PolicySearch.py 文件源码

def adam(self,cost, params, learning_rate=0.001, beta1=0.9,
             beta2=0.999, epsilon=1e-8):

        all_grads = T.grad(cost=cost, wrt=params)
        all_grads = total_norm_constraint(all_grads,10)

        grad_norm = T.sqrt(sum(map(lambda x: T.sqr(x).sum(), all_grads)))
        not_finite = T.or_(T.isnan(grad_norm), T.isinf(grad_norm))

        t_prev = theano.shared(utils.floatX(0.))
        updates = OrderedDict()

        t = t_prev + 1
        a_t = learning_rate*T.sqrt(1-beta2**t)/(1-beta1**t)

        for param, g_t in zip(params, all_grads):
            g_t = T.switch(not_finite, 0.1 * param,g_t)
            value = param.get_value(borrow=True)
            m_prev = theano.shared(np.zeros(value.shape, dtype=value.dtype),
                                   broadcastable=param.broadcastable)
            v_prev = theano.shared(np.zeros(value.shape, dtype=value.dtype),
                                   broadcastable=param.broadcastable)

            m_t = beta1*m_prev + (1-beta1)*g_t
            v_t = beta2*v_prev + (1-beta2)*g_t**2
            step = a_t*m_t/(T.sqrt(v_t) + epsilon)

            updates[m_prev] = m_t
            updates[v_prev] = v_t
            updates[param] = param - step

        updates[t_prev] = t
        return updates