def stream(batch_size=64):
path = '/hdd/musicnet/train_data'
pattern = '*.wav'
samplerate = zounds.SR22050()
feature_spec = {'spectrogram': (256, 128)}
feature_funcs = {'spectrogram': (spectrogram, (samplerate, ))}
bs = batch_stream(path, pattern, batch_size, feature_spec, 'spectrogram',
feature_funcs)
return bs
def repr_stream(repr_class):
path = '/hdd/musicnet/train_data'
pattern = '*.wav'
total_samples = 2**17
audio_channels = 1
feature_spec = {'audio': (total_samples, audio_channels)}
feature_funcs = {'audio': (audio, (samplerate, ))}
batch_size = 2
bs = batch_stream(path, pattern, batch_size, feature_spec, 'audio',
feature_funcs)
for samples, in bs:
rep = repr_class.from_audio(samples, samplerate)
yield rep
def stream(total_samples=8192, batch_size=32):
path = '/hdd/musicnet/train_data'
pattern = '*.wav'
samplerate = zounds.SR22050()
# total_samples = 8192
feature_spec = {'audio': (total_samples, 1)}
feature_funcs = {'audio': (audio, (samplerate, ))}
# batch_size = 32
bs = batch_stream(path, pattern, batch_size, feature_spec, 'audio',
feature_funcs)
for batch, in bs:
transformed = IdentityPhaseReovery.from_audio(batch, samplerate)
yield batch, transformed
import zounds
import torch
import numpy as np
from matplotlib import pyplot as plt
path = '/hdd/musicnet/train_data'
pattern = '*.wav'
total_samples = 2**17
samplerate = zounds.SR22050()
feature_spec = {'audio': (total_samples, 1)}
feature_funcs = {'audio': (audio, (samplerate, ))}
batch_size = 1
bs = batch_stream(path, pattern, batch_size, feature_spec, 'audio',
feature_funcs)
if __name__ == '__main__':
# app = zounds.ZoundsApp(locals=locals(), globals=globals())
# app.start_in_thread(9999)
# samples, = next(bs)
# samples = torch.from_numpy(samples)
# min_size = 2 ** (np.log2(total_samples) - 4)
# bands = fft_frequency_decompose(samples, min_size)
# samples = zounds.AudioSamples(samples.squeeze(), samplerate)
# input('Waiting...')
n_bands = 5
sr = samplerate
for i in range(n_bands):
start_hz = 0 if i == (n_bands - 1) else sr.nyquist / 2