Currently, AUTO and SUM_OVER_BATCH reduction operations on all losses fail when the batch size is dynamically defined. These reduction operations are very frequently used. Dynamic definition of batch size as well.
The problem comes from the LossesHelper class which seeks to divide the sum of the error by the size of the batch which is not dynamically obtained currently
Currently, AUTO and SUM_OVER_BATCH reduction operations on all losses fail when the batch size is dynamically defined. These reduction operations are very frequently used. Dynamic definition of batch size as well.
The problem comes from the LossesHelper class which seeks to divide the sum of the error by the size of the batch which is not dynamically obtained currently