def main(argv):
tf.logging.set_verbosity(FLAGS.log)
if FLAGS.acoustic_checkpoint_filename:
acoustic_checkpoint = os.path.join(
os.path.expanduser(FLAGS.acoustic_run_dir), 'train',
FLAGS.acoustic_checkpoint_filename)
else:
acoustic_checkpoint = tf.train.latest_checkpoint(
os.path.join(os.path.expanduser(FLAGS.acoustic_run_dir), 'train'))
hparams = tf_utils.merge_hparams(
constants.DEFAULT_HPARAMS, model.get_default_hparams())
hparams.parse(FLAGS.hparams)
transcription_session = initialize_session(acoustic_checkpoint, hparams)
for filename in argv[1:]:
tf.logging.info('Starting transcription for %s...', filename)
sequence_prediction = transcribe_audio(
transcription_session, filename, 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 get_default_hparams():
"""Returns the default hyperparameters.
Returns:
A tf.HParams object representing the default hyperparameters for the model.
"""
return tf_utils.merge_hparams(
constants.DEFAULT_HPARAMS,
tf.contrib.training.HParams(
activation_loss=False,
batch_size=8,
clip_norm=3,
combined_lstm_units=128,
frame_bidirectional=True,
frame_lstm_units=0,
learning_rate=0.0006,
min_duration_ms=0,
min_frame_occupancy_for_label=0.0,
normalize_audio=False,
onset_bidirectional=True,
onset_delay=0,
onset_length=32,
onset_lstm_units=128,
onset_mode='length_ms',
sample_rate=constants.DEFAULT_SAMPLE_RATE,
share_conv_features=False,
spec_fmin=30.0,
spec_hop_length=512,
spec_log_amplitude=True,
spec_n_bins=229,
spec_type='mel',
stop_activation_gradient=False,
stop_onset_gradient=False,
truncated_length=1500, # 48 seconds
weight_frame_and_activation_loss=False))
def main(argv):
tf.logging.set_verbosity(FLAGS.log)
if FLAGS.acoustic_checkpoint_filename:
acoustic_checkpoint = os.path.join(
os.path.expanduser(FLAGS.acoustic_run_dir), 'train',
FLAGS.acoustic_checkpoint_filename)
else:
acoustic_checkpoint = tf.train.latest_checkpoint(
os.path.join(os.path.expanduser(FLAGS.acoustic_run_dir), 'train'))
hparams = tf_utils.merge_hparams(constants.DEFAULT_HPARAMS,
model.get_default_hparams())
hparams.parse(FLAGS.hparams)
transcription_session = initialize_session(acoustic_checkpoint, hparams)
for filename in argv[1:]:
tf.logging.info('Starting transcription for %s...', filename)
sequence_prediction = transcribe_audio(transcription_session, filename,
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(unused_argv):
if FLAGS.acoustic_checkpoint_filename:
acoustic_checkpoint = os.path.join(
os.path.expanduser(FLAGS.acoustic_run_dir), 'train',
FLAGS.acoustic_checkpoint_filename)
else:
acoustic_checkpoint = tf.train.latest_checkpoint(
os.path.join(os.path.expanduser(FLAGS.acoustic_run_dir), 'train'))
run_dir = os.path.expanduser(FLAGS.run_dir)
hparams = tf_utils.merge_hparams(
constants.DEFAULT_HPARAMS, model.get_default_hparams())
hparams.parse(FLAGS.hparams)
# Batch size should always be 1 for inference.
hparams.batch_size = 1
tf.logging.info(hparams)
tf.gfile.MakeDirs(run_dir)
with tf.gfile.Open(os.path.join(run_dir, 'run_config.txt'), 'w') as f:
f.write(acoustic_checkpoint + '\n')
f.write(FLAGS.examples_path + '\n')
f.write(str(hparams) + '\n')
model_inference(
acoustic_checkpoint=acoustic_checkpoint,
hparams=hparams,
examples_path=FLAGS.examples_path,
run_dir=run_dir)
def main(unused_argv):
output_dir = os.path.expanduser(FLAGS.output_dir)
hparams = tf_utils.merge_hparams(
constants.DEFAULT_HPARAMS, model.get_default_hparams())
hparams.parse(FLAGS.hparams)
# Batch size should always be 1 for inference.
hparams.batch_size = 1
tf.logging.info(hparams)
tf.gfile.MakeDirs(output_dir)
summary_writer = tf.summary.FileWriter(logdir=output_dir)
with tf.Session():
run_config = '\n\n'.join([
'model_dir: ' + FLAGS.model_dir,
'checkpoint_path: ' + str(FLAGS.checkpoint_path),
'examples_path: ' + FLAGS.examples_path,
str(hparams),
])
run_config_summary = tf.summary.text(
'run_config',
tf.constant(run_config, name='run_config'),
collections=[])
summary_writer.add_summary(run_config_summary.eval())
if FLAGS.eval_loop:
assert not FLAGS.checkpoint_path
checkpoint_path = None
while True:
checkpoint_path = tf.contrib.training.wait_for_new_checkpoint(
FLAGS.model_dir, last_checkpoint=checkpoint_path)
model_inference(
model_dir=FLAGS.model_dir,
checkpoint_path=checkpoint_path,
hparams=hparams,
examples_path=FLAGS.examples_path,
output_dir=output_dir,
summary_writer=summary_writer,
master=FLAGS.master,
write_summary_every_step=False)
else:
model_inference(
model_dir=FLAGS.model_dir,
checkpoint_path=FLAGS.checkpoint_path,
hparams=hparams,
examples_path=FLAGS.examples_path,
output_dir=output_dir,
summary_writer=summary_writer,
master=FLAGS.master)
def init(self):
tf.logging.set_verbosity(self.log)
acoustic_checkpoint = tf.train.latest_checkpoint(
os.path.join(os.path.expanduser(self.acoustic_run_dir), 'train'))
default_hparams = tf_utils.merge_hparams(constants.DEFAULT_HPARAMS,
model.get_default_hparams())
default_hparams.parse(self.hparams)
self.transcription_session = self.initialize_session(
acoustic_checkpoint, default_hparams)
def main(unused_argv):
output_dir = os.path.expanduser(FLAGS.output_dir)
hparams = tf_utils.merge_hparams(constants.DEFAULT_HPARAMS,
model.get_default_hparams())
hparams.parse(FLAGS.hparams)
# Batch size should always be 1 for inference.
hparams.batch_size = 1
tf.logging.info(hparams)
tf.gfile.MakeDirs(output_dir)
summary_writer = tf.summary.FileWriter(logdir=output_dir)
with tf.Session():
run_config = '\n\n'.join([
'model_dir: ' + FLAGS.model_dir,
'checkpoint_path: ' + str(FLAGS.checkpoint_path),
'examples_path: ' + FLAGS.examples_path,
str(hparams),
])
run_config_summary = tf.summary.text('run_config',
tf.constant(run_config,
name='run_config'),
collections=[])
summary_writer.add_summary(run_config_summary.eval())
if FLAGS.eval_loop:
assert not FLAGS.checkpoint_path
checkpoint_path = None
while True:
checkpoint_path = tf.contrib.training.wait_for_new_checkpoint(
FLAGS.model_dir, last_checkpoint=checkpoint_path)
model_inference(model_dir=FLAGS.model_dir,
checkpoint_path=checkpoint_path,
hparams=hparams,
examples_path=FLAGS.examples_path,
output_dir=output_dir,
summary_writer=summary_writer,
master=FLAGS.master,
write_summary_every_step=False)
else:
model_inference(model_dir=FLAGS.model_dir,
checkpoint_path=FLAGS.checkpoint_path,
hparams=hparams,
examples_path=FLAGS.examples_path,
output_dir=output_dir,
summary_writer=summary_writer,
master=FLAGS.master)
def main(unused_argv):
tf.logging.set_verbosity(FLAGS.log)
tf.app.flags.mark_flags_as_required(['examples_path'])
run_dir = os.path.expanduser(FLAGS.run_dir)
hparams = tf_utils.merge_hparams(constants.DEFAULT_HPARAMS,
model.get_default_hparams())
