AlexNet+cifar（tensorflow）
RT（仅作记录）
[代码来自] (https://github.com/xi-mao/alexnet-cifar-10)
Alexnet
#coding=utf-8
import math
import tensorflow as tf
def print_activations(t):
print(t.op.name,'',t.get_shape().as_list) #get_shape获取一个TensorShape对象，然后通过as_list方法返回每一个维度数
def model():
_IMAGE_SIZE=32
_IMAGE_CHANNELS=3
_RESHAPE_SIZE=3*3*128
_NUM_CLASSES=10
parameters=[]
with tf.name_scope('data'):
x=tf.placeholder(tf.float32,shape=[None,_IMAGE_SIZE*_IMAGE_SIZE*_IMAGE_CHANNELS],name='images')
y=tf.placeholder(tf.float32,shape=[None,_NUM_CLASSES],name='Output')
images=tf.reshape(x,[-1,_IMAGE_SIZE,_IMAGE_SIZE,_IMAGE_CHANNELS],name='images')
print(images)
#conv1
#这里name_scope实际上是为了解决共享变量的问题，在name_scope下进行tf.Variable(name)
#如果name重名，会自动检测命名冲突进行处理
with tf.name_scope('conv1') as scope:
kernel=tf.Variable(tf.truncated_normal([5,5,3,64],dtype=tf.float32,
stddev=1e-1),name='weights')
#变量解释 [a,b,c,d]分别表示,1表示是否跳过一些样本，比如a=1时，就是从1，2，3...训
#跳过一些，a=2时就选择1，3，5...，b表示高方向滑动，c表示宽方向滑动，d表示通道滑动
#same表示当卷积核超出边界时会进行0填充
conv=tf.nn.conv2d(images,kernel,[1,1,1,1],padding='SAME')
biases=tf.Variable(tf.constant(0.0,shape=[64],dtype=tf.float32),
trainable=True,name='bias')
bias=tf.nn.bias_add(conv,biases)
conv1=tf.nn.relu(bias,name=scope) #这里返回的是一个tensor（一个张量类），但是这里的name=scope是什么意思？
print_activations(conv1)
tf.summary.histogram('Convolution_layers/conv1',conv1)
tf.summary.scalar('Convolution_layers/conver1',tf.nn.zero_fraction(conv1))
with tf.name_scope('lrn1') as scope:
lrn1=tf.nn.local_response_normalization(conv1,
alpha=1e-4,
beta=0.75,
depth_radius=2,
bias=2.0)#lrn
#pool1
pool1=tf.nn.max_pool(lrn1,ksize=[1,3,3,1],strides=[1,2,2,1],
padding='VALID',name='pool1')#池化
print_activations(pool1)#这有个问题 照理是先池化后lrn 可是这里是先lrn后池化
#conv2同上
with tf.name_scope('conv2') as scope:
kernel = tf.Variable(tf.truncated_normal([5, 5, 64, 64], dtype=tf.float32,
stddev=1e-1), name='weights')
conv = tf.nn.conv2d(pool1, kernel, [1, 1, 1, 1], padding='SAME')
biases = tf.Variable(tf.constant(0.0, shape=[64], dtype=tf.float32),
trainable=True, name='biases')
bias = tf.nn.bias_add(conv, biases)
conv2 = tf.nn.relu(bias, name=scope)
tf.summary.histogram('Convolution_layers/conv2',conv2)
tf.summary.scalar('Convolution_layers/conver2',tf.nn.zero_fraction(conv2))
print_activations(conv2)
#lrn2
with tf.name_scope('lrn2') as scope:
lrn2 = tf.nn.local_response_normalization(conv2,alpha=1e-4,beta=0.75,
depth_radius=2, bias=2.0)
# pool2
pool2 = tf.nn.max_pool(lrn2, ksize=[1, 3, 3, 1],strides=[1, 2, 2, 1],
padding='VALID',name='pool2')
print_activations(pool2)
#conv3少了池化和lrn
with tf.name_scope('conv3') as scope:
kernel =tf.Variable(tf.truncated_normal([3,3,64,128],dtype=tf.float32,
stddev=1e-1),name='weights')
conv=tf.nn.conv2d(pool2,kernel,[1,1,1,1],padding='SAME')
biases=tf.Variable(tf.constant(0.0,shape=[128],dtype=tf.float32),
trainable=True,name='biases')
bias=tf.nn.bias_add(conv,biases)
conv3=tf.nn.relu(bias,name=scope)
print_activations(conv3)
tf.summary.histogram('Convolution_layers/conv3',conv3)
tf.summary.scalar('Convolution_layers/conver3',tf.nn.zero_fraction(conv3))
#conv4同3
with tf.name_scope('conv4') as scope:
kernel =tf.Variable(tf.truncated_normal([3,3,128,128],dtype=tf.float32,
stddev=1e-1),name='weights')
conv=tf.nn.conv2d(conv3,kernel,[1,1,1,1],padding='SAME')
