def construct(args):
network = NetworkMelody(args)
with network.session.graph.as_default():
spectrogram_function, spectrogram_thumb, spectrogram_info = common.spectrograms(args)
# save spectrogram_thumb to hyperparams
args.spectrogram_thumb = spectrogram_thumb
hop_samples = args.frame_width*args.samplerate/44100
print("hop_samples", hop_samples)
def preload_fn(aa):
aa.annotation = datasets.Annotation.from_time_series(*aa.annotation, hop_samples=hop_samples)
aa.audio.load_resampled_audio(args.samplerate).load_spectrogram(spectrogram_function, spectrogram_thumb, spectrogram_info[2])
def dataset_transform(tf_dataset, dataset):
return tf_dataset.map(dataset.prepare_example, num_parallel_calls=args.threads).batch(args.batch_size_evaluation).prefetch(10)
def dataset_transform_train(tf_dataset, dataset):
return tf_dataset.shuffle(10**5).map(dataset.prepare_example, num_parallel_calls=args.threads).batch(args.batch_size).prefetch(10)
valid_hooks = [MetricsHook(), VisualOutputHook(False, False, True, True), SaveBestModelHook(args.logdir), CSVOutputWriterHook(), AdjustVoicingHook()]
train_dataset, test_datasets, validation_datasets = common.prepare_datasets(args.datasets, args, preload_fn, dataset_transform, dataset_transform_train, valid_hooks=valid_hooks)
network.construct(args, create_model, train_dataset.dataset.output_types, train_dataset.dataset.output_shapes, spectrogram_info=spectrogram_info)
return network, train_dataset, validation_datasets, test_datasets
def construct(args):
network = NetworkMelody(args)
with network.session.graph.as_default():
spectrogram_function, spectrogram_thumb, spectrogram_info = common.spectrograms(
args)
def preload_fn(aa):
aa.annotation = datasets.Annotation.from_time_series(
*aa.annotation, args.frame_width * args.samplerate / 44100)
aa.audio.load_resampled_audio(args.samplerate).load_spectrogram(
spectrogram_function, spectrogram_thumb, spectrogram_info[2])
def dataset_transform(tf_dataset, dataset):
return tf_dataset.map(dataset.prepare_example,
num_parallel_calls=args.threads).batch(
args.batch_size_evaluation).prefetch(10)
def dataset_transform_train(tf_dataset, dataset):
return tf_dataset.shuffle(
10**5).map(dataset.prepare_example,
num_parallel_calls=args.threads).batch(
args.batch_size).prefetch(10)
train_dataset, test_datasets, validation_datasets = common.prepare_datasets(
args.datasets, args, preload_fn, dataset_transform,
dataset_transform_train)
network.construct(args,
create_model,
train_dataset.dataset.output_types,
train_dataset.dataset.output_shapes,
spectrogram_info=spectrogram_info)
return network, train_dataset, validation_datasets, test_datasets
def construct(args):
network = NetworkMelody(args)
with network.session.graph.as_default():
def spec_function(audio, samplerate):
cqt_list = []
shapes = []
for h in HARMONICS:
cqt = librosa.cqt(
audio, sr=samplerate, hop_length=HOP_LENGTH, fmin=FMIN*float(h),
n_bins=N_BINS,
bins_per_octave=BINS_PER_OCTAVE
)
cqt_list.append(cqt)
shapes.append(cqt.shape)
shapes_equal = [s == shapes[0] for s in shapes]
if not all(shapes_equal):
print("NOT ALL", shapes_equal)
min_time = np.min([s[1] for s in shapes])
new_cqt_list = []
for i in range(len(cqt_list)):
new_cqt_list.append(cqt_list[i][:, :min_time])
cqt_list = new_cqt_list
log_hcqt = ((1.0/80.0) * librosa.core.amplitude_to_db(
np.abs(np.array(cqt_list)), ref=np.max)) + 1.0
return (log_hcqt*65535).astype(np.uint16)
# cqt = librosa.core.cqt(audio, samplerate, hop_length=HOP_LENGTH, fmin=FMIN, n_bins=N_BINS, bins_per_octave=BINS_PER_OCTAVE)
# # log scaling
# cqt = librosa.amplitude_to_db(np.abs(cqt), ref=np.max)
# # uint8 compression
# cqt = ((cqt/80+1)*255).astype(np.uint8)
# return cqt
spectrogram_thumb = "hcqt-fmin{}-oct{}-octbins{}-hop{}-db-uint16".format(FMIN, N_BINS/BINS_PER_OCTAVE, BINS_PER_OCTAVE, HOP_LENGTH)
def preload_fn(aa):
aa.annotation = datasets.Annotation.from_time_series(*aa.annotation, 512)
aa.audio.load_resampled_audio(args.samplerate).load_spectrogram(spec_function, spectrogram_thumb, HOP_LENGTH)
def dataset_transform(tf_dataset, dataset):
return tf_dataset.map(dataset.prepare_example, num_parallel_calls=args.threads).batch(args.batch_size_evaluation).prefetch(10)
def dataset_transform_train(tf_dataset, dataset):
return tf_dataset.shuffle(10**5).map(dataset.prepare_example, num_parallel_calls=args.threads).batch(args.batch_size).prefetch(10)
valid_hooks = [MetricsHook(write_estimations=True), VisualOutputHook(False, False, True, True), SaveBestModelHook(args.logdir), AdjustVoicingHook()]
