Bug about your IMTL code

[ZZX-11]

I think your IMTL code have bug.
if follow your code : parameters = itertools.chain(model.parameters(), IMTL.parameters())
and run list(IMTL.parameters()) will get a empty list. Also, after optimizer.step(), the s_t will not update.

The problem comes from you not initial self.register_parameter('s_t',nn.Parameter(torch.ones(self.num_losses+1).squeeze(),requires_grad=True))
in the init() function, therefore optimizer will not update s_t
I have changed your code like this :

[ZZX-11]

And when I use gradient mode, the NN can't get grad, and get updated

[JohnLaMaster]

Hi, thank you for pointing that out. I have fixed the code. When defining the module, you will now need to add the expected number of losses as an input. The README documentation was also updated.

Can you give me some details regarding the issue with the gradients? Was that dues to the initialization error, which sould now be resolved, or was that still an issue after your modification?

[ZZX-11]

Hi, I've find an better implementation of IMTL and other multitask algorithms. The MTL methods offered are plug-and-play, making them exceptionally convenient to use:

1. https://github.com/AvivNavon/nash-mtl/blob/main/methods/weight_methods.py
2. https://github.com/Cranial-XIX/FAMO/blob/main/methods/weight_methods.py