def loss(self, score, delta_per_class, target_delta_for_sample_roi,
bbox_bg_label_for_sample_roi):
assert isinstance(score, Variable)
assert isinstance(delta_per_class, Variable)
assert isinstance(target_delta_for_sample_roi, np.ndarray)
assert isinstance(bbox_bg_label_for_sample_roi, np.ndarray)
#---------- debug
target_delta_for_sample_roi = Variable(
torch.FloatTensor(target_delta_for_sample_roi))
bbox_bg_label_for_sample_roi = Variable(
torch.LongTensor(bbox_bg_label_for_sample_roi))
if torch.cuda.is_available():
target_delta_for_sample_roi = target_delta_for_sample_roi.cuda()
bbox_bg_label_for_sample_roi = bbox_bg_label_for_sample_roi.cuda()
n_sample = score.shape[0]
delta_per_class = delta_per_class.view(n_sample, -1, 4)
# get delta for roi w.r.t its corresponding bbox label
index = torch.arange(0, n_sample).long()
if torch.cuda.is_available():
index = index.cuda()
delta = delta_per_class[index, bbox_bg_label_for_sample_roi.data]
head_delta_loss = delta_loss(delta, target_delta_for_sample_roi,
bbox_bg_label_for_sample_roi, 1)
head_class_loss = F.cross_entropy(score, bbox_bg_label_for_sample_roi)
return head_delta_loss + head_class_loss
def loss(self, delta, score, anchor, gt_bbox, image_size):
#---------- debug
assert isinstance(delta, Variable)#this is a good way to check the conditions
assert isinstance(score, Variable)
assert isinstance(anchor, np.ndarray)
assert isinstance(gt_bbox, np.ndarray)
#---------- debug
target_delta, anchor_label = self.anchor_target_creator.make_anchor_target(anchor, gt_bbox, image_size)
target_delta = Variable(torch.FloatTensor(target_delta))
anchor_label = Variable(torch.LongTensor(anchor_label))
if torch.cuda.is_available():
target_delta, anchor_label = target_delta.cuda(), anchor_label.cuda()
rpn_delta_loss = delta_loss(delta, target_delta, anchor_label, 3)
rpn_class_loss = F.cross_entropy(score, anchor_label, ignore_index=-1) # ignore loss for label value -1
return rpn_delta_loss + rpn_class_loss