U-Net으로 시맨틱 세그멘테이션을 이용해 도로찾기

시맨틱 세그멘테이션으로 도로영역 찾기¶

---

1. 시맨틱 세그멘테이션 데이터셋¶

이미지, 데이터 가져오기¶

• mkdir -p ~/aiffel/semantic_segmentation/data
• wget https://s3.eu-central-1.amazonaws.com/avg-kitti/data_semantics.zip
• unzip data_semantics.zip -d ~/aiffel/semantic_segmentation/data

데이터 로더(data loader) 만들기¶

• pip install albumentations

• 참고자료

  • Keras Sequence에 기반한 Dataloader
  • Albumentation을 적용한 Keras sequence

In [1]:

#필요한 라이브러리를 로드합니다. 
import os
import math
import numpy as np
import tensorflow as tf

from PIL import Image
import matplotlib.pyplot as plt
from skimage.io import imread
from skimage.transform import resize
from glob import glob

from tensorflow.keras.models import *
from tensorflow.keras.layers import *
from tensorflow.keras.optimizers import *

print('슝=3')

슝=3

In [2]:

from albumentations import  HorizontalFlip, RandomSizedCrop, Compose, OneOf, Resize

def build_augmentation(is_train=True):
  if is_train:    # 훈련용 데이터일 경우
    return Compose([
                    HorizontalFlip(p=0.5),    # 50%의 확률로 좌우대칭
                    RandomSizedCrop(         # 50%의 확률로 RandomSizedCrop
                        min_max_height=(300, 370),
                        w2h_ratio=370/1242,
                        height=224,
                        width=224,
                        p=0.5
                        ),
                    Resize(              # 입력이미지를 224X224로 resize
                        width=224,
                        height=224
                        )
                    ])
  return Compose([      # 테스트용 데이터일 경우에는 224X224로 resize만 수행합니다. 
                Resize(
                    width=224,
                    height=224
                    )
                ])

In [3]:

import os

dir_path = os.getenv('HOME')+'/aiffel/semantic_segmentation/data/training'

augmentation = build_augmentation()
input_images = glob(os.path.join(dir_path, "image_2", "*.png"))

# 훈련 데이터셋에서 5개만 가져와 augmentation을 적용해 봅시다.  
plt.figure(figsize=(12, 20))
for i in range(5):
    image = imread(input_images[i]) 
    image_data = {"image":image}
    resized = augmentation(**image_data, is_train=False)
    processed = augmentation(**image_data)
    plt.subplot(5, 2, 2*i+1)
    plt.imshow(resized["image"])  # 왼쪽이 원본이미지
    plt.subplot(5, 2, 2*i+2)
    plt.imshow(processed["image"])  # 오른쪽이 augment된 이미지
  
plt.show()

In [4]:

class KittiGenerator(tf.keras.utils.Sequence):
  '''
  KittiGenerator는 tf.keras.utils.Sequence를 상속받습니다.
  우리가 KittiDataset을 원하는 방식으로 preprocess하기 위해서 Sequnce를 커스텀해 사용합니다.
  '''
  def __init__(self, 
               dir_path,
               batch_size=16,
               img_size=(224, 224, 3),
               output_size=(224, 224),
               is_train=True,
               augmentation=None):
    '''
    dir_path: dataset의 directory path입니다.
    batch_size: batch_size입니다.
    img_size: preprocess에 사용할 입력이미지의 크기입니다.
    output_size: ground_truth를 만들어주기 위한 크기입니다.
    is_train: 이 Generator가 학습용인지 테스트용인지 구분합니다.
    augmentation: 적용하길 원하는 augmentation 함수를 인자로 받습니다.
    '''
    self.dir_path = dir_path
    self.batch_size = batch_size
    self.is_train = is_train
    self.dir_path = dir_path
    self.augmentation = augmentation
    self.img_size = img_size
    self.output_size = output_size

    # load_dataset()을 통해서 kitti dataset의 directory path에서 라벨과 이미지를 확인합니다.
    self.data = self.load_dataset()

  def load_dataset(self):
    # kitti dataset에서 필요한 정보(이미지 경로 및 라벨)를 directory에서 확인하고 로드하는 함수입니다.
    # 이때 is_train에 따라 test set을 분리해서 load하도록 해야합니다.
    input_images = glob(os.path.join(self.dir_path, "image_2", "*.png"))
    label_images = glob(os.path.join(self.dir_path, "semantic", "*.png"))
    input_images.sort()
    label_images.sort()
    assert len(input_images) == len(label_images)
    data = [ _ for _ in zip(input_images, label_images)]

    if self.is_train:
      return data[:-30]
    return data[-30:]
    
  def __len__(self):
    # Generator의 length로서 전체 dataset을 batch_size로 나누고 소숫점 첫째자리에서 올림한 값을 반환합니다.
    return math.ceil(len(self.data) / self.batch_size)

  def __getitem__(self, index):
    # 입력과 출력을 만듭니다.
    # 입력은 resize및 augmentation이 적용된 input image이고 
    # 출력은 semantic label입니다.
    batch_data = self.data[
                           index*self.batch_size:
                           (index + 1)*self.batch_size
                           ]
    inputs = np.zeros([self.batch_size, *self.img_size])
    outputs = np.zeros([self.batch_size, *self.output_size])
        
    for i, data in enumerate(batch_data):
      input_img_path, output_path = data
      _input = imread(input_img_path)
      _output = imread(output_path)
      _output = (_output==7).astype(np.uint8)*1
      data = {
          "image": _input,
          "mask": _output,
          }
      augmented = self.augmentation(**data)
      inputs[i] = augmented["image"]/255
      outputs[i] = augmented["mask"]
      return inputs, outputs

  def on_epoch_end(self):
    # 한 epoch가 끝나면 실행되는 함수입니다. 학습중인 경우에 순서를 random shuffle하도록 적용한 것을 볼 수 있습니다.
    self.indexes = np.arange(len(self.data))
    if self.is_train == True :
      np.random.shuffle(self.indexes)
      return self.indexes

In [5]:

augmentation = build_augmentation()
test_preproc = build_augmentation(is_train=False)
        
train_generator = KittiGenerator(
    dir_path, 
    augmentation=augmentation,
)

test_generator = KittiGenerator(
    dir_path, 
    augmentation=test_preproc,
    is_train=False
)

시맨틱 세그멘테이션 모델¶

모델 구조 만들기¶

In [6]:

def unet(input_shape=(224, 224, 3)):
  inputs = Input(input_shape)

    #Contracting Path
  conv1 = Conv2D(64, 3, activation='relu', padding='same',kernel_initializer='he_normal')(inputs)
  conv1 = Conv2D(64, 3, activation='relu', padding='same',kernel_initializer='he_normal')(conv1)
  pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
  conv2 = Conv2D(128, 3, activation='relu', padding='same',kernel_initializer='he_normal')(pool1)
  conv2 = Conv2D(128, 3, activation='relu', padding='same',kernel_initializer='he_normal')(conv2)
  pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
  conv3 = Conv2D(256, 3, activation='relu', padding='same',kernel_initializer='he_normal')(pool2)
  conv3 = Conv2D(256, 3, activation='relu', padding='same',kernel_initializer='he_normal')(conv3)
  pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
  conv4 = Conv2D(512, 3, activation='relu', padding='same',kernel_initializer='he_normal')(pool3)
  conv4 = Conv2D(512, 3, activation='relu', padding='same',kernel_initializer='he_normal')(conv4)
  drop4 = Dropout(0.5)(conv4)
  pool4 = MaxPooling2D(pool_size=(2, 2))(drop4)

  conv5 = Conv2D(1024, 3, activation='relu', padding='same',kernel_initializer='he_normal')(pool4)  
  conv5 = Conv2D(1024, 3, activation='relu', padding='same',kernel_initializer='he_normal')(conv5)

    #Expanding Path
  drop5 = Dropout(0.5)(conv5)
  up6 = Conv2D(512, 2, activation='relu', padding='same',kernel_initializer='he_normal')(UpSampling2D(size = (2,2))(drop5)) 
  merge6 = concatenate([drop4,up6], axis = 3)
  conv6 = Conv2D(512, 3, activation='relu', padding='same',kernel_initializer='he_normal')(merge6)
  conv6 = Conv2D(512, 3, activation='relu', padding='same',kernel_initializer='he_normal')(conv6)
  up7 = Conv2D(256, 2, activation='relu', padding='same',kernel_initializer='he_normal')(UpSampling2D(size = (2,2))(conv6))
  merge7 = concatenate([conv3,up7], axis = 3)
  conv7 = Conv2D(256, 3, activation='relu', padding='same',kernel_initializer='he_normal')(merge7)
  conv7 = Conv2D(256, 3, activation='relu', padding='same',kernel_initializer='he_normal')(conv7)
  up8 = Conv2D(128, 2, activation='relu', padding='same',kernel_initializer='he_normal')(UpSampling2D(size = (2,2))(conv7))
  merge8 = concatenate([conv2,up8], axis = 3)
  conv8 = Conv2D(128, 3, activation='relu', padding='same',kernel_initializer='he_normal')(merge8)
  conv8 = Conv2D(128, 3, activation='relu', padding='same',kernel_initializer='he_normal')(conv8)
  up9 = Conv2D(64, 2, activation='relu', padding='same',kernel_initializer='he_normal')(UpSampling2D(size = (2,2))(conv8))
  merge9 = concatenate([conv1,up9], axis = 3)
  conv9 = Conv2D(64, 3, activation='relu', padding='same',kernel_initializer='he_normal')(merge9)
  conv9 = Conv2D(64, 3, activation='relu', padding='same',kernel_initializer='he_normal')(conv9)  
  conv9 = Conv2D(2, 3, activation='relu', padding='same',kernel_initializer='he_normal')(conv9)     
  conv10 = Conv2D(1, 1, activation='sigmoid')(conv9)

  model = Model(inputs = inputs, outputs = conv10)
  return model

모델 학습하기¶

In [ ]:

unet = build_model()
unet.compile(optimizer = Adam(lr = 1e-4), loss = 'binary_crossentropy')
unet.fit_generator(
     generator=train_generator,
     validation_data=test_generator,
     steps_per_epoch=len(train_generator),
     epochs=100,
 )

In [7]:

model_path = dir_path + '/unet.h5'
#unet.save(model_path)  #학습한 모델을 저장해 주세요.

시맨틱 세그멘테이션 모델 시각화¶

• model = tf.keras.models.load_model(model_path) 저장된 모델 불러오기

In [8]:

unet = tf.keras.models.load_model(model_path)

In [9]:

def get_output(model, preproc, image_path, output_path):
     origin_img = imread(image_path)
     data = {"image":origin_img}
     processed = preproc(**data)
     output = model(np.expand_dims(processed["image"]/255,axis=0))
     output = (output[0].numpy()>0.5).astype(np.uint8).squeeze(-1)*255  #0.5라는 threshold를 변경하면 도로인식 결과범위가 달라집니다.
     output = Image.fromarray(output)
     background = Image.fromarray(origin_img).convert('RGBA')
     output = output.resize((origin_img.shape[1], origin_img.shape[0])).convert('RGBA')
     output = Image.blend(background, output, alpha=0.5)
     output.show()
     return output

In [10]:

def calculate_iou_score(target, prediction):
    intersection = np.logical_and(target, prediction)
    union = np.logical_or(target, prediction)
    iou_score = float(np.sum(intersection)) / float(np.sum(union))
    print('IoU : %f' % iou_score )
    return iou_score

In [11]:

def get_output(model, preproc, image_path, output_path, label_path):
    origin_img = imread(image_path)
    data = {"image":origin_img}
    processed = preproc(**data)
    output = model(np.expand_dims(processed["image"]/255,axis=0))
    output = (output[0].numpy()>=0.5).astype(np.uint8).squeeze(-1)*255  #0.5라는 threshold를 변경하면 도로인식 결과범위가 달라집니다.
    prediction = output/255   # 도로로 판단한 영역
    
    output = Image.fromarray(output)
    background = Image.fromarray(origin_img).convert('RGBA')
    output = output.resize((origin_img.shape[1], origin_img.shape[0])).convert('RGBA')
    output = Image.blend(background, output, alpha=0.5)
    output.show()   # 도로로 판단한 영역을 시각화!
     
    if label_path:   
        label_img = imread(label_path)
        label_data = {"image":label_img}
        label_processed = preproc(**label_data)
        label_processed = label_processed["image"]
        target = (label_processed == 7).astype(np.uint8)*1   # 라벨에서 도로로 기재된 영역

        return output, prediction, target
    else:
        return output, prediction, _

In [45]:

# 완성한 뒤에는 시각화한 결과를 눈으로 확인해봅시다!
i = 2   # i값을 바꾸면 테스트용 파일이 달라집니다. 
output, prediction, target = get_output(
     unet, 
     test_preproc,
     image_path=dir_path + f'/image_2/00{str(i).zfill(4)}_10.png',
     output_path=dir_path + f'./result_{str(i).zfill(3)}.png',
     label_path=dir_path + f'/semantic/00{str(i).zfill(4)}_10.png'
 )

calculate_iou_score(target, prediction)

IoU : 0.622965

Out[45]:

0.6229652927219556

In [13]:

plt.figure(figsize=(8, 16))
for i in range(5) :
    data = next(iter(train_generator))
    
    plt.subplot(5, 2, 2*i+1)
    plt.imshow(data[0][0, :, :])
    plt.subplot(5, 2, 2*i+2)
    plt.imshow(data[1][0, :, :])
plt.show()

