一、學(xué)習(xí)路線

個(gè)人感覺對(duì)于任何一個(gè)深度學(xué)習(xí)庫，如mxnet、tensorflow、theano、caffe等，基本上我都采用同樣的一個(gè)學(xué)習(xí)流程，大體流程如下：

(1)訓(xùn)練階段：數(shù)據(jù)打包-》網(wǎng)絡(luò)構(gòu)建、訓(xùn)練-》模型保存-》可視化查看損失函數(shù)、驗(yàn)證精度

(2)測(cè)試階段：模型加載-》測(cè)試圖片讀取-》預(yù)測(cè)顯示結(jié)果

(3)移植階段：量化、壓縮加速-》微調(diào)-》C++移植打包-》上線

這邊我就以tensorflow為例子，講解整個(gè)流程的大體架構(gòu)，完成一個(gè)深度學(xué)習(xí)項(xiàng)目所需要熟悉的過程代碼。

二、訓(xùn)練、測(cè)試階段

1、tensorflow打包數(shù)據(jù)

這一步對(duì)于tensorflow來說，也可以直接自己在線讀?。?jpg圖片、標(biāo)簽文件等，然后通過phaceholder變量，把數(shù)據(jù)送入網(wǎng)絡(luò)中，進(jìn)行計(jì)算。

不過這種效率比較低，對(duì)于大規(guī)模訓(xùn)練數(shù)據(jù)來說，我們需要一個(gè)比較高效的方式，tensorflow建議我們采用tfrecoder進(jìn)行高效數(shù)據(jù)讀取。學(xué)習(xí)tensorflow一定要學(xué)會(huì)tfrecoder文件寫入、讀取，具體示例代碼如下：

[python]view plaincopy#coding=utf-8 #tensorflow高效數(shù)據(jù)讀取訓(xùn)練 importtensorflowastf importcv2 #把train.txt文件格式，每一行：圖片路徑名類別標(biāo)簽 #獎(jiǎng)數(shù)據(jù)打包，轉(zhuǎn)換成tfrecords格式，以便后續(xù)高效讀取 defencode_to_tfrecords(lable_file,data_root,new_name='data.tfrecords',resize=None): writer=tf.python_io.TFRecordWriter(data_root+'/'+new_name) num_example=0 withopen(lable_file,'r')asf: forlinf.readlines(): l=l.split() image=cv2.imread(data_root+"/"+l[0]) ifresizeisnotNone: image=cv2.resize(image,resize)#為了 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 printlable_file,"樣本數(shù)據(jù)量：",num_example writer.close() #讀取tfrecords文件 defdecode_from_tfrecords(filename,num_epoch=None): filename_queue=tf.train.string_input_producer([filename],num_epochs=num_epoch)#因?yàn)橛械挠?xùn)練數(shù)據(jù)過于龐大，被分成了很多個(gè)文件，所以第一個(gè)參數(shù)就是文件列表名參數(shù) 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':tf.FixedLenFeature([],tf.int64), 'image':tf.FixedLenFeature([],tf.string), 'label':tf.FixedLenFeature([],tf.int64) }) label=tf.cast(example['label'],tf.int32) image=tf.decode_raw(example['image'],tf.uint8) image=tf.reshape(image,tf.pack([ tf.cast(example['height'],tf.int32), tf.cast(example['width'],tf.int32), tf.cast(example['nchannel'],tf.int32)])) #label=example['label'] returnimage,label #根據(jù)隊(duì)列流數(shù)據(jù)格式，解壓出一張圖片后，輸入一張圖片，對(duì)其做預(yù)處理、及樣本隨機(jī)擴(kuò)充 defget_batch(image,label,batch_size,crop_size): #數(shù)據(jù)擴(kuò)充變換 distorted_image=tf.random_crop(image,[crop_size,crop_size,3])#隨機(jī)裁剪 distorted_image=tf.image.random_flip_up_down(distorted_image)#上下隨機(jī)翻轉(zhuǎn) #distorted_image=tf.image.random_brightness(distorted_image,max_delta=63)#亮度變化 #distorted_image=tf.image.random_contrast(distorted_image,lower=0.2,upper=1.8)#對(duì)比度變化 #生成batch #shuffle_batch的參數(shù)：capacity用于定義shuttle的范圍，如果是對(duì)整個(gè)訓(xùn)練數(shù)據(jù)集，獲取batch，那么capacity就應(yīng)該夠大 #保證數(shù)據(jù)打的足夠亂 images,label_batch=tf.train.shuffle_batch([distorted_image,label],batch_size=batch_size, num_threads=16,capacity=50000,min_after_dequeue=10000) #images,label_batch=tf.train.batch([distorted_image,label],batch_size=batch_size) #調(diào)試顯示 #tf.image_summary('images',images) returnimages,tf.reshape(label_batch,[batch_size]) #這個(gè)是用于測(cè)試階段，使用的get_batch函數(shù) defget_test_batch(image,label,batch_size,crop_size): #數(shù)據(jù)擴(kuò)充變換 distorted_image=tf.image.central_crop(image,39./45.) distorted_image=tf.random_crop(distorted_image,[crop_size,crop_size,3])#隨機(jī)裁剪 images,label_batch=tf.train.batch([distorted_image,label],batch_size=batch_size) returnimages,tf.reshape(label_batch,[batch_size]) #測(cè)試上面的壓縮、解壓代碼 deftest(): encode_to_tfrecords("data/train.txt","data",(100,100)) image,label=decode_from_tfrecords('data/data.tfrecords') batch_image,batch_label=get_batch(image,label,3)#batch生成測(cè)試 init=tf.initialize_all_variables() withtf.Session()assession: session.run(init) coord=tf.train.Coordinator() threads=tf.train.start_queue_runners(coord=coord) forlinrange(100000):#每run一次，就會(huì)指向下一個(gè)樣本，一直循環(huán) #image_np,label_np=session.run([image,label])#每調(diào)用run一次，那么 '''''cv2.imshow("temp",image_np) cv2.waitKey()''' #printlabel_np #printimage_np.shape batch_image_np,batch_label_np=session.run([batch_image,batch_label]) printbatch_image_np.shape printbatch_label_np.shape coord.request_stop()#queue需要關(guān)閉，否則報(bào)錯(cuò) coord.join(threads) #test()

