Question

我正在尝试使用以下自定义损失函数和指标来训练UNet在keras中进行图像分割：

def dice_coef(y_true, y_pred):
    '''
    Params: y_true -- the labeled mask corresponding to an rgb image
            y_pred -- the predicted mask of an rgb image
    Returns: dice_coeff -- A metric that accounts for precision and recall
                           on the scale from 0 - 1. The closer to 1, the
                           better.
    Citation (MIT License): https://github.com/jocicmarko/
                            ultrasound-nerve-segmentation/blob/
                            master/train.py
    '''
    y_true_f = K.flatten(y_true)
    y_pred_f = K.flatten(y_pred)
    intersection = K.sum(y_true_f * y_pred_f)
    smooth = 1.0
    return (2.0*intersection+smooth)/(K.sum(y_true_f)+K.sum(y_pred_f)+smooth)


def dice_coef_loss(y_true, y_pred):
    '''
    Params: y_true -- the labeled mask corresponding to an rgb image
            y_pred -- the predicted mask of an rgb image
    Returns: 1 - dice_coeff -- a negation of the dice coefficient on
                               the scale from 0 - 1. The closer to 0, the
                               better.
    Citation (MIT License): https://github.com/jocicmarko/
                            ultrasound-nerve-segmentation/blob/
                            master/train.py
    '''
    return 1-dice_coef(y_true, y_pred)

在对加载到RAM中的图像子集进行训练时，它可以正常工作。但是，当使用flow_from_directory和fit_generator从整个数据集中逐步加载一批图像时，我会得到负损失，并且骰子系数大于1。这是生成器的代码：

import keras
from keras.preprocessing.image import ImageDataGenerator

image_datagen = ImageDataGenerator()
mask_datagen = ImageDataGenerator()

seed = 1

image_generator = image_datagen.flow_from_directory(
    "train/flow_from_dir/256/dicom/",
    color_mode="grayscale",
    batch_size=32,
    class_mode=None,
    seed=seed)
mask_generator = image_datagen.flow_from_directory(
    "train/flow_from_dir/256/mask/",
    color_mode="grayscale",
    batch_size=32,
    class_mode=None,
    seed=seed)

train_generator = zip(image_generator, mask_generator)

model.compile(Adam(0.001), loss=unet_utils.dice_coef_loss, metrics=[unet_utils.dice_coef])

history = model.fit_generator(train_generator, steps_per_epoch=10712/32, epochs=300, verbose=1,
                    validation_data=None,
                    callbacks=[PlotLossesKeras()])

Answer 1

由于@today的洞察力，我意识到图像和遮罩都被加载为值为0到255的数组。因此，我添加了预处理功能以对其进行归一化，从而解决了我的问题：

image_datagen = ImageDataGenerator(preprocessing_function=lambda x: x/255)
mask_datagen = ImageDataGenerator(preprocessing_function=lambda x: x/255)

使用flow_from_directory时的荒谬损失和指标值

1 个答案: