How to add a dense layer after a 2d convolutional layer in a convolutional autoencoder?

Question

I am trying to implement a convolutional autoencoder with a dense layer at the bottleneck to do some dimensional reduction. I have seen two approaches for this, which aren't particularly scalable. The first was to introduce 2 dense layers (one at the bottleneck and one before & after that has the same number of nodes as the conv2d layer that precedes the dense layer in the encoder section:

input_image_shape=(200,200,3)
encoding_dims = 20

encoder = Sequential()
encoder.add(InputLayer(input_image_shape))
encoder.add(Conv2D(32, (3,3), activation="relu, padding="same"))
encoder.add(MaxPooling2D((2), padding="same"))
encoder.add(Flatten())
encoder.add(Dense(32*100*100, activation="relu"))
encoder.add(Dense(encoding_dims, activation="relu"))

#The decoder
decoder = Sequential()
decoder.add(InputLayer((encoding_dims,)))
decoder.add(Dense(32*100*100, activation="relu"))
decoder.add(Reshape((100, 100, 32)))
decoder.add(UpSampling2D(2))
decoder.add(Conv2D(3, (3,3), activation="sigmoid", padding="same"))

It's easy to see why this approach blows up as there are two densely connected layers with (32100100) nodes each or more or in that ballpark which is nuts.

Another approach I have found which makes sense for b/w images such as the MNIST stuff is to introduce an arbitrary number of encoding dimensions and reshape it (https://medium.com/analytics-vidhya/building-a-convolutional-autoencoder-using-keras-using-conv2dtranspose-ca403c8d144e). The following chunk of code is copied from the link, I claim no credit for it:

#ENCODER
inp = Input((28, 28,1))
e = Conv2D(32, (3, 3), activation='relu')(inp)
e = MaxPooling2D((2, 2))(e)
e = Conv2D(64, (3, 3), activation='relu')(e)
e = MaxPooling2D((2, 2))(e)
e = Conv2D(64, (3, 3), activation='relu')(e)
l = Flatten()(e)
l = Dense(49, activation='softmax')(l)
#DECODER
d = Reshape((7,7,1))(l)
d = Conv2DTranspose(64,(3, 3), strides=2, activation='relu', padding='same')(d)
d = BatchNormalization()(d)
d = Conv2DTranspose(64,(3, 3), strides=2, activation='relu', padding='same')(d)
d = BatchNormalization()(d)
d = Conv2DTranspose(32,(3, 3), activation='relu', padding='same')(d)
decoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(d)

So, is there a more rigorous way of adding a dense layer after a 2d convolutional layer?

Answer 1

For me, this worked perfectly. I encoded with conv2d and dense and then I flatten I and reshape in the decoder after the dense layer so the encoder and decoder are symmetrical. The only difference is that in my case I use images (224,224,1)

# create encoder
# 28,28 -> 1st conv2d (filter 3x3,relu activation, padding, strides == 'jumps')
self.encoder = tf.keras.Sequential([layers.Input(shape=(224,224,1)),
                                    layers.Conv2D(16,kernel_size=3,activation='relu',padding='same',strides=2),                
                                    layers.Conv2D(8,kernel_size=3,activation='relu',padding='same',strides=2),
                                    layers.Conv2D(4,kernel_size=3,activation='relu',padding='same',strides=2),
                                    layers.Flatten(),
                                    layers.Dense(units=3136,activation='sigmoid')]) # (28,28,4)

# deconvolution -> decoding 
self.decoder = tf.keras.Sequential([layers.Input(shape=(3136)),
                                    layers.Dense(units=3136,activation='sigmoid'),
                                    layers.Reshape((28,28,4)),
                                    layers.Conv2DTranspose(4,kernel_size=3,activation='relu',padding='same',strides=2),
                                    layers.Conv2DTranspose(8,kernel_size=3,strides=2,activation='relu',padding='same'),
                                    layers.Conv2DTranspose(16,kernel_size=3,strides=2,activation='relu',padding='same'),
                                    layers.Conv2D(1,kernel_size=(3,3),activation='sigmoid',padding='same')])