具有不同输入的集成模型（预计会看到 2 个数组）

Question

我训练了 2 个模型。

第一个模型是 UNet：

print(model_unet.summary())

__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_4 (InputLayer)            (None, 128, 128, 1)  0                                            
__________________________________________________________________________________________________
conv2d_26 (Conv2D)              (None, 128, 128, 32) 320         input_4[0][0]                    
__________________________________________________________________________________________________
conv2d_27 (Conv2D)              (None, 128, 128, 32) 9248        conv2d_26[0][0]  
.....
.....
conv2d_44 (Conv2D)              (None, 128, 128, 1)  33          zero_padding2d_4[0][0]           
==================================================================================================
Total params: 7,846,081
Trainable params: 7,846,081
Non-trainable params: 0

其次是 ResNet：

print(model_resnet.summary())

__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_3 (InputLayer)            (None, 128, 128, 3)  0                                            
__________________________________________________________________________________________________
conv1_pad (ZeroPadding2D)       (None, 134, 134, 3)  0           input_3[0][0]                    
....
....
conv2d_25 (Conv2D)              (None, 128, 128, 3)  99          zero_padding2d_3[0][0]           
==================================================================================================
Total params: 24,186,915
Trainable params: 24,133,795
Non-trainable params: 53,120

UNet 有 1 个通道（灰色），ResNet 有 3 个通道。

然后，我正在尝试创建一个集成模型：

def ensemble(models, models_input):

    outputs = [model(models_input[idx]) for idx, model in enumerate(models)]
    x = Average()(outputs)

    model_inputs = [model for model in models_input]
    model = Model(model_inputs, x)

    return model

models = [model_unet, model_resnet]
models_input = [Input((128,128,1)), Input((128,128, 3))]

ensemble_model = ensemble(models, models_input)

当我尝试预测验证数据时：

pred_val = ensemble_model.predict(X_val)

我收到错误消息：

Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 2 array(s), but instead got the following list of 1 arrays: [array([[[[0.46755977],
         [0.52268691],
         [0.52766109],
         ....

X_val.shape is : (800, 128, 128, 1)

我认为问题出在渠道上，但我不知道如何解决这个问题。

Answer 1

如果你的训练数据是灰度图像并且考虑到你的 ResNet 模型将输入 RGB 图像，那么你应该问问自己你想如何从灰度到 RGB？一种答案是将灰度图像重复 3 次以获得 RBG 图像。然后，您可以轻松地定义一个模型，其中一个输入层获取您的灰度图像并将它们相应地提供给您定义的模型：

from keras import backend as K

input_image = Input(shape=(128,128,1))

unet_out = model_unet(input_image)
rgb_image = Lambda(lambda x: K.repeat_elements(x, 3, -1))(input_image)
resnet_out = model_resnet(rgb_image)

output = Average()([unet_out, resnet_out])

ensemble_model = Model(input_image, output)

然后你可以很容易地用一个输入数组调用predict：

pred_val = ensemble_model.predict(X_val)

此解决方案的一种替代方法是使用您在问题中使用的解决方案。但是，您首先需要将图像从灰度转换为 RGB，然后将两个数组传递给 predict 方法：

X_val_rgb = np.repeat(X_val, 3, -1)

pred_val = ensemble_model.predict([X_val, X_val_rgb])

具有不同输入的集成模型（预计会看到 2 个数组）

ensemble model with different inputs (Expected to see 2 array(s))

python

machine-learning

deep-learning

ensemble-learning

keras