[Tensorflow] [自定义损失函数]错误：请确保所有操作都定义了渐变

Question

我正在尝试编写一个自定义损失函数，如下所示：

# Compute the mahalanobis distance
def maha(x,u,S):
    x_minus_u_transpose = tf.transpose(x - u)
    inv_S = tf.linalg.inv(S)
    x_minus_u = x - u
    left_term = tf.matmul(x_minus_u, inv_S)
    D_square = tf.matmul(left_term, x_minus_u_transpose)
    return D_square 

# Form a matrix from the prediction and compute the mahalanobis distance as loss
def custom_loss_function(vhmx,vhmy,glx,gly,l10,l11):
    z = tf.zeros([32,1])
    def loss(y_true,y_pred):
    # 32 is the batch size
        mDistance = 0
        for i in range(32):
            u = tf.Variable([glx,gly],validate_shape=False,dtype = tf.float32)
            x = tf.Variable([vhmx,vhmy],validate_shape=False,dtype = tf.float32)
            predLowerTriangularMatrix = tf.Variable([[y_pred[i],z[i]],[l10[i],l11[i]]],tf.float32)
            predMatrix = tf.reshape(predLowerTriangularMatrix,[2,2])
            predCovarianceMatrix = tf.matmul(predMatrix,tf.transpose(predMatrix))
            mDistance = mDistance + maha(x,u,predCovarianceMatrix)
        return ((1-mDistance)*(1-mDistance))/32
    return loss

当我尝试拟合模型时，出现此错误。

发生异常：ValueError 变量具有None用于渐变。请确保您所有的操作都定义了渐变（即可区分）。没有渐变的常见操作：K.argmax，K.round，K.eval。

有人可以帮我调试一下吗？