder(tf.string, [None])
self.dataset = data.provide_batch(examples=self.examples,
preprocess_examples=True,
params=self.hparams,
is_training=False,
shuffle_examples=False,
skip_n_initial_records=0)
self.estimator = train_util.create_estimator(self.config.model_fn,
self.checkpoint_dir,
self.hparams)
self.iterator = tf.data.make_initializable_iterator(self.dataset)
self.next_record = self.iterator.get_next()
def run(argv, config_map, data_fn):
"""Create transcriptions."""
tf.logging.set_verbosity(FLAGS.log)
config = config_map[FLAGS.config]
hparams = config.hparams
# For this script, default to not using cudnn.
hparams.use_cudnn = False
hparams.parse(FLAGS.hparams)
hparams.batch_size = 1
hparams.truncated_length_secs = 0
with tf.Graph().as_default():
examples = tf.placeholder(tf.string, [None])
dataset = data_fn(examples=examples,
preprocess_examples=True,
params=hparams,
is_training=False,
shuffle_examples=False,
skip_n_initial_records=0)
estimator = train_util.create_estimator(
config.model_fn, os.path.expanduser(FLAGS.model_dir), hparams)
iterator = dataset.make_initializable_iterator()
next_record = iterator.get_next()
with tf.Session() as sess:
sess.run([
tf.initializers.global_variables(),
tf.initializers.local_variables()
])
for filename in argv[1:]:
tf.logging.info('Starting transcription for %s...', filename)
# The reason we bounce between two Dataset objects is so we can use
# the data processing functionality in data.py without having to
# construct all the Example protos in memory ahead of time or create
# a temporary tfrecord file.
tf.logging.info('Processing file...')
sess.run(
iterator.initializer, {
examples: [
create_example(filename,
FLAGS.load_audio_with_librosa)
]
})
def transcription_data(params):
del params
return tf.data.Dataset.from_tensors(sess.run(next_record))
input_fn = infer_util.labels_to_features_wrapper(
transcription_data)
tf.logging.info('Running inference...')
checkpoint_path = None
if FLAGS.checkpoint_path:
checkpoint_path = os.path.expanduser(FLAGS.checkpoint_path)
prediction_list = list(
estimator.predict(input_fn,
checkpoint_path=checkpoint_path,
yield_single_examples=False))
assert len(prediction_list) == 1
sequence_prediction = music_pb2.NoteSequence.FromString(
prediction_list[0]['sequence_predictions'][0])
midi_filename = filename + FLAGS.transcribed_file_suffix + '.midi'
midi_io.sequence_proto_to_midi_file(sequence_prediction,
midi_filename)
tf.logging.info('Transcription written to %s.', midi_filename)
def transcribe(audio, sr, cuda=False):
"""
Google sucks and want to use audio path (raw wav) instead of decoded
samples loosing in decoupling between file format and DSP
input audio and sample rate, output mat like asmd with (pitch, ons, offs, velocity)
"""
# simple hack because google sucks... in this way we can accept audio data
# already loaded and keep our reasonable interface (and decouple i/o
# from processing)
original_google_sucks = audio_io.wav_data_to_samples
audio_io.wav_data_to_samples = google_sucks
audio = np.array(audio)
config = configs.CONFIG_MAP['onsets_frames']
hparams = config.hparams
hparams.use_cudnn = cuda
hparams.batch_size = 1
examples = tf.placeholder(tf.string, [None])
dataset = data.provide_batch(examples=examples,
preprocess_examples=True,
params=hparams,
is_training=False,
shuffle_examples=False,
skip_n_initial_records=0)
estimator = train_util.create_estimator(config.model_fn, CHECKPOINT_DIR,
hparams)
iterator = dataset.make_initializable_iterator()
next_record = iterator.get_next()
example_list = list(
audio_label_data_utils.process_record(wav_data=audio,
sample_rate=sr,
ns=music_pb2.NoteSequence(),
example_id="fakeid",
min_length=0,
max_length=-1,
allow_empty_notesequence=True,
load_audio_with_librosa=False))
assert len(example_list) == 1
to_process = [example_list[0].SerializeToString()]
sess = tf.Session()
sess.run([
tf.initializers.global_variables(),
tf.initializers.local_variables()
])
sess.run(iterator.initializer, {examples: to_process})
def transcription_data(params):
del params
return tf.data.Dataset.from_tensors(sess.run(next_record))
# put back the original function (it still writes and reload... stupid
# though
audio_io.wav_data_to_samples = original_google_sucks
input_fn = infer_util.labels_to_features_wrapper(transcription_data)
prediction_list = list(
estimator.predict(input_fn, yield_single_examples=False))
assert len(prediction_list) == 1
notes = music_pb2.NoteSequence.FromString(
prediction_list[0]['sequence_predictions'][0]).notes
out = np.empty((len(notes), 4))
for i, note in enumerate(notes):
out[i] = [note.pitch, note.start_time, note.end_time, note.velocity]
return out
# 하이퍼 파라미터 설정
config = configs.CONFIG_MAP['onsets_frames']
hparams = config.hparams
hparams.use_cudnn = False
hparams.batch_size = 1
# Placeholder 설정
examples = tf.placeholder(tf.string, [None])
# 배치 생성
dataset = data.provide_batch(examples=examples, preprocess_examples=True,
params=hparams, is_training=False, shuffle_examples=False,
skip_n_initial_records=0)
# Estimator 생성
estimator = train_util.create_estimator(
config.model_fn, CHECKPOINT_DIR, hparams)
# Iterator 생성
iterator = dataset.make_initializable_iterator()
next_record = iterator.get_next()
# 세션 생성
sess = tf.Session()
# 세션 초기화
sess.run([tf.initializers.global_variables(), tf.initializers.local_variables()])
# 모델 데이터로 Datasets 만들기
def input_fn(params):
del params
return tf.data.Dataset.from_tensors(sess.run(next_record))
example = tf.placeholder(tf.string, [None])
# Logging
tf.logging.info('model_dir=%s', MODEL_DIR)
tf.logging.info('checkpoint_path=%s', 'checkpoint')
# 배치 생성
dataset = data.provide_batch(examples=example,
preprocess_examples=True,
params=hparams,
is_training=False,
shuffle_examples=False,
skip_n_initial_records=0)
# Estimator
estimator = train_util.create_estimator(config.model_fn, MODEL_DIR, hparams)
# 배치를 순환하는 이터레이터
iterator = dataset.make_initializable_iterator()
next_record = iterator.get_next()
sess = tf.Session()
#전역변수 및 로컬변수 초기화
sess.run(
[tf.initializers.global_variables(),
tf.initializers.local_variables()])
#tensor로부터 사용하려는 데이터를 불러와 Dataset 인스턴스 생성
def input_fn(params):
def main(input, output):
MAESTRO_CHECKPOINT_DIR = '/data/maestro/train'
config = configs.CONFIG_MAP['onsets_frames']
hparams = config.hparams
hparams.use_cudnn = False
hparams.batch_size = 1
checkpoint_dir = MAESTRO_CHECKPOINT_DIR
examples = tf.placeholder(tf.string, [None])
dataset = data.provide_batch(
examples=examples,
preprocess_examples=True,
params=hparams,
is_training=False,
shuffle_examples=False,
skip_n_initial_records=0)
estimator = train_util.create_estimator(
config.model_fn, checkpoint_dir, hparams)
iterator = dataset.make_initializable_iterator()
next_record = iterator.get_next()