# Command line flags override any of the preceding hyperparameter values.
hparams.parse(FLAGS.hparams)
run(hparams, run_dir)
def main(unused_argv):
tf.logging.set_verbosity(FLAGS.log)
tf.app.flags.mark_flags_as_required(['examples_path'])
run_dir = os.path.expanduser(FLAGS.run_dir)
hparams = tf_utils.merge_hparams(constants.DEFAULT_HPARAMS,
model.get_default_hparams())
# Command line flags override any of the preceding hyperparameter values.
hparams.parse(FLAGS.hparams)
run(hparams, run_dir)
def main(argv):
tf.logging.set_verbosity(FLAGS.log)
hparams = tf_utils.merge_hparams(constants.DEFAULT_HPARAMS,
model.get_default_hparams())
hparams.parse(FLAGS.hparams)
for filename in argv[1:]:
tf.logging.info('Generating spectrogram for %s...', filename)
spec = create_spec(filename, hparams)
spec_filename = filename + '.json'
with tf.gfile.Open(spec_filename, 'w') as f:
f.write(json.dumps(spec.tolist()))
tf.logging.info('Wrote spectrogram json to %s.', spec_filename)
def main(argv):
tf.logging.set_verbosity(FLAGS.log)
hparams = tf_utils.merge_hparams(
constants.DEFAULT_HPARAMS, model.get_default_hparams())
hparams.parse(FLAGS.hparams)
for filename in argv[1:]:
tf.logging.info('Generating spectrogram for %s...', filename)
spec = create_spec(filename, hparams)
spec_filename = filename + '.json'
with tf.gfile.Open(spec_filename, 'w') as f:
f.write(json.dumps(spec.tolist()))
tf.logging.info('Wrote spectrogram json to %s.', spec_filename)
def get_default_hparams():
"""Returns the default hyperparameters.
Returns:
A tf.HParams object representing the default hyperparameters for the model.
"""
return tf_utils.merge_hparams(
constants.DEFAULT_HPARAMS,
tf.contrib.training.HParams(
activation_loss=False,
batch_size=8,
clip_norm=3,
combined_lstm_units=384,
frame_bidirectional=False,
frame_lstm_units=0,
learning_rate=0.0006,
decay_steps=10000,
decay_rate=0.98,
min_duration_ms=0,
min_frame_occupancy_for_label=0.0,
normalize_audio=False,
onset_bidirectional=False,
onset_delay=0,
onset_length=32,
onset_lstm_units=384,
velocity_lstm_units=0,
onset_mode='length_ms',
sample_rate=constants.DEFAULT_SAMPLE_RATE,
share_conv_features=False,
spec_fmin=30.0,
spec_hop_length=512,
spec_log_amplitude=True,
spec_mel_htk=True,
spec_n_bins=229,
spec_type='mel',
stop_activation_gradient=False,
stop_onset_gradient=False,
truncated_length=1500, # 48 seconds
weight_frame_and_activation_loss=True))
def main(unused_argv):
tf.logging.set_verbosity(FLAGS.log)
if FLAGS.acoustic_checkpoint_filename:
acoustic_checkpoint = os.path.join(
os.path.expanduser(FLAGS.acoustic_run_dir), 'train',
FLAGS.acoustic_checkpoint_filename)
else:
acoustic_checkpoint = tf.train.latest_checkpoint(
os.path.join(os.path.expanduser(FLAGS.acoustic_run_dir), 'train'))
run_dir = os.path.expanduser(FLAGS.run_dir)
hparams = tf_utils.merge_hparams(constants.DEFAULT_HPARAMS,
model.get_default_hparams())
hparams.parse(FLAGS.hparams)
tf.gfile.MakeDirs(run_dir)
model_inference(acoustic_checkpoint=acoustic_checkpoint,
