loss_t grows when training M-net?

[Jason-xin]

loss_t grows when training M-net?

[vlordier]

You can split Trimap and Matting networks by freezing each, and use requires_grad = False in train.py

For instance, after loading the model :

# depending on the training phase, freeze model part
if args.train_phase == 't_net':
	for name, child in model.named_children():
		if name is 'm_net':	
			child.requires_grad = False

if args.train_phase == 'm_net':
	for name, child in model.named_children():
		if name is 'm_net':
			child.requires_grad = False

also, you can change the loss to focus on each part of the network

if args.train_phase == 't_net':
	loss = L_t

if args.train_phase == 'm_net':
	loss = L_p

if args.train_phase == 'end_to_end':
		loss = L_p      + 0.01*L_t