深度學習(五十六)tensorflow專案構建流程
tensorflow專案構建流程
微博:黃錦池-hjimce qq:1393852684
一、構建路線
個人感覺對於任何一個深度學習庫,如mxnet、tensorflow、theano、caffe等,基本上我都採用同樣的一個學習流程,大體流程如下:
(1)訓練階段:資料打包-》網路構建、訓練-》模型儲存-》視覺化檢視損失函式、驗證精度
(2)測試階段:模型載入-》測試圖片讀取-》預測顯示結果
(3)移植階段:量化、壓縮加速-》微調-》C++移植打包-》上線
這邊我就以tensorflow為例子,講解整個流程的大體架構,完成一個深度學習專案所需要熟悉的過程程式碼。
二、訓練、測試階段
1、tensorflow打包資料
這一步對於tensorflow來說,也可以直接自己線上讀取:.jpg圖片、標籤檔案等,然後通過phaceholder變數,把資料送入網路中,進行計算。
不過這種效率比較低,對於大規模訓練資料來說,我們需要一個比較高效的方式,tensorflow建議我們採用tfrecoder進行高效資料讀取。學習tensorflow一定要學會tfrecoder檔案寫入、讀取,具體示例程式碼如下:
#coding=utf-8 #tensorflow高效資料讀取訓練 import tensorflow as tf import cv2 #把train.txt檔案格式,每一行:圖片路徑名 類別標籤 #獎資料打包,轉換成tfrecords格式,以便後續高效讀取 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)#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 print lable_file,"樣本資料量:",num_example writer.close() #讀取tfrecords檔案 def decode_from_tfrecords(filename,num_epoch=None): filename_queue=tf.train.string_input_producer([filename],num_epochs=num_epoch)#因為有的訓練資料過於龐大,被分成了很多個檔案,所以第一個引數就是檔案列表名引數 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'] return image,label #根據佇列流資料格式,解壓出一張圖片後,輸入一張圖片,對其做預處理、及樣本隨機擴充 def get_batch(image, label, batch_size,crop_size): #資料擴充變換 distorted_image = tf.random_crop(image, [crop_size, crop_size, 3])#隨機裁剪 distorted_image = tf.image.random_flip_up_down(distorted_image)#上下隨機翻轉 #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)#對比度變化 #生成batch #shuffle_batch的引數:capacity用於定義shuttle的範圍,如果是對整個訓練資料集,獲取batch,那麼capacity就應該夠大 #保證資料打的足夠亂 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) # 除錯顯示 #tf.image_summary('images', images) return images, tf.reshape(label_batch, [batch_size]) #這個是用於測試階段,使用的get_batch函式 def get_test_batch(image, label, batch_size,crop_size): #資料擴充變換 distorted_image=tf.image.central_crop(image,39./45.) distorted_image = tf.random_crop(distorted_image, [crop_size, crop_size, 3])#隨機裁剪 images, label_batch=tf.train.batch([distorted_image, label],batch_size=batch_size) return images, tf.reshape(label_batch, [batch_size]) #測試上面的壓縮、解壓程式碼 def test(): 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 生成測試 init=tf.initialize_all_variables() with tf.Session() as session: session.run(init) coord = tf.train.Coordinator() threads = tf.train.start_queue_runners(coord=coord) for l in range(100000):#每run一次,就會指向下一個樣本,一直迴圈 #image_np,label_np=session.run([image,label])#每呼叫run一次,那麼 '''cv2.imshow("temp",image_np) cv2.waitKey()''' #print label_np #print image_np.shape batch_image_np,batch_label_np=session.run([batch_image,batch_label]) print batch_image_np.shape print batch_label_np.shape coord.request_stop()#queue需要關閉,否則報錯 coord.join(threads) #test()
2、網路架構與訓練
經過上面的資料格式處理,接著我們只要寫一寫網路結構、網路優化方法,把資料搞進網路中就可以了,具體示例程式碼如下:
#coding=utf-8 import tensorflow as tf from data_encoder_decoeder import encode_to_tfrecords,decode_from_tfrecords,get_batch,get_test_batch import cv2 import os class network(object): def __init__(self): with tf.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()), } with tf.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)) } def inference(self,images): # 向量轉為矩陣 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') # 全連線層1,先把特徵圖轉為向量 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'] return fc2 def inference_test(self,images): # 向量轉為矩陣 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') # 全連線層1,先把特徵圖轉為向量 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'] return fc2 #計算softmax交叉熵損失函式 def sorfmax_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 return self.cost #梯度下降 def optimer(self,loss,lr=0.001): train_optimizer = tf.train.GradientDescentOptimizer(lr).minimize(loss) return train_optimizer def train(): 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 生成測試 #網路連結,訓練所用 net=network() inf=net.inference(batch_image) loss=net.sorfmax_loss(inf,batch_label) opti=net.optimer(loss) #驗證集所用 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() with tf.Session() as session: session.run(init) coord = tf.train.Coordinator() threads = tf.train.start_queue_runners(coord=coord) max_iter=100000 iter=0 if os.path.exists(os.path.join("model",'model.ckpt')) is True: tf.train.Saver(max_to_keep=None).restore(session, os.path.join("model",'model.ckpt')) while iter<max_iter: loss_np,_,label_np,image_np,inf_np=session.run([loss,opti,batch_label,batch_image,inf]) #print image_np.shape #cv2.imshow(str(label_np[0]),image_np[0]) #print label_np[0] #cv2.waitKey() #print label_np if iter%50==0: print 'trainloss:',loss_np if iter%500==0: accuracy_np=session.run([accuracy]) print '***************test accruacy:',accuracy_np,'*******************' tf.train.Saver(max_to_keep=None).save(session, os.path.join('model','model.ckpt')) iter+=1 coord.request_stop()#queue需要關閉,否則報錯 coord.join(threads) train()
3、視覺化顯示
(1)首先再原始碼中加入需要跟蹤的變數:
tf.scalar_summary("cost_function", loss)#損失函式值
(2)然後定義執行操作:merged_summary_op = tf.merge_all_summaries()
(3)再session中定義儲存路徑:summary_writer = tf.train.SummaryWriter('log', session.graph)
(4)然後再session執行的時候,儲存:
summary_str,loss_np,_=session.run([merged_summary_op,loss,opti])
summary_writer.add_summary(summary_str, iter)
(5)最後只要訓練完畢後,直接再終端輸入命令:
python /usr/local/lib/python2.7/dist-packages/tensorflow/tensorboard/tensorboard.py --logdir=log
然後開啟瀏覽器網址:http://0.0.0.0:6006
即可觀訓練曲線。
4、測試階段
測試階段主要是直接通過載入圖模型、讀取引數等,然後直接通過tensorflow的相關函式,進行呼叫,而不需要網路架構相關的程式碼;通過記憶體feed_dict的方式,對相關的輸入節點賦予相關的資料,進行前向傳導,並獲取相關的節點數值。
#coding=utf-8
import tensorflow as tf
import os
import cv2
def load_model(session,netmodel_path,param_path):
new_saver = tf.train.import_meta_graph(netmodel_path)
new_saver.restore(session, param_path)
x= tf.get_collection('test_images')[0]#在訓練階段需要呼叫tf.add_to_collection('test_images',test_images),儲存之
y = tf.get_collection("test_inf")[0]
batch_size = tf.get_collection("batch_size")[0]
return x,y,batch_size
def load_images(data_root):
filename_queue = tf.train.string_input_producer(data_root)
image_reader = tf.WholeFileReader()
key,image_file = image_reader.read(filename_queue)
image = tf.image.decode_jpeg(image_file)
return image, key
def test(data_root="data/race/cropbrown"):
image_filenames=os.listdir(data_root)
image_filenames=[(data_root+'/'+i) for i in image_filenames]
#print cv2.imread(image_filenames[0]).shape
