当前位置:   article > 正文

深度学习(五十六)tensorflow项目构建流程_tensorflow各种框架搭建

tensorflow各种框架搭建

tensorflow项目构建流程

博客http://blog.csdn.net/hjimce

微博黄锦池-hjimce   qq:1393852684

一、构建路线

个人感觉对于任何一个深度学习库,如mxnet、tensorflow、theano、caffe等,基本上我都采用同样的一个学习流程,大体流程如下:

(1)训练阶段:数据打包-》网络构建、训练-》模型保存-》可视化查看损失函数、验证精度

(2)测试阶段:模型加载-》测试图片读取-》预测显示结果

(3)移植阶段:量化、压缩加速-》微调-》C++移植打包-》上线

这边我就以tensorflow为例子,讲解整个流程的大体架构,完成一个深度学习项目所需要熟悉的过程代码。

二、训练、测试阶段

1、tensorflow打包数据

这一步对于tensorflow来说,也可以直接自己在线读取:.jpg图片、标签文件等,然后通过phaceholder变量,把数据送入网络中,进行计算。

不过这种效率比较低,对于大规模训练数据来说,我们需要一个比较高效的方式,tensorflow建议我们采用tfrecoder进行高效数据读取。学习tensorflow一定要学会tfrecoder文件写入、读取,具体示例代码如下:

  1. #coding=utf-8
  2. #tensorflow高效数据读取训练
  3. import tensorflow as tf
  4. import cv2
  5. #把train.txt文件格式,每一行:图片路径名 类别标签
  6. #奖数据打包,转换成tfrecords格式,以便后续高效读取
  7. def encode_to_tfrecords(lable_file,data_root,new_name='data.tfrecords',resize=None):
  8. writer=tf.python_io.TFRecordWriter(data_root+'/'+new_name)
  9. num_example=0
  10. with open(lable_file,'r') as f:
  11. for l in f.readlines():
  12. l=l.split()
  13. image=cv2.imread(data_root+"/"+l[0])
  14. if resize is not None:
  15. image=cv2.resize(image,resize)#为了
  16. height,width,nchannel=image.shape
  17. label=int(l[1])
  18. example=tf.train.Example(features=tf.train.Features(feature={
  19. 'height':tf.train.Feature(int64_list=tf.train.Int64List(value=[height])),
  20. 'width':tf.train.Feature(int64_list=tf.train.Int64List(value=[width])),
  21. 'nchannel':tf.train.Feature(int64_list=tf.train.Int64List(value=[nchannel])),
  22. 'image':tf.train.Feature(bytes_list=tf.train.BytesList(value=[image.tobytes()])),
  23. 'label':tf.train.Feature(int64_list=tf.train.Int64List(value=[label]))
  24. }))
  25. serialized=example.SerializeToString()
  26. writer.write(serialized)
  27. num_example+=1
  28. print lable_file,"样本数据量:",num_example
  29. writer.close()
  30. #读取tfrecords文件
  31. def decode_from_tfrecords(filename,num_epoch=None):
  32. filename_queue=tf.train.string_input_producer([filename],num_epochs=num_epoch)#因为有的训练数据过于庞大,被分成了很多个文件,所以第一个参数就是文件列表名参数
  33. reader=tf.TFRecordReader()
  34. _,serialized=reader.read(filename_queue)
  35. example=tf.parse_single_example(serialized,features={
  36. 'height':tf.FixedLenFeature([],tf.int64),
  37. 'width':tf.FixedLenFeature([],tf.int64),
  38. 'nchannel':tf.FixedLenFeature([],tf.int64),
  39. 'image':tf.FixedLenFeature([],tf.string),
  40. 'label':tf.FixedLenFeature([],tf.int64)
  41. })
  42. label=tf.cast(example['label'], tf.int32)
  43. image=tf.decode_raw(example['image'],tf.uint8)
  44. image=tf.reshape(image,tf.pack([
  45. tf.cast(example['height'], tf.int32),
  46. tf.cast(example['width'], tf.int32),
  47. tf.cast(example['nchannel'], tf.int32)]))
  48. #label=example['label']
  49. return image,label
  50. #根据队列流数据格式,解压出一张图片后,输入一张图片,对其做预处理、及样本随机扩充
  51. def get_batch(image, label, batch_size,crop_size):
  52. #数据扩充变换
  53. distorted_image = tf.random_crop(image, [crop_size, crop_size, 3])#随机裁剪
  54. distorted_image = tf.image.random_flip_up_down(distorted_image)#上下随机翻转
  55. #distorted_image = tf.image.random_brightness(distorted_image,max_delta=63)#亮度变化
  56. #distorted_image = tf.image.random_contrast(distorted_image,lower=0.2, upper=1.8)#对比度变化
  57. #生成batch
  58. #shuffle_batch的参数:capacity用于定义shuttle的范围,如果是对整个训练数据集,获取batch,那么capacity就应该够大
  59. #保证数据打的足够乱
  60. images, label_batch = tf.train.shuffle_batch([distorted_image, label],batch_size=batch_size,
  61. num_threads=16,capacity=50000,min_after_dequeue=10000)
  62. #images, label_batch=tf.train.batch([distorted_image, label],batch_size=batch_size)
  63. # 调试显示
  64. #tf.image_summary('images', images)
  65. return images, tf.reshape(label_batch, [batch_size])
  66. #这个是用于测试阶段,使用的get_batch函数
  67. def get_test_batch(image, label, batch_size,crop_size):
  68. #数据扩充变换
  69. distorted_image=tf.image.central_crop(image,39./45.)
  70. distorted_image = tf.random_crop(distorted_image, [crop_size, crop_size, 3])#随机裁剪
  71. images, label_batch=tf.train.batch([distorted_image, label],batch_size=batch_size)
  72. return images, tf.reshape(label_batch, [batch_size])
  73. #测试上面的压缩、解压代码
  74. def test():
  75. encode_to_tfrecords("data/train.txt","data",(100,100))
  76. image,label=decode_from_tfrecords('data/data.tfrecords')
  77. batch_image,batch_label=get_batch(image,label,3)#batch 生成测试
  78. init=tf.initialize_all_variables()
  79. with tf.Session() as session:
  80. session.run(init)
  81. coord = tf.train.Coordinator()
  82. threads = tf.train.start_queue_runners(coord=coord)
  83. for l in range(100000):#每run一次,就会指向下一个样本,一直循环
  84. #image_np,label_np=session.run([image,label])#每调用run一次,那么
  85. '''cv2.imshow("temp",image_np)
  86. cv2.waitKey()'''
  87. #print label_np
  88. #print image_np.shape
  89. batch_image_np,batch_label_np=session.run([batch_image,batch_label])
  90. print batch_image_np.shape
  91. print batch_label_np.shape
  92. coord.request_stop()#queue需要关闭,否则报错
  93. coord.join(threads)
  94. #test()

