Question

我正在尝试在Python中使用自动差异化来实现成本函数。我遇到了一些困难，因为我无法区分NumPy数组和autograd ArrayBox;错误是

ValueError: setting an array element with a sequence.

我使用的代码是：

import autograd.numpy as np
from autograd import grad 

n = 11 # Number of rows from original dataset
data = np.transpose(mcn_observed[:n+1, :])
# mcn_observed is simply a dataset imported with Pandas and
# converted to a NumPy array. 5000 rows and 10 columns

def cost(w):
    # Cost function for autograd

    output = np.empty((10,n))
    output[:, 0] = 0.1 * np.ones((10))

    W = w[:, :10] # Matrix of weights
    k = w[:, 10] # Vector of bias

    for i in range(1, n):
        # Euler's method for ODE
        lhs = np.dot(W, output[:, i-1]) - k*output[:, i-1]
        lhs = 0.1 * lhs
        output[:,i] = output[:, i-1] + lhs.reshape(output[:, i-1].shape)

    # Take difference between simulated and observed data
    r = np.asarray(output - data)
    return sum(sum(r))

# Initialize matrix of weights and bias
W = np.random.rand(10,11) + 1.0
for j in range(10):
    print("Iteration: {}".format(j))
    g = grad(cost)(W)
    W = W - 0.01 * g # Gradient descent

错误似乎在output[:,i] = output[:, i-1] + lhs.reshape(output[:, i-1].shape)。

Answer 1

我遇到了同样的问题。尝试将autograd ArrayBox分配给numpy数组时，似乎会发生这种情况。目前对我来说没有任何意义...

这是导致此问题的非常简单的代码：

import autograd.numpy as np
from autograd import grad


def fun(x):

    a = np.zeros((1,))
    a[0] = x**2

    return a


if __name__ == '__main__':

    print(fun(1.0))
    gradfun = grad(fun)
    print(gradfun(1.0))

将NumPy数组和ArrayBox添加在一起

1 个答案: