keras.engine.training.Model.compile()参数-- metric= = dice_loss -- dice_loss不同

Question

我用keras编写ResUnet模型，但是当我训练模型时，我使用以下代码

m.compile(optimizer= sgd, loss = Dice_coef_loss, metrics=[Dice_coef, Dice_coef_loss])显示不同。

下面的代码：

G  = MyResUnet.MyLiverDataGenerator.imageSegmentationGenerator( tarin_image_xmls ,  train_batch_size,  n_classes , 
input_height , input_width , input_channel, output_height , output_width , mode ='train')
sgd =keras.optimizers.SGD(lr=0.001, momentum=0.9, decay=0.0005, nesterov=True)

m = Models.ResUnet.ResUnet( n_classes , input_height=input_height, 
input_width=input_width  ,input_channel=input_channel )  # 这边没问题
m.summary()



m.compile(optimizer= sgd, loss = Dice_coef_loss, metrics=[Dice_coef, Dice_coef_loss])
m.fit_generator(G, samples_per_epoch = steps_per_epochs , epochs = 30,verbose = 1,
                # validation_data = G_valid, nb_val_samples = steps_per_epochs_valid,
                callbacks =  [modelCheckpoint,tensorboard],initial_epoch=0)

代码可以运行和训练，但显示不正确。[设置batch_size =1]

Epoch 1/30

    1/37739 [..............................] - ETA: 94:40:45 - loss: 2.6937 - Dice_coef: 0.1626 - Dice_coef_loss: 0.8374
    2/37739 [..............................] - ETA: 48:02:00 - loss: 2.6776 - Dice_coef: 0.1787 - Dice_coef_loss: 0.8213
    3/37739 [..............................] - ETA: 32:27:13 - loss: 2.7020 - Dice_coef: 0.1543 - Dice_coef_loss: 0.8457
    4/37739 [..............................] - ETA: 24:39:58 - loss: 2.7336 - Dice_coef: 0.1227 - Dice_coef_loss: 0.8773
    5/37739 [..............................] - ETA: 19:59:52 - loss: 2.7566 - Dice_coef: 0.0997 - Dice_coef_loss: 0.9003
    6/37739 [..............................] - ETA: 16:53:33 - loss: 2.7729 - Dice_coef: 0.0834 - Dice_coef_loss: 0.9166
    7/37739 [..............................] - ETA: 14:40:23 - loss: 2.7419 - Dice_coef: 0.1144 - Dice_coef_loss: 0.8856
    8/37739 [..............................] - ETA: 12:59:53 - loss: 2.7560 - Dice_coef: 0.1003 - Dice_coef_loss: 0.8997

损失需要与dice_coef_loss相同！

Answer 1

我找到了解决方案的问题。我让所有的conv添加kernel_regularizer = l2(0.001)，但是dice_loss不是1，最终限制为0，kernel_regularizer会影响这个