프로젝트 : 개선된 U-Net 모델 만들기¶

-U-Net ++

In [27]:

def build_UNet_Plus(input_shape=(224, 224, 3), n = 16) :
    
    filters = [n, n*2, n*4, n*8, n*16]
    
    inputs = Input(input_shape)
    
    #conv0_0
    conv001 = Conv2D(filters=filters[0], kernel_size=3, padding='same', kernel_initializer='he_normal', name='conv0_0_1')(inputs)
    bn001 = BatchNormalization()(conv001)
    act001 = Activation('relu')(bn001)
    conv002 = Conv2D(filters=filters[0], kernel_size=3, padding='same', kernel_initializer='he_normal', name='conv0_0_2')(act001)
    bn002 = BatchNormalization()(conv002)
    conv00 = Activation('relu')(bn002)
    
    #conv1_0
    conv00_down = MaxPooling2D(strides=2)(conv00)
    conv101 = Conv2D(filters[1], 3, padding='same', kernel_initializer='he_normal', name='conv1_0_1')(conv00_down)
    bn101 = BatchNormalization()(conv101)
    act101 = Activation('relu')(bn101)
    conv102 = Conv2D(filters[1], 3, padding='same', kernel_initializer='he_normal', name='conv1_0_2')(act101)
    bn102 = BatchNormalization()(conv102)
    conv10 = Activation('relu')(bn102)
    
    #conv2_0
    conv10_down = MaxPooling2D(strides=2)(conv10)
    conv201 = Conv2D(filters[2], 3, padding='same', kernel_initializer='he_normal', name='conv2_0_1')(conv10_down)
    bn201 = BatchNormalization()(conv201)
    act201 = Activation('relu')(bn201)
    conv202 = Conv2D(filters[2], 3, padding='same', kernel_initializer='he_normal', name='conv2_0_2')(act201)
    bn202 = BatchNormalization()(conv202)
    conv20 = Activation('relu')(bn202)
    
    #conv3_0
    conv20_down = MaxPooling2D(strides=2)(conv20)
    conv301 = Conv2D(filters[3], 3, padding='same', kernel_initializer='he_normal', name='conv3_0_1')(conv20_down)
    bn301 = BatchNormalization()(conv301)
    act301 = Activation('relu')(bn301)
    conv302 = Conv2D(filters[3], 3, padding='same', kernel_initializer='he_normal', name='conv3_0_2')(act301)
    bn302 = BatchNormalization()(conv302)
    conv30 = Activation('relu')(bn302)
    
    #conv4_0
    conv30_down = MaxPooling2D(strides=2)(conv30)
    conv401 = Conv2D(filters[4], 3, padding='same', kernel_initializer='he_normal', name='conv4_0_1')(conv30_down)
    bn401 = BatchNormalization()(conv401)
    act401 = Activation('relu')(bn401)
    conv402 = Conv2D(filters[4], 3, padding='same', kernel_initializer='he_normal', name='conv4_0_2')(act401)
    bn402 = BatchNormalization()(conv402)
    conv40 = Activation('relu')(bn402)
    
    
    
    #conv0_1
    conv10_up = UpSampling2D()(conv10)
    conv01_in = Concatenate()([conv00, conv10_up])
    
    conv011 = Conv2D(filters[0], 3, padding='same', kernel_initializer='he_normal', name='conv0_1_1')(conv01_in)
    bn011 = BatchNormalization()(conv011)
    act011 = Activation('relu')(bn011)
    conv012 = Conv2D(filters[0], 3, padding='same', kernel_initializer='he_normal', name='conv0_1_2')(act011)
    bn012 = BatchNormalization()(conv012)
    conv01 = Activation('relu')(bn012)
    
    #conv1_1
    conv20_up = UpSampling2D()(conv20)
    conv11_in = Concatenate()([conv10, conv20_up])
    
    conv111 = Conv2D(filters[1], 3, padding='same', kernel_initializer='he_normal', name='conv1_1_1')(conv11_in)
    bn111 = BatchNormalization()(conv111)
    act111 = Activation('relu')(bn111)
    conv112 = Conv2D(filters[1], 3, padding='same', kernel_initializer='he_normal', name='conv1_1_2')(act111)
    bn112 = BatchNormalization()(conv112)
    conv11 = Activation('relu')(bn112)
    
    #conv2_1
    conv30_up = UpSampling2D()(conv30)
    conv21_in = Concatenate()([conv20, conv30_up])
    
    conv211 = Conv2D(filters[2], 3, padding='same', kernel_initializer='he_normal', name='conv2_1_1')(conv21_in)
    bn211 = BatchNormalization()(conv211)
    act211 = Activation('relu')(bn211)
    conv212 = Conv2D(filters[2], 3, padding='same', kernel_initializer='he_normal', name='conv2_1_2')(act211)
    bn212 = BatchNormalization()(conv212)
    conv21 = Activation('relu')(bn212)
    
    #conv3_1
    conv40_up = UpSampling2D()(conv40)
    conv31_in = Concatenate()([conv30, conv40_up])
    
    conv311 = Conv2D(filters[3], 3, padding='same', kernel_initializer='he_normal', name='conv3_1_1')(conv31_in)
    bn311 = BatchNormalization()(conv311)
    act311 = Activation('relu')(bn311)
    conv312 = Conv2D(filters[3], 3, padding='same', kernel_initializer='he_normal', name='conv3_1_2')(act311)
    bn312 = BatchNormalization()(conv312)
    conv31 = Activation('relu')(bn312)
    
    
    
    #conv0_2
    conv11_up = UpSampling2D()(conv11)
    conv02_in = Concatenate()([conv00, conv01, conv11_up])
    
    conv021 = Conv2D(filters[0], 3, padding='same', kernel_initializer='he_normal', name='conv0_2_1')(conv02_in)
    bn021 = BatchNormalization()(conv021)
    act021 = Activation('relu')(bn021)
    conv022 = Conv2D(filters[0], 3, padding='same', kernel_initializer='he_normal', name='conv0_2_2')(act021)
    bn022 = BatchNormalization()(conv022)
    conv02 = Activation('relu')(bn022)
    
    #conv1_2
    conv21_up = UpSampling2D()(conv21)
    conv12_in = Concatenate()([conv10, conv11, conv21_up])
    
    conv121 = Conv2D(filters[1], 3, padding='same', kernel_initializer='he_normal', name='conv1_2_1')(conv12_in)
    bn121 = BatchNormalization()(conv121)
    act121 = Activation('relu')(bn121)
    conv122 = Conv2D(filters[1], 3, padding='same', kernel_initializer='he_normal', name='conv1_2_2')(act121)
    bn122 = BatchNormalization()(conv122)
    conv12 = Activation('relu')(bn122)
    
