pythonkerasplaidml

Remove Dustbin within Keras CNN


I have the following problem, I want to remove a "Dustbin" from the output of one of the layers in my Keras model.

The Code without the dustbin removal looks like this and works:

def create_detector_network():
    input = Input(shape=(128, 128, 512))
    x = Conv2D(128, kernel_size=3, strides=1, name='detect_1', padding='same')(input)
    x = BatchNormalization()(x)
    x = Conv2D(65, kernel_size=1, strides=1, name='detect_2')(x)
    x = BatchNormalization()(x)
    x = Activation('softmax')(x)
    x = keras.layers.UpSampling2D(size=(8, 8), data_format=None, interpolation='nearest')(x)
    x = Conv2D(1, kernel_size=1, strides=1, name='reduce_dim')(x)
    return Model(input, x)

However, if I add the removal to the network:

def create_detector_network():
    input = Input(shape=(128, 128, 512))
    x = Conv2D(128, kernel_size=3, strides=1, name='detect_1', padding='same')(input)
    x = BatchNormalization()(x)
    x = Conv2D(65, kernel_size=1, strides=1, name='detect_2')(x)
    x = BatchNormalization()(x)
    x = Activation('softmax')(x)
    x = Lambda(lambda x: x[:, :, :-1], output_shape= (128, 128, 64))(x)  #x[:, :, :-1] <------
    x = keras.layers.UpSampling2D(size=(8, 8), data_format=None, interpolation='nearest')(x)
    x = Conv2D(1, kernel_size=1, strides=1, name='reduce_dim')(x)
    return Model(input, x)

I get the following model.summary() output, where the dimension after the lambda layer increases to 65 again:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_38 (InputLayer)        (None, 128, 128, 512)     0         
_________________________________________________________________
detect_1 (Conv2D)            (None, 128, 128, 128)     589952    
_________________________________________________________________
batch_normalization_37 (Batc (None, 128, 128, 128)     512       
_________________________________________________________________
detect_2 (Conv2D)            (None, 128, 128, 65)      8385      
_________________________________________________________________
batch_normalization_38 (Batc (None, 128, 128, 65)      260       
_________________________________________________________________
activation_10 (Activation)   (None, 128, 128, 65)      0         
_________________________________________________________________
lambda_6 (Lambda)            (None, 128, 128, 64)      0         
_________________________________________________________________
up_sampling2d_18 (UpSampling (None, 1024, 1016, 65)    0         
_________________________________________________________________
reduce_dim (Conv2D)          (None, 1024, 1016, 1)     66        
=================================================================

Can anybody explain why this is happening and how to solve it?


Solution

  • on my machine works correctly (TF 2.2). I modify the lambda to take care also of batch dimension

    def create_detector_network():
        inp = Input(shape=(128, 128, 512))
         x = Conv2D(128, kernel_size=3, strides=1, name='detect_1', padding='same')(inp)
         x = BatchNormalization()(x)
         x = Conv2D(65, kernel_size=1, strides=1, name='detect_2')(x)
         x = BatchNormalization()(x)
         x = Activation('softmax')(x)
         x = Lambda(lambda x: x[:,:,:,:-1])(x) 
         x = UpSampling2D(size=(8, 8), data_format=None, interpolation='nearest')(x)
         x = Conv2D(1, kernel_size=1, strides=1, name='reduce_dim')(x)
    
         return Model(inp, x)
    

    this is the summary

    _________________________________________________________________
    Layer (type)                 Output Shape              Param   
    =================================================================
    input_33 (InputLayer)        [(None, 128, 128, 512)]   0         
    _________________________________________________________________
    detect_1 (Conv2D)            (None, 128, 128, 128)     589952    
    _________________________________________________________________
    batch_normalization_14 (Batc (None, 128, 128, 128)     512       
    _________________________________________________________________
    detect_2 (Conv2D)            (None, 128, 128, 65)      8385      
    _________________________________________________________________
    batch_normalization_15 (Batc (None, 128, 128, 65)      260       
    _________________________________________________________________
    activation_7 (Activation)    (None, 128, 128, 65)      0         
    _________________________________________________________________
    lambda_7 (Lambda)            (None, 128, 128, 64)      0         
    _________________________________________________________________
    up_sampling2d_7 (UpSampling2 (None, 1024, 1024, 64)    0         
    _________________________________________________________________
    reduce_dim (Conv2D)          (None, 1024, 1024, 1)     65        
    =================================================================