Preparation

In this example, we prepare training/validation data for melody/vocal separation task.

Before running bunch of codes below, please download DSD100 dataset from the DSD100 download page and say it DSD100.zip.

Initialize Sampler for train

from chimeranet.datasets import DSD100Melody, DSD100Vocal
from chimeranet import DatasetSampler, SyncSampler

dsd100_path = 'DSD100.zip'
duration = 2.
sr = 16000

# Dataset
vocal_dataset = DSD100Vocal(dsd100_path, dev=True, test=False)
# Sampler
vocal_sampler = DatasetSampler(vocal_dataset)
# augmentation
vocal_sampler.amplitude_factor = (-0.5, 0.5)
vocal_sampler.stretch_factor = (-0.1, 0.1)
vocal_sampler.shift_factor = (-0.1, 0.1)

# do same things to melody channel
melody_dataset = DSD100Melody(dsd100_path, dev=True, test=False)
melody_sampler = DatasetSampler(melody_dataset)
melody_sampler.amplitude_factor = (-0.5, 0.5)
melody_sampler.stretch_factor = (-0.1, 0.1)
melody_sampler.shift_factor = (-0.1, 0.1)

# combine two channels together
sampler = SyncSampler(vocal_sampler, melody_sampler)
sampler.duration = duration
sampler.samplerate = sr

Sample audio and make train data

import librosa
from chimeranet import to_training_data
n_fft = 512
hop_length = 128
n_mels = 150
sample_size = 32

T = librosa.time_to_frames(duration, sr, hop_length, n_fft)
F = n_mels

samples = sampler.sample(sample_size, n_jobs=4)
x_train, y_train = to_training_data(
    samples, T, F, sr=sr, n_fft=n_fft, hop_length=hop_length, n_jobs=4,
)

Save train data

import h5py
with h5py.File('example-dataset-train.hdf5', 'w') as f:
    f.create_dataset('x', data=x_train)
    f.create_dataset('y/mask', data=y_train['mask'])
    f.create_dataset('y/embedding', data=y_train['embedding'])

Show one of the training data

%matplotlib inline
import matplotlib.pyplot as plt

idx = 0
mixture = x_train[idx].transpose((1, 0))
mask_e = y_train['embedding'][idx].transpose((2, 1, 0))
mask_m = y_train['mask'][idx].transpose((2, 1, 0))

fig = plt.figure(figsize=(12, 8))
ax = fig.add_subplot(2, 3, 1)
ax.title.set_text('Mixture')
ax.imshow(mixture, origin='lower', aspect='auto')
for i, (mask, name) in enumerate(zip(mask_e, ('Vocal', 'Melody')), 1):
    ax = fig.add_subplot(2, 3, 1+i)
    ax.title.set_text(name)
    ax.imshow(mask, origin='lower', aspect='auto', vmin=0, vmax=1)
for i, (mask, name) in enumerate(zip(mask_m, ('Vocal', 'Melody')), 1):
    ax = fig.add_subplot(2, 3, 4+i)
    ax.title.set_text(name)
    ax.imshow(mask, origin='lower', aspect='auto', vmin=0, vmax=1)

Validation data

Like training data, validation data can be made in almost same way.

# vocal channel
vocal_dataset = DSD100Vocal(dsd100_path, dev=False, test=True)
vocal_sampler = DatasetSampler(vocal_dataset)
vocal_sampler.amplitude_factor = (-0.5, 0.5)
vocal_sampler.stretch_factor = (-0.1, 0.1)
vocal_sampler.shift_factor = (-0.1, 0.1)

# melody channel
melody_dataset = DSD100Melody(dsd100_path, dev=False, test=True)
melody_sampler = DatasetSampler(melody_dataset)
melody_sampler.amplitude_factor = (-0.5, 0.5)
melody_sampler.stretch_factor = (-0.1, 0.1)
melody_sampler.shift_factor = (-0.1, 0.1)

# combine two channels together
sampler = SyncSampler(vocal_sampler, melody_sampler)
sampler.duration = duration
sampler.samplerate = sr

# make validation data
samples = sampler.sample(sample_size, n_jobs=4)
x_validation, y_validation = to_training_data(
    samples, T, F, sr=sr, n_fft=n_fft, hop_length=hop_length, n_jobs=4,
)

# save validation data
with h5py.File('example-dataset-validation.hdf5', 'w') as f:
    f.create_dataset('x', data=x_train)
    f.create_dataset('y/mask', data=y_validation['mask'])
    f.create_dataset('y/embedding', data=y_validation['embedding'])

[1] A. Liutkus et al., “The 2016 Signal Separation Evaluation Campaign,” in Latent Variable Analysis and Signal Separation - 12th International Conference, {LVA/ICA} 2015, Liberec, Czech Republic, August 25-28, 2015, Proceedings, 2017, pp. 323–332.