Resetting hidden state during refinement

[qzed]

It looks looks you're resetting the hidden-state during each RAFT-based refinement step. Is this done on purpose?

unimatch/unimatch/unimatch.py

Lines 317 to 322 in 0dfa361

	net, inp = torch.chunk(proj, chunks=2, dim=1)
	net = torch.tanh(net)
	inp = torch.relu(inp)
	net, up_mask, residual_flow = self.refine(net, inp, correlation, flow.clone(),

[haofeixu]

No, just based on intuitions :)

[qzed]

Thanks! Feel free to close this issue (although I think it could be nice to have a comparison whether this leads to some improvement).