统计了上面LeNet5的改进形的参数量:
from functools import reduce
from operator import mul
def get_num_params():
num_params = 0
for variable in tf.trainable_variables():
shape = variable.get_shape()
num_params += reduce(mul, [dim.value for dim in shape], 1)
return num_params
print(get_num_params())
结果:
(3, 3, 1, 32) 288
(32,) 32
(3, 3, 32, 64) 18432
(64,) 64
(3136, 1024) 3211264
(1024,) 1024
(1024, 10) 10240
(10,) 10
3241354
可见,参数最多的在全连接层,当然运算量最大在前面卷积层。
下面我把卷积修改为3*1和1*3组合,第二个卷积面变为32全连接层变512节点,结果如下:
(3, 1, 1, 32) 96
(32,) 32
(1, 3, 32, 1) 96
(32,) 32
(3, 1, 32, 64) 6144
(64,) 64
(1, 3, 64, 32) 6144
(32,) 32
(1568, 512) 802816
(512,) 512
(512, 10) 5120
(10,) 10
821098
训练了100个epoch,准确率:0.9907
这样参数降低了75%左右,运算量也降低了75%左右,性能略微下跌。very good!
2018-3-16,这里感觉还是有点问题,前面卷积运算和n个平面同时卷积,应该运算量变大了,得找个只做一个平面卷积的函数。后面找到了回来修改吧。
完整代码:
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
#mnist已经作为官方的例子,做好了数据下载,分割,转浮点等一系列工作,源码在tensorflow源码中都可以找到
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
# 配置每个 GPU 上占用的内存的比例
# 没有GPU直接sess = tf.Session()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
#每个批次的大小
batch_size = 100
#定义训练轮数据
train_epoch = 100
#定义每n轮输出一次
test_epoch_n = 1
#计算一共有多少批次
n_batch = mnist.train.num_examples // batch_size
print("batch_size="+str(batch_size)+"n_batch="+str(n_batch))
#CNN有很多层,直接写了一样的初始化函数
#初始化权值
def weight_variable(shape):
initial = tf.truncated_normal(shape,stddev=0.1)#生成一个截断的正态分布
return tf.Variable(initial)
#初始化偏置
def bias_variable(shape):
initial = tf.constant(0.1,shape=shape)
return tf.Variable(initial)
#卷积层
def conv2d(x,W):
#x input tensor of shape `[batch, in_height, in_width, in_channels]`
#W filter / kernel tensor of shape [filter_height, filter_width, in_channels, out_channels]
#`strides[0] = strides[3] = 1`. strides[1]代表x方向的步长,strides[2]代表y方向的步长
#padding: A `string` from: `"SAME", "VALID"`
#x是图像矩阵,W是卷积和,strides表示步长,padding表示填充方式
#print('x=',x,'W=',W)
return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding='SAME')
#池化层,最大池化
def max_pool_2x2(x):
#ksize [1,x,y,1]
return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')
#定义两个placeholder
x = tf.placeholder(tf.float32,[None,784])
y = tf.placeholder(tf.float32,[None,10])
#改变x的格式转为4D的向量[batch, in_height, in_width, in_channels]
#batch=-1表示为不限长度
x_image = tf.reshape(x,[-1,28,28,1])
#初始化第一个卷积层的权值和偏置
W_conv1_A = weight_variable([3,1,1,32])#3*1采样窗口,32个卷积核从1个平面抽取特征
b_conv1_A = bias_variable([32])#每一个卷积核一个偏置值
h_conv1_A = tf.nn.relu( conv2d(x_image,W_conv1_A) + b_conv1_A )
W_conv1_B = weight_variable([1,3,32,1])#3*1采样窗口,32个卷积核从1个平面抽取特征
b_conv1_B = bias_variable([32])#每一个卷积核一个偏置值
h_conv1_B = tf.nn.relu( conv2d(h_conv1_A,W_conv1_B) + b_conv1_B )
h_pool1 = max_pool_2x2(h_conv1_B)#进行max-pooling
#初始化第二个卷积层的权值和偏置
W_conv2_A = weight_variable([3,1,32,64])#3*1采样窗口,64个卷积核从32个平面抽取特征
b_conv2_A = bias_variable([64])#每一个卷积核一个偏置值