2、网络架构与训练

经过上面的数据格式处理,接着我们只要写一写网络结构、网络优化方法,把数据搞进网络中就可以了,具体示例代码如下:

  1. #coding=utf-8
  2. import tensorflow as tf
  3. from data_encoder_decoeder import encode_to_tfrecords,decode_from_tfrecords,get_batch,get_test_batch
  4. import cv2
  5. import os
  6. class network(object):
  7. def __init__(self):
  8. with tf.variable_scope("weights"):
  9. self.weights={
  10. #39*39*3->36*36*20->18*18*20
  11. 'conv1':tf.get_variable('conv1',[4,4,3,20],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  12. #18*18*20->16*16*40->8*8*40
  13. 'conv2':tf.get_variable('conv2',[3,3,20,40],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  14. #8*8*40->6*6*60->3*3*60
  15. 'conv3':tf.get_variable('conv3',[3,3,40,60],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  16. #3*3*60->120
  17. 'fc1':tf.get_variable('fc1',[3*3*60,120],initializer=tf.contrib.layers.xavier_initializer()),
  18. #120->6
  19. 'fc2':tf.get_variable('fc2',[120,6],initializer=tf.contrib.layers.xavier_initializer()),
  20. }
  21. with tf.variable_scope("biases"):
  22. self.biases={
  23. 'conv1':tf.get_variable('conv1',[20,],initializer=tf.constant_initializer(value=0.0, dtype=tf.float32)),
  24. 'conv2':tf.get_variable('conv2',[40,],initializer=tf.constant_initializer(value=0.0, dtype=tf.float32)),
  25. 'conv3':tf.get_variable('conv3',[60,],initializer=tf.constant_initializer(value=0.0, dtype=tf.float32)),
  26. 'fc1':tf.get_variable('fc1',[120,],initializer=tf.constant_initializer(value=0.0, dtype=tf.float32)),
  27. 'fc2':tf.get_variable('fc2',[6,],initializer=tf.constant_initializer(value=0.0, dtype=tf.float32))
  28. }
  29. def inference(self,images):
  30. # 向量转为矩阵
  31. images = tf.reshape(images, shape=[-1, 39,39, 3])# [batch, in_height, in_width, in_channels]
  32. images=(tf.cast(images,tf.float32)/255.-0.5)*2#归一化处理
  33. #第一层
  34. conv1=tf.nn.bias_add(tf.nn.conv2d(images, self.weights['conv1'], strides=[1, 1, 1, 1], padding='VALID'),
  35. self.biases['conv1'])
  36. relu1= tf.nn.relu(conv1)
  37. pool1=tf.nn.max_pool(relu1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
  38. #第二层
  39. conv2=tf.nn.bias_add(tf.nn.conv2d(pool1, self.weights['conv2'], strides=[1, 1, 1, 1], padding='VALID'),
  40. self.biases['conv2'])
  41. relu2= tf.nn.relu(conv2)
  42. pool2=tf.nn.max_pool(relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
  43. # 第三层
  44. conv3=tf.nn.bias_add(tf.nn.conv2d(pool2, self.weights['conv3'], strides=[1, 1, 1, 1], padding='VALID'),
  45. self.biases['conv3'])
  46. relu3= tf.nn.relu(conv3)
  47. pool3=tf.nn.max_pool(relu3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
  48. # 全连接层1,先把特征图转为向量
  49. flatten = tf.reshape(pool3, [-1, self.weights['fc1'].get_shape().as_list()[0]])
  50. drop1=tf.nn.dropout(flatten,0.5)
  51. fc1=tf.matmul(drop1, self.weights['fc1'])+self.biases['fc1']
  52. fc_relu1=tf.nn.relu(fc1)
  53. fc2=tf.matmul(fc_relu1, self.weights['fc2'])+self.biases['fc2']
  54. return fc2
  55. def inference_test(self,images):
  56. # 向量转为矩阵
  57. images = tf.reshape(images, shape=[-1, 39,39, 3])# [batch, in_height, in_width, in_channels]
  58. images=(tf.cast(images,tf.float32)/255.-0.5)*2#归一化处理
  59. #第一层
  60. conv1=tf.nn.bias_add(tf.nn.conv2d(images, self.weights['conv1'], strides=[1, 1, 1, 1], padding='VALID'),
  61. self.biases['conv1'])
  62. relu1= tf.nn.relu(conv1)
