python - 从 U-NET 获得的预测值有误差

问题描述

我已经构建了一个 U-Net 架构，用于从时间序列数据中预测一个值。

我的 X_train 大小为 (500,1024)，Y_train 为 (500) 我的 X_train 大小为 (100,1024)，Y_test 为 (100) 当我进行预测时，我得到一个大小为 (1,1024) 的数组。我期望输出有一个值。

我不明白为什么会这样。

def UNetDS64(length, n_channel=1):

    
    x = 64

    inputs = Input((length, n_channel))
    conv1 = Conv1D(x,3, activation='relu', padding='same')(inputs)
    conv1 = BatchNormalization()(conv1)
    conv1 = Conv1D(x,3, activation='relu', padding='same')(conv1)
    conv1 = BatchNormalization()(conv1)
    pool1 = MaxPooling1D(pool_size=2)(conv1)

    conv2 = Conv1D(x*2,3, activation='relu', padding='same')(pool1)
    conv2 = BatchNormalization()(conv2)
    conv2 = Conv1D(x*2,3, activation='relu', padding='same')(conv2)
    conv2 = BatchNormalization()(conv2)
    pool2 = MaxPooling1D(pool_size=2)(conv2)

    conv3 = Conv1D(x*4,3, activation='relu', padding='same')(pool2)
    conv3 = BatchNormalization()(conv3)
    conv3 = Conv1D(x*4,3, activation='relu', padding='same')(conv3)
    conv3 = BatchNormalization()(conv3)
    pool3 = MaxPooling1D(pool_size=2)(conv3)

    conv4 = Conv1D(x*8,3, activation='relu', padding='same')(pool3)
    conv4 = BatchNormalization()(conv4)
    conv4 = Conv1D(x*8,3, activation='relu', padding='same')(conv4)
    conv4 = BatchNormalization()(conv4)
    pool4 = MaxPooling1D(pool_size=2)(conv4)

    conv5 = Conv1D(x*16, 3, activation='relu', padding='same')(pool4)
    conv5 = BatchNormalization()(conv5)
    conv5 = Conv1D(x*16, 3, activation='relu', padding='same')(conv5)
    conv5 = BatchNormalization()(conv5)
    
    level4 = Conv1D(1, 1, name="level4")(conv5)

    up6 = concatenate([UpSampling1D(size=2)(conv5), conv4], axis=2)
    conv6 = Conv1D(x*8, 3, activation='relu', padding='same')(up6)
    conv6 = BatchNormalization()(conv6)
    conv6 = Conv1D(x*8, 3, activation='relu', padding='same')(conv6)
    conv6 = BatchNormalization()(conv6)
    
    level3 = Conv1D(1, 1, name="level3")(conv6)

    up7 = concatenate([UpSampling1D(size=2)(conv6), conv3], axis=2)
    conv7 = Conv1D(x*4, 3, activation='relu', padding='same')(up7)
    conv7 = BatchNormalization()(conv7)
    conv7 = Conv1D(x*4, 3, activation='relu', padding='same')(conv7)
    conv7 = BatchNormalization()(conv7)
    
    level2 = Conv1D(1, 1, name="level2")(conv7)

    up8 = concatenate([UpSampling1D(size=2)(conv7), conv2], axis=2)
    conv8 = Conv1D(x*2, 3, activation='relu', padding='same')(up8)
    conv8 = BatchNormalization()(conv8)
    conv8 = Conv1D(x*2, 3, activation='relu', padding='same')(conv8)
    conv8 = BatchNormalization()(conv8)
    
    level1 = Conv1D(1, 1, name="level1")(conv8)

    up9 = concatenate([UpSampling1D(size=2)(conv8), conv1], axis=2)
    conv9 = Conv1D(x, 3, activation='relu', padding='same')(up9)
    conv9 = BatchNormalization()(conv9)
    conv9 = Conv1D(x, 3, activation='relu', padding='same')(conv9)
    conv9 = BatchNormalization()(conv9)

    out = Dense(1,name="out")(conv9)

    model = Model(inputs=[inputs], outputs=[out])
    
    

    return model
model_dict['UNetDS64'] = UNetDS64

    #model_dict['MultiResUNet1D'] = MultiResUNet1D



    mdlName1 = 'UNetDS64'                                       # approximation network
    mdlName2 = 'MultiResUNet1D'                                 # refinement network
    
    length = 1024                                               # length of the signal

    try:                                                        # create directory to save models
        os.makedirs('models')
    except:
        pass

    try:                                                        # create directory to save training history
        os.makedirs('History')
    except:
        pass

                                                                    # 10 fold cross validation
    for foldname in range(1):

        print('----------------')
        print('Training Fold {}'.format(foldname+1))
        print('----------------')
                                                                                            # loading training data


        #Y_train = prepareLabel(Y_train)                                         # prepare labels for training deep supervision
        
        #Y_val = prepareLabel(Y_val)                                             # prepare labels for training deep supervision
    

    
        mdl1 = model_dict[mdlName1](length)             # create approximation network

                                                                            # loss = mae, with deep supervision weights
        mdl1.compile(loss='mean_absolute_error',optimizer='adam',metrics=['mean_squared_error'])                                                         


        checkpoint1_ = ModelCheckpoint(os.path.join('models','{}_model1_fold{}.h5'.format(mdlName1,foldname)), verbose=1, monitor='val_loss',save_best_only=True, mode='auto')  
                                                                        # train approximation network for 100 epochs
        history1 = mdl1.fit(norm_x_train,norm_y_train,epochs=10,batch_size=32,validation_data=(norm_x_test,norm_y_test),callbacks=[checkpoint1_],verbose=1)

标签： pythonmachine-learningneural-networkartificial-intelligenceunity3d-unet

python - 从 U-NET 获得的预测值有误差

问题描述

解决方案

推荐阅读