2、網(wǎng)絡(luò)架構(gòu)與訓(xùn)練

經(jīng)過上面的數(shù)據(jù)格式處理，接著我們只要寫一寫網(wǎng)絡(luò)結(jié)構(gòu)、網(wǎng)絡(luò)優(yōu)化方法，把數(shù)據(jù)搞進(jìn)網(wǎng)絡(luò)中就可以了，具體示例代碼如下：

[python]view plaincopy#coding=utf-8 importtensorflowastf fromdata_encoder_decoederimportencode_to_tfrecords,decode_from_tfrecords,get_batch,get_test_batch importcv2 importos classnetwork(object): def__init__(self): withtf.variable_scope("weights"): self.weights={ #39*39*3->36*36*20->18*18*20 'conv1':tf.get_variable('conv1',[4,4,3,20],initializer=tf.contrib.layers.xavier_initializer_conv2d()), #18*18*20->16*16*40->8*8*40 'conv2':tf.get_variable('conv2',[3,3,20,40],initializer=tf.contrib.layers.xavier_initializer_conv2d()), #8*8*40->6*6*60->3*3*60 'conv3':tf.get_variable('conv3',[3,3,40,60],initializer=tf.contrib.layers.xavier_initializer_conv2d()), #3*3*60->120 'fc1':tf.get_variable('fc1',[3*3*60,120],initializer=tf.contrib.layers.xavier_initializer()), #120->6 'fc2':tf.get_variable('fc2',[120,6],initializer=tf.contrib.layers.xavier_initializer()), } withtf.variable_scope("biases"): self.biases={ 'conv1':tf.get_variable('conv1',[20,],initializer=tf.constant_initializer(value=0.0,dtype=tf.float32)), 'conv2':tf.get_variable('conv2',[40,],initializer=tf.constant_initializer(value=0.0,dtype=tf.float32)), 'conv3':tf.get_variable('conv3',[60,],initializer=tf.constant_initializer(value=0.0,dtype=tf.float32)), 'fc1':tf.get_variable('fc1',[120,],initializer=tf.constant_initializer(value=0.0,dtype=tf.float32)), 'fc2':tf.get_variable('fc2',[6,],initializer=tf.constant_initializer(value=0.0,dtype=tf.float32)) } definference(self,images): #向量轉(zhuǎn)為矩陣 images=tf.reshape(images,shape=[-1,39,39,3])#[batch,in_height,in_width,in_channels] images=(tf.cast(images,tf.float32)/255.-0.5)*2#歸一化處理 #第一層 conv1=tf.nn.bias_add(tf.nn.conv2d(images,self.weights['conv1'],strides=[1,1,1,1],padding='VALID'), self.biases['conv1']) relu1=tf.nn.relu(conv1) pool1=tf.nn.max_pool(relu1,ksize=[1,2,2,1],strides=[1,2,2,1],padding='VALID') #第二層 conv2=tf.nn.bias_add(tf.nn.conv2d(pool1,self.weights['conv2'],strides=[1,1,1,1],padding='VALID'), self.biases['conv2']) relu2=tf.nn.relu(conv2) pool2=tf.nn.max_pool(relu2,ksize=[1,2,2,1],strides=[1,2,2,1],padding='VALID') #第三層 conv3=tf.nn.bias_add(tf.nn.conv2d(pool2,self.weights['conv3'],strides=[1,1,1,1],padding='VALID'), self.biases['conv3']) relu3=tf.nn.relu(conv3) pool3=tf.nn.max_pool(relu3,ksize=[1,2,2,1],strides=[1,2,2,1],padding='VALID') #全連接層1，先把特征圖轉(zhuǎn)為向量 flatten=tf.reshape(pool3,[-1,self.weights['fc1'].get_shape().as_list()[0]]) drop1=tf.nn.dropout(flatten,0.5) fc1=tf.matmul(drop1,self.weights['fc1'])+self.biases['fc1'] fc_relu1=tf.nn.relu(fc1) fc2=tf.matmul(fc_relu1,self.weights['fc2'])+self.biases['fc2'] returnfc2 definference_test(self,images): #向量轉(zhuǎn)為矩陣 images=tf.reshape(images,shape=[-1,39,39,3])#[batch,in_height,in_width,in_channels] images=(tf.cast(images,tf.float32)/255.