log_prob (custom used in RL)

def log_prob(self, value, pre_tanh_value=None):
    """

    :param value: some value, x
    :param pre_tanh_value: arctanh(x)
    :return:
    """

    if pre_tanh_value is None:
        pre_tanh_value = self.atanh(value)

    return self.normal.log_prob(pre_tanh_value) - torch.log(
        1 - value * value + self.epsilon
    )


###################################################################

def forward(self, obs, reparameterize=True, return_log_prob=True):
    log_prob = None

    tanh_normal = self.actor(obs,reparameterize=reparameterize,)

    if return_log_prob:
        if reparameterize is True:
            action, pre_tanh_value = tanh_normal.rsample(
                return_pretanh_value=True
            )
        else:
            action, pre_tanh_value = tanh_normal.sample(
                return_pretanh_value=True
            )

        log_prob = tanh_normal.log_prob(
            action,
            pre_tanh_value=pre_tanh_value
        )


        log_prob = log_prob.sum(dim=1, keepdim=True) # get the entropy of the actions
    else:
        if reparameterize is True:
            action = tanh_normal.rsample()
        else:
            action = tanh_normal.sample()
    return action, log_prob

from：offlinerl/neorl

标签：pre,log,tanh,value,custom,RL,True,prob
来源： https://www.cnblogs.com/leifzhang/p/16200628.html