Question

我尝试使用U-net进行语义分割问题。遮罩图像是二进制的。但是当训练时，我发现我的损失是负数。在这里我用binary_crossentropy损失。这是我的代码：

X_train = X_train /255
y_train = y_train /255
X_val = X_val/255
y_val = y_val/255

所有类型均为np.float32

然后我使用imageDataGenerator扩展图像，代码如下：

def image_augmentation(X_train,y_train):
    # Set hyper parameters for the model.
    data_gen_args = dict(featurewise_center=True,
                         featurewise_std_normalization=True,
                         rotation_range=90.,
                         width_shift_range=0.1,
                         height_shift_range=0.1,
                         zoom_range=0.2,
                         horizontal_flip=True, 
                         vertical_flip=True)
    image_datagen = ImageDataGenerator(**data_gen_args)
    mask_datagen = ImageDataGenerator(**data_gen_args)

    seed = 42
    image_datagen.fit(X_train, augment=True, seed=seed)
    mask_datagen.fit(y_train, augment=True, seed=seed)

    image_generator = image_datagen.flow(
                         X_train,batch_size=8,
                         seed=seed)

    mask_generator = mask_datagen.flow(
                         y_train, batch_size=8,
                         seed=seed)

    while True:
         yield(image_generator.next(),mask_generator.next())


train_generator = image_augmentation(X_train,y_train)

pat_init = 50
pat = pat_init
learning_rate = 1e-4
##change the model weight you want
file_path = "./model_v1/improvement-{epoch:02d}-{val_my_iou_metric:.5f}.hdf5"
checkpoint = ModelCheckpoint(file_path,monitor = 'val_my_iou_metric',verbose=1,save_best_only=True,mode='max')
reduce_lr = ReduceLROnPlateau(monitor='val_loss', mode = 'auto',factor=0.5, patience=5, min_lr=1e-9, verbose=1)
model.compile(loss='binary_crossentropy', optimizer=Adam(lr=learning_rate), metrics=[my_iou_metric])

# Use the image data Augment below to achieve better result
model.fit_generator(
        train_generator,steps_per_epoch=2000,epochs=300,
        validation_data=(X_val, y_val), verbose=1,
        callbacks=[checkpoint,reduce_lr]
        )

“我的网络”的最后一层定义如下：

output = Conv2D(1,activation='sigmoid',
                            kernel_size=(1,1), 
                            padding='same',
                            data_format='channels_last')(x)

我真的很好奇为什么会这样？ sigmoid函数的输出在0到1之间吗？

如果您有任何想法，请与我讨论。非常感谢！

Answer 1

samplewise_center=True,
samplewise_std_normalization=True

在imagedatagenerator中

即使我对输入进行归一化，binary_corssentropy损失为负

1 个答案: