def loss(self, mask_pred, mask_targets, labels):
loss = dict()
if self.class_agnostic:
loss_mask = mask_cross_entropy(mask_pred, mask_targets,
torch.zeros_like(labels))
else:
loss_mask = mask_cross_entropy(mask_pred, mask_targets, labels)
loss['loss_mask'] = loss_mask
return loss
def loss(self, mask_pred, mask_targets, labels):
loss = dict()
if self.class_agnostic:
loss_mask = mask_cross_entropy(mask_pred, mask_targets,
torch.zeros_like(labels))
#loss_mask2 = mask_volume_loss(mask_pred, mask_targets, torch.zeros_like(labels))
else:
loss_mask = mask_cross_entropy(mask_pred, mask_targets, labels)
#loss_mask2 = mask_volume_loss(mask_pred, mask_targets, labels) * 100
# print("loss:", (loss_mask.item(), loss_mask2.item()))
# loss['loss_mask'] = loss_mask + loss_mask2
loss['loss_mask'] = loss_mask
return loss
def loss_aux(self, mask_pred, mask_targets, labels, alpha=0.25):
loss = dict()
mask_pred_level0 = mask_pred[0::4,:]
mask_pred_level1 = mask_pred[1::4,:]
mask_pred_level2 = mask_pred[2::4,:]
mask_pred_level3 = mask_pred[3::4,:]
if self.class_agnostic:
loss_mask_level0 = mask_cross_entropy(mask_pred_level0, mask_targets,
torch.zeros_like(labels))
loss_mask_level1 = mask_cross_entropy(mask_pred_level1, mask_targets,
torch.zeros_like(labels))
loss_mask_level2 = mask_cross_entropy(mask_pred_level2, mask_targets,
torch.zeros_like(labels))
loss_mask_level3 = mask_cross_entropy(mask_pred_level3, mask_targets,
torch.zeros_like(labels))
else:
loss_mask_level0 = mask_cross_entropy(mask_pred_level0, mask_targets, labels)
loss_mask_level1 = mask_cross_entropy(mask_pred_level1, mask_targets, labels)
loss_mask_level2 = mask_cross_entropy(mask_pred_level2, mask_targets, labels)
loss_mask_level3 = mask_cross_entropy(mask_pred_level3, mask_targets, labels)
loss['loss_mask_level0'] = loss_mask_level0 * alpha
loss['loss_mask_level1'] = loss_mask_level1 * alpha
loss['loss_mask_level2'] = loss_mask_level2 * alpha
loss['loss_mask_level3'] = loss_mask_level3 * alpha
return loss
def loss(self, mask_pred, mask_targets, iou_pred, iou_targets, labels):
loss = dict()
if self.class_agnostic:
loss_mask = mask_cross_entropy(mask_pred, mask_targets,
torch.zeros_like(labels))
iou_targets = self.mask_iou(mask_pred, mask_targets,
torch.zeros_like(labels))
# use smooth L1 loss instead of L2 loss.
# iou_loss = smooth_l1_loss(iou_pred, iou_targets, reduction='mean')
else:
loss_mask = mask_cross_entropy(mask_pred, mask_targets, labels)
iou_targets = self.mask_iou(mask_pred, mask_targets, labels)
# iou_loss = smooth_l1_loss()
loss_iou = smooth_l1_loss(iou_pred, iou_targets, labels)
loss['loss_mask'] = loss_mask
loss['loss_maskiou'] = loss_iou
return loss
def loss(self, mask_pred, mask_targets, labels):
loss = dict()
loss_mask = mask_cross_entropy(mask_pred, mask_targets, labels)
loss['loss_mask'] = loss_mask
return loss