  63. pool1=tf.nn.max_pool(relu1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
  64. #第二层
  65. conv2=tf.nn.bias_add(tf.nn.conv2d(pool1, self.weights['conv2'], strides=[1, 1, 1, 1], padding='VALID'),
  66. self.biases['conv2'])
  67. relu2= tf.nn.relu(conv2)
  68. pool2=tf.nn.max_pool(relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
  69. # 第三层
  70. conv3=tf.nn.bias_add(tf.nn.conv2d(pool2, self.weights['conv3'], strides=[1, 1, 1, 1], padding='VALID'),
  71. self.biases['conv3'])
  72. relu3= tf.nn.relu(conv3)
  73. pool3=tf.nn.max_pool(relu3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
  74. # 全连接层1,先把特征图转为向量
  75. flatten = tf.reshape(pool3, [-1, self.weights['fc1'].get_shape().as_list()[0]])
  76. fc1=tf.matmul(flatten, self.weights['fc1'])+self.biases['fc1']
  77. fc_relu1=tf.nn.relu(fc1)
  78. fc2=tf.matmul(fc_relu1, self.weights['fc2'])+self.biases['fc2']
  79. return fc2
  80. #计算softmax交叉熵损失函数
  81. def sorfmax_loss(self,predicts,labels):
  82. predicts=tf.nn.softmax(predicts)
  83. labels=tf.one_hot(labels,self.weights['fc2'].get_shape().as_list()[1])
  84. loss =-tf.reduce_mean(labels * tf.log(predicts))# tf.nn.softmax_cross_entropy_with_logits(predicts, labels)
  85. self.cost= loss
  86. return self.cost
  87. #梯度下降
  88. def optimer(self,loss,lr=0.001):
  89. train_optimizer = tf.train.GradientDescentOptimizer(lr).minimize(loss)
  90. return train_optimizer
  91. def train():
  92. encode_to_tfrecords("data/train.txt","data",'train.tfrecords',(45,45))
  93. image,label=decode_from_tfrecords('data/train.tfrecords')
  94. batch_image,batch_label=get_batch(image,label,batch_size=50,crop_size=39)#batch 生成测试
  95. #网络链接,训练所用
  96. net=network()
  97. inf=net.inference(batch_image)
  98. loss=net.sorfmax_loss(inf,batch_label)
  99. opti=net.optimer(loss)
  100. #验证集所用
  101. encode_to_tfrecords("data/val.txt","data",'val.tfrecords',(45,45))
  102. test_image,test_label=decode_from_tfrecords('data/val.tfrecords',num_epoch=None)
  103. test_images,test_labels=get_test_batch(test_image,test_label,batch_size=120,crop_size=39)#batch 生成测试
  104. test_inf=net.inference_test(test_images)
  105. correct_prediction = tf.equal(tf.cast(tf.argmax(test_inf,1),tf.int32), test_labels)
  106. accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
  107. init=tf.initialize_all_variables()
  108. with tf.Session() as session:
  109. session.run(init)
  110. coord = tf.train.Coordinator()
  111. threads = tf.train.start_queue_runners(coord=coord)
  112. max_iter=100000
  113. iter=0
  114. if os.path.exists(os.path.join("model",'model.ckpt')) is True:
  115. tf.train.Saver(max_to_keep=None).restore(session, os.path.join("model",'model.ckpt'))
  116. while iter<max_iter:
  117. loss_np,_,label_np,image_np,inf_np=session.run([loss,opti,batch_label,batch_image,inf])
  118. #print image_np.shape
  119. #cv2.imshow(str(label_np[0]),image_np[0])
  120. #print label_np[0]
  121. #cv2.waitKey()
  122. #print label_np
  123. if iter%50==0:
  124. print 'trainloss:',loss_np
  125. if iter%500==0:
  126. accuracy_np=session.run([accuracy])
  127. print '***************test accruacy:',accuracy_np,'*******************'
  128. tf.train.Saver(max_to_keep=None).save(session, os.path.join('model','model.ckpt'))
  129. iter+=1
  130. coord.request_stop()#queue需要关闭,否则报错
  131. coord.join(threads)
  132. train()

