在训练图像分割网络UNET期间，验证损失并没有减少，但保持不变

问题描述

我正在训练来自figshare的脑肿瘤数据集上的UNET图像分割网络，训练得很好，训练损失和训练骰子得分都在相应地改变，或者在具有相同的语气，具有验证丢失和验证骰子得分。毫无疑问过度拟合。但是在大约 40个时代之后，绩效指标均未得到改善。它在损失0.58和骰子得分0.47 左右切换。 该如何解决？请建议我。 以下是我的UNET网络-

def unet(pretrained_weights = None,input_size = (512,512,3)):
    inputs = Input(input_size)
    
    conv1 = Convolution2D(64,3,activation = 'relu',padding = 'same',kernel_initializer = 'he_normal')(inputs)
    conv1 = BatchNormalization()(conv1)
    #conv1 = Dropout(0.2)(conv1)
    conv1 = Convolution2D(64,kernel_initializer = 'he_normal')(conv1)
    conv1 = BatchNormalization()(conv1)
    #conv1 = Dropout(0.2)(conv1)
    
    pool1 = MaxPooling2D(pool_size=(2,2))(conv1)
    
    conv2 = Convolution2D(128,kernel_initializer = 'he_normal')(pool1)
    conv2 = BatchNormalization()(conv2)
    #conv2 = Dropout(0.1)(conv2)
    conv2 = Convolution2D(128,kernel_initializer = 'he_normal')(conv2)
    conv2 = BatchNormalization()(conv2)
    #conv2 = Dropout(0.1)(conv2)
    
    pool2 = MaxPooling2D(pool_size=(2,2))(conv2)
    
    conv3 = Convolution2D(256,kernel_initializer = 'he_normal')(pool2)
    conv3 = BatchNormalization()(conv3)
    #conv3 = Dropout(0.1)(conv3)
    conv3 = Convolution2D(256,kernel_initializer = 'he_normal')(conv3)
    conv3 = BatchNormalization()(conv3)
    #conv3 = Dropout(0.1)(conv3)
    
    pool3 = MaxPooling2D(pool_size=(2,2))(conv3)
    
    conv4 = Convolution2D(512,kernel_initializer = 'he_normal')(pool3)
    conv4 = BatchNormalization()(conv4)
    #conv4 = Dropout(0.1)(conv4)
    conv4 = Convolution2D(512,kernel_initializer = 'he_normal')(conv4)
    conv4 = BatchNormalization()(conv4)
    #conv4 = Dropout(0.5)(conv4)
        
    pool4 = MaxPooling2D(pool_size=(2,2))(conv4)

    conv5 = Convolution2D(1024,kernel_initializer = 'he_normal')(pool4)
    conv5 = BatchNormalization()(conv5)
    #conv5 = Dropout(0.1)(conv5)
    conv5 = Convolution2D(1024,kernel_initializer = 'he_normal')(conv5)
    conv5 = BatchNormalization()(conv5)
    #conv5 = Dropout(0.5)(conv5)
    

    up6 = Convolution2D(512,2,kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv5))
    merge6 = concatenate([conv4,up6],axis = 3)
    
    conv6 = Convolution2D(512,kernel_initializer = 'he_normal')(merge6)
    conv6 = BatchNormalization()(conv6)
    #conv6 = Dropout(0.1)(conv6)
    conv6 = Convolution2D(512,kernel_initializer = 'he_normal')(conv6)
    conv6 = BatchNormalization()(conv6)
    #conv6 = Dropout(0.1)(conv6)

    up7 = Convolution2D(256,2))(conv6))
    merge7 = concatenate([conv3,up7],axis = 3)
    
    conv7 = Convolution2D(256,kernel_initializer = 'he_normal')(merge7)
    conv7 = BatchNormalization()(conv7)
    #conv7 = Dropout(0.1)(conv7)
    conv7 = Convolution2D(256,kernel_initializer = 'he_normal')(conv7)
    conv7 = BatchNormalization()(conv7)
    #conv7 = Dropout(0.1)(conv7)

    up8 = Convolution2D(128,2))(conv7))
    merge8 = concatenate([conv2,up8],axis = 3)
    
    conv8 = Convolution2D(128,kernel_initializer = 'he_normal')(merge8)
    conv8 = BatchNormalization()(conv8)
    #conv8 = Dropout(0.1)(conv8)
    conv8 = Convolution2D(128,kernel_initializer = 'he_normal')(conv8)
    conv8 = BatchNormalization()(conv8)
    #conv8 = Dropout(0.1)(conv8)

    up9 = Convolution2D(64,2))(conv8))
    merge9 = concatenate([conv1,up9],axis = 3)
    
    conv9 = Convolution2D(64,kernel_initializer = 'he_normal')(merge9)
    conv9 = BatchNormalization()(conv9)
    #conv9 = Dropout(0.2)(conv9)
    conv9 = Convolution2D(64,kernel_initializer = 'he_normal')(conv9)
    conv9 = BatchNormalization()(conv9)
    #conv9 = Dropout(0.2)(conv9)
    
    conv9 = Convolution2D(2,kernel_initializer = 'he_normal')(conv9)
    conv9 = BatchNormalization()(conv9)
    #conv9 = Dropout(0.2)(conv9)
    
    conv10 = Convolution2D(1,1,activation = 'sigmoid')(conv9)

    model = Model(input = inputs,output = conv10)

      
    #model.summary()

    if(pretrained_weights):
        model.load_weights(pretrained_weights)

    return model

回调详细信息已初始化。凝视LR = 1e-4

callbacks = [EarlyStopping(monitor='val_loss',mode="min",patience=30,verbose=1,min_delta=1e-4),ReduceLROnPlateau(monitor='val_loss',factor=0.1,patience=8,verbose=1),ModelCheckpoint(monitor='val_loss',filepath='weights/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-{epoch:03d}-{dice_coef:.6f}--{val_loss:.6f}.hdf5',save_weights_only=True,CSVLogger('weights/anmol/1/UNET_mixed_loss_monitor_DC_new.csv')]

我用户定义的骰子得分和损失函数。我在这里用过dice_coef_loss。

    def dice_coef(y_true,y_pred,smooth=1):
        y_true_f = K.flatten(y_true)
        y_pred_f = K.flatten(y_pred)
        intersection = K.sum(y_true_f * y_pred_f)
        return (2. * intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)
    
    def dice_loss(y_true,y_pred):
        loss = 1 - dice_coef(y_true,y_pred)
        return loss
    
    def dice_coef_loss(y_true,y_pred):
        return binary_crossentropy(y_true,y_pred) + dice_loss(y_true,y_pred)

使用2605张图像进行训练，使用306张图像进行验证。 Train_image Train_mask

    img_size = 512

image_args = dict(seed=seed,batch_size=2,shuffle=True,class_mode=None,target_size=(img_size,img_size),color_mode='rgb')

mask_args = dict(seed=seed,color_mode='grayscale')

DIR = 'raw/brain/'
image = 'images'
masks = 'masks'

# combine generators into one which yields image and masks
train_generator = zip(image_datagen.flow_from_directory(**image_args,directory=DIR+'train_'+image),mask_datagen.flow_from_directory(**mask_args,directory=DIR+'train_'+masks))

validation_generator = zip(image_datagen.flow_from_directory(**image_args,directory=DIR+'validation_'+image),directory=DIR+'validation_'+masks))

model.fit_generator(train_generator,steps_per_epoch=1302,epochs=100,validation_data=validation_generator,validation_steps=153,callbacks=callbacks)

一些培训日志如下所示

Epoch 00041：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-041-0.466533--0.511900.hdf5 时代42/100 1302/1302 [==============================]-1063s 817ms / step-损失：0.5939-dice_coef：0.4658-val_loss ：0.5076-val_dice_coef：0.5430

Epoch 00042：将模型保存到权重/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-042-0.465990--0.507603.hdf5 时代43/100 1302/1302 [==============================]-1062s 816ms / step-损失：0.5928-dice_coef：0.4678-val_loss ：0.5191-val_dice_coef：0.5270

Epoch 00043：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-043-0.467685--0.519115.hdf5 时代44/100 1302/1302 [==============================]-1063s 817ms / step-损失：0.5966-dice_coef：0.4632-val_loss ：0.5158-val_dice_coef：0.5364