hparams=hparams,
examples_path=FLAGS.examples_path,
run_dir=run_dir)
def main(unused_argv):
tf.logging.set_verbosity(FLAGS.log)
if FLAGS.acoustic_checkpoint_filename:
acoustic_checkpoint = os.path.join(
os.path.expanduser(FLAGS.acoustic_run_dir), 'train',
FLAGS.acoustic_checkpoint_filename)
else:
acoustic_checkpoint = tf.train.latest_checkpoint(
os.path.join(os.path.expanduser(FLAGS.acoustic_run_dir), 'train'))
run_dir = os.path.expanduser(FLAGS.run_dir)
hparams = tf_utils.merge_hparams(
constants.DEFAULT_HPARAMS, model.get_default_hparams())
hparams.parse(FLAGS.hparams)
tf.gfile.MakeDirs(run_dir)
model_inference(
acoustic_checkpoint=acoustic_checkpoint,
hparams=hparams,
examples_path=FLAGS.examples_path,
run_dir=run_dir)
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 proxy_find_library(lib):
if lib == 'fluidsynth':
return 'libfluidsynth.so.1'
else:
return orig_find_library(lib)
ctypes.util.find_library = proxy_find_library
CHECKPOINT_DIR = './train/train_50002' ##todo
acoustic_checkpoint = tf.train.latest_checkpoint(CHECKPOINT_DIR)
print('acoustic_checkpoint=' + acoustic_checkpoint)
hparams = tf_utils.merge_hparams(constants.DEFAULT_HPARAMS,
model.get_default_hparams())
with tf.Graph().as_default():
examples = tf.placeholder(tf.string, [None])
num_dims = constants.MIDI_PITCHES
batch, iterator = data.provide_batch(batch_size=1,
examples=examples,
hparams=hparams,
is_training=False,
truncated_length=0)
model.get_model(batch, hparams, is_training=False)
session = tf.Session()
DEFAULT_MIN_FRAME_OCCUPANCY_FOR_LABEL = 0.0
DEFAULT_JITTER_AMOUNT_MS = 0
DEFAULT_MIN_DURATION_MS = 0
DEFAULT_BACKWARD_SHIFT_AMOUNT_MS = 0
DEFAULT_BIDIRECTIONAL = True
DEFAULT_ONSET_OVERLAP = True
DEFAULT_OFFSET_LENGTH = 100
DEFAULT_AUDIO_HPARAMS = tf.contrib.training.HParams(
sample_rate=DEFAULT_SAMPLE_RATE,
spec_type=DEFAULT_SPEC_TYPE,
spec_mel_htk=DEFAULT_SPEC_MEL_HTK,
spec_log_amplitude=DEFAULT_SPEC_LOG_AMPLITUDE,
spec_hop_length=DEFAULT_SPEC_HOP_LENGTH,
spec_n_bins=DEFAULT_SPEC_N_BINS,
spec_fmin=DEFAULT_SPEC_FMIN,
cqt_bins_per_octave=DEFAULT_CQT_BINS_PER_OCTAVE,
onset_length=DEFAULT_ONSET_LENGTH,
offset_length=DEFAULT_OFFSET_LENGTH,
onset_mode=DEFAULT_ONSET_MODE,
onset_delay=DEFAULT_ONSET_DELAY,
min_frame_occupancy_for_label=DEFAULT_MIN_FRAME_OCCUPANCY_FOR_LABEL,
jitter_amount_ms=DEFAULT_JITTER_AMOUNT_MS,
min_duration_ms=DEFAULT_MIN_DURATION_MS,
backward_shift_amount_ms=DEFAULT_BACKWARD_SHIFT_AMOUNT_MS,
bidirectional=DEFAULT_BIDIRECTIONAL,
onset_overlap=DEFAULT_ONSET_OVERLAP)
DEFAULT_HPARAMS = tf_utils.merge_hparams(
DEFAULT_AUDIO_HPARAMS, audio_transform.DEFAULT_AUDIO_TRANSFORM_HPARAMS)
DEFAULT_HPARAMS = tf_utils.merge_hparams(
audio_transform.DEFAULT_AUDIO_TRANSFORM_HPARAMS,
contrib_training.HParams(
eval_batch_size=1,
predict_batch_size=1,
shuffle_buffer_size=64,
sample_rate=16000,
spec_type='mel',
spec_mel_htk=True,
spec_log_amplitude=True,
spec_hop_length=512,
spec_n_bins=229,
spec_fmin=30.0, # A0
cqt_bins_per_octave=36,
truncated_length_secs=0.0,
