def evaluate_model_all(test_guesses, test_labels):
'''
Run standard Mirex evaluations on all test data
Inputs:
1d np.array containing all predictions made by the model
1d np.array containing all ground truth labels
Outputs:
Dict holding results of all evaluations
'''
print('Running conversions...')
ref_freq = hr.note_to_hz_zeros(test_labels) # And back to Hz!
est_freq = hr.note_to_hz_zeros(test_guesses)
ref_cent = mel_eval.hz2cents(ref_freq) # Then to cents...
est_cent = mel_eval.hz2cents(est_freq)
ref_voicing = mel_eval.freq_to_voicing(ref_freq)[1] # And voicings!
est_voicing = mel_eval.freq_to_voicing(est_freq)[1]
print('Evaluating voicing...')
vx_recall, vx_false_alarm = mel_eval.voicing_measures(
ref_voicing, est_voicing)
print('Evaluating pitch...')
raw_pitch = mel_eval.raw_pitch_accuracy(ref_voicing,
ref_cent,
est_voicing,
est_cent,
cent_tolerance=50)
print('Evaluating chroma...')
raw_chroma = mel_eval.raw_chroma_accuracy(ref_voicing,
ref_cent,
est_voicing,
est_cent,
cent_tolerance=50)
print('Evaluating overall accuracy...')
overall_accuracy = mel_eval.overall_accuracy(ref_voicing,
ref_cent,
est_voicing,
est_cent,
cent_tolerance=50)
metrics = {
'vx_recall': vx_recall,
'vx_false_alarm ': vx_false_alarm,
'raw_pitch': raw_pitch,
'raw_chroma': raw_chroma,
'overall_accuracy': overall_accuracy
}
for m, v in metrics.items(): # Python2 is iteritems I think
print(m, ':', v)
return metrics
def format_annotation(new_times, annot_times, annot_freqs):
""" Format an annotation file and resample to a uniform timebase.
Parameters
----------
new_times : np.array
Times to resample to
annot_times : np.array
Annotation time stamps
annot_freqs : np.array
Annotation frequency values
Returns
-------
ref_cent : np.array
Annotation frequencies in cents at the new timescale
ref_voicing : np.array
Annotation voicings at the new timescale
"""
ref_freq, ref_voicing = melody.freq_to_voicing(annot_freqs)
ref_cent = melody.hz2cents(ref_freq)
ref_cent, ref_voicing = melody.resample_melody_series(
annot_times, ref_cent, ref_voicing, new_times,
kind='linear'
)
return ref_cent, ref_voicing
def eval_from_hz(ref_hz, est_hz):
ref_cent = melody.hz2cents(ref_hz)
est_cent = melody.hz2cents(est_hz)
ref_voicing = np.array(ref_cent > 0, dtype=float)
est_voicing = np.array(est_cent > 0, dtype=float)
recall, false_alarm = melody.voicing_measures(ref_voicing, est_voicing)
return {
"recall":
recall,
"false_alarm":
false_alarm,
"rca50":
melody.raw_chroma_accuracy(ref_voicing,
ref_cent,
est_voicing,
est_cent,
cent_tolerance=50),
"rpa50":
melody.raw_pitch_accuracy(ref_voicing,
ref_cent,
est_voicing,
est_cent,
cent_tolerance=50),
"rpa25":
melody.raw_pitch_accuracy(ref_voicing,
ref_cent,
est_voicing,
est_cent,
cent_tolerance=25),
"rpa10":
melody.raw_pitch_accuracy(ref_voicing,
ref_cent,
est_voicing,
est_cent,
cent_tolerance=10),
"std50":
eval_pitch_std(ref_voicing,
ref_cent,
est_voicing,
est_cent,
cent_tolerance=50)
}
def matrix_parser(m):
x = np.zeros(shape=(m.shape[0], 2))
for i in range(len(m)):
if (np.sum(m[i]) != 0):
x[i][0] = 1
x[i][1] = midi2freq(np.argmax(m[i]) / 4 + 21)
x[:, 1] = melody.hz2cents(x[:, 1])
return x
def evaluate_model_melody(test_guesses, test_labels):
'''
Run standard pitch and chroma evaluations on all test data
Inputs:
1d np.array containing all predictions made by the model
1d np.array containing all ground truth labels
Outputs:
Dict holding results of all evaluations
'''
ref_freq = hr.note_to_hz_zeros(test_labels)
est_freq = hr.note_to_hz_zeros(test_guesses)
ref_cent = mel_eval.hz2cents(ref_freq)
est_cent = mel_eval.hz2cents(est_freq)
all_voiced = np.ones(len(ref_cent), dtype=bool)
print('Evaluating pitch...')
raw_pitch = mel_eval.raw_pitch_accuracy(all_voiced,
ref_cent,
all_voiced,
est_cent,
cent_tolerance=50)
print('Evaluating chroma...')
raw_chroma = mel_eval.raw_chroma_accuracy(all_voiced,
ref_cent,
all_voiced,
est_cent,
cent_tolerance=50)
metrics = {
'raw_pitch': raw_pitch,
'raw_chroma': raw_chroma,
}
for m, v in metrics.items():
print(m, ':', v)
return metrics
def format_contour_data(frequencies):
""" Convert contour frequencies to cents + voicing.
Parameters
----------
frequencies : np.array
Contour frequency values
Returns
-------
est_cent : np.array
Contour frequencies in cents
est_voicing : np.array
Contour voicings
"""
est_freqs, est_voicing = melody.freq_to_voicing(frequencies)
est_cents = melody.hz2cents(est_freqs)
return est_cents, est_voicing
def getf0CentsVoicingArrays(f0Values):
f0ValuesArray, voicingArray = melody.freq_to_voicing(f0Values)
f0CentsArray = melody.hz2cents(f0ValuesArray, base_frequency=10.0)
return f0CentsArray, voicingArray