to_process = []
def process(files):
for fn in files:
print('**\n\n', fn, '\n\n**')
with open(fn, 'rb', buffering=0) as f:
wav_data = f.read()
example_list = list(
audio_label_data_utils.process_record(
wav_data=wav_data,
ns=music_pb2.NoteSequence(),
example_id=fn,
min_length=0,
max_length=-1,
allow_empty_notesequence=True))
assert len(example_list) == 1
to_process.append(example_list[0].SerializeToString())
print('Processing complete for', fn)
sess = tf.Session()
sess.run([
tf.initializers.global_variables(),
tf.initializers.local_variables()
])
sess.run(iterator.initializer, {examples: to_process})
def transcription_data(params):
del params
return tf.data.Dataset.from_tensors(sess.run(next_record))
input_fn = infer_util.labels_to_features_wrapper(transcription_data)
#@title Run inference
prediction_list = list(
estimator.predict(
input_fn,
yield_single_examples=False))
assert len(prediction_list) == 1
# Ignore warnings caused by pyfluidsynth
import warnings
warnings.filterwarnings("ignore", category=DeprecationWarning)
sequence_prediction = music_pb2.NoteSequence.FromString(
prediction_list[0]['sequence_predictions'][0])
pathname = fn.split('/').pop()
print('**\n\n', pathname, '\n\n**')
midi_filename = '{outputs}/{file}.mid'.format(outputs=output,file=pathname)
midi_io.sequence_proto_to_midi_file(sequence_prediction, midi_filename)
files = ['{inputs}/{file}'.format(inputs=input, file=file) for file in os.listdir(input) if file.split('.').pop() == 'wav']
print('the files', files)
process(files)
def model_inference(model_fn,
model_dir,
checkpoint_path,
hparams,
examples_path,
output_dir,
summary_writer,
master,
preprocess_examples,
write_summary_every_step=True):
"""Runs inference for the given examples."""
tf.logging.info('model_dir=%s', model_dir)
tf.logging.info('checkpoint_path=%s', checkpoint_path)
tf.logging.info('examples_path=%s', examples_path)
tf.logging.info('output_dir=%s', output_dir)
estimator = train_util.create_estimator(
model_fn, model_dir, hparams, master=master)
with tf.Graph().as_default():
num_dims = constants.MIDI_PITCHES
dataset = data.provide_batch(
examples=examples_path,
preprocess_examples=preprocess_examples,
hparams=hparams,
is_training=False)
# Define some metrics.
(metrics_to_updates, metric_note_precision, metric_note_recall,
metric_note_f1, metric_note_precision_with_offsets,
metric_note_recall_with_offsets, metric_note_f1_with_offsets,
metric_note_precision_with_offsets_velocity,
metric_note_recall_with_offsets_velocity,
metric_note_f1_with_offsets_velocity, metric_frame_labels,
metric_frame_predictions) = infer_util.define_metrics(num_dims)
summary_op = tf.summary.merge_all()
if write_summary_every_step:
global_step = tf.train.get_or_create_global_step()
global_step_increment = global_step.assign_add(1)
else:
global_step = tf.constant(
estimator.get_variable_value(tf.GraphKeys.GLOBAL_STEP))
global_step_increment = global_step
iterator = dataset.make_initializable_iterator()
next_record = iterator.get_next()
with tf.Session() as sess:
sess.run([
tf.initializers.global_variables(),
tf.initializers.local_variables()
])
infer_times = []
num_frames = []
sess.run(iterator.initializer)
while True:
try:
record = sess.run(next_record)
except tf.errors.OutOfRangeError:
break
def input_fn(params):
del params
return tf.data.Dataset.from_tensors(record)
start_time = time.time()
