tensorflow搭建多层感知器网络(MLP)
最近在学tensorflow，在调试的过程中无意实现了一个小的网络搭建，感觉代码还比较清晰，就记录一下，希望能帮助初学的同学。
使用的是python 3.0， tf 1.0
下面的代码搭建一个多层感知器网络，示例学的是简单的函数如 f=2*x+3y，f=x^2+3*y，然后还包括loss可视化，最终会保存在文件夹visual_dir中
要想查看可视化的结果
tensorboard --logdir=your_path/visual_dir
注意如果再windows 下用tensorboard，需要先定位到盘符比如你保存的目录在e盘
e:
tensorboard --logdir=your_path/visual_dir
另外网络有一些没有用的变量，希望大家进行辨别。
# -*- coding:utf-8 -*
import os ; import sys
os.chdir( os.path.split( os.path.realpath( sys.argv[0] ) )[0] )
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
import tensorflow as tf
from tensorflow import flags
import numpy as np
import numpy
from numpy import random as nr
true=True
false=False
FLAGS=flags.FLAGS
FLAGS.visual_trian=true
FLAGS.visual_savedir="visual_dir/"
input_szie=2
output_szie=1
# 每一层节点个个数
node_num=[200,200,200,200,output_szie]
layer_num=5
active_types=[0,0,1,1,1]
active_type=1
active_func=["sigmoid","liner"]
def sigmoid(x):
out=tf.sigmoid(x)
return out
def connected(x,w,b,scope):
out=tf.nn.xw_plus_b(x,w,b,name=scope)
if active_func[active_type]=="sigmoid":
out=sigmoid(out)
return out
def get_w_b(indim,outdim,name):
args = [0., 1e-2, [indim,outdim]]
w_init = np.random.normal(*args)
args = [0., 1e-2, [outdim]]
b_init = np.random.normal(*args)
w_init=tf.constant_initializer(w_init)
b_init=tf.constant_initializer(b_init)
w=tf.get_variable(name+"_w",
shape=[indim,outdim],
dtype=tf.float32,
initializer=w_init)
b=tf.get_variable(name+"_b",
shape=[outdim],
dtype=tf.float32,
initializer=b_init)
return [w,b]
def compute_loss(truth,net_out):
diff = truth - net_out
loss = tf.nn.l2_loss(diff)
return loss
def get_input_target():
# 目前我这里的函数是 f=x^2+3*y x，y要 归一化到 0-1 才行
# 否则很难训的
max_batches = 50000
batch_num = 64
batch_cnt = 0
seen = 0
hd_data=[None]*2
while batch_cnt < max_batches:
hd_data[0]=[]
hd_data[1]=[]
for i in range(batch_num):
x, y = nr.rand(2) #* 1000
f = x * x + 3 * y
#x += nr.randn()
#y += nr.randn()
hd_data[0] += [ [x,y] ]
hd_data[1] += [ [f] ]
fimg_batch = numpy.array(hd_data[0])
labels_batch = numpy.array(hd_data[1])
datum = dict({"input_data": fimg_batch, "truth_target": labels_batch})
batch_cnt += 1
seen += batch_num
yield [seen, datum]
input_data=tf.placeholder(tf.float32, [None,input_szie], 'input_data')
truth_target=tf.placeholder(tf.float32, [None,output_szie], 'truth_target')
# 开始构建网络
x=input_data
pre_dim=input_szie
for i in range(layer_num):
active_type=active_types[i]
now_dim=node_num[i]
w,b=get_w_b(pre_dim,now_dim,str(i+1))
x=connected(x,w,b,"conn-"+str(i+1))
pre_dim=now_dim
net_out=x
loss_op=compute_loss(truth_target,net_out)
optimizer=tf.train.AdamOptimizer(0.001)
#optimizer=tf.train.GradientDescentOptimizer(0.001)
gradient =optimizer.compute_gradients(loss_op)
train_op = optimizer.apply_gradients(gradient)
sess=tf.Session()
init_op = tf.global_variables_initializer()
init_local_op = tf.local_variables_initializer()
sess.run(init_op)
sess.run(init_local_op)
loss_mva = None
placeholders=dict()
placeholders["input_data"] = input_data
placeholders["truth_target"] = truth_target
''' 可视化有关
'''
if FLAGS.visual_trian:
summary_writer = tf.summary.FileWriter( \
FLAGS.visual_savedir, tf.get_default_graph())
tf.summary.scalar('sse_l2', loss_op)
merged = tf.summary.merge_all()
batches = get_input_target()
for i, (seen,datum) in enumerate(batches):
# 各种 holder的填入
feed_dict = {
placeholders[key]: datum[key] for key in datum }
fetches = [train_op, loss_op ]
# 加上可视信息
if FLAGS.visual_trian == True:
merged_id=len(fetches)
fetches+=[merged]
fetched = sess.run(fetches, feed_dict)
if FLAGS.visual_trian == True:
summary_writer.add_summary(fetched[merged_id], i)
loss = fetched[1]
grad = sess.run(gradient, feed_dict)
pre_grad=numpy.max([numpy.max(x) for x in grad])
print("max gradient: ", pre_grad)
if loss_mva is None: loss_mva = loss
loss_mva = .9 * loss_mva + .1 * loss
form = 'step {} - seen {} - loss {} - loss_mva {}'
print(form.format(i, seen, loss, loss_mva))