Epoch 00044：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-044-0.463308--0.515760.hdf5 时代45/100 1302/1302 [==============================]-1064s 817ms / step-损失：0.5892-dice_coef：0.4702-val_loss ：0.4993-val_dice_coef：0.5507

Epoch 00045：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-045-0.470134--0.499294.hdf5 时代46/100 1302/1302 [==============================]-1063s 816ms / step-损失：0.5960-dice_coef：0.4636-val_loss ：0.5166-val_dice_coef：0.5329

Epoch 00046：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-046-0.463810--0.516552.hdf5 时代47/100 1302/1302 [==============================]-1065s 818ms / step-损失：0.5920-dice_coef：0.4672-val_loss ：0.5062-val_dice_coef：0.5427

Epoch 00047：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-047-0.467146--0.506242.hdf5 时代48/100 1302/1302 [==============================]-1063s 816ms / step-损失：0.5938-dice_coef：0.4657-val_loss ：0.5239-val_dice_coef：0.5277

Epoch 00048：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-048-0.465866--0.523923.hdf5 时代49/100 1302/1302 [==============================]-1064s 817ms / step-损失：0.5962-dice_coef：0.4639-val_loss ：0.5035-val_dice_coef：0.5434

Epoch 00049：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-049-0.463924--0.503518.hdf5 时代50/100 1302/1302 [==============================]-1063s 816ms / step-损失：0.5854-dice_coef：0.4743-val_loss ：0.5463-val_dice_coef：0.5066

Epoch 00050：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-050-0.474530--0.546343.hdf5 时代51/100 1302/1302 [==============================]-1063s 816ms / step-损失：0.5840-dice_coef：0.4749-val_loss ：0.5146-val_dice_coef：0.5360

Epoch 00051：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-051-0.475072--0.514581.hdf5 时代52/100 1302/1302 [==============================]-1064s 817ms / step-损失：0.5852-dice_coef：0.4742-val_loss ：0.5257-val_dice_coef：0.5256

Epoch 00052：将模型保存到权重/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-052-0.474234--0.525729.hdf5 时代53/100 1302/1302 [==============================]-1065s 818ms / step-损失：0.5857-dice_coef：0.4736-val_loss ：0.5157-val_dice_coef：0.5315

Epoch 00053：ReduceLROnPlateau将学习率降低到9.999999747378752e-07。

Epoch 00053：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-053-0.473557--0.515651.hdf5 时代54/100 1302/1302 [==============================]-1065s 818ms / step-损失：0.5852-dice_coef：0.4737-val_loss ：0.5067-val_dice_coef：0.5421

Epoch 00054：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-054-0.473682--0.506671.hdf5 时代55/100 1302/1302 [==============================]-1065s 818ms / step-损失：0.5903-dice_coef：0.4696-val_loss ：0.4910-val_dice_coef：0.5571

Epoch 00055：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-055-0.469478--0.491024.hdf5 时代56/100 1302/1302 [==============================]-1065s 818ms / step-损失：0.5876-dice_coef：0.4711-val_loss ：0.5154-val_dice_coef：0.5340

Epoch 00056：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-056-0.471110--0.515441.hdf5 时代57/100 1302/1302 [==============================]-1064s 817ms / step-损失：0.5897-dice_coef：0.4703-val_loss ：0.5263-val_dice_coef：0.5258

Epoch 00057：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-057-0.470255--0.526310.hdf5 时代58/100 1302/1302 [==============================]-1064s 817ms / step-损失：0.5849-dice_coef：0.4741-val_loss ：0.5067-val_dice_coef：0.5451

Epoch 00058：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-058-0.474262--0.506664.hdf5 时代59/100 1302/1302 [==============================]-1062s 816ms / step-损失：0.5816-dice_coef：0.4769-val_loss ：0.5160-val_dice_coef：0.5348

Epoch 00059：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-059-0.476830--0.516005.hdf5 时代60/100 1302/1302 [==============================]-1064s 817ms / step-损失：0.5891-dice_coef：0.4709-val_loss ：0.5179-val_dice_coef：0.5318

Epoch 00060：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-060-0.470746--0.517893.hdf5 时代61/100 1302/1302 [==============================]-1065s 818ms / step-损失：0.5873-dice_coef：0.4727-val_loss ：0.5064-val_dice_coef：0.5431

