def lab2target_single(lab):
if lab == "N" or lab == "X":
return C.LABEL_N
root, _, bass = chord.encode(lab, reduce_extended_chords=True)
_, quality, _, _ = chord.split(lab, reduce_extended_chords=True)
target = quality_triad_map[quality] * 12 + root
return target
def encode_chordseq_hierarchical(seq):
vec = np.zeros((len(seq), 6), dtype=np.int32)
for i in range(len(seq)):
lab = seq[i]
if (lab != "N"):
root, bitmap, bass = chord.encode(lab,
reduce_extended_chords=False)
_, quality, scale, _ = chord.split(lab,
reduce_extended_chords=False)
vec[i, 0] = quality_triad_map[quality] * 12 + root
vec[i, 1] = (root + bass) % 12
vec[i, 2] = quality_sevenths_map[quality]
vec[i, 3] = quality_ninths_map[quality]
vec[i, 4] = quality_elevenths_map[quality]
vec[i, 5] = quality_thirteenths_map[quality]
"""
for sc in scale:
tone = chord.scale_degree_to_semitone(sc)
if tone in [9,10,11]:
vec[i,2] = tone - 9
elif tone in [13,14,15]:
vec[i,3] = tone - 13
elif tone in [17,18]:
vec[i,4] = tone - 17
elif tone in [20,21]:
vec[i,5] = tone - 20
"""
elif lab == "N" or lab == "X":
vec[i, :] = [C.LABEL_N, 12, 3, 3, 2, 2]
return vec
def evaluate_feature(feat, labstr):
root, bitmap, bass = chord.encode(labstr, reduce_extended_chords=True)
bitmap = bitmap.astype(np.float32)
vec_bass = np.zeros(12, dtype=np.float32)
vec_bass[(root + bass) % 12] = 1
vec_template = np.concatenate((vec_bass, bitmap))
score = 1 - cosine(feat, vec_template)
return score
def shift_label(l, shift):
if l in ["N", "X"]:
return l
if shift == 0:
return l
l_split = chord.split(l)
l_root, _, _ = chord.encode(l)
l_root = (l_root + shift) % 12
l_root_str = notes[l_root]
return chord.join(l_root_str, l_split[1], l_split[2], l_split[3])
def generate_templates(vocab=157):
chord_labels = []
for qual in labels.QUALITIES[vocab][:-1]:
for root in labels.ROOTS:
chord_labels.append("%s:%s" % (root, qual))
chord_labels.append("N")
templates = np.zeros([vocab, 12])
for idx, label in enumerate(chord_labels):
root, semitones, bass = chord_eval.encode(label)
templates[idx, :] = chord_eval.rotate_bitmap_to_root(semitones, root)
templates[-1, :] = 0.0001
return templates
def lab2target(seq, module=np):
seq_target = []
for lab in seq:
if lab == "N":
target = C.LABEL_N
else:
_, quality, _, _ = chord.split(lab, reduce_extended_chords=True)
root, _, bass = chord.encode(lab, reduce_extended_chords=True)
quality_id = quality_triad_map[quality]
target = quality_id * 12 + root
seq_target.append(target)
return module.array(seq_target, dtype=np.int32)
def filter_label_seq(labs, shift=0):
new_seq = []
for l in labs:
if l == "N" or l == "X":
l_new = l
else:
sp = chord.split(l)
root, _, _ = chord.encode(l)
if shift != 0:
root_new = (root + shift) % 12
sp[0] = notes[root_new]
l_new = chord.join(sp[0], sp[1], None, sp[3])
new_seq.append(l_new)
return new_seq
def template_matching_matrix(feat, lablist):
templates = []
for lab in lablist:
root, bitmap, bass = chord.encode(lab, reduce_extended_chords=True)
bitmap = bitmap.astype(np.float32)
vec_bass = np.zeros(12, dtype=np.float32)