# TODO(fjord): This is a hack that allows us to keep using our existing
# infer/scoring code with a tf.Estimator model. Ideally, we should
# move things around so that we can use estimator.evaluate, which will
# also be more efficient because it won't have to restore the checkpoint
# for every example.
prediction_list = list(
estimator.predict(
input_fn,
checkpoint_path=checkpoint_path,
yield_single_examples=False))
assert len(prediction_list) == 1
input_features = record[0]
input_labels = record[1]
filename = input_features.sequence_id[0]
note_sequence = music_pb2.NoteSequence.FromString(
input_labels.note_sequence[0])
labels = input_labels.labels[0]
frame_probs = prediction_list[0]['frame_probs'][0]
frame_predictions = prediction_list[0]['frame_predictions'][0]
onset_predictions = prediction_list[0]['onset_predictions'][0]
velocity_values = prediction_list[0]['velocity_values'][0]
offset_predictions = prediction_list[0]['offset_predictions'][0]
if not FLAGS.require_onset:
onset_predictions = None
if not FLAGS.use_offset:
offset_predictions = None
sequence_prediction = sequences_lib.pianoroll_to_note_sequence(
frame_predictions,
frames_per_second=data.hparams_frames_per_second(hparams),
min_duration_ms=0,
min_midi_pitch=constants.MIN_MIDI_PITCH,
onset_predictions=onset_predictions,
offset_predictions=offset_predictions,
velocity_values=velocity_values)
end_time = time.time()
infer_time = end_time - start_time
infer_times.append(infer_time)
num_frames.append(frame_predictions.shape[0])
tf.logging.info(
'Infer time %f, frames %d, frames/sec %f, running average %f',
infer_time, frame_predictions.shape[0],
frame_predictions.shape[0] / infer_time,
np.sum(num_frames) / np.sum(infer_times))
tf.logging.info('Scoring sequence %s', filename)
def shift_notesequence(ns_time):
return ns_time + hparams.backward_shift_amount_ms / 1000.
sequence_label = sequences_lib.adjust_notesequence_times(
note_sequence, shift_notesequence)[0]
infer_util.score_sequence(
sess,
global_step_increment,
metrics_to_updates,
metric_note_precision,
metric_note_recall,
metric_note_f1,
metric_note_precision_with_offsets,
metric_note_recall_with_offsets,
metric_note_f1_with_offsets,
metric_note_precision_with_offsets_velocity,
metric_note_recall_with_offsets_velocity,
metric_note_f1_with_offsets_velocity,
metric_frame_labels,
metric_frame_predictions,
frame_labels=labels,
sequence_prediction=sequence_prediction,
frames_per_second=data.hparams_frames_per_second(hparams),
sequence_label=sequence_label,
sequence_id=filename)
if write_summary_every_step:
# Make filenames UNIX-friendly.
filename_safe = filename.decode('utf-8').replace('/', '_').replace(
':', '.')
output_file = os.path.join(output_dir, filename_safe + '.mid')
tf.logging.info('Writing inferred midi file to %s', output_file)
midi_io.sequence_proto_to_midi_file(sequence_prediction, output_file)
label_output_file = os.path.join(output_dir,
filename_safe + '_label.mid')
tf.logging.info('Writing label midi file to %s', label_output_file)
midi_io.sequence_proto_to_midi_file(sequence_label, label_output_file)
# Also write a pianoroll showing acoustic model output vs labels.
pianoroll_output_file = os.path.join(output_dir,
filename_safe + '_pianoroll.png')
tf.logging.info('Writing acoustic logit/label file to %s',
pianoroll_output_file)
with tf.gfile.GFile(pianoroll_output_file, mode='w') as f:
scipy.misc.imsave(
f,
infer_util.posterior_pianoroll_image(
frame_probs,
sequence_prediction,
labels,
overlap=True,
frames_per_second=data.hparams_frames_per_second(hparams)))
summary = sess.run(summary_op)
summary_writer.add_summary(summary, sess.run(global_step))
summary_writer.flush()
if not write_summary_every_step:
# Only write the summary variables for the final step.
summary = sess.run(summary_op)
summary_writer.add_summary(summary, sess.run(global_step))
summary_writer.flush()
def model_inference(model_fn, model_dir, checkpoint_path, data_fn, hparams,
examples_path, output_dir, summary_writer, master,
preprocess_examples, shuffle_examples):
"""Runs inference for the given examples."""