3、可视化显示

(1)首先再源码中加入需要跟踪的变量:

tf.scalar_summary("cost_function", loss)#损失函数值
(2)然后定义执行操作:

merged_summary_op = tf.merge_all_summaries()
(3)再session中定义保存路径:
summary_writer = tf.train.SummaryWriter('log', session.graph)

(4)然后再session执行的时候,保存:

  1. summary_str,loss_np,_=session.run([merged_summary_op,loss,opti])
  2. summary_writer.add_summary(summary_str, iter)

(5)最后只要训练完毕后,直接再终端输入命令:

python /usr/local/lib/python2.7/dist-packages/tensorflow/tensorboard/tensorboard.py --logdir=log

然后打开浏览器网址:
http://0.0.0.0:6006

即可观训练曲线。

4、测试阶段

测试阶段主要是直接通过加载图模型、读取参数等,然后直接通过tensorflow的相关函数,进行调用,而不需要网络架构相关的代码;通过内存feed_dict的方式,对相关的输入节点赋予相关的数据,进行前向传导,并获取相关的节点数值。

  1. #coding=utf-8
  2. import tensorflow as tf
  3. import os
  4. import cv2
  5. def load_model(session,netmodel_path,param_path):
  6. new_saver = tf.train.import_meta_graph(netmodel_path)
  7. new_saver.restore(session, param_path)
  8. x= tf.get_collection('test_images')[0]#在训练阶段需要调用tf.add_to_collection('test_images',test_images),保存之
  9. y = tf.get_collection("test_inf")[0]
  10. batch_size = tf.get_collection("batch_size")[0]
  11. return x,y,batch_size
  12. def load_images(data_root):
  13. filename_queue = tf.train.string_input_producer(data_root)
  14. image_reader = tf.WholeFileReader()
  15. key,image_file = image_reader.read(filename_queue)
  16. image = tf.image.decode_jpeg(image_file)
  17. return image, key
  18. def test(data_root="data/race/cropbrown"):
  19. image_filenames=os.listdir(data_root)
  20. image_filenames=[(data_root+'/'+i) for i in image_filenames]
  21. #print cv2.imread(image_filenames[0]).shape
  22. #image,key=load_images(image_filenames)
  23. race_listsrc=['black','brown','white','yellow']
  24. with tf.Session() as session:
  25. coord = tf.train.Coordinator()
  26. threads = tf.train.start_queue_runners(coord=coord)
  27. x,y,batch_size=load_model(session,os.path.join("model",'model_ori_race.ckpt.meta'),
  28. os.path.join("model",'model_ori_race.ckpt'))
  29. predict_label=tf.cast(tf.argmax(y,1),tf.int32)
  30. print x.get_shape()
  31. for imgf in image_filenames:
  32. image=cv2.imread(imgf)
  33. image=cv2.resize(image,(76,76)).reshape((1,76,76,3))
  34. print "cv shape:",image.shape
  35. #cv2.imshow("t",image_np[:,:,::-1])
  36. y_np=session.run(predict_label,feed_dict = {x:image, batch_size:1})
  37. print race_listsrc[y_np]
  38. coord.request_stop()#queue需要关闭,否则报错
  39. coord.join(threads)


