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(acoustic_checkpoint, hparams, examples_path, run_dir):
"""Runs inference for the given examples."""
tf.logging.info('acoustic_checkpoint=%s', acoustic_checkpoint)
tf.logging.info('examples_path=%s', examples_path)
tf.logging.info('run_dir=%s', run_dir)
with tf.Graph().as_default():
num_dims = constants.MIDI_PITCHES
# Build the acoustic model within an 'acoustic' scope to isolate its
# variables from the other models.
with tf.variable_scope('acoustic'):
truncated_length = 0
if FLAGS.max_seconds_per_sequence:
truncated_length = int(
math.ceil((FLAGS.max_seconds_per_sequence *
data.hparams_frames_per_second(hparams))))
acoustic_data_provider, _ = data.provide_batch(
batch_size=1,
examples=examples_path,
hparams=hparams,
is_training=False,
truncated_length=truncated_length,
include_note_sequences=True)
_, _, data_labels, _, _ = model.get_model(
acoustic_data_provider, hparams, is_training=False)
# The checkpoints won't have the new scopes.
acoustic_variables = {
re.sub(r'^acoustic/', '', var.op.name): var
for var in slim.get_variables(scope='acoustic/')
}
acoustic_restore = tf.train.Saver(acoustic_variables)
onset_probs_flat = tf.get_default_graph().get_tensor_by_name(
'acoustic/onsets/onset_probs_flat:0')
frame_probs_flat = tf.get_default_graph().get_tensor_by_name(
'acoustic/frame_probs_flat:0')
offset_probs_flat = tf.get_default_graph().get_tensor_by_name(
'acoustic/offsets/offset_probs_flat:0')
velocity_values_flat = tf.get_default_graph().get_tensor_by_name(
'acoustic/velocity/velocity_values_flat:0')
# 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()
global_step = tf.contrib.framework.get_or_create_global_step()
global_step_increment = global_step.assign_add(1)
# Use a custom init function to restore the acoustic and language models
# from their separate checkpoints.
def init_fn(unused_self, sess):
acoustic_restore.restore(sess, acoustic_checkpoint)
scaffold = tf.train.Scaffold(init_fn=init_fn)
session_creator = tf.train.ChiefSessionCreator(
scaffold=scaffold, master=FLAGS.master)
with tf.train.MonitoredSession(session_creator=session_creator) as sess:
tf.logging.info('running session')
summary_writer = tf.summary.FileWriter(
logdir=run_dir, graph=sess.graph)
tf.logging.info('Inferring for %d batches',
acoustic_data_provider.num_batches)
infer_times = []
num_frames = []
for unused_i in range(acoustic_data_provider.num_batches):
start_time = time.time()
(labels, filenames, note_sequences, frame_probs, onset_probs,
offset_probs, velocity_values) = sess.run([
data_labels,
acoustic_data_provider.filenames,
acoustic_data_provider.note_sequences,
frame_probs_flat,
onset_probs_flat,
offset_probs_flat,
velocity_values_flat,
])
# We expect these all to be length 1 because batch size is 1.
assert len(filenames) == len(note_sequences) == 1
# These should be the same length and have been flattened.
assert len(labels) == len(frame_probs) == len(onset_probs)
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', filenames[0])
def shift_notesequence(ns_time):
return ns_time + hparams.backward_shift_amount_ms / 1000.
sequence_label = infer_util.score_sequence(
sess,
global_step_increment,
summary_op,
summary_writer,
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=sequences_lib.adjust_notesequence_times(
music_pb2.NoteSequence.FromString(note_sequences[0]),
shift_notesequence)[0],
sequence_id=filenames[0])
# Make filenames UNIX-friendly.
filename = filenames[0].decode('utf-8').replace('/', '_').replace(
':', '.')
output_file = os.path.join(run_dir, filename + '.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(run_dir, filename + '_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(run_dir,
filename + '_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_writer.flush()
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()
