Keras 中的特定丢失

Question

I would like to train an autoencoder by using only specific PARTS of a layer (the layer named FEATURES in the autoencoder example at the bottom of this question).

In my case, NOK pictures for a new product are very rare, but needed for training. The aim is generate NOK pictures from OK pictures (all examples I found did the opposite). The idea is to force learning OK-picture structure in features[0:nx] and learning NOK-picture structure (maybe from a similiar product) in features[nx:n] in order to use the NOK-features as parameters to generate NOK-pictures from OK-pictures.

Two ideas came to my mind using a non-random dropout

(1) keras.layers.Dropout(rate, noise_shape=None, seed=None) has the noise_shape argument, but I am not sure if it helps me as it only describes the shape. It would be perfect to be able to provide a mask consisting of {0,1} to apply on the layer in order to switch on/off specific nodes

(2) creating a custom layer (named MaskLayer below) which performs masking specific nodes of the layer eg as a tuple of {0,1}.

I have read this , but I do not think it applies (generate a layer by concatenating layers which can be freezed separately).

def autoEncGenerate0( imgSizeX=28, imgSizeY=28, imgDepth=1):  ####:
    ''' keras blog autoencoder'''
    input_img = Input(shape=(imgSizeX, imgSizeY, imgDepth)) 
    x = Conv2D(32, (3, 3), activation='relu', padding='same')(input_img)
    x = MaxPooling2D((4, 4), padding='same')(x)
    x = Conv2D(32, (3, 3), activation='relu', padding='same')(x)
    encoded0 = MaxPooling2D((8, 8), padding='same', name="FEATURES")(x) 
    encoded1 = MaskLayer(mask)(encoded0) # TO BE DONE (B2) masking layer parts
    x = Conv2D(32, (3, 3), activation='relu', padding='same')(encoded1)
    x = UpSampling2D((8, 8))(x)
    x = Conv2D(32, (3, 3), activation='relu', padding='same')(x)
    x = UpSampling2D((4, 4))(x)
    decoded = Conv2D( imgDepth, (3, 3), activation='sigmoid', padding='same')(x)
    autoencoder = Model(input_img, decoded)
    autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')
    return( autoencoder)

Thanks for hints.

Answer 1

There is trainable attribute that each instance of tf.keras.layer.Layer has which disables training of the variables of that layer. UpSampling2D doesn't have any variables so you CAN'T train it. What you want is to train the variables of the convolutional layer that comes before that upsampling layer.

You could do it like this:

# define architecture here
autoencoder = Model(input_img, decoded)
layers_names = [l.name for l in autoencoder.layers]
trainable_layer_index = layers_names.index('FEATURES') - 1
for i in range(len(autoencoder.layers)):
    if i != trainable_layer_index:
        autoencoder.layers[i].trainable = False
# compile here

NOTE that you compile the model AFTER you set layers to trainable/non-trainable.