def ingest_all():
data = [
zounds.InternetArchive('AOC11B'),
zounds.InternetArchive('Greatest_Speeches_of_the_20th_Century'),
zounds.InternetArchive('Kevin_Gates_-_By_Any_Means-2014'),
zounds.PhatDrumLoops()
]
for d in data:
zounds.ingest(d, Sound, multi_threaded=True)
from client import Client
import requests
import urllib
from io import BytesIO
from log import module_logger
from pathlib import Path
import zounds
from bot_helper import BotDriver
from urllib.parse import urlparse
logger = module_logger(__file__)
datasets = [
# Classical
zounds.InternetArchive('AOC11B'),
zounds.InternetArchive('CHOPINBallades-NEWTRANSFER'),
zounds.InternetArchive('beethoven_ingigong_850'),
zounds.InternetArchive('jamendo-086440'),
zounds.InternetArchive('The_Four_Seasons_Vivaldi-10361'),
# Pop
zounds.InternetArchive('02.LostInTheShadowsLouGramm'),
zounds.InternetArchive('08Scandalous'),
zounds.InternetArchive('09.JoyceKennedyDidntITellYou'),
zounds.InternetArchive('02.InThisCountryRobinZander'),
zounds.InternetArchive('PeterGabrielOutOut'),
zounds.InternetArchive('07.SpeedOfLightJoeSatriani'),
# Jazz
zounds.InternetArchive('Free_20s_Jazz_Collection'),
parser = argparse.ArgumentParser()
parser.add_argument(
'--epochs',
help='how many epochs (full passes over data) should the network train',
default=100,
type=int)
parser.add_argument(
'--force',
help='retrain the network, even if its already been trained',
action='store_true',
default=False)
args = parser.parse_args()
zounds.ingest(
zounds.InternetArchive('Greatest_Speeches_of_the_20th_Century'),
Sound,
multi_threaded=True)
def generate_training_and_test_set():
snds = list(Sound)
# get all sounds where Nixon is the speaker
nixon = filter(lambda snd: 'Nixon' in snd.meta['artist'], snds)
# get an equal number of speeches by anyone besides Nixon
not_nixon = filter(lambda snd: 'Nixon' not in snd.meta['artist'],
snds)[:len(nixon)]
for snd in nixon:
yield dict(data=snd.rasterized,
model.parameters(), lr=0.0002, betas=(0.5, 0.999)),
discriminator_optim_func=lambda model: optim.Adam(
model.parameters(), lr=0.00005, betas=(0.5, 0.999)),
latent_dimension=(LATENT_DIM, ),
epochs=500,
batch_size=64),
needs=scaled,
store=False)
pipeline = ff.PickleFeature(zounds.PreprocessingPipeline,
needs=(mu_law, scaled, network),
store=True)
if __name__ == '__main__':
zounds.ingest(zounds.InternetArchive('AOC11B'), Sound, multi_threaded=True)
if not GanPipeline.exists():
GanPipeline.process(docs=(snd.rasterized for snd in Sound))
gan = GanPipeline()
noise = np.random.normal(0, 1, (32, LATENT_DIM))
generated_samples = gan.pipeline.transform(noise)
# start up an in-browser REPL to interact with the results
app = zounds.ZoundsApp(model=Sound,
audio_feature=Sound.ogg,
visualization_feature=Sound.bark,
globals=globals(),
locals=locals())
app.start(8888)
type=str,
help='the internet archive id to use for training')
parser.add_argument('--epochs',
type=int,
help='the number of epochs to train the network')
parser.add_argument(
'--force-train',
action='store_true',
help='re-train the network, even if it has already been trained')
parser.add_argument('--categorical',
action='store_true',
help='use a categorical distribution of samples')
args = parser.parse_args()
if args.internet_archive_id:
zounds.ingest(zounds.InternetArchive(args.internet_archive_id),
Sound,
multi_threaded=True)
if args.categorical:
network = CategoricalAutoEncoder()
loss = CategoricalLoss()
synthesize = categorical_synthesize
pipeline_cls = CategoricalAutoEncoderPipeline
data_preprocessor = label_preprocessor = preprocess_categorical
batch_size = 16
else:
network = RawSamplesAutoEncoder()
loss = FrequencyBandLoss()
synthesize = raw_samples_synthesize
pipeline_cls = AutoEncoderPipeline
encoded = zounds.ArrayWithUnitsFeature(zounds.Learned,
learned=FreqAdaptiveAutoEncoder(),
needs=freq_adaptive,
store=False)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--internet-archive-id',
help='the internet archive id to process',
type=str,
required=False,
default='AOC11B')
args = parser.parse_args()
zounds.ingest(dataset=zounds.InternetArchive(args.internet_archive_id),
cls=Sound,
multi_threaded=True)
# train the pipeline, including the autoencoder
if not FreqAdaptiveAutoEncoder.exists():
FreqAdaptiveAutoEncoder.process(docs=(snd.freq_adaptive
for snd in Sound))
# get a reference to the trained pipeline
autoencoder = FreqAdaptiveAutoEncoder()
# get references to all the sounds. features are lazily
# loaded/evaluated, so this is a cheap operation
snds = list(Sound)
packed, [x.dimensions[0], zounds.IdentityDimension()])
@zounds.simple_lmdb_settings('hamming_index',
map_size=1e11,
user_supplied_id=True)
class Sound(BaseModel):
fake_hash = zounds.ArrayWithUnitsFeature(produce_fake_hash,
needs=BaseModel.fft,
store=True)
if __name__ == '__main__':
zounds.ingest(zounds.InternetArchive('Kevin_Gates_-_By_Any_Means-2014'),
Sound,
multi_threaded=True)
def web_url(doc, ts):
return doc.meta['web_url']
def total_duration(doc, ts):
return doc.fake_hash.dimensions[0].end / zounds.Seconds(1)
index = zounds.HammingIndex(Sound,
Sound.fake_hash,
path='fake_hash_index',
web_url=web_url,
total_duration=total_duration)