AttributeError: 'ArrayBox' object has no attribute 'exp'

Hi everyone,

I’m implementing an alpha-divergence minimisation algorithm with a pretty simple loss function that I need to differentiate. The problem, it seems, is that the function contains exp and log, which should be totally fine.

Moreover, if I isolate the exponential or the logarithm function I can compute its gradient normally, so I guess the problem may be in the context of my loss function.

The code looks like this:

def wake_sleep(inputs, inputs_f, S, theta, alpha):
    nn_params = [1e-1, [1, 10, 10, 1]]
    z = [init_random_params(nn_params) for _ in range(S)]
    init_theta = theta

    def bb_alpha_energy(theta, t):
        print(theta)
        m_a, s_a, sigma = theta
        a = np.random.normal(size=(S, S)) * s_a + m_a
        m_qgp = 1. / S * np.sum([g_nn(inputs, z_i) for z_i in z], axis=0)

        mc_q = []
        for a_k in a:
            m_k = []
            for z_s, a_s in zip(z, a_k):
                m_k.append((g_nn(inputs_f, z_s) - m_qgp) * a_s)
            m = (1. / np.sqrt(S - 1) * np.sum(m_k, axis = 0))
            mc_q.append(gaussian_pdf(inputs_f, m, sigma)**alpha)
        mc_q = np.mean(mc_q, axis = 0)

        alpha_energy = 1. / alpha * np.sum(np.log(mc_q)) - kl(m_a, s_a, 0., 1.)
        return alpha_energy

    grad_loss = grad(bb_alpha_energy)
    opt_theta = adam(grad_loss, init_theta, step_size=0.01, num_iters=500)
    return opt_theta

The failure occurs when the tracer goes through mc_q.append(gaussian_pdf(inputs_f, m, sigma)**alpha).

May it be because I’m mixing vector-output functions (such as gaussian_pdf) with my scalar-output function? How should I go about this?

Thank you, Alex

[j-towns]

Are you using import numpy as np? If so change to import autograd.numpy as np.

[jennyJRwong]

我也遇到了类似的问题，使用了几行简单得多的代码。下列：

def function(self, variable=None): result = np.exp(variable) return result 生成错误消息：

result = np.exp(variable)

AttributeError：“ ArrayBox”对象没有属性“ exp”

但是，只需将对np.exp的调用替换为

def function(self, variable=None): from math import e result = e**(variable) return result 运行良好。

how about the log() the same issue.