h_conv2_A = tf.nn.relu( conv2d(h_pool1,W_conv2_A) + b_conv2_A )
W_conv2_B = weight_variable([1,3,64,32])#3*1采样窗口,64个卷积核从32个平面抽取特征
b_conv2_B = bias_variable([32])#每一个卷积核一个偏置值
h_conv2_B = tf.nn.relu( conv2d(h_conv2_A,W_conv2_B) + b_conv2_B )
h_pool2 = max_pool_2x2(h_conv2_B)#进行max-pooling
#28*28的图片第一次卷积后还是28*28,第一次池化后变为14*14
#第二次卷积后为14*14,第二次池化后变为了7*7
#进过上面操作后得到64张7*7的平面
#初始化第一个全连接层的权值
W_fc1 = weight_variable([7*7*32,512])#上一层有7*7*64个神经元,全连接层有1024个神经元
b_fc1 = bias_variable([512])#1024个节点
#把池化层2的输出扁平化为1维
h_pool2_flat = tf.reshape(h_pool2,[-1,7*7*32])
#求第一个全连接层的输出
h_fc1 = tf.nn.relu( tf.matmul(h_pool2_flat,W_fc1) + b_fc1 )
#用keep_prob用来表示神经元输出的概率
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1,keep_prob)
#初始化第二个全连接层
W_fc2 = weight_variable([512,10])
b_fc2 = bias_variable([10])
#计算输出
prediction = tf.nn.softmax(tf.matmul(h_fc1_drop,W_fc2) + b_fc2)
#交叉熵代价函数
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y,logits=prediction))
#使用AdamOptimizer进行优化
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
#求准确率
correct_prediction = tf.equal(tf.argmax(prediction,1),tf.argmax(y,1))#argmax返回一维张量中最大的值所在的位置
accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
init = tf.global_variables_initializer()
sess.run(init)
MaxACC = 0#最好的ACC
saver = tf.train.Saver()
#训练n个epoch
for epoch in range(train_epoch):
for batch in range(n_batch):
batch_xs,batch_ys = mnist.train.next_batch(batch_size)
sess.run(train_step,feed_dict={x:batch_xs,y:batch_ys,keep_prob:0.5})
if(0==(epoch%test_epoch_n)):#每若干次预测test一次
test_acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels,keep_prob:1.0})
train_acc = sess.run(accuracy,feed_dict={x:mnist.train.images[:10000],y:mnist.train.labels[:10000],keep_prob:1.0})
print('epoch=',epoch,'ACC=',test_acc,'train acc =',train_acc)
if(test_acc>MaxACC):
MaxACC = test_acc
saver.save(sess, "Model/ModelLeNet.ckpt")
print('Save model! Now ACC=',MaxACC)
#关闭sess
sess.close()
#读取模型
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
saver.restore(sess, "./Model/ModelLeNet.ckpt") # 注意此处路径前添加"./"
print('Load Model OK!')
print('ACC=',sess.run(accuracy, feed_dict={x:mnist.test.images, y: mnist.test.labels,keep_prob:1.0}))
from functools import reduce
from operator import mul
def get_num_params():
num_params = 0
for variable in tf.trainable_variables():
shape = variable.get_shape()
print(shape,reduce(mul, [dim.value for dim in shape], 1))
num_params += reduce(mul, [dim.value for dim in shape], 1)
return num_params
print(get_num_params())
同样的道理,我还可以模仿inception,或者resnet中的各种网络设计。
例如加一个1*1的卷积并联:
W_conv2_C = weight_variable([1,1,16,32])#1*1采样窗口,32个卷积核从32个平面抽取特征
b_conv2_C = bias_variable([32])#每一个卷积核一个偏置值
h_conv2_C = tf.nn.relu( conv2d(h_conv2_A,W_conv2_C) + b_conv2_C )
h_pool2_C = max_pool_2x2(h_conv2_C)#进行max-pooling
h_pool2 = h_pool2+h_pool2_C#2个卷积相加
总之,TF在手,天下我有。