implementation of ExponentialMovingAverage is not correct

[FrankWork]

def get_ema_if_exists(v, gvs):
    name = v.name.split(':')[0]
    ema_name = name+'/ExponentialMovingAverage:0'
    ema_v = [v for v in gvs if v.name == ema_name]
    if len(ema_v) == 0:
        ema_v = [v]
    return ema_v[0]

def get_ema_vars(*vs):
    if tf.get_variable_scope().reuse:
        gvs = tf.global_variables()
        vs = [get_ema_if_exists(v, gvs) for v in vs]
    if len(vs) == 1:
        return vs[0]
    else:
        return vs

g, b = get_ema_vars(g, b) I think g, b is the original tensor, not the ema

[Newmu]

Yeah, this is dead code that isn't used in the release.