TensorFlow：学习参数获得NAN

Question

我尝试下面的简单线性reg梯度下降样本是我的代码...

def gardientDecent(features,target):
     alpha = tf.constant(0.001,dtype=tf.float64)
     length = tf.cast(tf.size(target),dtype=tf.float64)
     feature_one = features[:,0]
     features_Val = tf.transpose(features)

     dummy_theta = tf.zeros([2,1],dtype=tf.float64)

     dot_product = tf.multiply(features_Val, dummy_theta)
     diff = tf.subtract(dot_product, target)
     diff_one = tf.multiply(diff, feature_one)

     theta1 = tf.reduce_sum(diff)
     theta2 = tf.reduce_sum(diff_one)

     final1 = tf.div(theta1, length)
     final2 = tf.div(theta2, length)

     t1 = tf.subtract(theta1, tf.multiply(alpha, final1))
     t2 = tf.subtract(theta2, tf.multiply(alpha, final2))

     newArray = np.array([[0],[0]],dtype=np.float64)
     for i in range(1000):
            temp1,temp2 = vas.run([t1,t2],{dummy_theta:newArray})
            print i ,temp1,temp2
            newArray = np.array([[temp1],[temp2]],dtype=np.float64)
     print vas.run(tf.cast(newArray,dtype=tf.float64))

但是在经过几个步骤后运行时，我会得到像这样的Nan值......

请帮帮我......对于ML和Tensor流程来说都是新手...... TY提前

Answer 1

似乎是你从参数中减去的损失，而不是相对于θ的实际的损失梯度。您应该为grad1 = tf.gradients(tf.reduce_sum(diff), theta1)（theta2）调用diff_one相同的内容，然后从theta1和theta2中减去（带有alpha权重）：{{1}和t1 = tf.subtract(theta1, tf.multiply(alpha, grad1))

相同