Mirrored strategy distributed training.

Has anyone worked with the tf.distribute.MirroredStrategy? I seem to be getting low model accuracy when using it for image segmentation with Dice loss.

Also saw minor gains in timing when running with 8 gpu setup.