#image,key=load_images(image_filenames)
race_listsrc=['black','brown','white','yellow']
with tf.Session() as session:
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
x,y,batch_size=load_model(session,os.path.join("model",'model_ori_race.ckpt.meta'),
os.path.join("model",'model_ori_race.ckpt'))
predict_label=tf.cast(tf.argmax(y,1),tf.int32)
print x.get_shape()
for imgf in image_filenames:
image=cv2.imread(imgf)
image=cv2.resize(image,(76,76)).reshape((1,76,76,3))
print "cv shape:",image.shape
#cv2.imshow("t",image_np[:,:,::-1])
y_np=session.run(predict_label,feed_dict = {x:image, batch_size:1})
print race_listsrc[y_np]
coord.request_stop()#queue需要關閉,否則報錯
coord.join(threads)
4、移植階段
(1)一個演算法經過實驗階段後,接著就要進入移植商用,因此接著需要採用tensorflow的c api函式,直接進行預測推理,首先我們先把tensorflow編譯成連結庫,然後編寫cmake,呼叫tensorflow連結庫:
bazel build -c opt //tensorflow:libtensorflow.so
在bazel-bin/tensorflow目錄下會生成libtensorflow.so檔案
5、C++ API呼叫、cmake 編寫:
三、熟悉常用API
1、LSTM使用
import tensorflow.nn.rnn_cell
lstm = rnn_cell.BasicLSTMCell(lstm_size)#建立一個lstm cell單元類,隱藏層神經元個數為lstm_size
state = tf.zeros([batch_size, lstm.state_size])#一個序列隱藏層的狀態值
loss = 0.0
for current_batch_of_words in words_in_dataset:
output, state = lstm(current_batch_of_words, state)#返回值為隱藏層神經元的輸出
logits = tf.matmul(output, softmax_w) + softmax_b#matmul矩陣點乘
probabilities = tf.nn.softmax(logits)#softmax輸出
loss += loss_function(probabilities, target_words)
1、one-hot函式:
#ont hot 可以把訓練資料的標籤,直接轉換成one_hot向量,用於交叉熵損失函式
import tensorflow as tf
a=tf.convert_to_tensor([[1],[2],[4]])
b=tf.one_hot(a,5)
>>b的值為
[[[ 0. 1. 0. 0. 0.]]
[[ 0. 0. 1. 0. 0.]]
[[ 0. 0. 0. 0. 1.]]]
2、assign_sub
import tensorflow as tf
x = tf.Variable(10, name="x")
sub=x.assign_sub(3)#如果直接採用x.assign_sub,那麼可以看到x的值也會發生變化
init_op=tf.initialize_all_variables()
with tf.Session() as sess:
sess.run(init_op)
print sub.eval()
print x.eval()
可以看到輸入sub=x=7state_ops.assign_sub
採用state_ops的assign_sub也是同樣sub=x=7也就是說assign函式返回結果值的同時,變數本身的值也會被改變
3、變數檢視
#檢視所有的變數
for l in tf.all_variables():
print l.name
4、slice函式:
import cv2
import tensorflow as tf
#slice 函式可以用於切割子矩形圖片,引數矩形框的rect,begin=(minx,miny),size=(width,height)
minx=20
miny=30
height=100
width=200
image=tf.placeholder(dtype=tf.uint8,shape=(386,386,3))
rect_image=tf.slice(image,(miny,minx,0),(height,width,-1))
cvimage=cv2.imread("1.jpg")
cv2.imshow("cv2",cvimage[miny:(miny+height),minx:(minx+width),:])
with tf.Session() as sess:
tfimage=sess.run([rect_image],{image:cvimage})
cv2.imshow('tf',tfimage[0])
cv2.waitKey()
5、正太分佈隨機初始化
tf.truncated_normal
6、列印操作運算在硬體裝置資訊
tf.ConfigProto(log_device_placement=True)
7、變數域名的reuse:import tensorflow as tf
with tf.variable_scope('foo'):#在沒有啟用reuse的情況下,如果該變數還未被建立,那麼就建立該變數,如果已經建立過了,那麼就獲取該共享變數
v=tf.get_variable('v',[1])
with tf.variable_scope('foo',reuse=True):#如果啟用了reuse,那麼編譯的時候,如果get_variable沒有遇到一個已經建立的變數,是會出錯的
v1=tf.get_variable('v1',[1])
8、allow_soft_placement的使用:allow_soft_placement=True,允許當在程式碼中指定tf.device裝置,如果裝置找不到,那麼就採用預設的裝置。如果該引數設定為false,當裝置找不到的時候,會直接編譯不通過。
9、batch normalize呼叫:
tf.contrib.layers.batch_norm(x, decay=0.9, updates_collections=None, epsilon=self.epsilon, scale=True, scope=self.name)