来自keras_contrib问题的Densenet

Question

我正在尝试将keras_contrib中的Densenet用于我自己的具有维度(30k,2,96,96)的数据。

我的shape数据不能使用这个实现吗？它会给出以下错误和警告。

    Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            (None, 96, 96, 2)    0                                            
__________________________________________________________________________________________________
initial_conv2D (Conv2D)         (None, 96, 96, 16)   288         input_1[0][0]                    
__________________________________________________________________________________________________
dense_0_0_bn (BatchNormalizatio (None, 96, 96, 16)   64          initial_conv2D[0][0]             
__________________________________________________________________________________________________
activation_1 (Activation)       (None, 96, 96, 16)   0           dense_0_0_bn[0][0]               
__________________________________________________________________________________________________
dense_0_0_conv2D (Conv2D)       (None, 96, 96, 4)    576         activation_1[0][0]               
__________________________________________________________________________________________________
concatenate_1 (Concatenate)     (None, 96, 96, 20)   0           initial_conv2D[0][0]             
                                                                 dense_0_0_conv2D[0][0]           
__________________________________________________________________________________________________
final_bn (BatchNormalization)   (None, 96, 96, 20)   80          concatenate_1[0][0]              
__________________________________________________________________________________________________
activation_2 (Activation)       (None, 96, 96, 20)   0           final_bn[0][0]                   
__________________________________________________________________________________________________
dense_1 (Dense)                 (None, 96, 96, 2)    42          activation_2[0][0]               
==================================================================================================
Total params: 1,050
Trainable params: 978
Non-trainable params: 72
__________________________________________________________________________________________________
Finished compiling
/home/arka/anaconda2/envs/hyperas/lib/python3.6/site-packages/keras_preprocessing/image.py:1213: UserWarning: Expected input to be images (as Numpy array) following the data format convention "channels_last" (channels on axis 3), i.e. expected either 1, 3 or 4 channels on axis 3. However, it was passed an array with shape (39840, 96, 96, 2) (2 channels).
  ' channels).')
/home/arka/anaconda2/envs/hyperas/lib/python3.6/site-packages/keras_preprocessing/image.py:1437: UserWarning: NumpyArrayIterator is set to use the data format convention "channels_last" (channels on axis 3), i.e. expected either 1, 3, or 4 channels on axis 3. However, it was passed an array with shape (39840, 96, 96, 2) (2 channels).
  str(self.x.shape[channels_axis]) + ' channels).')
Traceback (most recent call last):
  File "keras_densenet.py", line 149, in <module>
    fit_model(X_train,y_train,X_val,y_val)
  File "keras_densenet.py", line 140, in fit_model
    verbose=2)
  File "/home/arka/anaconda2/envs/hyperas/lib/python3.6/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)
  File "/home/arka/anaconda2/envs/hyperas/lib/python3.6/site-packages/keras/engine/training.py", line 1415, in fit_generator
    initial_epoch=initial_epoch)
  File "/home/arka/anaconda2/envs/hyperas/lib/python3.6/site-packages/keras/engine/training_generator.py", line 140, in fit_generator
    val_x, val_y, val_sample_weight)
  File "/home/arka/anaconda2/envs/hyperas/lib/python3.6/site-packages/keras/engine/training.py", line 787, in _standardize_user_data
    exception_prefix='target')
  File "/home/arka/anaconda2/envs/hyperas/lib/python3.6/site-packages/keras/engine/training_utils.py", line 127, in standardize_input_data
    'with shape ' + str(data_shape))
ValueError: Error when checking target: expected dense_1 to have 4 dimensions, but got array with shape (7440, 2)

这就是我在这里调用Densenet函数的方式。至少可以告诉我这是不是可以使用两个通道的输入与这个Densenet函数，将是一个很大的帮助。

Answer 1

文档上说，它应该恰好有3个输入通道。我想你可以使用一个嵌入层，或者https://keras.io/applications/#densenet一个常量值的维度。

Answer 2

通过传递带有Classes=2和pooling='avg‘的include_top=True解决了此问题。说明:当include_top设置为True时，会在上面添加一个致密层，其中包含与激活函数一样多的softmax类。现在，密集层需要一维输入。在这一阶段，网络输出的是导致误差的四维张量。当使用池化作为“avg”时，它应用了全局平均池化，这折叠了维度并使其扁平，因此可以计算它。但就我个人而言，在这个阶段我更喜欢Flatten。需要为此编辑densenet代码。