train_dataset, test_datasets, validation_datasets = common.prepare_datasets(args.datasets, args, preload_fn, dataset_transform, dataset_transform_train, valid_hooks=valid_hooks)
network.construct(args, create_model, train_dataset.dataset.output_types, train_dataset.dataset.output_shapes)
return network, train_dataset, validation_datasets, test_datasets
def construct(args):
network = NetworkMelody(args)
with network.session.graph.as_default():
spectrogram_function, spectrogram_thumb, spectrogram_info = common.spectrograms(
args)
# save spectrogram_thumb to hyperparams
args.spectrogram_thumb = spectrogram_thumb
def preload_fn(aa):
annot_path, uid = aa.annotation
if uid.startswith("mdb_"):
uid = uid + "_mel4"
aa.annotation = datasets.Annotation.from_time_series(
annot_path,
uid,
hop_samples=args.frame_width * args.samplerate / 44100,
unique_mf0=True)
aa.audio.load_resampled_audio(args.samplerate).load_spectrogram(
spectrogram_function, spectrogram_thumb, spectrogram_info[2])
def dataset_transform(tf_dataset, dataset):
return tf_dataset.map(dataset.prepare_example,
num_parallel_calls=args.threads).batch(
args.batch_size_evaluation).prefetch(10)
def dataset_transform_train(tf_dataset, dataset):
return tf_dataset.shuffle(
10**5).map(dataset.prepare_example,
num_parallel_calls=args.threads).batch(
args.batch_size).prefetch(10)
train_dataset, test_datasets, validation_datasets = common.prepare_datasets(
args.datasets, args, preload_fn, dataset_transform,
dataset_transform_train)
if not args.voicing:
for vd in validation_datasets:
if not vd.name.startswith("small_"):
vd.hooks.append(AdjustVoicingHook())
network.construct(args,
create_model,
train_dataset.dataset.output_types,
train_dataset.dataset.output_shapes,
spectrogram_info=spectrogram_info)
return network, train_dataset, validation_datasets, test_datasets
def construct(args):
network = NetworkMelody(args)
with network.session.graph.as_default():
def preload_fn(aa):
aa.annotation = datasets.Annotation.from_time_series(*aa.annotation)
aa.audio.load_resampled_audio(args.samplerate)
def dataset_transform(tf_dataset, dataset):
return tf_dataset.map(dataset.prepare_example, num_parallel_calls=args.threads).batch(args.batch_size_evaluation).prefetch(10)
def dataset_transform_train(tf_dataset, dataset):
return tf_dataset.shuffle(10**5).map(dataset.prepare_example, num_parallel_calls=args.threads).filter(dataset.is_example_voiced).batch(args.batch_size).prefetch(10)
train_dataset, test_datasets, validation_datasets = common.prepare_datasets(args.datasets, args, preload_fn, dataset_transform, dataset_transform_train)
network.construct(args, create_model, train_dataset.dataset.output_types, train_dataset.dataset.output_shapes)
return network, train_dataset, validation_datasets, test_datasets
def construct(args):
network = NetworkMelody(args)
with network.session.graph.as_default():
def preload_fn(aa):
aa.annotation = datasets.Annotation.from_time_series(
*aa.annotation)
aa.audio.load_resampled_audio(args.samplerate)
# augment_audio_basa = datasets.Audio("/mnt/tera/jirka/V1/MatthewEntwistle_FairerHopes/MatthewEntwistle_FairerHopes_STEMS/MatthewEntwistle_FairerHopes_STEM_07.wav",
# "augment_low").load_resampled_audio(args.samplerate).slice(20, 30)
# augment_audio_perkuse = datasets.Audio("/mnt/tera/jirka/V1/MatthewEntwistle_FairerHopes/MatthewEntwistle_FairerHopes_STEMS/MatthewEntwistle_FairerHopes_STEM_08.wav",
# "augment_low").load_resampled_audio(args.samplerate).slice(20, 30)
# augment_audio = augment_audio_basa.samples*10 + augment_audio_perkuse.samples*10
def dataset_transform(tf_dataset, dataset):
return tf_dataset.map(dataset.prepare_example).batch(
args.batch_size_evaluation).prefetch(10)
# return tf_dataset.map(dataset.prepare_example).map(dataset.mix_example_with(augment_audio)).batch(args.batch_size_evaluation).prefetch(1)
def dataset_transform_train(tf_dataset, dataset):
return tf_dataset.shuffle(10**5).map(
dataset.prepare_example,
num_parallel_calls=4).filter(dataset.is_example_voiced).batch(
args.batch_size).prefetch(10)
train_dataset, test_datasets, validation_datasets = common.prepare_datasets(
args.datasets, args, preload_fn, dataset_transform,
dataset_transform_train)
# Add voicing hook to the validation dataset
for vd in validation_datasets:
if not vd.name.startswith("small_"):
vd.hooks.append(AdjustVoicingHook())
network.construct(args, create_model,
train_dataset.dataset.output_types,
train_dataset.dataset.output_shapes)
return network, train_dataset, validation_datasets, test_datasets