I'm very glad to see that torch.complex64 is default setting in this efficient toolbox, and issue #89 rarely happens when normally calling the functions. But when I built a dataset under torch.complex64, running by cuda, similar error like #89 raised.
I noticed that some of the functions in table_interp_adjoint under _nufft.interp automatically changes the data dtype. To overcome this little bug, I recommend adding .to(dtype) in line 685 to make sure the output of imag_exp() does not interfere original datatype.
If this bug has already been fixed in previous commits, i'm sorry for bothering.
I'm very glad to see that
torch.complex64is default setting in this efficient toolbox, and issue #89 rarely happens when normally calling the functions. But when I built a dataset undertorch.complex64, running by cuda, similar error like #89 raised.I noticed that some of the functions in
table_interp_adjointunder _nufft.interp automatically changes the data dtype. To overcome this little bug, I recommend adding.to(dtype)in line 685 to make sure the output ofimag_exp()does not interfere original datatype.If this bug has already been fixed in previous commits, i'm sorry for bothering.