Epoch 00061：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-061-0.472722--0.506373.hdf5 时代62/100 1302/1302 [==============================]-1064s 817ms / step-损失：0.5803-dice_coef：0.4793-val_loss ：0.5187-val_dice_coef：0.5319

Epoch 00062：将模型保存到权重/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-062-0.479199--0.518674.hdf5 时代63/100 1302/1302 [==============================]-1066s 819ms / step-损失：0.5843-dice_coef：0.4738-val_loss ：0.5052-val_dice_coef：0.5459

Epoch 00063：ReduceLROnPlateau将学习率降低到9.999999974752428e-08。

Epoch 00063：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-063-0.473731--0.505171.hdf5 时代64/100 1302/1302 [==============================]-1065s 818ms / step-损失：0.5859-dice_coef：0.4731-val_loss ：0.5064-val_dice_coef：0.5419

Epoch 00064：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-064-0.473008--0.506380.hdf5 时代65/100 1302/1302 [==============================]-1064s 817ms / step-损失：0.5836-dice_coef：0.4752-val_loss ：0.4997-val_dice_coef：0.5508

Epoch 00065：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-065-0.475424--0.499673.hdf5 时代66/100 1302/1302 [==============================]-1063s 817ms / step-损失：0.5932-dice_coef：0.4660-val_loss ：0.5168-val_dice_coef：0.5338

Epoch 00066：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-066-0.465829--0.516758.hdf5 时代67/100 1302/1302 [==============================]-1063s 816ms / step-损失：0.5820-dice_coef：0.4765-val_loss ：0.5179-val_dice_coef：0.5323

Epoch 00067：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-067-0.476715--0.517926.hdf5 时代68/100 1302/1302 [==============================]-1062s 816ms / step-损失：0.5912-dice_coef：0.4689-val_loss ：0.5125-val_dice_coef：0.5375

Epoch 00068：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-068-0.468950--0.512456.hdf5 时代69/100 1302/1302 [==============================]-1062s 816ms / step-损失：0.5820-dice_coef：0.4769-val_loss ：0.5282-val_dice_coef：0.5237

Epoch 00069：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-069-0.476976--0.528154.hdf5 时代70/100 1302/1302 [==============================]-1062s 816ms / step-损失：0.5845-dice_coef：0.4743-val_loss ：0.5204-val_dice_coef：0.5303

Epoch 00070：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-070-0.474195--0.5203​​56.hdf5 时代71/100 1302/1302 [==============================]-1063s 816ms / step-损失：0.5886-dice_coef：0.4708-val_loss ：0.5230-val_dice_coef：0.5270

Epoch 00071：ReduceLROnPlateau将学习率降低到1.0000000116860975e-08。

Epoch 00071：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-071-0.470715--0.523011.hdf5 时代72/100 1302/1302 [==============================]-1062s 816ms / step-损失：0.5837-dice_coef：0.4759-val_loss ：0.5216-val_dice_coef：0.5303

Epoch 00072：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-072-0.475787--0.521631.hdf5 时代73/100 1302/1302 [==============================]-1062s 815ms / step-损失：0.5804-dice_coef：0.4780-val_loss ：0.5333-val_dice_coef：0.5171

Epoch 00073：将模型保存到权重/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-073-0.478063--0.533321.hdf5 时代74/100 1302/1302 [==============================]-1065s 818ms / step-损失：0.5842-dice_coef：0.4747-val_loss ：0.5126-val_dice_coef：0.5393

Epoch 00074：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-074-0.474628--0.512649.hdf5 纪元75/100 1302/1302 [==============================]-1069s 821ms / step-损失：0.5836-dice_coef：0.4755-val_loss ：0.5103-val_dice_coef：0.5386

Epoch 00075：将模型保存到砝码/anmol/1/UNET_sigmoid_focus_DC_2605_R_B_t-075-0.475690--0.510267.hdf5 时代76/100 160/1302 [==> ...............................]-预计到达时间：15：02-损失：0.6069-骰子系数：0.4548

解决方法

暂无找到可以解决该程序问题的有效方法，小编努力寻找整理中！

如果你已经找到好的解决方法，欢迎将解决方案带上本链接一起发送给小编。

小编邮箱:dio#foxmail.com (将#修改为@）