def _decode_head_forward_train(self, x, img_metas, gt_semantic_seg_0,
gt_semantic_seg_1, gt_semantic_seg_2):
"""Run forward function and calculate loss for decode head in
training."""
losses = dict()
# class check
#temp_gt = gt_semantic_seg
#temp_gt[temp_gt==255]=-1
rank, world_size = get_dist_info()
if rank == 100:
print(gt_semantic_seg_2.max(), gt_semantic_seg_1.max(),
gt_semantic_seg_0.max())
loss_decode_dir = self.lane_dir_head.forward_train(
x, img_metas, gt_semantic_seg_2, self.train_cfg)
loss_decode_sty = self.lane_sty_head.forward_train(
x, img_metas, gt_semantic_seg_1, self.train_cfg)
loss_decode_typ = self.lane_typ_head.forward_train(
x, img_metas, gt_semantic_seg_0, self.train_cfg)
rank, world_size = get_dist_info()
if rank == 10:
print('pin_size', gt_semantic_seg.size(), 'pin_size')
print('\npin 0', gt_semantic_seg[:, :, :, 0].max(), '\npin 0')
print('\npin 1', gt_semantic_seg[:, :, :, 1].max(), '\npin 0')
print('\npin 2', gt_semantic_seg[:, :, :, 2].max(), '\npin 0')
losses.update(add_prefix(loss_decode_dir, 'decode_lane_dir'))
losses.update(add_prefix(loss_decode_sty, 'decode_lane_sty'))
losses.update(add_prefix(loss_decode_typ, 'decode_lane_typ'))
return losses
def _loss_regularization_forward_train(self):
"""Calculate regularization loss for model weight in training."""
losses = dict()
if isinstance(self.loss_regularization, nn.ModuleList):
for idx, regularize_loss in enumerate(self.loss_regularization):
loss_regularize = dict(
loss_regularize=regularize_loss(self.modules()))
losses.update(add_prefix(loss_regularize, f'regularize_{idx}'))
else:
loss_regularize = dict(
loss_regularize=self.loss_regularization(self.modules()))
losses.update(add_prefix(loss_regularize, 'regularize'))
return losses
def losses(self, seg_logit, seg_label):
"""Compute ``pam_cam``, ``pam``, ``cam`` loss."""
pam_cam_seg_logit, pam_seg_logit, cam_seg_logit = seg_logit
loss = dict()
loss.update(
add_prefix(
super(DAHead, self).losses(pam_cam_seg_logit, seg_label),
'pam_cam'))
loss.update(
add_prefix(
super(DAHead, self).losses(pam_seg_logit, seg_label), 'pam'))
loss.update(
add_prefix(
super(DAHead, self).losses(cam_seg_logit, seg_label), 'cam'))
return loss
def _auxiliary_head_forward_train(self, x, img_metas, gt_semantic_seg):
"""Run forward function and calculate loss for auxiliary head in
training."""
losses = dict()
if isinstance(self.auxiliary_head, nn.ModuleList):
for idx, aux_head in enumerate(self.auxiliary_head):
loss_aux = aux_head.forward_train(x, img_metas,
gt_semantic_seg,
self.train_cfg)
losses.update(add_prefix(loss_aux, f'aux_{idx}'))
else:
loss_aux = self.auxiliary_head.forward_train(
x, img_metas, gt_semantic_seg, self.train_cfg)
losses.update(add_prefix(loss_aux, 'aux'))
return losses
def _decode_head_forward_train_r(self, x, img_metas, gt_semantic_seg):
"""Run forward function and calculate loss for decode head in
training."""
losses = dict()
loss_decode = self.decode_head_r.forward_train(x, img_metas,
gt_semantic_seg,
self.train_cfg)
losses.update(add_prefix(loss_decode, 'decode_r'))
return losses
def _decode_head_forward_train_r(self, x, img_metas, gt_semantic_seg):
"""Run forward function and calculate loss for decode head in
training."""
losses = dict()
loss_decode = self.decode_head_r[0].forward_train(
x, img_metas, gt_semantic_seg, self.train_cfg)
losses.update(add_prefix(loss_decode, 'decode_r_0'))
for i in range(1, self.num_stages):
# forward test again, maybe unnecessary for most methods.
prev_outputs = self.decode_head_r[i - 1].forward_test(
x, img_metas, self.test_cfg)
loss_decode = self.decode_head_r[i].forward_train(
x, prev_outputs, img_metas, gt_semantic_seg, self.train_cfg)
losses.update(add_prefix(loss_decode, f'decode_r_{i}'))
return losses
def _auxiliary_head_forward_train_kd(self, x, T_x, img_metas, gt_semantic_seg):
"""Run forward function and calculate loss for auxiliary head in
training."""
with torch.no_grad():
# x_T = self.teacher.extract_feat(img)
self.teacher.eval()
Teacher_seg_logit = self.teacher._decode_head_forward_test(T_x, img_metas)
losses = dict()
if isinstance(self.auxiliary_head, nn.ModuleList):
for idx, aux_head in enumerate(self.auxiliary_head):
loss_aux = aux_head.forward_train_kd(x, img_metas,
gt_semantic_seg,
self.train_cfg,
Teacher_seg_logit)
losses.update(add_prefix(loss_aux, f'aux_{idx}'))
else:
loss_aux = self.auxiliary_head.forward_train_kd(
x, img_metas, gt_semantic_seg, self.train_cfg, Teacher_seg_logit)
losses.update(add_prefix(loss_aux, 'aux'))
return losses
def _decode_head_forward_train_kd(self, x, T_x, img_metas, gt_semantic_seg):
"""Run forward function and calculate loss for decode head in
training."""
with torch.no_grad():
# x_T = self.teacher.extract_feat(img)
self.teacher.eval()
Teacher_seg_logit = self.teacher._decode_head_forward_test(T_x, img_metas)
# seg_logits = self.teacher.forward(inputs)
losses = dict()
loss_decode = self.decode_head.forward_train_kd(x, img_metas,
gt_semantic_seg,
self.train_cfg,
Teacher_seg_logit
)
losses.update(add_prefix(loss_decode, 'decode'))
return losses
