Data Augment ------TensorFlow 訓練圖片處理
阿新 • • 發佈：2019-01-27
Deep Learning 是基於一個數據的分析模型或是資料的分析方法。資料在足夠大時，其分析或事更佳，資料更佳能分析出結果。如果收集的資料過於小或是資料收集小，這樣訓練出來的結果會讓其出乎意料。我在此是針對於CNN模型，其實該模型不是說一定需要大量資料才可以得到結果，其涉及到您訓練的模型-fine-tune，採集的資料是否夠優質或是資料之間的區分度等。方法1:其實我們如果涉及到一個大類中的資料進行在此的分類時，我們可以進行拆解不一樣區域而進行多模型的輸出，最後進行一個綜合判定，該方法自己在閱讀RCNN、Faster-RCNN等論文的後感，具體自己會在MIT的bird資料集上進行測試和驗證，具體後面來續寫。方法2:還有一種方法是遷移學習模型，主要遷移imageNet資料集上的訓練。該方法目前表現不錯，自己也用於測試過在Category101上，基於ResNet50，基本在15epoch上其表現已經很好了。方法3:細小顆粒區分模型訓練，其實也是正對其少量資料訓練非常不錯的，關注過近年來的論文，在細小顆粒上區分模型也是可以進行區分的，該方法目前還是需要完善。方法4: 適用Data Augment方法進行，圖片的翻轉等，實際我們都知道，圖片實際是一個矩陣，我們實際就是將矩陣裡資料部分變換了（所謂的翻轉等）。自己在學習過程中需要擴充資料寫了一個數據擴充的類：
def floatrange(start,stop,steps): return [start+float(i)*(stop-start)/(float(steps)-1) for i in range(steps)] class DataAugment: def __init__(self,path,shape,count=0,num1=150,num2=100,num3=100,num4=150,num5=100,num6=50,num7=200,num8=150): get_tensor=tf.gfile.FastGFile(path,'rb').read() self.image_data=tf.image.decode_jpeg(get_tensor) self.shape=shape self.count=count self.num1 = num1 self.num2 = num2 self.num3 = num3 self.num4 = num4 self.num5 = num5 self.num6 = num6 self.num7 = num7 self.num8 = num8 def Save_image(self,name,x): x=tf.image.encode_jpeg(x) with tf.Session() as sess : sess.run(tf.global_variables_initializer()) with tf.gfile.FastGFile(name,'wb') as save: save.write(x.eval()) #150 1 def get_image_central(self): for i in floatrange(0.4,1.0,self.num1): self.count=self.count+1 t=self.count name=dir+str(t)+'.jpg' contest=tf.image.central_crop(self.image_data,i) contest.set_shape(self.shape) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) self.Save_image(name,contest) #print ('1') #100 2 def get_image_random_tranpose(self): for i in range(self.num2): self.count=self.count+1 t=self.count name=dir+str(t)+'.jpg' contest=tf.image.transpose_image(self.image_data) contest.set_shape(self.shape) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) self.Save_image(name,contest) #print ('2') #100 3 def get_image_up_down(self): for i in range(self.num3): self.count=self.count+1 t=self.count name=dir+str(t)+'.jpg' contest=tf.image.random_flip_up_down(self.image_data) contest.set_shape(self.shape) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) self.Save_image(name,contest) #print('3') #150 4 def get_image_left_right(self): for i in range(self.num4): self.count=self.count+1 t=self.count name=dir+str(t)+'.jpg' contest=tf.image.random_flip_left_right(self.image_data) contest.set_shape(self.shape) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) self.Save_image(name,contest) #print('4') #100 5 def get_image_bright(self): for i in floatrange(-0.6,0.6,self.num5): self.count=self.count+1 t=self.count name=dir+str(t)+'.jpg' contest=tf.image.adjust_brightness(self.image_data,i) contest.set_shape(self.shape) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) self.Save_image(name,contest) #print('5') #050 6 def get_image_contrast(self): for i in range(self.num6): index=np.random.randint(-4,4) self.count=self.count+1 t=self.count name=dir+str(t)+'.jpg' contest=tf.image.adjust_contrast(self.image_data,index) contest.set_shape(self.shape) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) self.Save_image(name,contest) #print ('6') #200 7 def get_image_hue(self): for i in floatrange(0.0,0.5,self.num7): self.count=self.count+1 t=self.count name=dir+str(t)+'.jpg' contest=tf.image.random_hue(self.image_data,i) contest.set_shape(self.shape) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) self.Save_image(name,contest) #print ('7') #150 8 def get_image_sation(self): for i in range(self.num8): index=np.random.randint(-6,10) self.count=self.count+1 t=self.count name=dir+str(t)+'.jpg' contest=tf.image.adjust_saturation(self.image_data,index) contest.set_shape(self.shape) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) self.Save_image(name,contest) #print ('8') #'1' def get_count(self): return self.count def run(self): self.get_image_random_tranpose() self.get_image_left_right() self.get_image_central() self.get_image_bright() self.get_image_hue() self.get_image_contrast() self.get_image_sation() self.get_image_up_down() def run_1(self): #self.get_image_bright() self.get_image_random_tranpose() self.get_image_left_right() #self.get_image_contrast() #self.get_image_up_down() #self.get_image_central() #self.get_image_hue() self.get_image_hue() self.get_image_sation() #self.get_image_sation() def show(self): x=tf.image.encode_jpeg(self.image_data) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) plt.imread(x.eval()) plt.show() def get_image_path(path): list_name = [] if os.path.isdir(path): for file in os.listdir(path): if file.endswith('.jpg') and not file.startswith('.'): filename = path+'/'+file list_name.append(filename) return list_name def get_dir(dir): dirs = [] if os.path.isdir(dir): for path in os.listdir(dir): paths = dir +'/'+path if os.path.isdir(paths): dirs.append(paths) return dirs def make_dirs(path): name = 'file' for i in range(250): names = path +name +str(i) if not os._exists(names): os.makedirs(names)
呼叫用例為：
if __name__=='__main__': test=DataAugment('/Users/josen/Desktop/001_0003.jpg',[64,64,3],0,2,2,2,2,2,2,2,2) test.run_1() t=test.get_count() print ('the final data is ',str(t)) cout = 0 path = '/Users/josen/Desktop/256_ObjectCategories/'+input('dir:') dir = input('file-path:') list_name = get_image_path(path) for i in range(len(list_name)): test = DataAugment(list_name[i],[64,64,3],cout,2,2,2,2,2,2,2,2) test.run_1() cout = test.get_count() print(cout) root = '/Users/josen/Desktop/Object256/'+input('file:--') #list_dir = get_dir(root) #print(len(list_dir)) cout = 0 #for i in range(len(list_dir)): list_name = get_image_path(root) #dir = list_dir[i] +'/' #nfor = list_name.split('/')[4].split('.')[0] for j in range(len(list_name)): test = DataAugment(list_name[j],[64,64,3],cout,2,2,2,2,2,2,2,2) test.run_1() cout = test.get_count() if cout == 650: break print(cout)