    #conv2_2
    conv31_up = UpSampling2D()(conv31)
    conv22_in = Concatenate()([conv20, conv21, conv31_up])
    
    conv221 = Conv2D(filters[2], 3, padding='same', kernel_initializer='he_normal', name='conv2_2_1')(conv22_in)
    bn221 = BatchNormalization()(conv221)
    act221 = Activation('relu')(bn221)
    conv222 = Conv2D(filters[2], 3, padding='same', kernel_initializer='he_normal', name='conv2_2_2')(act221)
    bn222 = BatchNormalization()(conv222)
    conv22 = Activation('relu')(bn222)
    
    
    
    #conv0_3
    conv12_up = UpSampling2D()(conv12)
    conv03_in = Concatenate()([conv00, conv01, conv02, conv12_up])
    
    conv031 = Conv2D(filters[0], 3, padding='same', kernel_initializer='he_normal', name='conv0_3_1')(conv03_in)
    bn031 = BatchNormalization()(conv031)
    act031 = Activation('relu')(bn031)
    conv032 = Conv2D(filters[0], 3, padding='same', kernel_initializer='he_normal', name='conv0_3_2')(act031)
    bn032 = BatchNormalization()(conv032)
    conv03 = Activation('relu')(bn032)
    
    #conv1_3
    conv22_up = UpSampling2D()(conv22)
    conv13_in = Concatenate()([conv10, conv11, conv12, conv22_up])
    
    conv131 = Conv2D(filters[1], 3, padding='same', kernel_initializer='he_normal', name='conv1_3_1')(conv13_in)
    bn131 = BatchNormalization()(conv131)
    act131 = Activation('relu')(bn131)
    conv132 = Conv2D(filters[1], 3, padding='same', kernel_initializer='he_normal', name='conv1_3_2')(act131)
    bn132 = BatchNormalization()(conv132)
    conv13 = Activation('relu')(bn132)
    
    
    
    #conv0_4
    conv13_up = UpSampling2D()(conv13)
    conv04_in = Concatenate()([conv00, conv01, conv02, conv03, conv13_up])
    
    conv041 = Conv2D(filters[0], 3, padding='same', kernel_initializer='he_normal', name='conv0_4_1')(conv04_in)
    bn041 = BatchNormalization()(conv041)
    act041 = Activation('relu')(bn041)
    conv042 = Conv2D(filters[0], 3, padding='same', kernel_initializer='he_normal', name='conv0_4_2')(act041)
    bn042 = BatchNormalization()(conv042)
    conv04 = Activation('relu')(bn042)
    
    output = Conv2D(1, 1, activation='sigmoid', name='final')(conv04)
    
    model = Model(inputs = inputs, outputs = output)
    
    return model

In [28]:

unet_pp = build_UNet_Plus()
unet_pp.summary()

Model: "functional_5"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_3 (InputLayer)            [(None, 224, 224, 3) 0                                            
__________________________________________________________________________________________________
conv0_0_1 (Conv2D)              (None, 224, 224, 16) 448         input_3[0][0]                    
__________________________________________________________________________________________________
batch_normalization_60 (BatchNo (None, 224, 224, 16) 64          conv0_0_1[0][0]                  
__________________________________________________________________________________________________
activation_60 (Activation)      (None, 224, 224, 16) 0           batch_normalization_60[0][0]     
__________________________________________________________________________________________________
conv0_0_2 (Conv2D)              (None, 224, 224, 16) 2320        activation_60[0][0]              
__________________________________________________________________________________________________
batch_normalization_61 (BatchNo (None, 224, 224, 16) 64          conv0_0_2[0][0]                  
__________________________________________________________________________________________________
activation_61 (Activation)      (None, 224, 224, 16) 0           batch_normalization_61[0][0]     
__________________________________________________________________________________________________
max_pooling2d_8 (MaxPooling2D)  (None, 112, 112, 16) 0           activation_61[0][0]              
__________________________________________________________________________________________________
conv1_0_1 (Conv2D)              (None, 112, 112, 32) 4640        max_pooling2d_8[0][0]            
__________________________________________________________________________________________________
batch_normalization_62 (BatchNo (None, 112, 112, 32) 128         conv1_0_1[0][0]                  
__________________________________________________________________________________________________
activation_62 (Activation)      (None, 112, 112, 32) 0           batch_normalization_62[0][0]     
__________________________________________________________________________________________________
conv1_0_2 (Conv2D)              (None, 112, 112, 32) 9248        activation_62[0][0]              
__________________________________________________________________________________________________
batch_normalization_63 (BatchNo (None, 112, 112, 32) 128         conv1_0_2[0][0]                  
__________________________________________________________________________________________________
activation_63 (Activation)      (None, 112, 112, 32) 0           batch_normalization_63[0][0]     
__________________________________________________________________________________________________
max_pooling2d_9 (MaxPooling2D)  (None, 56, 56, 32)   0           activation_63[0][0]              
__________________________________________________________________________________________________
conv2_0_1 (Conv2D)              (None, 56, 56, 64)   18496       max_pooling2d_9[0][0]            
__________________________________________________________________________________________________
batch_normalization_64 (BatchNo (None, 56, 56, 64)   256         conv2_0_1[0][0]                  
__________________________________________________________________________________________________
activation_64 (Activation)      (None, 56, 56, 64)   0           batch_normalization_64[0][0]     
__________________________________________________________________________________________________
conv2_0_2 (Conv2D)              (None, 56, 56, 64)   36928       activation_64[0][0]              
__________________________________________________________________________________________________
batch_normalization_65 (BatchNo (None, 56, 56, 64)   256         conv2_0_2[0][0]                  
__________________________________________________________________________________________________
activation_65 (Activation)      (None, 56, 56, 64)   0           batch_normalization_65[0][0]     
__________________________________________________________________________________________________
max_pooling2d_10 (MaxPooling2D) (None, 28, 28, 64)   0           activation_65[0][0]              
__________________________________________________________________________________________________
conv3_0_1 (Conv2D)              (None, 28, 28, 128)  73856       max_pooling2d_10[0][0]           
__________________________________________________________________________________________________
batch_normalization_66 (BatchNo (None, 28, 28, 128)  512         conv3_0_1[0][0]                  
__________________________________________________________________________________________________
activation_66 (Activation)      (None, 28, 28, 128)  0           batch_normalization_66[0][0]     
__________________________________________________________________________________________________
conv3_0_2 (Conv2D)              (None, 28, 28, 128)  147584      activation_66[0][0]              
__________________________________________________________________________________________________
batch_normalization_67 (BatchNo (None, 28, 28, 128)  512         conv3_0_2[0][0]                  
__________________________________________________________________________________________________
activation_67 (Activation)      (None, 28, 28, 128)  0           batch_normalization_67[0][0]     
__________________________________________________________________________________________________
max_pooling2d_11 (MaxPooling2D) (None, 14, 14, 128)  0           activation_67[0][0]              
__________________________________________________________________________________________________
conv4_0_1 (Conv2D)              (None, 14, 14, 256)  295168      max_pooling2d_11[0][0]           
__________________________________________________________________________________________________
batch_normalization_68 (BatchNo (None, 14, 14, 256)  1024        conv4_0_1[0][0]                  
__________________________________________________________________________________________________
activation_68 (Activation)      (None, 14, 14, 256)  0           batch_normalization_68[0][0]     
__________________________________________________________________________________________________
conv4_0_2 (Conv2D)              (None, 14, 14, 256)  590080      activation_68[0][0]              
__________________________________________________________________________________________________
batch_normalization_69 (BatchNo (None, 14, 14, 256)  1024        conv4_0_2[0][0]                  
__________________________________________________________________________________________________
activation_69 (Activation)      (None, 14, 14, 256)  0           batch_normalization_69[0][0]     
__________________________________________________________________________________________________
up_sampling2d_23 (UpSampling2D) (None, 28, 28, 256)  0           activation_69[0][0]              
__________________________________________________________________________________________________
up_sampling2d_22 (UpSampling2D) (None, 56, 56, 128)  0           activation_67[0][0]              
__________________________________________________________________________________________________
concatenate_23 (Concatenate)    (None, 28, 28, 384)  0           activation_67[0][0]              
                                                                 up_sampling2d_23[0][0]           
__________________________________________________________________________________________________
up_sampling2d_21 (UpSampling2D) (None, 112, 112, 64) 0           activation_65[0][0]              
__________________________________________________________________________________________________
concatenate_22 (Concatenate)    (None, 56, 56, 192)  0           activation_65[0][0]              
                                                                 up_sampling2d_22[0][0]           
__________________________________________________________________________________________________
conv3_1_1 (Conv2D)              (None, 28, 28, 128)  442496      concatenate_23[0][0]             
__________________________________________________________________________________________________
up_sampling2d_20 (UpSampling2D) (None, 224, 224, 32) 0           activation_63[0][0]              
__________________________________________________________________________________________________
concatenate_21 (Concatenate)    (None, 112, 112, 96) 0           activation_63[0][0]              
                                                                 up_sampling2d_21[0][0]           
__________________________________________________________________________________________________
conv2_1_1 (Conv2D)              (None, 56, 56, 64)   110656      concatenate_22[0][0]             
__________________________________________________________________________________________________
batch_normalization_76 (BatchNo (None, 28, 28, 128)  512         conv3_1_1[0][0]                  
__________________________________________________________________________________________________
concatenate_20 (Concatenate)    (None, 224, 224, 48) 0           activation_61[0][0]              
                                                                 up_sampling2d_20[0][0]           
__________________________________________________________________________________________________
conv1_1_1 (Conv2D)              (None, 112, 112, 32) 27680       concatenate_21[0][0]             
__________________________________________________________________________________________________
batch_normalization_74 (BatchNo (None, 56, 56, 64)   256         conv2_1_1[0][0]                  
__________________________________________________________________________________________________
activation_76 (Activation)      (None, 28, 28, 128)  0           batch_normalization_76[0][0]     
__________________________________________________________________________________________________
conv0_1_1 (Conv2D)              (None, 224, 224, 16) 6928        concatenate_20[0][0]             
__________________________________________________________________________________________________
batch_normalization_72 (BatchNo (None, 112, 112, 32) 128         conv1_1_1[0][0]                  
__________________________________________________________________________________________________
activation_74 (Activation)      (None, 56, 56, 64)   0           batch_normalization_74[0][0]     
__________________________________________________________________________________________________
conv3_1_2 (Conv2D)              (None, 28, 28, 128)  147584      activation_76[0][0]              
__________________________________________________________________________________________________
batch_normalization_70 (BatchNo (None, 224, 224, 16) 64          conv0_1_1[0][0]                  
__________________________________________________________________________________________________
activation_72 (Activation)      (None, 112, 112, 32) 0           batch_normalization_72[0][0]     
__________________________________________________________________________________________________
conv2_1_2 (Conv2D)              (None, 56, 56, 64)   36928       activation_74[0][0]              
__________________________________________________________________________________________________
batch_normalization_77 (BatchNo (None, 28, 28, 128)  512         conv3_1_2[0][0]                  
__________________________________________________________________________________________________
activation_70 (Activation)      (None, 224, 224, 16) 0           batch_normalization_70[0][0]     
__________________________________________________________________________________________________
conv1_1_2 (Conv2D)              (None, 112, 112, 32) 9248        activation_72[0][0]              
__________________________________________________________________________________________________
batch_normalization_75 (BatchNo (None, 56, 56, 64)   256         conv2_1_2[0][0]                  
__________________________________________________________________________________________________
activation_77 (Activation)      (None, 28, 28, 128)  0           batch_normalization_77[0][0]     
__________________________________________________________________________________________________
conv0_1_2 (Conv2D)              (None, 224, 224, 16) 2320        activation_70[0][0]              
