def SaveEstimatedLabelsFramewise(est, filename, feature=None):
f = open(filename, "w")
cur_label = est[0]
st = 0
ed = 1
for i in range(est.size):
if est[i] != cur_label:
ed = i
if feature is not None:
feat = feature[st:ed, :]
sign = voc.ChordSignature7thbass(cur_label,
feat,
sevenths=True,
inv=True)
else:
sign = voc.ChordSignature(cur_label)
text = "%.4f %.4f %s\n" % (float(st * const.H) / const.SR,
float(ed * const.H) / const.SR, sign)
f.write(text)
cur_label = est[i]
st = i
if feature is not None:
feat = feature[st:est.size, :]
text = "%.4f %.4f %s" % (float(st * const.H) / const.SR,
float(est.size * const.H) / const.SR,
voc.ChordSignature7thbass(cur_label, feat))
else:
text = "%.4f %.4f %s" % (float(st * const.H) / const.SR,
float(est.size * const.H) / const.SR,
voc.ChordSignature(cur_label))
f.write(text)
f.close()
def EvaluateChord(idx, verbose=True, sonify=False, cross=False):
lablist = np.array(find_files(const.PATH_CHORDLAB, ext=["lab",
"chords"]))[idx]
est_lablist = np.array(find_files(
const.PATH_ESTIMATE_CROSS, ext="lab"))[idx] if cross else find_files(
const.PATH_ESTIMATE, ext="lab")
scorelist_majmin = np.array([])
scorelist_sevenths = np.array([])
scorelist_majmininv = np.array([])
scorelist_seventhinv = np.array([])
durations = np.array([])
confmatrix = np.zeros((const.N_CHORDS, const.N_CHORDS))
song_durations = np.array([])
for labfile, estfile in zip(lablist, est_lablist):
(ref_intervals,
ref_labels) = mir_eval.io.load_labeled_intervals(labfile)
(est_intervals,
est_labels) = mir_eval.io.load_labeled_intervals(estfile)
est_intervals, est_labels = mir_eval.util.adjust_intervals(
est_intervals, est_labels, ref_intervals.min(),
ref_intervals.max(), mir_eval.chord.NO_CHORD,
mir_eval.chord.NO_CHORD)
(intervals, ref_labels,
est_labels) = mir_eval.util.merge_labeled_intervals(
ref_intervals, ref_labels, est_intervals, est_labels)
durations = mir_eval.util.intervals_to_durations(intervals)
comparisons_sevenths = mir_eval.chord.sevenths(ref_labels, est_labels)
comparisons_majmininv = mir_eval.chord.majmin_inv(
ref_labels, est_labels)
comparisons_seventhinv = mir_eval.chord.sevenths_inv(
ref_labels, est_labels)
comparisons_majmin = mir_eval.chord.majmin(ref_labels, est_labels)
score_majmin = mir_eval.chord.weighted_accuracy(
comparisons_majmin, durations)
scorelist_majmin = np.append(scorelist_majmin, score_majmin)
score_sevenths = mir_eval.chord.weighted_accuracy(
comparisons_sevenths, durations)
scorelist_sevenths = np.append(scorelist_sevenths, score_sevenths)
score_majmininv = mir_eval.chord.weighted_accuracy(
comparisons_majmininv, durations)
scorelist_majmininv = np.append(scorelist_majmininv, score_majmininv)
score_seventhinv = mir_eval.chord.weighted_accuracy(
comparisons_seventhinv, durations)
scorelist_seventhinv = np.append(scorelist_seventhinv,
score_seventhinv)
if verbose:
print("%s --- %.3f" % (labfile.split('/')[-1], score_majmin))
for i in range(len(ref_labels)):
confmatrix[voc.GetChordIDSign(ref_labels[i]),
voc.GetChordIDSign(est_labels[i])] += durations[i]
song_durations = np.append(song_durations, np.sum(durations))
return scorelist_majmin, scorelist_sevenths, scorelist_majmininv, scorelist_seventhinv, confmatrix, song_durations
def TrainConvnetExtractorDeepChroma(trainidx,
epoch=20,
saveas="convnet.model"):
cqtfilelist = np.array(find_files(const.PATH_HCQT, ext="npy"))[trainidx]
labfilelist = np.array(
find_files(const.PATH_CHORDLAB, ext=["lab", "chords"]))[trainidx]
#midifilelist = find_files(const.PATH_MIDI,ext="mid")[:filecnt]
config.train = True
config.enable_backprop = True
convnet = networks.ConvnetFeatExtractor()
model = networks.ConvnetPredictor(convnet)
model.to_gpu(0)
opt = optimizers.MomentumSGD()
opt.setup(model)
print("DeepChroma Convnet Training...")
print("start epochs...")