max_expected_train_example_len=0,
onset_length=32,
offset_length=32,
onset_mode='length_ms',
onset_delay=0,
min_frame_occupancy_for_label=0.0,
jitter_amount_ms=0,
min_duration_ms=0,
backward_shift_amount_ms=0,
velocity_scale=80.0,
velocity_bias=10.0,
drum_data_map='',
drum_prediction_map='',
velocity_loss_weight=1.0,
splice_n_examples=0,
viterbi_decoding=False,
viterbi_alpha=0.5))
DEFAULT_MIN_DURATION_MS = 0
DEFAULT_BACKWARD_SHIFT_AMOUNT_MS = 0
DEFAULT_BIDIRECTIONAL = True
DEFAULT_ONSET_OVERLAP = True
DEFAULT_OFFSET_LENGTH = 100
DEFAULT_AUDIO_HPARAMS = tf.contrib.training.HParams(
sample_rate=DEFAULT_SAMPLE_RATE,
spec_type=DEFAULT_SPEC_TYPE,
spec_mel_htk=DEFAULT_SPEC_MEL_HTK,
spec_log_amplitude=DEFAULT_SPEC_LOG_AMPLITUDE,
spec_hop_length=DEFAULT_SPEC_HOP_LENGTH,
spec_n_bins=DEFAULT_SPEC_N_BINS,
spec_fmin=DEFAULT_SPEC_FMIN,
cqt_bins_per_octave=DEFAULT_CQT_BINS_PER_OCTAVE,
normalize_audio=DEFAULT_NORMALIZE_AUDIO,
crop_training_sequence_to_notes=DEFAULT_CROP_TRAINING_SEQUENCE_TO_NOTES,
onset_length=DEFAULT_ONSET_LENGTH,
offset_length=DEFAULT_OFFSET_LENGTH,
onset_mode=DEFAULT_ONSET_MODE,
onset_delay=DEFAULT_ONSET_DELAY,
min_frame_occupancy_for_label=DEFAULT_MIN_FRAME_OCCUPANCY_FOR_LABEL,
jitter_amount_ms=DEFAULT_JITTER_AMOUNT_MS,
min_duration_ms=DEFAULT_MIN_DURATION_MS,
backward_shift_amount_ms=DEFAULT_BACKWARD_SHIFT_AMOUNT_MS,
bidirectional=DEFAULT_BIDIRECTIONAL,
onset_overlap=DEFAULT_ONSET_OVERLAP)
DEFAULT_HPARAMS = tf_utils.merge_hparams(
DEFAULT_AUDIO_HPARAMS, audio_transform.DEFAULT_AUDIO_TRANSFORM_HPARAMS)
import tensorflow as tf_head
Config = collections.namedtuple('Config', ('model_fn', 'hparams'))
DEFAULT_HPARAMS = tf_utils.merge_hparams(
audio_transform.DEFAULT_AUDIO_TRANSFORM_HPARAMS,
tf_head.contrib.training.HParams(
eval_batch_size=1,
predict_batch_size=1,
shuffle_buffer_size=64,
sample_rate=16000,
spec_type='mel',
spec_mel_htk=True,
spec_log_amplitude=True,
spec_hop_length=512,
spec_n_bins=229,
spec_fmin=30.0, # A0
cqt_bins_per_octave=36,
truncated_length_secs=0.0,
max_expected_train_example_len=0,
onset_length=32,
offset_length=32,
onset_mode='length_ms',
onset_delay=0,
min_frame_occupancy_for_label=0.0,
jitter_amount_ms=0,
min_duration_ms=0,
backward_shift_amount_ms=0))
CONFIG_MAP = {}
CONFIG_MAP['onsets_frames'] = Config(
import tensorflow as tf
Config = collections.namedtuple('Config', ('model_fn', 'hparams'))
DEFAULT_HPARAMS = tf_utils.merge_hparams(
audio_transform.DEFAULT_AUDIO_TRANSFORM_HPARAMS,
tf.contrib.training.HParams(
eval_batch_size=1,
predict_batch_size=1,
sample_rate=16000,
spec_type='mel',
spec_mel_htk=True,
spec_log_amplitude=True,
spec_hop_length=512,
spec_n_bins=229,
spec_fmin=30.0, # A0
cqt_bins_per_octave=36,
truncated_length_secs=0,
max_expected_train_example_len=0,
onset_length=32,
offset_length=32,
onset_mode='length_ms',
onset_delay=0,
min_frame_occupancy_for_label=0.0,
jitter_amount_ms=0,
min_duration_ms=0,
backward_shift_amount_ms=0))
CONFIG_MAP = {}
CONFIG_MAP['onsets_frames'] = Config(