__________________________________________________________________________________________________
batch_normalization_73 (BatchNo (None, 112, 112, 32) 128         conv1_1_2[0][0]                  
__________________________________________________________________________________________________
activation_75 (Activation)      (None, 56, 56, 64)   0           batch_normalization_75[0][0]     
__________________________________________________________________________________________________
up_sampling2d_26 (UpSampling2D) (None, 56, 56, 128)  0           activation_77[0][0]              
__________________________________________________________________________________________________
batch_normalization_71 (BatchNo (None, 224, 224, 16) 64          conv0_1_2[0][0]                  
__________________________________________________________________________________________________
activation_73 (Activation)      (None, 112, 112, 32) 0           batch_normalization_73[0][0]     
__________________________________________________________________________________________________
up_sampling2d_25 (UpSampling2D) (None, 112, 112, 64) 0           activation_75[0][0]              
__________________________________________________________________________________________________
concatenate_26 (Concatenate)    (None, 56, 56, 256)  0           activation_65[0][0]              
                                                                 activation_75[0][0]              
                                                                 up_sampling2d_26[0][0]           
__________________________________________________________________________________________________
activation_71 (Activation)      (None, 224, 224, 16) 0           batch_normalization_71[0][0]     
__________________________________________________________________________________________________
up_sampling2d_24 (UpSampling2D) (None, 224, 224, 32) 0           activation_73[0][0]              
__________________________________________________________________________________________________
concatenate_25 (Concatenate)    (None, 112, 112, 128 0           activation_63[0][0]              
                                                                 activation_73[0][0]              
                                                                 up_sampling2d_25[0][0]           
__________________________________________________________________________________________________
conv2_2_1 (Conv2D)              (None, 56, 56, 64)   147520      concatenate_26[0][0]             
__________________________________________________________________________________________________
concatenate_24 (Concatenate)    (None, 224, 224, 64) 0           activation_61[0][0]              
                                                                 activation_71[0][0]              
                                                                 up_sampling2d_24[0][0]           
__________________________________________________________________________________________________
conv1_2_1 (Conv2D)              (None, 112, 112, 32) 36896       concatenate_25[0][0]             
__________________________________________________________________________________________________
batch_normalization_82 (BatchNo (None, 56, 56, 64)   256         conv2_2_1[0][0]                  
__________________________________________________________________________________________________
conv0_2_1 (Conv2D)              (None, 224, 224, 16) 9232        concatenate_24[0][0]             
__________________________________________________________________________________________________
batch_normalization_80 (BatchNo (None, 112, 112, 32) 128         conv1_2_1[0][0]                  
__________________________________________________________________________________________________
activation_82 (Activation)      (None, 56, 56, 64)   0           batch_normalization_82[0][0]     
__________________________________________________________________________________________________
batch_normalization_78 (BatchNo (None, 224, 224, 16) 64          conv0_2_1[0][0]                  
__________________________________________________________________________________________________
activation_80 (Activation)      (None, 112, 112, 32) 0           batch_normalization_80[0][0]     
__________________________________________________________________________________________________
conv2_2_2 (Conv2D)              (None, 56, 56, 64)   36928       activation_82[0][0]              
__________________________________________________________________________________________________
activation_78 (Activation)      (None, 224, 224, 16) 0           batch_normalization_78[0][0]     
__________________________________________________________________________________________________
conv1_2_2 (Conv2D)              (None, 112, 112, 32) 9248        activation_80[0][0]              
__________________________________________________________________________________________________
batch_normalization_83 (BatchNo (None, 56, 56, 64)   256         conv2_2_2[0][0]                  
__________________________________________________________________________________________________
conv0_2_2 (Conv2D)              (None, 224, 224, 16) 2320        activation_78[0][0]              
__________________________________________________________________________________________________
batch_normalization_81 (BatchNo (None, 112, 112, 32) 128         conv1_2_2[0][0]                  
__________________________________________________________________________________________________
activation_83 (Activation)      (None, 56, 56, 64)   0           batch_normalization_83[0][0]     
__________________________________________________________________________________________________
batch_normalization_79 (BatchNo (None, 224, 224, 16) 64          conv0_2_2[0][0]                  
__________________________________________________________________________________________________
activation_81 (Activation)      (None, 112, 112, 32) 0           batch_normalization_81[0][0]     
__________________________________________________________________________________________________
up_sampling2d_28 (UpSampling2D) (None, 112, 112, 64) 0           activation_83[0][0]              
__________________________________________________________________________________________________
activation_79 (Activation)      (None, 224, 224, 16) 0           batch_normalization_79[0][0]     
__________________________________________________________________________________________________
up_sampling2d_27 (UpSampling2D) (None, 224, 224, 32) 0           activation_81[0][0]              
__________________________________________________________________________________________________
concatenate_28 (Concatenate)    (None, 112, 112, 160 0           activation_63[0][0]              
                                                                 activation_73[0][0]              
                                                                 activation_81[0][0]              
                                                                 up_sampling2d_28[0][0]           
__________________________________________________________________________________________________
concatenate_27 (Concatenate)    (None, 224, 224, 80) 0           activation_61[0][0]              
                                                                 activation_71[0][0]              
                                                                 activation_79[0][0]              
                                                                 up_sampling2d_27[0][0]           
__________________________________________________________________________________________________
conv1_3_1 (Conv2D)              (None, 112, 112, 32) 46112       concatenate_28[0][0]             
__________________________________________________________________________________________________
conv0_3_1 (Conv2D)              (None, 224, 224, 16) 11536       concatenate_27[0][0]             
__________________________________________________________________________________________________
batch_normalization_86 (BatchNo (None, 112, 112, 32) 128         conv1_3_1[0][0]                  
__________________________________________________________________________________________________
batch_normalization_84 (BatchNo (None, 224, 224, 16) 64          conv0_3_1[0][0]                  
__________________________________________________________________________________________________
activation_86 (Activation)      (None, 112, 112, 32) 0           batch_normalization_86[0][0]     
__________________________________________________________________________________________________
activation_84 (Activation)      (None, 224, 224, 16) 0           batch_normalization_84[0][0]     
__________________________________________________________________________________________________
conv1_3_2 (Conv2D)              (None, 112, 112, 32) 9248        activation_86[0][0]              
__________________________________________________________________________________________________
conv0_3_2 (Conv2D)              (None, 224, 224, 16) 2320        activation_84[0][0]              
__________________________________________________________________________________________________
batch_normalization_87 (BatchNo (None, 112, 112, 32) 128         conv1_3_2[0][0]                  
__________________________________________________________________________________________________
batch_normalization_85 (BatchNo (None, 224, 224, 16) 64          conv0_3_2[0][0]                  
__________________________________________________________________________________________________
activation_87 (Activation)      (None, 112, 112, 32) 0           batch_normalization_87[0][0]     
__________________________________________________________________________________________________
activation_85 (Activation)      (None, 224, 224, 16) 0           batch_normalization_85[0][0]     
__________________________________________________________________________________________________
up_sampling2d_29 (UpSampling2D) (None, 224, 224, 32) 0           activation_87[0][0]              
__________________________________________________________________________________________________
concatenate_29 (Concatenate)    (None, 224, 224, 96) 0           activation_61[0][0]              
                                                                 activation_71[0][0]              
                                                                 activation_79[0][0]              
                                                                 activation_85[0][0]              
                                                                 up_sampling2d_29[0][0]           
__________________________________________________________________________________________________
conv0_4_1 (Conv2D)              (None, 224, 224, 16) 13840       concatenate_29[0][0]             
__________________________________________________________________________________________________
batch_normalization_88 (BatchNo (None, 224, 224, 16) 64          conv0_4_1[0][0]                  
__________________________________________________________________________________________________
activation_88 (Activation)      (None, 224, 224, 16) 0           batch_normalization_88[0][0]     
__________________________________________________________________________________________________
conv0_4_2 (Conv2D)              (None, 224, 224, 16) 2320        activation_88[0][0]              
__________________________________________________________________________________________________
batch_normalization_89 (BatchNo (None, 224, 224, 16) 64          conv0_4_2[0][0]                  
__________________________________________________________________________________________________
activation_89 (Activation)      (None, 224, 224, 16) 0           batch_normalization_89[0][0]     
__________________________________________________________________________________________________
final (Conv2D)                  (None, 224, 224, 1)  17          activation_89[0][0]              
==================================================================================================
Total params: 2,297,441
Trainable params: 2,293,793
Non-trainable params: 3,648
__________________________________________________________________________________________________

In [29]:

gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
  # Restrict TensorFlow to only allocate 1GB of memory on the first GPU
  try:
    tf.config.experimental.set_virtual_device_configuration(
        gpus[0],
        [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=6144)])
    logical_gpus = tf.config.experimental.list_logical_devices('GPU')
    print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs")
  except RuntimeError as e:
    # Virtual devices must be set before GPUs have been initialized
    print(e)

Virtual devices cannot be modified after being initialized

In [30]:

unet_pp.compile(
    optimizer = Adam(lr=1e-4), loss = 'binary_crossentropy')

unet_pp.fit_generator(
    generator=train_generator,
    validation_data=test_generator,
    steps_per_epoch=len(train_generator),
    epochs=100
)

model_path = dir_path + '/unet_pp.h5'
unet_pp.save(model_path)  #학습한 모델을 저장해 주세요.

Epoch 1/100
 2/11 [====>.........................] - ETA: 1s - loss: 0.6303WARNING:tensorflow:Callbacks method `on_train_batch_end` is slow compared to the batch time (batch time: 0.1030s vs `on_train_batch_end` time: 0.2016s). Check your callbacks.
11/11 [==============================] - 3s 307ms/step - loss: 0.5503 - val_loss: 0.6677
Epoch 2/100
11/11 [==============================] - 3s 276ms/step - loss: 0.4594 - val_loss: 0.6182
Epoch 3/100
11/11 [==============================] - 3s 298ms/step - loss: 0.3954 - val_loss: 0.5686
Epoch 4/100
11/11 [==============================] - 3s 303ms/step - loss: 0.3503 - val_loss: 0.5203
Epoch 5/100
11/11 [==============================] - 3s 308ms/step - loss: 0.3402 - val_loss: 0.4828
Epoch 6/100
11/11 [==============================] - 3s 287ms/step - loss: 0.3191 - val_loss: 0.4569
Epoch 7/100
11/11 [==============================] - 3s 285ms/step - loss: 0.2980 - val_loss: 0.4325
Epoch 8/100
11/11 [==============================] - 3s 282ms/step - loss: 0.2806 - val_loss: 0.4140
Epoch 9/100
11/11 [==============================] - 3s 280ms/step - loss: 0.2853 - val_loss: 0.3961
Epoch 10/100
11/11 [==============================] - 3s 278ms/step - loss: 0.2731 - val_loss: 0.3802
Epoch 11/100
11/11 [==============================] - 3s 278ms/step - loss: 0.2588 - val_loss: 0.3646
Epoch 12/100
11/11 [==============================] - 3s 278ms/step - loss: 0.2499 - val_loss: 0.3533
Epoch 13/100
11/11 [==============================] - 3s 278ms/step - loss: 0.2660 - val_loss: 0.3443
Epoch 14/100
11/11 [==============================] - 3s 278ms/step - loss: 0.2470 - val_loss: 0.3355
Epoch 15/100
11/11 [==============================] - 3s 278ms/step - loss: 0.2344 - val_loss: 0.3208
Epoch 16/100
11/11 [==============================] - 3s 278ms/step - loss: 0.2314 - val_loss: 0.3142
Epoch 17/100
11/11 [==============================] - 3s 279ms/step - loss: 0.2332 - val_loss: 0.3098
Epoch 18/100
11/11 [==============================] - 3s 279ms/step - loss: 0.2270 - val_loss: 0.2968
Epoch 19/100
11/11 [==============================] - 3s 279ms/step - loss: 0.2167 - val_loss: 0.2866
Epoch 20/100
11/11 [==============================] - 3s 279ms/step - loss: 0.2057 - val_loss: 0.2806
Epoch 21/100
11/11 [==============================] - 3s 279ms/step - loss: 0.2033 - val_loss: 0.2729
Epoch 22/100
11/11 [==============================] - 3s 279ms/step - loss: 0.2038 - val_loss: 0.2612
Epoch 23/100
11/11 [==============================] - 3s 280ms/step - loss: 0.2130 - val_loss: 0.2573
Epoch 24/100
11/11 [==============================] - 3s 280ms/step - loss: 0.2050 - val_loss: 0.2499
Epoch 25/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1926 - val_loss: 0.2371
Epoch 26/100
11/11 [==============================] - 3s 280ms/step - loss: 0.2019 - val_loss: 0.2375
Epoch 27/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1864 - val_loss: 0.2337
Epoch 28/100
11/11 [==============================] - 3s 281ms/step - loss: 0.1952 - val_loss: 0.2265
Epoch 29/100
11/11 [==============================] - 3s 281ms/step - loss: 0.1903 - val_loss: 0.2213
Epoch 30/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1824 - val_loss: 0.2184
Epoch 31/100
11/11 [==============================] - 3s 281ms/step - loss: 0.2046 - val_loss: 0.2163
Epoch 32/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1974 - val_loss: 0.2095
Epoch 33/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1864 - val_loss: 0.1972
Epoch 34/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1973 - val_loss: 0.2059
Epoch 35/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1790 - val_loss: 0.1989
Epoch 36/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1812 - val_loss: 0.1968
Epoch 37/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1894 - val_loss: 0.1968
Epoch 38/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1747 - val_loss: 0.1938
Epoch 39/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1823 - val_loss: 0.1863
Epoch 40/100
11/11 [==============================] - 3s 283ms/step - loss: 0.1760 - val_loss: 0.1866
Epoch 41/100
11/11 [==============================] - 3s 281ms/step - loss: 0.1796 - val_loss: 0.1852
Epoch 42/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1755 - val_loss: 0.1854
Epoch 43/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1672 - val_loss: 0.1826
Epoch 44/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1736 - val_loss: 0.1813
Epoch 45/100
11/11 [==============================] - 3s 281ms/step - loss: 0.1699 - val_loss: 0.1787
Epoch 46/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1625 - val_loss: 0.1784
Epoch 47/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1547 - val_loss: 0.1751
Epoch 48/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1575 - val_loss: 0.1700
Epoch 49/100
11/11 [==============================] - 3s 281ms/step - loss: 0.1540 - val_loss: 0.1700
Epoch 50/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1532 - val_loss: 0.1638
Epoch 51/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1604 - val_loss: 0.1639
Epoch 52/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1498 - val_loss: 0.1574
Epoch 53/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1477 - val_loss: 0.1588
Epoch 54/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1529 - val_loss: 0.1632
Epoch 55/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1479 - val_loss: 0.1465
Epoch 56/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1482 - val_loss: 0.1425
Epoch 57/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1424 - val_loss: 0.1490
Epoch 58/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1495 - val_loss: 0.1426
Epoch 59/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1472 - val_loss: 0.1400
Epoch 60/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1423 - val_loss: 0.1371
Epoch 61/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1443 - val_loss: 0.1375
Epoch 62/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1391 - val_loss: 0.1337
Epoch 63/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1383 - val_loss: 0.1387
Epoch 64/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1378 - val_loss: 0.1345
Epoch 65/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1371 - val_loss: 0.1349
Epoch 66/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1397 - val_loss: 0.1350
Epoch 67/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1407 - val_loss: 0.1278
Epoch 68/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1363 - val_loss: 0.1247
Epoch 69/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1315 - val_loss: 0.1307
Epoch 70/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1269 - val_loss: 0.1475
Epoch 71/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1359 - val_loss: 0.1284
Epoch 72/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1332 - val_loss: 0.1260
Epoch 73/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1280 - val_loss: 0.1249
Epoch 74/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1276 - val_loss: 0.1217
Epoch 75/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1285 - val_loss: 0.1199
Epoch 76/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1250 - val_loss: 0.1208
Epoch 77/100
11/11 [==============================] - 3s 281ms/step - loss: 0.1249 - val_loss: 0.1369
Epoch 78/100
11/11 [==============================] - 3s 281ms/step - loss: 0.1281 - val_loss: 0.1291
Epoch 79/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1219 - val_loss: 0.1176
Epoch 80/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1208 - val_loss: 0.1223
Epoch 81/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1176 - val_loss: 0.1163
Epoch 82/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1191 - val_loss: 0.1120
Epoch 83/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1195 - val_loss: 0.1099
Epoch 84/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1154 - val_loss: 0.1239
Epoch 85/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1139 - val_loss: 0.1157
Epoch 86/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1127 - val_loss: 0.1090
Epoch 87/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1109 - val_loss: 0.1128
Epoch 88/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1077 - val_loss: 0.1108
Epoch 89/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1094 - val_loss: 0.1043
Epoch 90/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1052 - val_loss: 0.1034
Epoch 91/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1200 - val_loss: 0.1050
Epoch 92/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1042 - val_loss: 0.1074
Epoch 93/100
11/11 [==============================] - 3s 279ms/step - loss: 0.1051 - val_loss: 0.1124
Epoch 94/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1071 - val_loss: 0.1065
Epoch 95/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1094 - val_loss: 0.1014
Epoch 96/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1031 - val_loss: 0.1110
Epoch 97/100
11/11 [==============================] - 3s 280ms/step - loss: 0.1017 - val_loss: 0.1060
Epoch 98/100
11/11 [==============================] - 3s 280ms/step - loss: 0.0974 - val_loss: 0.1054
Epoch 99/100
11/11 [==============================] - 3s 280ms/step - loss: 0.0971 - val_loss: 0.1054
Epoch 100/100
11/11 [==============================] - 3s 280ms/step - loss: 0.0968 - val_loss: 0.1048