tf.logging.info('model_dir=%s', model_dir)
tf.logging.info('checkpoint_path=%s', checkpoint_path)
tf.logging.info('examples_path=%s', examples_path)
tf.logging.info('output_dir=%s', output_dir)
estimator = train_util.create_estimator(model_fn,
model_dir,
hparams,
master=master)
transcription_data = functools.partial(
data_fn,
examples=examples_path,
preprocess_examples=preprocess_examples,
is_training=False,
shuffle_examples=shuffle_examples,
skip_n_initial_records=0)
input_fn = infer_util.labels_to_features_wrapper(transcription_data)
start_time = time.time()
infer_times = []
num_frames = []
file_num = 0
all_metrics = collections.defaultdict(list)
for predictions in estimator.predict(input_fn,
checkpoint_path=checkpoint_path,
yield_single_examples=False):
# Remove batch dimension for convenience.
for k in predictions.keys():
if predictions[k].shape[0] != 1:
raise ValueError(
'All predictions must have batch size 1, but shape of '
'{} was: {}'.format(k, +predictions[k].shape[0]))
predictions[k] = predictions[k][0]
end_time = time.time()
infer_time = end_time - start_time
infer_times.append(infer_time)
num_frames.append(predictions['frame_predictions'].shape[0])
tf.logging.info(
'Infer time %f, frames %d, frames/sec %f, running average %f',
infer_time, num_frames[-1], num_frames[-1] / infer_time,
np.sum(num_frames) / np.sum(infer_times))
tf.logging.info('Scoring sequence %s', predictions['sequence_ids'])
sequence_prediction = music_pb2.NoteSequence.FromString(
predictions['sequence_predictions'])
sequence_label = music_pb2.NoteSequence.FromString(
predictions['sequence_labels'])
# Make filenames UNIX-friendly.
filename_chars = predictions['sequence_ids'].decode('utf-8')
filename_chars = [c if c.isalnum() else '_' for c in filename_chars]
filename_safe = ''.join(filename_chars).rstrip()
filename_safe = '{:04d}_{}'.format(file_num, filename_safe[:200])
file_num += 1
output_file = os.path.join(output_dir, filename_safe + '.mid')
tf.logging.info('Writing inferred midi file to %s', output_file)
midi_io.sequence_proto_to_midi_file(sequence_prediction, output_file)
label_output_file = os.path.join(output_dir,
filename_safe + '_label.mid')
tf.logging.info('Writing label midi file to %s', label_output_file)
midi_io.sequence_proto_to_midi_file(sequence_label, label_output_file)
# Also write a pianoroll showing acoustic model output vs labels.
pianoroll_output_file = os.path.join(output_dir,
filename_safe + '_pianoroll.png')
tf.logging.info('Writing acoustic logit/label file to %s',
pianoroll_output_file)
with tf.gfile.GFile(pianoroll_output_file, mode='w') as f:
scipy.misc.imsave(
f,
infer_util.posterior_pianoroll_image(
predictions['frame_probs'], predictions['frame_labels']))
# Update histogram and current scalar for metrics.
with tf.Graph().as_default(), tf.Session().as_default():
for k, v in predictions.items():
if not k.startswith('metrics/'):
continue
all_metrics[k].extend(v)
histogram_name = 'histogram/' + k
metric_summary = tf.summary.histogram(histogram_name,
tf.constant(
all_metrics[k],
name=histogram_name),
collections=[])
summary_writer.add_summary(metric_summary.eval(),
global_step=file_num)
scalar_name = k
metric_summary = tf.summary.scalar(scalar_name,
tf.constant(
np.mean(all_metrics[k]),
name=scalar_name),
collections=[])
summary_writer.add_summary(metric_summary.eval(),
global_step=file_num)
summary_writer.flush()
start_time = time.time()
# Write final mean values for all metrics.
with tf.Graph().as_default(), tf.Session().as_default():
for k, v in all_metrics.items():
final_scalar_name = 'final/' + k
metric_summary = tf.summary.scalar(final_scalar_name,
tf.constant(
np.mean(all_metrics[k]),
name=final_scalar_name),
collections=[])
summary_writer.add_summary(metric_summary.eval())
summary_writer.flush()
start_time = time.time()
def model_inference(model_dir,
checkpoint_path,
hparams,
examples_path,
output_dir,
summary_writer,
write_summary_every_step=True):
"""Runs inference for the given examples."""