biases=tf.Variable(tf.constant(0.0,shape=[128],dtype=tf.float32),
trainable=True,name='biases')
bias=tf.nn.bias_add(conv,biases)
conv4=tf.nn.relu(bias,name=scope)
print_activations(conv4)
tf.summary.histogram('Convolution_layers/conv4',conv4)
tf.summary.scalar('Convolution_layers/conver4',tf.nn.zero_fraction(conv4))
#conv5
with tf.name_scope('conv5') as scope:
kernel =tf.Variable(tf.truncated_normal([3,3,128,128],dtype=tf.float32,
stddev=1e-1),name='weights')
conv=tf.nn.conv2d(conv4,kernel,[1,1,1,1],padding='SAME')
biases=tf.Variable(tf.constant(0.0,shape=[128],dtype=tf.float32),
trainable=True,name='biases')
bias=tf.nn.bias_add(conv,biases)
conv5=tf.nn.relu(bias,name=scope)
print_activations(conv5)
tf.summary.histogram('Convolution_layers/conv5',conv5)
tf.summary.scalar('Convolution_layers/conver5',tf.nn.zero_fraction(conv5))
#pool5
pool5=tf.nn.max_pool(conv5,ksize=[1,3,3,1],strides=[1,2,2,1],
padding='VALID',name='pool5')
print_activations(pool5)
#fully_connected1
with tf.name_scope('fully_connected1') as scope:
reshape=tf.reshape(pool5,[-1,_RESHAPE_SIZE])
dim=reshape.get_shape()[1].value
weights =tf.Variable(tf.truncated_normal([dim,384],dtype=tf.float32,
stddev=1e-1),name='weights')
print_activations(weights)
biases=tf.Variable(tf.constant(0.0,shape=[384],dtype=tf.float32),
trainable=True,name='biases')
local3=tf.nn.relu(tf.matmul(reshape,weights)+biases,name=scope)
print_activations(local3)
tf.summary.histogram('Fully connected layers/fc1',local3)
tf.summary.scalar('Fully connected layers/fc1',tf.nn.zero_fraction(local3))
#fully_connected2
with tf.name_scope('fully_connected') as scope:
weights =tf.Variable(tf.truncated_normal([384,192],dtype=tf.float32,
stddev=1e-1),name='weights')
print_activations(weights)
biases=tf.Variable(tf.constant(0.0,shape=[192],dtype=tf.float32),
trainable=True,name='biases')
local4=tf.nn.relu(tf.matmul(local3,weights)+biases,name=scope)
print_activations(local4)
tf.summary.histogram('Fully connected layers/fc2',local4)
tf.summary.scalar('Fully connected layers/fc4',tf.nn.zero_fraction(local4))
#output
with tf.name_scope('output') as scope:
weights =tf.Variable(tf.truncated_normal([192,_NUM_CLASSES],dtype=tf.float32,
stddev=1e-1),name='weights')
print_activations(weights)
biases=tf.Variable(tf.constant(0.0,shape=[_NUM_CLASSES],dtype=tf.float32),
trainable=True,name='biases')
softmax_linear=tf.add(tf.matmul(local4,weights),biases,name=scope)
tf.summary.histogram('Fully connected layers/output',softmax_linear)
global_step=tf.Variable(initial_value=0,name='global_step',trainable=False)
y_pred_cls=tf.argmax(softmax_linear,axis=1)
return x,y,softmax_linear,global_step,y_pred_cls
train
#coding=utf-8
import numpy as np
import tensorflow as tf
from sklearn.metrics import confusion_matrix
from time import time
from alexnet import model
from data import get_data_set
train_x,train_y,tain_l=get_data_set("train")#train_x训练图片，train_y label tain_l表示classes
test_x,test_y,test_l=get_data_set("test")
x,y,output,global_step,y_pred_cls=model()#x表示image y表示输出 ，output表示softmax值 y_pred_cls表示分类值
_IMG_SIZE = 32
_NUM_CHANNELS = 3
_BATCH_SIZE = 128
_CLASS_SIZE = 10
_ITERATION = 30000
_SAVE_PATH = "tensorboard/cifar-10/" #先创建好这些文件
_SAVE_BOARD_PATH="tensorboard/board/"
loss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=output,labels=y))