In [40]:

def get_output(model, preproc, image_path, output_path):
     origin_img = imread(image_path)
     data = {"image":origin_img}
     processed = preproc(**data)
     output = model(np.expand_dims(processed["image"]/255,axis=0))
     output = (output[0].numpy()>0.5).astype(np.uint8).squeeze(-1)*255  #0.5라는 threshold를 변경하면 도로인식 결과범위가 달라집니다.
     output = Image.fromarray(output)
     background = Image.fromarray(origin_img).convert('RGBA')
     output = output.resize((origin_img.shape[1], origin_img.shape[0])).convert('RGBA')
     output = Image.blend(background, output, alpha=0.5)
     output.show()
     return output
 

# 완성한 뒤에는 시각화한 결과를 눈으로 확인해봅시다!
i = 2    # i값을 바꾸면 테스트용 파일이 달라집니다. 
get_output(
     unet, 
     test_preproc,
     image_path=dir_path + f'/image_2/00{str(i).zfill(4)}_10.png',
     output_path=dir_path + f'./result_{str(i).zfill(3)}.png'
 )

Out[40]:

In [41]:

# 완성한 뒤에는 시각화한 결과를 눈으로 확인해봅시다!
i = 2   # i값을 바꾸면 테스트용 파일이 달라집니다. 
get_output(
     unet_pp, 
     test_preproc,
     image_path=dir_path + f'/image_2/00{str(i).zfill(4)}_10.png',
     output_path=dir_path + f'./result_{str(i).zfill(3)}.png'
 )

Out[41]:

In [42]:

def get_output(model, preproc, image_path, output_path, label_path):
    origin_img = imread(image_path)
    data = {"image":origin_img}
    processed = preproc(**data)
    output = model(np.expand_dims(processed["image"]/255,axis=0))
    output = (output[0].numpy()>=0.5).astype(np.uint8).squeeze(-1)*255  #0.5라는 threshold를 변경하면 도로인식 결과범위가 달라집니다.
    prediction = output/255   # 도로로 판단한 영역
    
    output = Image.fromarray(output)
    background = Image.fromarray(origin_img).convert('RGBA')
    output = output.resize((origin_img.shape[1], origin_img.shape[0])).convert('RGBA')
    output = Image.blend(background, output, alpha=0.5)
    output.show()   # 도로로 판단한 영역을 시각화!
     
    if label_path:   
        label_img = imread(label_path)
        label_data = {"image":label_img}
        label_processed = preproc(**label_data)
        label_processed = label_processed["image"]
        target = (label_processed == 7).astype(np.uint8)*1   # 라벨에서 도로로 기재된 영역

        return output, prediction, target
    else:
        return output, prediction, _

In [43]:

def calculate_iou_score(target, prediction):
    intersection = np.logical_and(target, prediction)
    union = np.logical_or(target, prediction)
    iou_score = float(np.sum(intersection)) / float(np.sum(union))
    print('IoU : %f' % iou_score )
    return iou_score

In [44]:

# 완성한 뒤에는 시각화한 결과를 눈으로 확인해봅시다!
i = 2   # i값을 바꾸면 테스트용 파일이 달라집니다. 
output, prediction, target = get_output(
     unet_pp, 
     test_preproc,
     image_path=dir_path + f'/image_2/00{str(i).zfill(4)}_10.png',
     output_path=dir_path + f'./result_{str(i).zfill(3)}.png',
     label_path=dir_path + f'/semantic/00{str(i).zfill(4)}_10.png'
 )

calculate_iou_score(target, prediction)

IoU : 0.674780

Out[44]:

0.674779843444227

정리¶

• 시각화로 보았을때 U-Net은 자전거를 인식하지못하고 도로로 인식한 반면, U-Net++는 자전거를 인식한 모습을 볼 수 있습니다.
• IoU 수치로 비교해보았을 땐, 1번 파일비교시 U-Net++가 0.12, 2번파일 비교시 0.05정도 성능이 좋아짐을 확인할 수 있었습니다.