tf.logging.info('model_dir=%s', model_dir)
tf.logging.info('checkpoint_path=%s', checkpoint_path)
tf.logging.info('examples_path=%s', examples_path)
tf.logging.info('output_dir=%s', output_dir)
estimator = train_util.create_estimator(model_dir, hparams)
with tf.Graph().as_default():
num_dims = constants.MIDI_PITCHES
if FLAGS.max_seconds_per_sequence:
truncated_length = int(
math.ceil((FLAGS.max_seconds_per_sequence *
data.hparams_frames_per_second(hparams))))
else:
truncated_length = 0
dataset = data.provide_batch(batch_size=1,
examples=examples_path,
hparams=hparams,
is_training=False,
truncated_length=truncated_length)
# Define some metrics.
(metrics_to_updates, metric_note_precision, metric_note_recall,
metric_note_f1, metric_note_precision_with_offsets,
metric_note_recall_with_offsets, metric_note_f1_with_offsets,
metric_note_precision_with_offsets_velocity,
metric_note_recall_with_offsets_velocity,
metric_note_f1_with_offsets_velocity, metric_frame_labels,
metric_frame_predictions) = infer_util.define_metrics(num_dims)
summary_op = tf.summary.merge_all()
if write_summary_every_step:
global_step = tf.train.get_or_create_global_step()
global_step_increment = global_step.assign_add(1)
else:
global_step = tf.constant(
estimator.get_variable_value(tf.GraphKeys.GLOBAL_STEP))
global_step_increment = global_step
iterator = dataset.make_initializable_iterator()
next_record = iterator.get_next()
with tf.Session() as sess:
sess.run([
tf.initializers.global_variables(),
tf.initializers.local_variables()
])
infer_times = []
num_frames = []
sess.run(iterator.initializer)
while True:
try:
record = sess.run(next_record)
except tf.errors.OutOfRangeError:
break
def input_fn():
return tf.data.Dataset.from_tensors(record)
start_time = time.time()
# TODO(fjord): This is a hack that allows us to keep using our existing
# infer/scoring code with a tf.Estimator model. Ideally, we should
# move things around so that we can use estimator.evaluate, which will
# also be more efficient because it won't have to restore the checkpoint
# for every example.
prediction_list = list(
estimator.predict(input_fn,
checkpoint_path=checkpoint_path,
yield_single_examples=False))
assert len(prediction_list) == 1
input_features = record[0]
input_labels = record[1]
filename = input_features.sequence_id[0]
note_sequence = music_pb2.NoteSequence.FromString(
input_labels.note_sequence[0])
labels = input_labels.labels[0]
frame_probs = prediction_list[0]['frame_probs_flat']
onset_probs = prediction_list[0]['onset_probs_flat']
velocity_values = prediction_list[0]['velocity_values_flat']
offset_probs = prediction_list[0]['offset_probs_flat']
frame_predictions = frame_probs > FLAGS.frame_threshold
if FLAGS.require_onset:
onset_predictions = onset_probs > FLAGS.onset_threshold
else:
onset_predictions = None
if FLAGS.use_offset:
offset_predictions = offset_probs > FLAGS.offset_threshold
else:
offset_predictions = None
sequence_prediction = sequences_lib.pianoroll_to_note_sequence(
frame_predictions,
frames_per_second=data.hparams_frames_per_second(hparams),
min_duration_ms=0,
min_midi_pitch=constants.MIN_MIDI_PITCH,
onset_predictions=onset_predictions,
offset_predictions=offset_predictions,
velocity_values=velocity_values)
end_time = time.time()
infer_time = end_time - start_time
infer_times.append(infer_time)
num_frames.append(frame_probs.shape[0])
tf.logging.info(
'Infer time %f, frames %d, frames/sec %f, running average %f',
infer_time, frame_probs.shape[0],
frame_probs.shape[0] / infer_time,
np.sum(num_frames) / np.sum(infer_times))
tf.logging.info('Scoring sequence %s', filename)
def shift_notesequence(ns_time):
return ns_time + hparams.backward_shift_amount_ms / 1000.
