6D reconstruction leads to nans when using dipole with default cheetah tracking

[jp-ga]

6D reconstruction leads to nans when using dipole with default cheetah tracking. Workaround is leaving Dipole tracking_method="bmadx", but we should look into why the default tracking leads to nans.

gpsr/gpsr/modeling.py

Line 125 in 3644c54

tracking_method="bmadx",

[seongyeolk]

I also experienced nan several times for the reconstruction even with the tracking method to be bmadx.

I tried to find the source and I got the nan case if the batch size is not the multiple number of total number of datasets. And especially it happens often when I used the function "normalize_images" (I also saw that the recon process went back to normal if I turned off normalization function and use the arbitrary batch size number).

Currently I do not clearly know why there is a relationship between batch size and nan.