def MeExt(filepath,
model_type='vocal',
model_path='./pretrain_model/MSnet_vocal',
GPU=True,
mode='std'):
if 'std' in mode:
data, CenFreq, time_arr = cfp_process(filepath, model_type=model_type)
elif 'fast' in mode:
data, CenFreq, time_arr = cfp_process(filepath,
model_type=model_type,
sr=22050,
hop=512)
print('Melody extraction with Melodic Segnet ...')
if 'vocal' in model_type:
Net = model.MSnet_vocal()
elif 'melody' in model_type:
Net = model.MSnet_melody()
else:
print(
"Error: Wrong type of model. Please assign model_type = 'vocal' or 'melody'"
)
return None
if GPU:
Net.cuda()
else:
Net.cpu()
Net.float()
Net.eval()
Net.load_state_dict(torch.load(model_path))
frames = data.shape[-1]
if frames > 5120:
seg_x = seg(data)
seg_y = []
for batch_x in seg_x:
batch_x = batch_x[np.newaxis, :]
batch_x = torch.from_numpy(batch_x).float()
if GPU:
batch_x = batch_x.cuda()
pred_y, emb = Net(batch_x)
pred_y = pred_y.cpu().detach().numpy()
seg_y.append(pred_y)
pred_y = iseg(seg_y)
else:
batch_x = data[np.newaxis, :]
batch_x = torch.from_numpy(batch_x).float()
if GPU:
batch_x = batch_x.cuda()
pred_y, emb = Net(batch_x)
pred_y = pred_y.cpu().detach().numpy()
est_arr = est(pred_y, CenFreq, time_arr)
return est_arr
def main(data_folder, model_type, output_folder):
if 'vocal' in model_type:
train_ids, val_ids, test_ids = get_split_lists_vocal(data_folder)
xlist = []
ylist = []
for chunk_id in train_ids:
filepath = data_folder + '/audio/synth_mix_' + chunk_id + '.wav'
data, CenFreq, time_arr = cfp_process(filepath, model_type=model_type, sr=44100, hop=256)
ypath = data_folder + '/annotations/melody/synth_mix_' + chunk_id + '.csv'
lpath = data_folder + '/annotations/activations/synth_mix_' + chunk_id + '.lab'
ref_arr = select_vocal_track(ypath, lpath)
gt_map = seq2map(ref_arr[:, 1], CenFreq)
xlist, ylist = batchize(data, gt_map, xlist, ylist, size=430)
xlist = np.array(xlist)
ylist = np.array(ylist)
hf = h5py.File('./data/train_vocal.h5', 'w')
hf.create_dataset('x', data=xlist)
hf.create_dataset('y', data=ylist)
hf.close()
xlist = []
ylist = []
for chunk_id in val_ids:
filepath = data_folder + '/audio/synth_mix_' + chunk_id + '.wav'
data, CenFreq, time_arr = cfp_process(filepath, model_type=model_type, sr=44100, hop=256)
data = batchize_val(data)
ypath = data_folder + '/annotations/melody/synth_mix_' + chunk_id + '.csv'
lpath = data_folder + '/annotations/activations/synth_mix_' + chunk_id + '.lab'
ref_arr = select_vocal_track(ypath, lpath)
xlist.append(data)
ylist.append(ref_arr)
with open(output_folder+'/val_x_vocal.pickle', 'wb') as fp:
pickle.dump(xlist, fp)
with open(output_folder+'/val_y_vocal.pickle', 'wb') as fp:
pickle.dump(ylist, fp)
elif 'melody' in model_type:
train_ids, val_ids, test_ids = get_split_lists()
xlist = []
ylist = []
for chunk_id in train_ids:
filepath = data_folder + '/audio/synth_mix_' + chunk_id + '.wav'
data, CenFreq, time_arr = cfp_process(filepath, model_type=model_type, sr=44100, hop=256)
ypath = data_folder + '/annotations/melody/synth_mix_' + chunk_id + '.csv'
ref_arr = csv2ref(ypath)
gt_map = seq2map(ref_arr[:,1], CenFreq)
xlist, ylist = batchize(data, gt_map, xlist, ylist, size=430)
xlist = np.array(xlist)
ylist = np.array(ylist)
hf = h5py.File(output_folder+'/train.h5', 'w')
hf.create_dataset('x', data=xlist)
hf.create_dataset('y', data=ylist)
hf.close()
xlist = []
ylist = []
for chunk_id in val_ids:
filepath = data_folder + '/audio/synth_mix_' + chunk_id + '.wav'
data, CenFreq, time_arr = cfp_process(filepath, model_type=model_type, sr=44100, hop=256)
data = batchize_val(data)
