keras_segmentation 튜토리얼 따라하기

Section4 딥러닝 과제

<Segmentation, Object Detection 레포지토리 찾아서 따라해보기>

 

github에서 segmentation으로 검색해서 가장 먼저 나온 레포

'awesome semantic segmentation'(https://github.com/mrgloom/awesome-semantic-segmentation)에서 따라해볼 만한 레포지토리를 찾았다. 이 레포에는 다양한 딥러닝 모델을 활용한 segmentation 레포 정리되어 있다. 여기 있는 것만 다 따라해봐도 많이 배울 듯..

mrgloom/awesome-semantic-segmentation

---

내가 따라할 레포지토리는keras_segmentation_python_example(https://github.com/divamgupta/image-segmentation-keras)이다. 간단하고 친절해보여서 선택했는데 처음부터 바로 돌아가진 않았다. 역시 환경문제가 계속 발생했다. 

AttributeError: module 'keras.utils' has no attribute 'get_file'
AttributeError: module 'tensorflow_core.compat.v2' has no attribute '__internal__' site:stackoverflow.com

이런 식으로 계속 AttributeError가 발생했는데 제대로 된 버전의 텐서플로우와 케라스를 설치해주지 않아서 발생하는 것이었다. 

코랩을 돌리기 전에

!pip install keras==2.4.3
!pip install tensorflow==2.4.1
!apt-get install -y libsm6 libxext6 libxrender-dev
!pip install opencv-python

이렇게 깃헙 레포에서 추천하는 버전을 다시 깔아주니까 깔끔하게 돌아갔다. 

---

내가 따라한 튜토리얼은 아래와 같은 길거리 이미지를 이용하여 이미지를 분류하는 것이었다.  

간단한 튜토리얼이라서 특별한 전처리 과정은 없었다. 

 

from keras_segmentation.models.unet import vgg_unet
model = vgg_unet(n_classes=50 ,  input_height=320, input_width=640)

pre-trained Unet 모델을 활용했다.  

모델의 은닉층을 살펴보면 다음과 같다. convolution 층, maxpooling 층이 반복되다가 zeropadding, batch normalization 이후 upsampling, concatenate, zeropadding, batch normalization이 반복되는 Unet의 형태를 보여준다. 

