就是取前几个epoch的weight的平均值，可以缓解微调时的灾难性遗忘（因为新数据引导，模型权重逐渐，偏离训练时学到的数据分布，忘记之前学好的先验知识）

class EMA():def __init__(self, model, decay):self.model = modelself.decay = decay  # decay rateself.shadow = {}  # old weightself.backup = {}  # new weightdef register(self):  # deep copy weight for initfor name, param in self.model.named_parameters():if param.requires_grad:self.shadow[name] = param.data.clone()def update(self):  # ema：average weight for trainfor name, param in self.model.named_parameters():if param.requires_grad:assert name in self.shadownew_average = (1.0 - self.decay) * param.data + self.decay * self.shadow[name]self.shadow[name] = new_average.clone()def apply_shadow(self):  # load old weight for eval beginfor name, param in self.model.named_parameters():if param.requires_grad:assert name in self.shadowself.backup[name] = param.dataparam.data = self.shadow[name]def restore(self):  # load new weight for eval endfor name, param in self.model.named_parameters():if param.requires_grad:assert name in self.backupparam.data = self.backup[name]self.backup = {}# 初始化
ema = EMA(model, 0.999)
ema.register()# 训练过程中，更新完参数后，同步update shadow weights
def train():optimizer.step()ema.update()# eval前，apply shadow weights；eval之后，恢复原来模型的参数
def evaluate():ema.apply_shadow()# evaluateema.restore()