first fix to make training resumption possible, but this will slow down the training by 5-10s per epoch (if model improves)
