TensorFlow Learning to use routes

Copyright NOTICE: This article for Bo Master hjimce original article, the original address is http://blog.csdn.net/hjimce/article/details/51899683.

I. Course of study

Personal feeling for any deep learning library, such as Mxnet, TensorFlow, Theano, Caffe, and so on, basically I use the same learning process, the general process is as follows:

(1) Training stage : Data Packaging-"network construction, training-" model preservation-"visual view of loss function, verification accuracy

(2) test phase : Model load-"test picture read-" forecast display results

(3) Transplant phase : quantification, Compression acceleration-"fine-tuning-" C + + porting package-"on-line

This way, I'll take tensorflow as an example, to explain the overall structure of the process, to complete a deep learning project needs to be familiar with the process code.

second, training, testing phase

1, tensorflow packaging data

This step for TensorFlow, you can also directly read online:. jpg pictures, tag files, etc., and then through the Phaceholder variable, the data into the network, to calculate.

But this is less efficient, and for large-scale training data we need a more efficient way, and TensorFlow recommends that we use Tfrecoder for efficient data reading. Learn TensorFlow must learn to Tfrecoder file write, read, the specific sample code is as follows:

[Python]  View Plain  copy   #coding =utf-8   #tensorflow高效数据读取训练    import tensorflow& nbsp;as tf   import cv2      #把train. txt file format, each line: Picture path name     Category Tags    #奖数据打包, converted to tfrecords format for subsequent efficient reads    def encode_to_tfrecords (Lable_file,data_root, New_name= ' Data.tfrecords ', Resize=none):       writer=tf.python_io. Tfrecordwriter (data_root+ '/' +new_name)        num_example=0        with open (Lable_file, ' R ')  as f:            for l in f.readlines ():                l=l.split ()                 image=cv2.imread (data_root+ "/" +l[0])                 if resize is not none:                    image=cv2.resize (image,resize) #为了                 height,width,nchannel=image.shape                  label=int (l[1])                    example= Tf.train.Example (Features=tf.train.features (feature={                     ' height ': tf.train.Feature (Int64_list=tf.train.int64list ( Value=[height]),                     ' width ': tf.train.Feature (Int64_list=tf.train.int64list (value=[width)),                     ' Nchannel ': Tf.train.Feature (Int64_list=tf.train.int64list ( Value=[nchannel]),                     ' image ': Tf.train.Feature (Bytes_list=tf.train.byteslist (Value=[image.tobytes ())),                     ' label ': Tf.train.Feature (Int64_list=tf.train.int64list (Value=[label))                 }))                 serialized=example. Serializetostring ()                 Writer.write (serialized)                 num_example+=1       print lable_file, "Sample Data Volume:", Num_example       writer.close ()    #读取tfrecords文件    def decode_from_ Tfrecords (filename,num_epoch=none):       filename_queue=tf.train.string_input_ Producer ([Filename],num_epochs=num_epoch) #因为有的训练数据过于庞大, is divided into a number of files, so the first parameter is the file list name parameter         READER=TF. Tfrecordreader ()        _,serialized=reader.read (filename_queue)         example=tf.parse_single_example (serialized,features={             ' height ': tf. Fixedlenfeature ([],tf.int64),            ' width ': tf. Fixedlenfeature ([],tf.int64),            ' Nchannel '