Better failure message when loss explodes during FP16 training

        overflow = DynamicLossScaler.has_overflow(grad_norm)

        self.scaler.update_scale(overflow)

        if overflow:

            if self.scaler.loss_scale <= self.args.min_loss_scale:

                raise Exception((

                    'Minimum loss scale reached ({}). Your loss is probably exploding. '

                    'Try lowering the learning rate, using gradient clipping or '

                    'increasing the batch size.'

                ).format(self.args.min_loss_scale))

            raise OverflowError('setting loss scale to: ' + str(self.scaler.loss_scale))

        return grad_norm

                       help='learning rate shrink factor for annealing, lr_new = (lr * lr_shrink)')

    group.add_argument('--min-lr', default=1e-5, type=float, metavar='LR',

                       help='minimum learning rate')

    group.add_argument('--min-loss-scale', default=1e-4, type=float, metavar='D',

                       help='minimum loss scale (for FP16 training)')

    return group