# Inputs and Placeholders
x = tf.placeholder(tf.float32, shape=[None, 30])
y_ = tf.placeholder(tf.float32)
# Inference
W_1 = tf.Variable(tf.zeros([30,50]))
b_1 = tf.Variable(tf.zeros([50]))
layer_1 = tf.sigmoid(tf.add(tf.matmul(x, W_1), b_1))
W_2 = tf.Variable(tf.zeros([50,25]))
b_2 = tf.Variable(tf.zeros([25]))
layer_1_value = tf.add(tf.matmul(layer_1, W_2), b_2)
layer_2 = tf.nn.softmax(layer_1_value)
# W_3 is a fixed weight matrix.
w_linear = np.array([item + 5 for item in xrange(50, 300, 10)])
W_3 = tf.Variable(w_linear, trainable=False)
y = tf.reduce_sum(tf.multiply(tf.to_float(W_3), layer_2))
# Loss
mean_square = tf.losses.mean_squared_error(y_, y)
loss = tf.reduce_mean(mean_square, name='square_mean')
# Training
tf.summary.scalar('loss', loss)
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
global_step = tf.Variable(0, name='global_step', trainable=False)
train_op = optimizer.minimize(loss, global_step=global_step)
在这个网络中，我试图得到layer2（softmax层）输出的加权平均值之和。最后的线性层不参与培训。
有谁知道为什么损失在第一个时代之后没有改变？
('Epoch:0001', 'cost=2499180.068965517')
('Epoch:0002', 'cost=2335760.384482760')
('Epoch:0003', 'cost=2335760.384482760')
('Epoch:0004', 'cost=2335760.384482760')
('Epoch:0005', 'cost=2335760.384482760')
('Epoch:0006', 'cost=2335760.384482760')
('Epoch:0007', 'cost=2335760.384482760')
('Epoch:0008', 'cost=2335760.384482760')
('Epoch:0009', 'cost=2335760.384482760')