我的DNN对所有测试数据（tensorflow）返回相同的预测结果

Question

我试图在我的数据上运行此代码以进行回归。看起来网络可以预测第一个测试数据，但所有其他预测与第一个测试数据相同。第一个函数为初始化生成随机权重。预测变量的数量是54，输出的数量是4.这里是我的代码：

def init_weights(shape): 
    weights = tf.random_uniform(shape, -2,2) 
    return tf.Variable(weights) 

def forwardprop(X, w, b, sig): 
    if sig==1: 
     yhat = tf.sigmoid(tf.add(tf.matmul(X, w),b)) 
    else: 
     yhat = tf.add(tf.matmul(X, w),0.) 
return yhat 

def main(itr,starter_learning_rate):  

    x_size = train_X.shape[1] 
    h_size = 4 
    y_size = train_y.shape[1] 

    X = tf.placeholder("float", shape = [None, x_size]) 
    y = tf.placeholder("float", shape = [None, y_size]) 

    w_1 = init_weights((x_size, h_size)) 
    b_1 = tf.constant(1.) 

    w_2 = init_weights((h_size, y_size)) 
    b_2 = tf.constant(1.) 

    yhat_1 = forwardprop(X, w_1, b_1, 1)  
    yhat = forwardprop(yhat_1, w_2, b_2, 0) 


    n_samples = train_X.shape[0] 
    cost = tf.reduce_sum(tf.pow(yhat-y, 2))/(2*n_samples) 
    updates = tf.train.GradientDescentOptimizer(starter_learning_rate).minimize(cost) 


    sess = tf.Session() 
    init = tf.global_variables_initializer() 
    sess.run(init) 

    for epoch in range(itr): 
     sess.run(updates, feed_dict={X: train_X, y: train_y}) 
     train_err = train_y - sess.run(yhat, feed_dict={X: train_X, y: train_y}) 
     train_accuracy = np.mean(train_err**2) 

     test_err = test_y - sess.run(yhat, feed_dict={X: test_X, y: test_y}) 
     test_accuracy =np.mean(test_err**2) 


    print(sess.run(yhat, feed_dict={X: test_X, y: test_y})) 
    sess.close() 
if __name__ == '__main__': 
    main(itr=10000,starter_learning_rate=0.001) 

Answer 1

标准矢量的标准化是解决方案。

Answer 2

检查test_X的数据，也许所有的text_X记录都是一样的。