知识总结
（1）再次注意summary的使用
（2）x = rdm.rand(dataset_size, 2) y_ = [[x1**2 + x2**2] for (x1, x2) in x]这里的问题要注意
（3）注意batch时，全部先按照一套W进行前向传播，这时候在进行正则化时，加的是同一套W，然后反向传播改变W值，进行下一轮前向传播
代码如下
import tensorflow as tf
import numpy as np
from numpy.random import RandomState
rdm = RandomState(1)
dataset_size = 128
x = rdm.rand(dataset_size, 2)
y_ = [[x1**2 + x2**2] for (x1, x2) in x]
def get_weight(shape, alpha, name):
with tf.variable_scope("get_variable" + name):
var = tf.get_variable(name, shape, tf.float32, initializer=tf.truncated_normal_initializer(0.01))
tf.add_to_collection("losses", tf.contrib.layers.l2_regularizer(alpha)(var))
return var
with tf.name_scope("generate_value"):
xs = tf.placeholder(tf.float32, [None, 2], name="x_input")
ys = tf.placeholder(tf.float32, [None, 1], name="y_output")
batch_size = 8
layers_dimension = [2 ,10, 10, 10 ,1]
n_layers = len(layers_dimension)
in_dimension = layers_dimension[0]
cur_layer = xs
for i in range(1, n_layers):
out_dimension = layers_dimension[i]
with tf.variable_scope("layer%d" % i):
weights = get_weight([in_dimension, out_dimension], 0.001, "layers")
biases = tf.get_variable("biases", [out_dimension], tf.float32, tf.constant_initializer(0.0))
cur_layer = tf.matmul(cur_layer, weights) + biases
cur_layer = tf.nn.relu(cur_layer)
in_dimension = layers_dimension[i]
with tf.name_scope("loss_op"):
mse_loss = tf.reduce_mean(tf.square(ys - cur_layer))
tf.add_to_collection("losses", mse_loss)
loss = tf.add_n(tf.get_collection("losses"))
tf.summary.scalar("loss", loss)
train_op = tf.train.GradientDescentOptimizer(0.01).minimize(loss)
merged = tf.summary.merge_all()
init = tf.global_variables_initializer()
with tf.Session() as sess:
writer = tf.summary.FileWriter("path/", tf.get_default_graph())
sess.run(init)
for i in range(5000):
start = i*batch_size % dataset_size
end = min(start+batch_size, dataset_size)
if i % 50 == 0:
result = sess.run(merged, feed_dict={xs: x, ys: y_})
writer.add_summary(result, global_step=i)
if i % 500 ==0:
loss_op = sess.run(loss, feed_dict={xs: x, ys: y_})
print("After %d training, loss is %g" % (i, loss_op))
_ = sess.run(train_op, feed_dict={xs: x[start:end], ys: y_[start:end]})
writer.close()
