0
训练的时候成本不变的我是一个总的新秀,并试图使用tensorflow解决多输入多输出的问题。然而,培训的过程中,权重和网络的成本是不变的。下面是一些主要的代码,任何建议,将不胜感激!重量和tensorflow
learning_rate = 0.01
training_epoch = 2000
batch_size = 100
display_step = 1
# place holder for graph input
x = tf.placeholder("float64", [None, 14])
y = tf.placeholder("float64", [None, 8])
# model weights
w_1 = tf.Variable(tf.zeros([14, 11], dtype = tf.float64))
w_2 = tf.Variable(tf.zeros([11, 8], dtype = tf.float64))
# construct a model
h_in = tf.matmul(x, w_1)
h_out = tf.nn.relu(h_in)
o_in = tf.matmul(h_out, w_2)
o_out = tf.nn.relu(o_in)
# cost: mean square error
cost = tf.reduce_sum(tf.pow((o_out - y), 2))
# optimizer
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
# initializer
init = tf.global_variables_initializer()
# launch the graph
with tf.Session() as sess:
sess.run(init)
for epoch in range(training_epoch):
pos = 0;
# loop over all batches
if pos < train_input_array.shape[0]:
# get the next batch
batch_i = []
batch_o = []
for i in range(pos, pos + batch_size):
batch_i.append(train_input_array[i].tolist())
batch_o.append(train_output_array[i].tolist())
np.array(batch_i)
np.array(batch_o)
pos += batch_size;
sess.run(optimizer, feed_dict = {x: batch_i, y: batch_o})
print sess.run(w_2[0])
if (epoch + 1) % display_step == 0:
c = sess.run(cost, feed_dict = {x: batch_i, y: batch_o})
print("Epoch: ", "%04d" % (epoch + 1), "cost: ", "{:.9f}".format(c))
我改成了reduce_mean,但损失仍然是不变的。 – Dennis
您可以尝试使用momentm优化,数目较多的 –
谢谢你的建议的参数,我改变了权重变量初始化从tf.zero到tf.random_normal,损失终于开始下降。 – Dennis