add mask to loss

[marouaneamz]

this pull request adds option of the mask to the loss which allows not to take into account images in the loss if they do not have labels for certain task

[piercus]

@marouaneamz please see inline comments

[marouaneamz]

I try to adapt the pipeline if there is no mask , but I can't do your method because we add the labels with the first if (if we haven't hardcoded the tasks here ), and in this block we add the labels and the maks we go through the whole results list and not all of them at once

[piercus]

Suggested change

	for k, v in results.items():
	if k.endswith('_img_label'):
	task = k[:-10]
	gt_label[task] = to_tensor(v)
	elif k.endswith('_mask'):
	task = k[:-5]
	gt_mask[task] = v
	for k, v in results.items():
	if k.endswith('_img_label'):
	task = k[:-10]
	if task not in gt_label:
	gt_label[task] = dict()
	gt_label[task]['label'] = to_tensor(v)
	elif k.endswith('_mask'):
	task = k[:-5]
	if task not in gt_label:
	gt_label[task] = dict()
	gt_label[task]['mask'] = v