Experiment reproduction issue with updated modules about s4 HOT 5 CLOSED

Nice catch! Can you pull and try again?

from s4.

I happen to be getting a different issue now

TypeError: __init__() got an unexpected keyword argument 'hurwitz'

I can see that this is one of the parameters in the S4 block that was removed in one of the recent updates to the project, so I would assume that this might pop up in some other experiments as well if they weren't updated.

from s4.

You're right, I just pushed another fix for this.

Actually, I can't test right now on my local environment. The model API did change very slightly in the latest release and I wasn't able to test the Sashimi experiments, so there may be some small bugs in the configs. These should only be very small tweaks though in the model and configs. If you could figure out the necessary changes to get the model to run and report back, that would be extremely helpful :)

from s4.

I think your most recent commit fixed the issues with the experiments by just commenting out the hurwitz parameter in the config files.

The stand-alone SaShiMi module still has an issue where in setup_rnn, the mode is being passed to the submodules in the SaShiMi model on lines 443-444

for module in self.modules():
    if hasattr(module, 'setup_step'): module.setup_step(mode)

but this is passed to the standalone S4 module which doesn't take in the mode as a parameter. I just changed the signature of each setup_step function (in S4 and HippoSSKernel) to pass in the mode downwards into the SSKernelNPLR module for the time being. This is a pretty hacky fix though so I'm sure you might want to look into if there's a better way to handle this later. This leads to another error

torch._C._LinAlgError: linalg.solve: (Batch element 0): The diagonal element 1 is zero, the solve could not be completed because the input matrix is singular.

which I didn't run into beforehand when I tried python sashimi.py prior to the modifications to the standalone S4 kernel implementation.

from s4.

V3 should have fixed any potential issues here. Feel free to re-open an issue if you find any further problems.

from s4.