def main(filepath, model_type, output_dir, gpu_index, evaluate, mode):
songname = filepath.split('/')[-1].split('.')[0]
model_path = './MSnet/pretrain_model/MSnet_'+str(model_type)
if gpu_index is not None:
with torch.cuda.device(gpu_index):
est_arr = MeExt(filepath, model_type=model_type, model_path=model_path, GPU=True, mode=mode)
else:
est_arr = MeExt(filepath, model_type=model_type, model_path=model_path, GPU=False, mode=mode)
if not os.path.exists(output_dir):
os.makedirs(output_dir)
print('Save the result in '+output_dir+'/'+songname+'.txt')
np.savetxt(output_dir+'/'+songname+'.txt', est_arr)
if evaluate is not None:
import pandas as pd
from MSnet.utils import melody_eval
if 'csv' in evaluate:
ycsv = pd.read_csv(evaluate, names = ["time", "freq"])
gtt = ycsv['time'].values
gtf = ycsv['freq'].values
ref_arr = np.concatenate((gtt[:,None], gtf[:,None]), axis=1)
elif 'txt' in evaluate:
ref_arr = np.loadtxt(evaluate)
else:
print("Error: Wrong type of ground truth. The file must be '.txt' or '.csv' ")
return None
eval_arr = melody_eval(ref_arr, est_arr)
print(songname, ' | VR: {:.2f}% VFA: {:.2f}% RPA: {:.2f}% RCA: {:.2f}% OA: {:.2f}%'.format(
eval_arr[0], eval_arr[1], eval_arr[2], eval_arr[3], eval_arr[4]))
def main(data_dir, model_type, output_dir, gpu_index, dataset='Mdb_vocal'):
model_path = './MSnet/pretrain_model/MSnet_' + str(model_type)
if 'Mdb_vocal' in dataset:
_, _, songlist = get_split_lists_vocal()
elif 'Mdb_melody2' in dataset:
_, _, songlist = get_split_lists()
else:
print('Error: Wrong type of dataset, Must be Mdb_vocal or Mdb_melody2')
if not os.path.exists(output_dir):
os.makedirs(output_dir)
with open(output_dir + '/' + str(dataset) + '_result.csv', 'w',
newline='') as csvfile:
writer = csv.writer(csvfile)
writer.writerow(['Songname', 'VR', 'VFA', 'RPA', 'RCA', 'OA'])
avg_arr = [0, 0, 0, 0, 0]
for songname in songlist:
if 'Mdb_vocal' in dataset:
filepath = data_dir + '/Audio/' + songname + '/' + songname + '_MIX.wav'
ypath = data_dir + '/Annotations/Melody_Annotations/MELODY2/' + songname + '_MELODY2.csv'
lpath = data_dir + 'Annotations/Instrument_Activations/SOURCEID/' + songname + '_SOURCEID.lab'
ref_arr = select_vocal_track(ypath, lpath)
elif 'Mdb_melody2' in dataset:
filepath = data_dir + '/Audio/' + songname + '/' + songname + '_MIX.wav'
ypath = data_dir + '/Annotations/Melody_Annotations/MELODY2/' + songname + '_MELODY2.csv'
ref_arr = csv2ref(ypath)
else:
print('No add dataset')
if gpu_index is not None:
with torch.cuda.device(gpu_index):
est_arr = MeExt(filepath,
model_type=model_type,
model_path=model_path,
GPU=True,
mode='std',
gid=gpu_index)
else:
est_arr = MeExt(filepath,
model_type=model_type,
model_path=model_path,
GPU=False,
mode='std')
# np.savetxt(output_dir+'/'+songname+'.txt', est_arr)
eval_arr = melody_eval(ref_arr, est_arr)
avg_arr += eval_arr
writer.writerow([
songname, eval_arr[0], eval_arr[1], eval_arr[2], eval_arr[3],
eval_arr[4]
])
avg_arr /= len(songlist)
writer.writerow([
'Avg', avg_arr[0], avg_arr[1], avg_arr[2], avg_arr[3], avg_arr[4]
])