sequence_label = sequences_lib.adjust_notesequence_times(
note_sequence, shift_notesequence)[0]
infer_util.score_sequence(
sess,
global_step_increment,
metrics_to_updates,
metric_note_precision,
metric_note_recall,
metric_note_f1,
metric_note_precision_with_offsets,
metric_note_recall_with_offsets,
metric_note_f1_with_offsets,
metric_note_precision_with_offsets_velocity,
metric_note_recall_with_offsets_velocity,
metric_note_f1_with_offsets_velocity,
metric_frame_labels,
metric_frame_predictions,
frame_labels=labels,
sequence_prediction=sequence_prediction,
frames_per_second=data.hparams_frames_per_second(hparams),
sequence_label=sequence_label,
sequence_id=filename)
if write_summary_every_step:
# Make filenames UNIX-friendly.
filename_safe = filename.decode('utf-8').replace(
'/', '_').replace(':', '.')
output_file = os.path.join(output_dir,
filename_safe + '.mid')
tf.logging.info('Writing inferred midi file to %s',
output_file)
midi_io.sequence_proto_to_midi_file(
sequence_prediction, output_file)
label_output_file = os.path.join(
output_dir, filename_safe + '_label.mid')
tf.logging.info('Writing label midi file to %s',
label_output_file)
midi_io.sequence_proto_to_midi_file(
sequence_label, label_output_file)
# Also write a pianoroll showing acoustic model output vs labels.
pianoroll_output_file = os.path.join(
output_dir, filename_safe + '_pianoroll.png')
tf.logging.info('Writing acoustic logit/label file to %s',
pianoroll_output_file)
with tf.gfile.GFile(pianoroll_output_file, mode='w') as f:
scipy.misc.imsave(
f,
infer_util.posterior_pianoroll_image(
frame_probs,
sequence_prediction,
labels,
overlap=True,
frames_per_second=data.
hparams_frames_per_second(hparams)))
summary = sess.run(summary_op)
summary_writer.add_summary(summary, sess.run(global_step))
summary_writer.flush()
if not write_summary_every_step:
# Only write the summary variables for the final step.
summary = sess.run(summary_op)
summary_writer.add_summary(summary, sess.run(global_step))
summary_writer.flush()
checkpoint_dir = MAESTRO_CHECKPOINT_DIR
elif model_type.startswith('E-GMD'):
config = configs.CONFIG_MAP['drums']
hparams = config.hparams
hparams.batch_size = 1
checkpoint_dir = EGMD_CHECKPOINT_DIR
else:
raise ValueError('Unknown Model Type')
examples = tf.placeholder(tf.string, [None])
dataset = data.provide_batch(examples=examples,
preprocess_examples=True,
params=hparams,
is_training=False,
shuffle_examples=False,
skip_n_initial_records=0)
estimator = train_util.create_estimator(config.model_fn, checkpoint_dir,
hparams)
iterator = tf.data.make_initializable_iterator(dataset)
next_record = iterator.get_next()
def transcription_data(params):
del params
return tf.data.Dataset.from_tensors(sess.run(next_record))
"""# Upload Audio
Run the following cell to upload audio files.
