Python tf神经网络二分类，最后结果都是一样的

神经网络input是140个样本 1296维度的特征值，label前60是[1.0, 0.0] 后面是[0.0, 1.0]，想来实现二分类，可是我用softmax交叉熵+L2正则化，训练出来的output全是一个相同的1*2矩阵。

有没有大神可以看出来错在哪里  T_T

    xs = tf.placeholder(tf.float32,shape=(None, 36*36))
    ys = tf.placeholder(tf.float32,shape=(None,2))

    layer1_node = 150 # 第一层隐含层神经元个数

    # first layer inference
    L1_Weights = tf.Variable(tf.random_normal([36*36,layer1_node], stddev=0.1, seed=1))
    L1_biases = tf.Variable(tf.zeros([1,layer1_node]) + 0.1)
    L1_out = tf.matmul(xs, L1_Weights) + L1_biases

    L1_out = tf.nn.relu(L1_out)

    # second layer inference
    L2_Weights = tf.Variable(tf.random_normal([layer1_node,2], stddev=0.1, seed=1))
    L2_biases = tf.Variable(tf.zeros([1,2]) + 0.1)
    out = tf.matmul(L1_out, L2_Weights) + L2_biases

    # cross entropy
    out_pro = tf.nn.log_softmax(out)
    cross_entropy = -tf.reduce_sum(ys*out_pro)
    cross_entropy_mean = tf.reduce_mean(cross_entropy)

    # regularization
    regularizer = tf.contrib.layers.l2_regularizer(0.1)
    regularization = regularizer(L1_Weights) + regularizer(L2_Weights)

    # loss
    loss = cross_entropy_mean + regularization

    # backward
    train_step = tf.train.GradientDescentOptimizer(0.01).minimize(loss)  # 基础梯度下降法
    init = tf.global_variables_initializer()

    with tf.Session() as sess:
        sess.run(init)
        for i in range(20000):
            sess.run(train_step,feed_dict={xs:recon_imgs,ys:label})
            if i % 50 == 0:
                print(sess.run(loss,feed_dict={xs:recon_imgs,ys:label}))
               
        print(sess.run(out,feed_dict={xs:recon_imgs,ys:label}))

output：[[-0.09252657 0.29252994]，…… 全是一样的。。

梯度下降试着用过动量法，但是不收敛。。