S = []
T = []
for cqtfile, labfile in zip(cqtfilelist, labfilelist):
cqt = np.load(cqtfile)
spec = utils.PreprocessSpec(cqt[:const.CQT_H, :, :])
targ = voc.LoadChromaTarget(labfile)
minlen = min([cqt.shape[1], targ.shape[0]])
S.append(spec[:, :minlen, :])
T.append(targ[:minlen, :])
S = np.concatenate(S, axis=1)
T = np.concatenate(T, axis=0)
assert (S.shape[1] == T.shape[0])
for ep in range(epoch):
sum_loss = 0
randidx = np.random.randint(0,
S.shape[1] - const.CONV_TRAIN_SEQLEN - 1,
S.shape[1] // const.CONV_TRAIN_SEQLEN * 4)
for i in range(0, randidx.size - const.CONV_TRAIN_BATCH,
const.CONV_TRAIN_BATCH):
x_batch = np.stack([
S[:, randidx[j]:randidx[j] + const.CONV_TRAIN_SEQLEN, :]
for j in range(i, i + const.CONV_TRAIN_BATCH)
])
t_batch = np.stack([
T[randidx[j]:randidx[j] + const.CONV_TRAIN_SEQLEN, :]
for j in range(i, i + const.CONV_TRAIN_BATCH)
])
x_in = cp.asarray(x_batch)
t_in = cp.asarray(t_batch)
model.cleargrads()
loss = model(x_in, t_in)
loss.backward()
opt.update()
sum_loss += loss.data
convnet.save(saveas)
print("epoch: %d/%d loss:%.04f" %
(ep + 1, epoch, sum_loss / const.CONV_TRAIN_BATCH))
convnet.save(saveas)
def LoadLabelArr(labfile, hop_size=const.H):
f = open(labfile)
line = f.readline()
labarr = np.zeros(800 * const.SR // hop_size, dtype="int32")
while line != "" and line.isspace() == False:
items = line.split()
st = int(round(float(items[0]) * const.SR / hop_size))
ed = int(round(float(items[1]) * const.SR / hop_size))
lab = voc.GetChordIDSign(items[2])
labarr[st:ed] = lab
line = f.readline()
return labarr[:ed]
def ConfMatrix(idx, cross=False):
lablist = np.array(find_files(const.PATH_CHORDLAB, ext=["lab",
"chords"]))[idx]
est_lablist = np.array(find_files(
const.PATH_ESTIMATE_CROSS, ext="lab"))[idx] if cross else find_files(
const.PATH_ESTIMATE, ext="lab")
durations = np.array([])
confmatrix_chord = np.zeros((const.N_CHORDS, const.N_CHORDS))
confmatrix_quality = np.zeros((3, 3))
confmatrix_bass = np.zeros((3, 3))
for labfile, estfile in zip(lablist, est_lablist):
(ref_intervals,
ref_labels) = mir_eval.io.load_labeled_intervals(labfile)
(est_intervals,
est_labels) = mir_eval.io.load_labeled_intervals(estfile)
est_intervals, est_labels = mir_eval.util.adjust_intervals(
est_intervals, est_labels, ref_intervals.min(),
ref_intervals.max(), mir_eval.chord.NO_CHORD,
mir_eval.chord.NO_CHORD)
(intervals, ref_labels,
est_labels) = mir_eval.util.merge_labeled_intervals(
ref_intervals, ref_labels, est_intervals, est_labels)
durations = mir_eval.util.intervals_to_durations(intervals)
for i in range(len(ref_labels)):
confmatrix_chord[voc.GetChordIDSign(ref_labels[i]),
voc.GetChordIDSign(est_labels[i])] += durations[i]
ref_b = voc.GetBassFromSign(ref_labels[i])
est_b = voc.GetBassFromSign(est_labels[i])
if voc.TriadIsequal(
ref_labels[i],
est_labels[i]) and (ref_b >= 0) and (est_b >= 0):
confmatrix_bass[ref_b, est_b] += durations[i]
ref_q = voc.GetQualityFromSign(ref_labels[i])
est_q = voc.GetQualityFromSign(est_labels[i])
if voc.TriadIsequal(
ref_labels[i],
est_labels[i]) and (ref_q >= 0) and (est_q >= 0):
confmatrix_quality[ref_q, est_q] += durations[i]
return confmatrix_chord, confmatrix_quality, confmatrix_bass
def TrainTranscribeDNNChord(idx, epoch=20, saveas="dnn_deepchroma.model"):
cqtfilelist = np.array(find_files(const.PATH_HCQT, ext="npy"))[idx]
chordlablist = np.array(
find_files(const.PATH_CHORDLAB, ext=["lab", "chords"]))[idx]
featurelist = []
targetlist = []
chainer.config.train = True
chainer.config.enable_backprop = True
for cqtfile, labfile in zip(cqtfilelist, chordlablist):
cqt = np.load(cqtfile)[0, :, :]
chroma = voc.LoadChromaTarget(labfile)
min_sz = min([cqt.shape[0], chroma.shape[0]])
cqt = utils.Embed(utils.PreprocessSpec(cqt[:min_sz]), size=7)
chroma = chroma[:min_sz]
featurelist.append(cqt)
targetlist.append(chroma.astype(np.int32))
featurelist = np.concatenate(featurelist)
targetlist = np.concatenate(targetlist)
itemcnt = targetlist.shape[0]
print("DNN Training begin...")
dnn = networks.FeatureDNN()
dnn.train = True
model = networks.DNNModel(predictor=dnn)
model.to_gpu()
opt = optimizers.AdaDelta()
opt.setup(model)
for ep in range(epoch):
randidx = np.random.permutation(itemcnt)
sumloss = 0.0
for i in range(0, itemcnt, const.DNN_TRAIN_BATCH):
X = cp.asarray(featurelist[randidx[i:i + const.DNN_TRAIN_BATCH]])
T = cp.asarray(targetlist[randidx[i:i + const.DNN_TRAIN_BATCH]])
opt.update(model, X, T)
sumloss += model.loss.data * const.DNN_TRAIN_BATCH
print("epoch %d/%d loss=%.3f" % (ep + 1, epoch, sumloss / itemcnt))
dnn.save(saveas)