vec_bass[(root + bass) % 12] = 1
vec_template = np.concatenate((vec_bass, bitmap))
templates.append(vec_template)
templates = np.stack(templates, axis=1)
xp = C.cupy.get_array_module(feat)
prob_mat = xp.matmul(feat, xp.asarray(templates))
prob_mat_logz = xp.log(xp.sum(xp.exp(prob_mat), axis=1, keepdims=True))
prob_mat -= prob_mat_logz
return prob_mat
def LoadChromaTarget(labfile):
f = open(labfile)
line = f.readline()
labarr = np.zeros((600 * const.SR // const.H, 12), dtype="int32")
bassarr = np.zeros(labarr.shape, dtype="int32")
while line != "" and line.isspace() == False:
items = line.split()
st = int(round(float(items[0]) * const.SR / const.H))
ed = int(round(float(items[1]) * const.SR / const.H))
lab = items[2]
root, bitmap, bass = evalchord.encode(lab)
chroma = evalchord.rotate_bitmap_to_root(bitmap, root)
bassnumber = (root + bass) % 12
labarr[st:ed, :] = chroma
bassarr[st:ed, bassnumber] = 1
line = f.readline()
return np.concatenate((bassarr[:ed], labarr[:ed]),
axis=1).astype(np.float32)
def template_matching_matrix_along_class(feat):
templates = []
lablist = []
lablist.extend(notes)
lablist.extend([n + ":min" for n in notes])
lablist.append("N")
for lab in lablist:
root, bitmap, bass = chord.encode(lab, reduce_extended_chords=True)
bitmap = bitmap.astype(np.float32)
vec_bass = np.zeros(12, dtype=np.float32)
vec_bass[(root + bass) % 12] = 1
vec_template = np.concatenate((vec_bass, bitmap))
templates.append(vec_template)
templates = np.stack(templates, axis=1)
xp = C.cupy.get_array_module(feat)
prob_mat = xp.exp(xp.matmul(feat, xp.asarray(templates)))
prob_mat_sum = xp.sum(prob_mat, axis=1, keepdims=True)
prob_mat = prob_mat / prob_mat_sum
return prob_mat
ref_intervals, ref_labels, est_intervals, est_labels)
durations = mir_eval.util.intervals_to_durations(intervals)
ref_labels_id = encode_chordseq_hierarchical(ref_labels)
est_labels_id = encode_chordseq_hierarchical(est_labels)
for i in range(len(ref_labels)):
if durations[i] > 0.5 and ref_labels_id[
i, 0] == ref and est_labels_id[i, 0] == est:
list_examples.append((labfile, ref, est, intervals[i]))
return list_examples
template_matrix = np.zeros((C.N_VOCABULARY_TRIADS, 36), dtype=np.float32)
lablist = class2label(range(C.N_VOCABULARY_TRIADS))
for i in range(C.N_VOCABULARY_TRIADS):
root, bitmap, bass = chord.encode(lablist[i], reduce_extended_chords=True)
template_matrix[i, root] = 1
template_matrix[i, 12:24] = np.roll(bitmap, root)
template_matrix_zero_center = np.array(template_matrix)
template_matrix_zero_center[:, :
24] = template_matrix_zero_center[:, :24] * 4 - 2
"""
template_matrix = np.zeros((C.N_VOCABULARY_TRIADS,C.N_DIMS_FEAT),dtype=np.float32) + 1e-2
lablist = class2label(range(C.N_VOCABULARY_TRIADS))
for i in range(C.N_VOCABULARY_TRIADS):
root,bitmap,bass = chord.encode(lablist[i],reduce_extended_chords=True)
templ = np.roll(bitmap,root)
template_matrix[i,:] = np.concatenate([templ for _ in range(C.N_DIMS_FEAT//12)])
"""
label_shifter = np.zeros((C.N_VOCABULARY_TRIADS, C.N_VOCABULARY_TRIADS,