ypath = data_folder + '/annotations/melody/synth_mix_' + chunk_id + '.csv'
ref_arr = csv2ref(ypath)
xlist.append(data)
ylist.append(ref_arr)
with open(output_folder+'/val_x.pickle', 'wb') as fp:
pickle.dump(xlist, fp)
with open(output_folder+'/val_y.pickle', 'wb') as fp:
pickle.dump(ylist, fp)
def main(data_folder, model_type, output_folder):
if 'vocal' in model_type:
train_songlist, val_songlist, test_songlist = getlist_mdb_vocal()
xlist = []
ylist = []
for songname in train_list:
filepath = data_folder + '/Audio/' + songname + '/' + songname + '_MIX.wav'
data, CenFreq, time_arr = cfp_process(filepath,
model_type=model_type,
sr=44100,
hop=256)
ypath = data_folder + '/Annotations/Melody_Annotations/MELODY2/' + songname + '_MELODY2.csv'
lpath = data_folder + '/Annotations/Instrument_Activations/SOURCEID/' + songname + '_SOURCEID.lab'
ref_arr = select_vocal_track(ypath, lpath)
gt_map = seq2map(ref_arr[:, 1], CenFreq)
xlist, ylist = batchize(data, gt_map, xlist, ylist, size=430)
xlist = np.array(xlist)
ylist = np.array(ylist)
hf = h5py.File('./data/train_vocal.h5', 'w')
hf.create_dataset('x', data=xlist)
hf.create_dataset('y', data=ylist)
hf.close()
xlist = []
ylist = []
for songname in val_songlist:
filepath = data_folder + '/Audio/' + songname + '/' + songname + '_MIX.wav'
data, CenFreq, time_arr = cfp_process(filepath,
model_type=model_type,
sr=44100,
hop=256)
data = batchize_val(data)
ypath = data_folder + '/Annotations/Melody_Annotations/MELODY2/' + songname + '_MELODY2.csv'
lpath = data_folder + '/Annotations/Instrument_Activations/SOURCEID/' + songname + '_SOURCEID.lab'
ref_arr = select_vocal_track(ypath, lpath)
xlist.append(data)
ylist.append(ref_arr)
with open(output_folder + '/val_x_vocal.pickle', 'wb') as fp:
pickle.dump(xlist, fp)
with open(output_folder + '/val_y_vocal.pickle', 'wb') as fp:
pickle.dump(ylist, fp)
elif 'melody' in model_type:
train_songlist, val_songlist, test_songlist = getlist_mdb()
xlist = []
ylist = []
for songname in train_list:
filepath = data_folder + '/Audio/' + songname + '/' + songname + '_MIX.wav'
data, CenFreq, time_arr = cfp_process(filepath,
model_type=model_type,
sr=44100,
hop=256)
ypath = data_folder + '/Annotations/Melody_Annotations/MELODY2/' + songname + '_MELODY2.csv'
ref_arr = csv2ref(ypath)
gt_map = seq2map(ref_arr[:, 1], CenFreq)
xlist, ylist = batchize(data, gt_map, xlist, ylist, size=430)
xlist = np.array(xlist)
ylist = np.array(ylist)
hf = h5py.File(output_folder + '/train.h5', 'w')
hf.create_dataset('x', data=xlist)
hf.create_dataset('y', data=ylist)
hf.close()
xlist = []
ylist = []
for songname in val_songlist:
filepath = data_folder + '/Audio/' + songname + '/' + songname + '_MIX.wav'
data, CenFreq, time_arr = cfp_process(filepath,
model_type=model_type,
sr=44100,
hop=256)
data = batchize_val(data)
ypath = data_folder + '/Annotations/Melody_Annotations/MELODY2/' + songname + '_MELODY2.csv'
ref_arr = csv2ref(ypath)
xlist.append(data)
ylist.append(ref_arr)
with open(output_folder + '/val_x.pickle', 'wb') as fp:
pickle.dump(xlist, fp)
with open(output_folder + '/val_y.pickle', 'wb') as fp:
pickle.dump(ylist, fp)
import numpy as np
import wave as we
from MSnet.cfp import cfp_process
import os
rootdir = "./raw/"
npz_dir = './train_NPZ/'
list1 = os.listdir(rootdir)
wav_list = []
for name in list1:
wav_list.append(name.split('.')[0])
for namelist in wav_list:
print(namelist)
filepath = rootdir + '/' + namelist + '.wav'
data, CenFreq_useless, time_arr = cfp_process(filepath, model_type='vocal')
np_save_name = npz_dir + namelist + ".npz"
np.savez(np_save_name,
data=data,
CenFreq=CenFreq_useless,
time_arr=time_arr)