请提出你的问题 Please ask your question

paddle version 2.2.2, 动态图
model训练中，维护一个model的exponential moving average，伪代码如下：
问题：加入update_ema_model后，训练的速度约为原来的1/2-1/3。尝试了在此基础上额外加入注解部分的consist loss，速度不会更慢。可见主要是update_ema_model使得速度降低，这是什么原因呢？

model = MyModel()
ema_model = MyModel()
freeze_model(ema_model)
for each iteration:
    output = model(input)
    loss = calculate_loss(output, labels)
    #####consist loss########
    # ema_output = ema_model(input)
    # loss = loss + calculate_consistent_loss(output, ema_output)
    ##############
    loss.backward()
    optimizer.step()
    optimizer.clear_grad()

    update_ema_model(ema_model, model, factor)

def freeze_model(model):
    for p in model.parameters():
        p.stop_gradient=True

@paddle.no_grad()
def update_ema_model(ema_model, model, factor):
    for (param_name, param_value), (ema_param_name, ema_param_value) in \
        zip(model.named_parameters(), ema_model.named_parameters()):
        assert param_name == ema_param_name, \
            'name not match for model: {} and ema_model: {}'.format(param_name, ema_param_name)
        new_ema = (1 - factor) * param_value + factor * ema_param_value
        ema_param_value.set_value(new_ema)

这种直接操作model参数使得速度变慢的现象在我另一个实验中也有出现。另一个实验是在基础loss上加了一个loss，目标是使当前model的参数值趋近于给定的tensor，该loss定义如下，速度也是大约是原来的1/2-1/3

def regularize_by_one(model, param_dict, omiga_dict):
    loss = 0
    for param_name, param_value in model.named_parameters():
        if param_name in param_dict:
            loss = loss + paddle.sum(omiga_dict[param_name] * ((param_value - param_dict[param_name])**2))

    return loss