#计算loss softmax_cross_entropy_with_logits参数（神经网络最后一层的输出，实际的标签）
optimizer=tf.train.RMSPropOptimizer(learning_rate=1e-3).minimize(loss,global_step=global_step)
#RMS算法优化器用于优化loss具体见http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf
correct_prediction=tf.equal(y_pred_cls,tf.argmax(y,axis=1))#将y_pred_cls，tf.argmax(y,axis=1)对比bool值
accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))#计算准确值 tf.cast是转换类型函数 将correct_prediction转换成tf.float32
tf.summary.scalar('loss',loss)
#用来显示标量信息，其格式为：
#tf.summary.scalar(tags, values, collections=None, name=None)
tf.summary.scalar("Accyracy/train",accuracy)
tf.summary.histogram('histogram',accuracy)
#用来显示直方图信息，其格式为：
#tf.summary.histogram(tags, values, collections=None, name=None)
saver=tf.train.Saver()
sess=tf.Session()
merged=tf.summary.merge_all()
#merge_all 可以将所有summary全部保存到磁盘，以便tensorboard显示。
train_writer=tf.summary.FileWriter(_SAVE_BOARD_PATH,sess.graph)#指定一个文件用来保存图。
sess.run(tf.global_variables_initializer())
#sess_path=saver.save(sess,_SAVE_PATH)
#try:
# print("Trying to restore last checkpoint..... ")
# last_chk_path=tf.train.latest_checkpoint(checkpoint_dir=_SAVE_PATH) #将变量保存在此路径
# saver.restore(sess,save_path=last_chk_path)
# print("Restored checkpoint from:",last_chk_path)
#except:
# print("Failed to restore checkpoint.Initializing variables instead")
# sess.run(tf.global_variables_initializer())
def train(num_iterations):
for i in range(num_iterations):
randidx=np.random.randint(len(train_x),size=_BATCH_SIZE) #此处返回的是小于len（train）的离散均匀分布，总共有128个
batch_xs=train_x[randidx]
batch_ys=train_y[randidx]
start_time=time()
i_global,_=sess.run([global_step,optimizer],feed_dict={x:batch_xs,y:batch_ys})
duration=time()-start_time
if(i_global%10==0)or(i==num_iterations-1):#每十次或者最后次时
_loss,batch_acc=sess.run([loss,accuracy],feed_dict={x:batch_xs,y:batch_ys})
msg= "Glo bal Step: {0:>6}, accuracy: {1:>6.1%}, loss = {2:.2f} ({3:.1f} examples/sec, {4:.2f} sec/batch)"
print(msg.format(i_global, batch_acc, _loss, _BATCH_SIZE / duration, duration))
resultmerged=sess.run(merged,feed_dict={x:batch_xs,y:batch_ys})
train_writer.add_summary(resultmerged,i_global)#tensorboard保存
if (i_global%100==0)or(i==num_iterations-1):#每一百次或者最后次时
acc=predict_test()
print('test accuracy is:')
print(acc)
saver.save(sess,save_path=_SAVE_PATH,global_step=global_step)
print("Saved checkpoint")
def predict_test(show_confusion_matrix=False):
i=0
predicted_class=np.zeros(shape=len(test_x),dtype=np.int)#返回一个新的数组，用零填充
print('test_x的长度：')
print(len(test_x))
while i<len(test_x):
j=min(i+_BATCH_SIZE,len(test_x))
batch_xs=test_x[i:j,:]
#batch_xs是128*3072的大小，最后一个是16*3072
batch_ys=test_y[i:j,:]
predicted_class[i:j]=sess.run(y_pred_cls,feed_dict={x:batch_xs,y:batch_ys})
i=j
correct=(np.argmax(test_y,axis=1)==predicted_class)#取test_y中每个最大值的索引号
acc=correct.mean()*100
correct_numbers=correct.sum()
print("Accuracy on Test-Set:{0:.2f}%({1}/{2})".format(acc,correct_numbers,len(test_x)))
if show_confusion_matrix is True:
cm=confusion_matrix(y_true=np.argmax(test_y,axis=1),y_pred=predicted_class)#混淆矩阵
for i in range(_CLASS_SIZE):
class_name="({}){}".format(i,test_l[i])
print (cm[i:],class_name)
class_numbers=["({0})".format(i) for i in range(_CLASS_SIZE)]
print("".join(class_numbers))
return acc
if _ITERATION!=0:
train(_ITERATION)
sess.close()
剩下两个文件不做标注了。