深度学习--使用自制数据在改进Unet上训练时loss，accuracy不变

Question

训练数据用的是自制数据，如下所示，左图是输入数据，右图是输出标签

优化器选择adam，损失函数选择MSE，训练数据只有15张图片，测试图片有2张。

刚开始选择标准Unet训练数据，没有任何问题，后来在每层Unet中加入ResNet后loss从训练一开始就开始震荡。学习速率从1e-5~0.1之间都试过，batchsize从4~7之间也试过，但这个问题一直存在。

loss、accuracy都一直在震荡，情况如下图所示

想请教大佬们这种情况该怎么解决？

加入ResNet的网络结构代码如下

def Res_block(cov_input, filter_size):

    con_shortcut = Conv2D(filter_size, (1, 1), padding = 'same', kernel_initializer = 'he_normal')(cov_input)

    con1 = Conv2D(filter_size, (1, 1), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(cov_input)
    con1 = BatchNormalization()(con1)
    con2 = Conv2D(filter_size, (3, 3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(con1)
    con2 = BatchNormalization()(con2)
    con3 = Conv2D(filter_size, (1, 1), padding = 'same', kernel_initializer = 'he_normal')(con2)

    con = Add()([con3, con_shortcut])
    con = Activation("relu")(con)
    cov_out = BatchNormalization()(con)

    return cov_out

def Unet(input_height = 224, input_width = 224, loss_names = "binary_crossentropy", optimizer_names = "adam"):

    main_inputs = Input(shape = (input_height, input_width, 6))

    # encode

     # 第一层
    conv1 = Res_block(main_inputs, 32)

    conv1 = Res_block(conv1, 32)

    pool1 = MaxPooling2D(pool_size = (2, 2), strides = (2, 2))(conv1)

    # 第二层
    conv2 = Res_block(pool1, 64)

    conv2 = Res_block(conv2, 64)

    pool2 = MaxPooling2D(pool_size = (2, 2), strides = (2, 2))(conv2)

    # 第三层
    conv3 = Res_block(pool2, 128)

    conv3 = Res_block(conv3, 128)

    pool3 = MaxPooling2D(pool_size = (2, 2), strides = (2, 2))(conv3)

    # 第四层
    conv4 = Res_block(pool3, 256)

    conv4 = Res_block(conv4, 256)

    pool4 = MaxPooling2D(pool_size = (2, 2), strides = (2, 2))(conv4)

    # 第五层
    conv5 = Res_block(pool4, 512)

    conv5 = Res_block(conv5, 512)

    pool5 = MaxPooling2D(pool_size = (2, 2), strides = (2, 2))(conv5)

    # 第六层
    conv6 = Res_block(pool5, 1024)

    conv6 = Res_block(conv6, 1024)

    # decode

    # 第七层
    up7 = (UpSampling2D((2, 2)))(conv6)
    up7 = Conv2D(512, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(up7)
    conv7 = concatenate([up7, conv5], axis = -1)

    conv7 = Res_block(conv7, 512)

    conv7 = Res_block(conv7, 512)

    # 第八层
    up8 = (UpSampling2D((2, 2)))(conv7)
    up8 = Conv2D(256, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(up8)
    conv8 = concatenate([up8, conv4], axis = -1)

    conv8 = Res_block(conv8, 256)

    conv8 = Res_block(conv8, 256)

    # 第九层
    up9 = (UpSampling2D((2, 2)))(conv8)
    up9 = Conv2D(128, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(up9)
    conv9 = concatenate([up9, conv3], axis = -1)

    conv9 = Res_block(conv9, 128)

    conv9 = Res_block(conv9, 128)

    # 第十层
    up10 = (UpSampling2D((2, 2)))(conv9)
    up10 = Conv2D(64, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(up10)
    conv10 = concatenate([up10, conv2], axis = -1)

    conv10 = Res_block(conv10, 64)

    conv10 = Res_block(conv10, 64)

    # 第十一层
    up11 = (UpSampling2D((2, 2)))(conv10)
    up11 = Conv2D(32, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(up11)
    conv11 = (concatenate([up11, conv1], axis = -1))

    conv11 = Res_block(conv11, 32)

    conv11 = Res_block(conv11, 32)

    # 输出层
    o = Conv2D(32, (3, 3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv11)
    o = BatchNormalization()(o)

    o = Conv2D(32, (3, 3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(o)
    o = BatchNormalization()(o)

    o = Conv2D(16, (3, 3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(o)
    o = BatchNormalization()(o)

    o = Conv2D(16, (3, 3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(o)
    o = BatchNormalization()(o)

    o = Conv2D(8, (3, 3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(o)
    o = Conv2D(4, (3, 3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(o)
    o = Conv2D(2, (3, 3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(o)
    o = Conv2D(1, (3, 3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(o)
    o = Conv2D(1, (1, 1), activation = 'relu', kernel_initializer = 'he_normal')(o)

    model = Model(inputs = main_inputs, outputs = o)

Answer 1

说明你的数据样本不足，或者网络模型太复杂，学不动了，导致反复在一个值附近振荡。