Why does the L1 loss keep fluctuating during the training of RealESRNetModel+MSRResNet? 