Multitask learning with uncertainty ?

[kakusikun]

Here, it looks like that the uncertainty is used to learn multitask.

BUT, I can NOT find that the parameters is updated by any optimizer ...
I just found that the instance of loss is made here.

Could you point out where the uncertainty is learned during training?

[ifzhang]

We set the parameters here:

FairMOT/src/lib/trains/mot.py

Line 35 in 12f50ca

self.s_det = nn.Parameter(-1.85 * torch.ones(1))

[kakusikun]

Thanks for your quick reply.
I know the uncertainty is set as a parameter.
But it needs to be added to the optimizer to be updated not just make it a parameter.
Still the question, where it is updated.

[ifzhang]

I use the parameters in the loss below:

FairMOT/src/lib/trains/mot.py

Line 76 in 4cdb349

loss = torch.exp(-self.s_det) * det_loss + torch.exp(-self.s_id) * id_loss + (self.s_det + self.s_id)

[kakusikun]

I mean something like

FairMOT/src/train.py

Line 44 in 4cdb349

optimizer = torch.optim.Adam(model.parameters(), opt.lr)

could you point out the same action on uncertainty like above.

[ifzhang]

Thank you very much for your question! I have added the parameters to optimizer and fixed the bug.

[whut2962575697]

I found that you added a new param_group in the optimizer, so the optimizer will fail when it is resumed

[whut2962575697]

maybe you should change the position of resume code

[ifzhang]

Thanks, I will fix the bug.