4、移植阶段

(1)一个算法经过实验阶段后,接着就要进入移植商用,因此接着需要采用tensorflow的c api函数,直接进行预测推理,首先我们先把tensorflow编译成链接库,然后编写cmake,调用tensorflow链接库:

bazel build -c opt //tensorflow:libtensorflow.so

bazel-bin/tensorflow目录下会生成libtensorflow.so文件

5、C++ API调用、cmake 编写:

三、熟悉常用API

1、LSTM使用

  1. import tensorflow.nn.rnn_cell
  2. lstm = rnn_cell.BasicLSTMCell(lstm_size)#创建一个lstm cell单元类,隐藏层神经元个数为lstm_size
  3. state = tf.zeros([batch_size, lstm.state_size])#一个序列隐藏层的状态值
  4. loss = 0.0
  5. for current_batch_of_words in words_in_dataset:
  6. output, state = lstm(current_batch_of_words, state)#返回值为隐藏层神经元的输出
  7. logits = tf.matmul(output, softmax_w) + softmax_b#matmul矩阵点乘
  8. probabilities = tf.nn.softmax(logits)#softmax输出
  9. loss += loss_function(probabilities, target_words)


1、one-hot函数:

  1. #ont hot 可以把训练数据的标签,直接转换成one_hot向量,用于交叉熵损失函数
  2. import tensorflow as tf
  3. a=tf.convert_to_tensor([[1],[2],[4]])
  4. b=tf.one_hot(a,5)

>>b的值为
  1. [[[ 0. 1. 0. 0. 0.]]
  2. [[ 0. 0. 1. 0. 0.]]
  3. [[ 0. 0. 0. 0. 1.]]]

2、assign_sub

  1. import tensorflow as tf
  2. x = tf.Variable(10, name="x")
  3. sub=x.assign_sub(3)#如果直接采用x.assign_sub,那么可以看到x的值也会发生变化
  4. init_op=tf.initialize_all_variables()
  5. with tf.Session() as sess:
  6. sess.run(init_op)
  7. print sub.eval()
  8. print x.eval()
可以看到输入sub=x=7

state_ops.assign_sub
采用state_ops的assign_sub也是同样sub=x=7

也就是说assign函数返回结果值的同时,变量本身的值也会被改变
3、变量查看

  1. #查看所有的变量
  2. for l in tf.all_variables():
  3. print l.name

4、slice函数:

  1. import cv2
  2. import tensorflow as tf
  3. #slice 函数可以用于切割子矩形图片,参数矩形框的rect,begin=(minx,miny),size=(width,height)
  4. minx=20
  5. miny=30
  6. height=100
  7. width=200
  8. image=tf.placeholder(dtype=tf.uint8,shape=(386,386,3))
  9. rect_image=tf.slice(image,(miny,minx,0),(height,width,-1))
  10. cvimage=cv2.imread("1.jpg")
  11. cv2.imshow("cv2",cvimage[miny:(miny+height),minx:(minx+width),:])
  12. with tf.Session() as sess:
  13. tfimage=sess.run([rect_image],{image:cvimage})
  14. cv2.imshow('tf',tfimage[0])
  15. cv2.waitKey()

5、正太分布随机初始化

tf.truncated_normal

6、打印操作运算在硬件设备信息

tf.ConfigProto(log_device_placement=True)
7、变量域名的reuse:
  1. import tensorflow as tf
  2. with tf.variable_scope('foo'):#在没有启用reuse的情况下,如果该变量还未被创建,那么就创建该变量,如果已经创建过了,那么就获取该共享变量
  3. v=tf.get_variable('v',[1])
  4. with tf.variable_scope('foo',reuse=True):#如果启用了reuse,那么编译的时候,如果get_variable没有遇到一个已经创建的变量,是会出错的
  5. v1=tf.get_variable('v1',[1])

8、allow_soft_placement的使用:allow_soft_placement=True,允许当在代码中指定tf.device设备,如果设备找不到,那么就采用默认的设备。如果该参数设置为false,当设备找不到的时候,会直接编译不通过。

9、batch normalize调用:

tf.contrib.layers.batch_norm(x, decay=0.9, updates_collections=None, epsilon=self.epsilon, scale=True, scope=self.name)







声明:本文内容由网友自发贡献,不代表【wpsshop博客】立场,版权归原作者所有,本站不承担相应法律责任。如您发现有侵权的内容,请联系我们。转载请注明出处:https://www.wpsshop.cn/w/笔触狂放9/article/detail/650016
推荐阅读
相关标签
  

闽ICP备14008679号