def test_metric(references, estimates, is_framewise, is_sources, nb_win,
nb_hop):
metrics.bss_eval(references,
estimates,
framewise_filters=is_framewise,
bsseval_sources_version=is_sources,
window=nb_win,
hop=nb_hop)
def test_silent_input(references, estimates, is_framewise, is_sources, nb_win,
nb_hop):
estimates = np.zeros(references.shape)
with pytest.raises(ValueError):
metrics.bss_eval(references,
estimates,
framewise_filters=is_framewise,
bsseval_sources_version=is_sources,
window=nb_win,
hop=nb_hop)
def eval_estimated(reference, ref_labels, estimated, est_labels, sampling_rate, win, hop, mode, compute_permutation=True):
if len(reference) != len(estimated):
raise Exception("You need the same amount of references as estimates!")
reference_np, estimated_np = pad_or_truncate(reference, estimated)
SDR, ISR, SIR, SAR, _ = bss_eval(
reference_np,
estimated_np,
compute_permutation=compute_permutation,
window=int(win*sampling_rate),
hop=int(hop*sampling_rate),
framewise_filters=(mode == "v3"),
bsseval_sources_version=False
)
source_results = []
for i, (ref_label, est_labels) in enumerate(zip(ref_labels, est_labels)):
source_results.append({
"reference_name": ref_label,
"estimated_name": est_labels,
"SDR": SDR[i].tolist(),
"SIR": SIR[i].tolist(),
"ISR": ISR[i].tolist(),
"SAR": SAR[i].tolist()
})
return source_results
def test_window(self, audios, rate):
from museval.metrics import bss_eval
if len(audios) != 2:
raise ValueError('BSSEval needs a reference and a test signals.')
result = bss_eval(reference_sources=audios[0].T,
estimated_sources=audios[1].T,
window=self.bss_window * rate,
hop=self.bss_hop * rate)
return {'sdr': result[0], 'isr': result[1], 'sar': result[3]}
def test_empty_input(is_framewise, is_sources, nb_win, nb_hop):
inputs = [np.array([]), np.array([])]
with pytest.warns(UserWarning):
output = metrics.bss_eval(*inputs,
framewise_filters=is_framewise,
bsseval_sources_version=is_sources,
window=nb_win,
hop=nb_hop)
assert np.allclose(output, np.array([]))
import numpy as np
from museval.metrics import bss_eval
from spleeter.audio.adapter import get_default_audio_adapter
parser = argparse.ArgumentParser()
parser.add_argument('original')
parser.add_argument('estimate')
args = parser.parse_args()
print("load audio")
audio_loader = get_default_audio_adapter()
sample_rate = 44100
tracks = []
estimates = []
for i in ['bass', 'drums', 'other', 'vocals']:
filename = os.path.join(args.original, i+'.wav')
track, _ = audio_loader.load(filename, sample_rate=sample_rate)
a = np.sum(track, axis=1)
print(np.all(a == 0))
tracks.append(track + 1e-7)
filename2 = os.path.join(args.estimate, i+'.wav')
estimate, _ = audio_loader.load(filename2, sample_rate=sample_rate)
estimates.append(estimate + 1e-7)
tracks = np.array(tracks)
estimates = np.array(estimates)
print(tracks.shape)
print(estimates.shape)
ans = bss_eval(tracks, estimates)
SDR = ans[0]
print(np.average(SDR, axis=1))