def _match_to_lbl(self, proposals, bbx, cat, match):
cls_lbl = []
bbx_lbl = []
for i, (proposals_i, bbx_i, cat_i,
match_i) in enumerate(zip(proposals, bbx, cat, match)):
if match_i is not None:
pos = match_i >= 0
# Objectness labels
cls_lbl_i = proposals_i.new_zeros(proposals_i.size(0),
dtype=torch.long)
cls_lbl_i[pos] = cat_i[match_i[pos]] + 1 - self.num_stuff
# Bounding box regression labels
if pos.any().item():
bbx_lbl_i = calculate_shift(proposals_i[pos],
bbx_i[match_i[pos]])
bbx_lbl_i *= bbx_lbl_i.new(self.bbx_reg_weights)
else:
bbx_lbl_i = None
cls_lbl.append(cls_lbl_i)
bbx_lbl.append(bbx_lbl_i)
else:
cls_lbl.append(None)
bbx_lbl.append(None)
return PackedSequence(cls_lbl), PackedSequence(bbx_lbl)
def _match_to_lbl(self, proposals, bbx, cat, ids, msk, match):
cls_lbl = []
bbx_lbl = []
msk_lbl = []
for i, (proposals_i, bbx_i, cat_i, ids_i, msk_i,
match_i) in enumerate(zip(proposals, bbx, cat, ids, msk,
match)):
if match_i is not None:
pos = match_i >= 0
# Objectness labels
cls_lbl_i = proposals_i.new_zeros(proposals_i.size(0),
dtype=torch.long)
cls_lbl_i[pos] = cat_i[match_i[pos]] + 1 - self.num_stuff
# Bounding box regression labels
if pos.any().item():
bbx_lbl_i = calculate_shift(proposals_i[pos],
bbx_i[match_i[pos]])
bbx_lbl_i *= bbx_lbl_i.new(self.bbx_reg_weights)
iis_lbl_i = ids_i[match_i[pos]]
# Compute instance segmentation masks
msk_i = roi_sampling(
msk_i.unsqueeze(0),
proposals_i[pos],
msk_i.new_zeros(pos.long().sum().item()),
self.lbl_roi_size, #28*28
interpolation="nearest")
# Calculate mask segmentation labels
msk_lbl_i = (msk_i == iis_lbl_i.view(
-1, 1, 1, 1)).any(dim=1).to(torch.long)
if not self.void_is_background:
msk_lbl_i[(msk_i == 0).all(dim=1)] = -1
else:
bbx_lbl_i = None
msk_lbl_i = None
cls_lbl.append(cls_lbl_i)
bbx_lbl.append(bbx_lbl_i)
msk_lbl.append(msk_lbl_i)
else:
cls_lbl.append(None)
bbx_lbl.append(None)
msk_lbl.append(None)
return PackedSequence(cls_lbl), PackedSequence(
bbx_lbl), PackedSequence(msk_lbl)
def _match_to_lbl(anchors, bbx, match):
pos, neg = match >= 0, match == -1
# Objectness labels from matching tensor
obj_lbl = torch.full_like(match, -1)
obj_lbl[neg] = 0
obj_lbl[pos] = 1
# Bounding box regression labels from matching tensor
bbx_lbl = anchors.new_zeros(len(bbx), anchors.size(0), anchors.size(1))
for i, (pos_i, bbx_i, match_i) in enumerate(zip(pos, bbx, match)):
if pos_i.any():
bbx_lbl[i, pos_i] = calculate_shift(anchors[pos_i],
bbx_i[match_i[pos_i]])
return obj_lbl, bbx_lbl