-0.5)*2#歸一化處理 #第一層 conv1=tf.nn.bias_add(tf.nn.conv2d(images,self.weights['conv1'],strides=[1,1,1,1],padding='VALID'), self.biases['conv1']) relu1=tf.nn.relu(conv1) pool1=tf.nn.max_pool(relu1,ksize=[1,2,2,1],strides=[1,2,2,1],padding='VALID') #第二層 conv2=tf.nn.bias_add(tf.nn.conv2d(pool1,self.weights['conv2'],strides=[1,1,1,1],padding='VALID'), self.biases['conv2']) relu2=tf.nn.relu(conv2) pool2=tf.nn.max_pool(relu2,ksize=[1,2,2,1],strides=[1,2,2,1],padding='VALID') #第三層 conv3=tf.nn.bias_add(tf.nn.conv2d(pool2,self.weights['conv3'],strides=[1,1,1,1],padding='VALID'), self.biases['conv3']) relu3=tf.nn.relu(conv3) pool3=tf.nn.max_pool(relu3,ksize=[1,2,2,1],strides=[1,2,2,1],padding='VALID') #全連接層1，先把特征圖轉(zhuǎn)為向量 flatten=tf.reshape(pool3,[-1,self.weights['fc1'].get_shape().as_list()[0]]) fc1=tf.matmul(flatten,self.weights['fc1'])+self.biases['fc1'] fc_relu1=tf.nn.relu(fc1) fc2=tf.matmul(fc_relu1,self.weights['fc2'])+self.biases['fc2'] returnfc2 #計(jì)算softmax交叉熵?fù)p失函數(shù) defsorfmax_loss(self,predicts,labels): predicts=tf.nn.softmax(predicts) labels=tf.one_hot(labels,self.weights['fc2'].get_shape().as_list()[1]) loss=-tf.reduce_mean(labels*tf.log(predicts))#tf.nn.softmax_cross_entropy_with_logits(predicts,labels) self.cost=loss returnself.cost #梯度下降 defoptimer(self,loss,lr=0.001): train_optimizer=tf.train.GradientDescentOptimizer(lr).minimize(loss) returntrain_optimizer deftrain(): encode_to_tfrecords("data/train.txt","data",'train.tfrecords',(45,45)) image,label=decode_from_tfrecords('data/train.tfrecords') batch_image,batch_label=get_batch(image,label,batch_size=50,crop_size=39)#batch生成測(cè)試 #網(wǎng)絡(luò)鏈接,訓(xùn)練所用 net=network() inf=net.inference(batch_image) loss=net.sorfmax_loss(inf,batch_label) opti=net.optimer(loss) #驗(yàn)證集所用 encode_to_tfrecords("data/val.txt","data",'val.tfrecords',(45,45)) test_image,test_label=decode_from_tfrecords('data/val.tfrecords',num_epoch=None) test_images,test_labels=get_test_batch(test_image,test_label,batch_size=120,crop_size=39)#batch生成測(cè)試 test_inf=net.inference_test(test_images) correct_prediction=tf.equal(tf.cast(tf.argmax(test_inf,1),tf.int32),test_labels) accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) init=tf.initialize_all_variables() withtf.Session()assession: session.run(init) coord=tf.train.Coordinator() threads=tf.train.start_queue_runners(coord=coord) max_iter=100000 iter=0 ifos.path.exists(os.path.join("model",'model.ckpt'))isTrue: tf.train.Saver(max_to_keep=None).restore(session,os.path.join("model",'model.ckpt')) whileiter<max_iter: loss_np,_,label_np,image_np,inf_np=session.run([loss,opti,batch_label,batch_image,inf]) #printimage_np.shape #cv2.imshow(str(label_np[0]),image_np[0]) #printlabel_np[0] #cv2.waitKey() #printlabel_np ifiter%50==0: print'trainloss:',loss_np ifiter%500==0: accuracy_np=session.run([accuracy]) print'***************testaccruacy:',accuracy_np,'*******************' tf.train.Saver(max_to_keep=None).save(session,os.path.join('model','model.ckpt')) iter+=1 coord.request_stop()#queue需要關(guān)閉，否則報(bào)錯(cuò) coord.join(threads) train()

3、可視化顯示

(1)首先再源碼中加入需要跟蹤的變量：

[python]view plaincopy