"""
# pianoAudio = 'misty.wav'
pianoAudio = 'lalala.wav'
def main(argv):
tf.logging.set_verbosity(FLAGS.log)
hparams = tf_utils.merge_hparams(constants.DEFAULT_HPARAMS,
model.get_default_hparams())
# For this script, default to not using cudnn.
hparams.use_cudnn = False
hparams.parse(FLAGS.hparams)
hparams.batch_size = 1
with tf.Graph().as_default():
examples = tf.placeholder(tf.string, [None])
dataset = data.provide_batch(batch_size=1,
examples=examples,
hparams=hparams,
is_training=False,
truncated_length=0)
estimator = train_util.create_estimator(
os.path.expanduser(FLAGS.model_dir), hparams)
iterator = dataset.make_initializable_iterator()
next_record = iterator.get_next()
with tf.Session() as sess:
sess.run([
tf.initializers.global_variables(),
tf.initializers.local_variables()
])
for filename in argv[1:]:
tf.logging.info('Starting transcription for %s...', filename)
# The reason we bounce between two Dataset objects is so we can use
# the data processing functionality in data.py without having to
# construct all the Example protos in memory ahead of time or create
# a temporary tfrecord file.
tf.logging.info('Processing file...')
sess.run(iterator.initializer,
{examples: [create_example(filename)]})
def input_fn():
return tf.data.Dataset.from_tensors(sess.run(next_record))
tf.logging.info('Running inference...')
checkpoint_path = None
if FLAGS.checkpoint_path:
checkpoint_path = os.path.expanduser(FLAGS.checkpoint_path)
prediction_list = list(
estimator.predict(input_fn,
checkpoint_path=checkpoint_path,
yield_single_examples=False))
assert len(prediction_list) == 1
sequence_prediction = transcribe_audio(prediction_list[0],
hparams,
FLAGS.frame_threshold,
FLAGS.onset_threshold)
midi_filename = filename + '.midi'
midi_io.sequence_proto_to_midi_file(sequence_prediction,
midi_filename)
tf.logging.info('Transcription written to %s.', midi_filename)
def main(argv):
tf.logging.set_verbosity(FLAGS.log)
config = configs.CONFIG_MAP[FLAGS.config]
hparams = config.hparams
# For this script, default to not using cudnn.
hparams.use_cudnn = False
hparams.parse(FLAGS.hparams)
hparams.batch_size = 1
hparams.truncated_length_secs = 0
with tf.Graph().as_default():
examples = tf.placeholder(tf.string, [None])
dataset = data.provide_batch(
examples=examples,
preprocess_examples=True,
hparams=hparams,
is_training=False)
estimator = train_util.create_estimator(config.model_fn,
os.path.expanduser(FLAGS.model_dir),
hparams)
iterator = dataset.make_initializable_iterator()
next_record = iterator.get_next()
with tf.Session() as sess:
sess.run([
tf.initializers.global_variables(),
tf.initializers.local_variables()
])
for filename in argv[1:]:
tf.logging.info('Starting transcription for %s...', filename)
# The reason we bounce between two Dataset objects is so we can use
# the data processing functionality in data.py without having to
# construct all the Example protos in memory ahead of time or create
# a temporary tfrecord file.
tf.logging.info('Processing file...')
sess.run(iterator.initializer, {examples: [create_example(filename)]})
def input_fn(params):
del params
return tf.data.Dataset.from_tensors(sess.run(next_record))
tf.logging.info('Running inference...')
checkpoint_path = None
if FLAGS.checkpoint_path:
checkpoint_path = os.path.expanduser(FLAGS.checkpoint_path)
prediction_list = list(
estimator.predict(
input_fn,
checkpoint_path=checkpoint_path,
yield_single_examples=False))
assert len(prediction_list) == 1
sequence_prediction = transcribe_audio(prediction_list[0], hparams)
midi_filename = filename + '.midi'
midi_io.sequence_proto_to_midi_file(sequence_prediction, midi_filename)
tf.logging.info('Transcription written to %s.', midi_filename)
