def __init__(self,
sr=SR,
n_fft=NFFT,
hop_length=HOP,
n_mels=NMEL,
n_bins=NBIN,
mode='mel'):
self.sr = sr
self.n_fft = n_fft
self.hop_length = hop_length
self.n_mels = n_mels
self.n_bins = n_bins
if mode == 'mel':
self.spectrogram = Spectrogram.MelSpectrogram(
sr=sr,
n_fft=n_fft,
n_mels=n_mels,
hop_length=hop_length,
fmin=20,
fmax=11000)
elif mode == 'cqt':
self.spectrogram = Spectrogram.CQT(sr=sr,
hop_length=hop_length,
fmin=22.5,
n_bins=n_bins,
bins_per_octave=24,
pad_mode='constant')
import nnAudio.Spectrogram as Spec
from plots import plot_cqt
from parameters import *
if USE_CQT:
cqt_layer = Spec.CQT(sr=FS,
hop_length=HOP_LENGTH,
fmin=F_MIN,
n_bins=N_BINS,
bins_per_octave=BINS_PER_OCTAVE,
norm=NORM,
pad_mode='constant',
window=WINDOW)
cqt_layer.to(DEVICE)
def cqt(signal, numpy=True, db=True):
time_array = np.arange(np.ceil(
signal.size / HOP_LENGTH).astype(int)) / (FS / HOP_LENGTH)
signal_tensor = torch.tensor(signal, device=DEVICE, dtype=torch.float)
cqt_tensor = cqt_layer(signal_tensor, normalization_type='wrap')
if db:
cqt_tensor = 20 * torch.log10(cqt_tensor + EPS)
if numpy:
cqt_array = cqt_tensor.cpu().numpy()[0, :, :]
torch.cuda.empty_cache()
return cqt_array, time_array
else:
print("Using CPU")
device = "cpu"
aug = Augment(Chords())
config = yaml.load(open("./config/config.yaml"))
sr = config['preprocess']['sample_rate']
hop_size = config['preprocess']['hop_size']
window_size = config['preprocess']['window_size']
song_hz = config['preprocess']['song_hz']
save_dir = config['preprocess']['save_dir']
cqt_layer = Spectrogram.CQT(device=device,
sr=sr,
hop_length=hop_size,
fmin=220,
fmax=None,
n_bins=108,
bins_per_octave=24,
norm=1,
window='hann',
center=True,
pad_mode='reflect')
p = Preprocess(sr, hop_size, song_hz, window_size, save_dir, aug, cqt_layer)
num_epochs = config['model'].get('num_epochs')
def get_data():
datasets = {
"isophonics-beetles": {
"mp3": config['preprocess']['data_path'] + "/beetles_albums",
"labels":
