Cuda out of memory error for ecalib and pics about bart

[siatjiasen]

Hi,

I got same out of memory error previously. Dr. Martin Uecker told me to try 'rsense' . Currently, 'rsense' comes with limited regularization options, thus my solution is:

1. perform 1D ifft along Readout dimensions.
2. using gpu pics to recon 288x236x6 data (repeat 320 times).

The last dimensions 6 is channel dimension for parallel imaging, that is using 6 receivers of different sensitivities to do sampling. Thus you can't slim down the 4th dimension from 6 to 1.

Best wishes,

[jtamir]

I haven't tested the GPU implementation of ecalib, so I can't comment on it.

As @siatjiasen said, you are running out of memory. If you are just interested in seeing if it's possible to run on the GPU, you could try to coil-compress the data to a smaller number of channels first using the bart cc command. But you will probably lose some accuracy.

If you want to go the slice-by-slice route and require more advanced options than are currently available through bart rsense, you can call bart pics in a loop. I suggest using the Grasp script as a reference.

In the future we will probably merge rsense and pics.