def prepare_gt_as_class(gt_locs):
assert len(gt_locs) <= 1
if len(gt_locs) == 1:
azi, ele = apkit.vec2ae(gt_locs[0])
return int((azi + np.pi) // (2 * np.pi / _N_CLASSES))
else:
return 0 # assume one and only one source
def load_gt_cont(path):
with open(path) as s:
gt = pickle.load(s)
ae = np.array([apkit.vec2ae(p) for p, _, _, _, _, _, in gt[4]])
assert len(ae) == 1
return ae[0, 0]
def prepare_gt_as_heatmap(gt_locs):
hmap = np.zeros(_N_CLASSES)
aindex = np.arange(_N_CLASSES, dtype='float32') \
* 2 * np.pi / _N_CLASSES
for l in gt_locs:
azi, ele = apkit.vec2ae(l)
d = np.remainder(aindex - azi, 2 * np.pi)
d[d > np.pi] -= 2 * np.pi
h = np.exp(-d**2 / _SIGMA**2)
hmap = np.maximum(hmap, h)
return hmap.astype('float32', copy=False)
def prepare_gt_as_ordered_posterior(gt_locs):
assert len(gt_locs) <= 2
labels = []
for l in gt_locs:
azi, ele = apkit.vec2ae(l)
labels.append(
int(azi / (2 * np.pi / _OPOST_N_DIR) + 0.5) % _OPOST_N_DIR)
for _ in xrange(2 - len(labels)):
labels.append(_OPOST_N_DIR)
labels = sorted(labels)
post = np.zeros(2 * (_OPOST_N_DIR + 1))
post[labels[0]] = 1.0
post[labels[1] + _OPOST_N_DIR + 1] = 1.0
return post.astype('float32', copy=False)
def prepare_gtf_location_and_snsc(gt):
res = np.zeros((2, _N_CLASSES), dtype='float32')
if len(gt) > 0:
ctb = np.zeros((len(gt), _N_CLASSES))
for i, (l, _, _) in enumerate(gt):
azi, ele = apkit.vec2ae(l)
d = np.remainder(_AZI_DIRECTIONS - azi, 2 * np.pi)
d[d > np.pi] -= 2 * np.pi
ctb[i] = np.exp(-d**2 / _SIGMA**2)
stype = np.array([t for _, t, _ in gt], dtype='float32')
# localization map (likelihood)
res[0] = np.max(ctb, axis=0)
# speech/non-speech map
res[1] = stype[np.argmax(ctb, axis=0)]
return res
def _gt_azimuth_distance(src_doa, ref_doas):
azi, ele = apkit.vec2ae(src_doa)
d = np.remainder(ref_doas - azi, 2 * np.pi)
d[d > np.pi] -= 2 * np.pi
return d
def prepare_gt_as_posterior(gt_locs):
post = np.zeros(_N_CLASSES)
for l in gt_locs:
azi, ele = apkit.vec2ae(l)
post[int(azi / (2 * np.pi / _N_CLASSES) + 0.5) % _N_CLASSES] = 1.0
return post.astype('float32', copy=False)