Model: "model_3"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            [(None, 320, 640, 3) 0                                            
__________________________________________________________________________________________________
block1_conv1 (Conv2D)           (None, 320, 640, 64) 1792        input_1[0][0]                    
__________________________________________________________________________________________________
block1_conv2 (Conv2D)           (None, 320, 640, 64) 36928       block1_conv1[0][0]               
__________________________________________________________________________________________________
block1_pool (MaxPooling2D)      (None, 160, 320, 64) 0           block1_conv2[0][0]               
__________________________________________________________________________________________________
block2_conv1 (Conv2D)           (None, 160, 320, 128 73856       block1_pool[0][0]                
__________________________________________________________________________________________________
block2_conv2 (Conv2D)           (None, 160, 320, 128 147584      block2_conv1[0][0]               
__________________________________________________________________________________________________
block2_pool (MaxPooling2D)      (None, 80, 160, 128) 0           block2_conv2[0][0]               
__________________________________________________________________________________________________
block3_conv1 (Conv2D)           (None, 80, 160, 256) 295168      block2_pool[0][0]                
__________________________________________________________________________________________________
block3_conv2 (Conv2D)           (None, 80, 160, 256) 590080      block3_conv1[0][0]               
__________________________________________________________________________________________________
block3_conv3 (Conv2D)           (None, 80, 160, 256) 590080      block3_conv2[0][0]               
__________________________________________________________________________________________________
block3_pool (MaxPooling2D)      (None, 40, 80, 256)  0           block3_conv3[0][0]               
__________________________________________________________________________________________________
block4_conv1 (Conv2D)           (None, 40, 80, 512)  1180160     block3_pool[0][0]                
__________________________________________________________________________________________________
block4_conv2 (Conv2D)           (None, 40, 80, 512)  2359808     block4_conv1[0][0]               
__________________________________________________________________________________________________
block4_conv3 (Conv2D)           (None, 40, 80, 512)  2359808     block4_conv2[0][0]               
__________________________________________________________________________________________________
block4_pool (MaxPooling2D)      (None, 20, 40, 512)  0           block4_conv3[0][0]               
__________________________________________________________________________________________________
zero_padding2d (ZeroPadding2D)  (None, 22, 42, 512)  0           block4_pool[0][0]                
__________________________________________________________________________________________________
conv2d (Conv2D)                 (None, 20, 40, 512)  2359808     zero_padding2d[0][0]             
__________________________________________________________________________________________________
batch_normalization (BatchNorma (None, 20, 40, 512)  2048        conv2d[0][0]                     
__________________________________________________________________________________________________
up_sampling2d (UpSampling2D)    (None, 40, 80, 512)  0           batch_normalization[0][0]        
__________________________________________________________________________________________________
concatenate (Concatenate)       (None, 40, 80, 768)  0           up_sampling2d[0][0]              
                                                                 block3_pool[0][0]                
__________________________________________________________________________________________________
zero_padding2d_1 (ZeroPadding2D (None, 42, 82, 768)  0           concatenate[0][0]                
__________________________________________________________________________________________________
conv2d_1 (Conv2D)               (None, 40, 80, 256)  1769728     zero_padding2d_1[0][0]           
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 40, 80, 256)  1024        conv2d_1[0][0]                   
__________________________________________________________________________________________________
up_sampling2d_1 (UpSampling2D)  (None, 80, 160, 256) 0           batch_normalization_1[0][0]      
__________________________________________________________________________________________________
concatenate_1 (Concatenate)     (None, 80, 160, 384) 0           up_sampling2d_1[0][0]            
                                                                 block2_pool[0][0]                
__________________________________________________________________________________________________
zero_padding2d_2 (ZeroPadding2D (None, 82, 162, 384) 0           concatenate_1[0][0]              
__________________________________________________________________________________________________
conv2d_2 (Conv2D)               (None, 80, 160, 128) 442496      zero_padding2d_2[0][0]           
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 80, 160, 128) 512         conv2d_2[0][0]                   
__________________________________________________________________________________________________
up_sampling2d_2 (UpSampling2D)  (None, 160, 320, 128 0           batch_normalization_2[0][0]      
__________________________________________________________________________________________________
concatenate_2 (Concatenate)     (None, 160, 320, 192 0           up_sampling2d_2[0][0]            
                                                                 block1_pool[0][0]                
__________________________________________________________________________________________________
zero_padding2d_3 (ZeroPadding2D (None, 162, 322, 192 0           concatenate_2[0][0]              
__________________________________________________________________________________________________
seg_feats (Conv2D)              (None, 160, 320, 64) 110656      zero_padding2d_3[0][0]           
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 160, 320, 64) 256         seg_feats[0][0]                  
__________________________________________________________________________________________________
conv2d_3 (Conv2D)               (None, 160, 320, 50) 28850       batch_normalization_3[0][0]      
__________________________________________________________________________________________________
reshape (Reshape)               (None, 51200, 50)    0           conv2d_3[0][0]                   
__________________________________________________________________________________________________
activation (Activation)         (None, 51200, 50)    0           reshape[0][0]                    
==================================================================================================
Total params: 12,350,642
Trainable params: 12,348,722
Non-trainable params: 1,920
__________________________________________________________________________________________________

epoch 5번을 돌린 결과는 loss: 0.2675 - accuracy: 0.9123로 엄청 높은 성능을 보인다. 

test 파일에 있는 그림을 넣어보면 다음과 같은 결과를 보여준다. 위의 이미지가 아래 이미지처럼 분류가 되고 라벨을 넣으면 각각의 색이 어떤 라벨인지도 보여준다. 

튜토리얼이라 결과가 잘 나온 것 같다. 다음에는 주어진 데이터 말고 다른 데이터로 실험을 해봐야겠다. 

---

마지막으로 해당 깃헙 레포를 설명하는 블로그 글

https://divamgupta.com/image-segmentation/2019/06/06/deep-learning-semantic-segmentation-keras.html

A Beginner’s guide to Deep Learning based Semantic Segmentation using Keras