How to get encoder and decoder parts of autoencoder in order to stack them?

Hey,

I want to represent 128x128 images in a 1x64 vector, and for that I want to use autoencoders.

I have trained autoencoders in stages following the example in "Train Stacked Autoencoders for Image Classificatio".

However, I dont want to build a classifier, I want to build an autoencoder to embedd images.

(I tried build my own autoencoder with CNNs without using the built in autoencoder functions but I can't get good results for some reason)

Following the example I created two autoencoders, and added another one of my own instead of the softmax output and trained them:

encoder1 : 128x128 ->256-> 128x128

encoder2 : 256->128->256

encoder3 : 128->64->128

But how do I combine them?

If I use the "stack" function on these encoders like in the example:

stackednet = stack(encoder1,encoder2,encoder3)

I get this transformation :

128x128 -> 256 -> 128 -> 64

but what I want is this:

128x128 -> 256 -> 128 -> 64 ->128 -> 256 -> 128x128

So basicaly I need to stack this:

stackednet = stack(encoder1-encoder,encoder2-encoder,encoder3-encoder,encoder3-decoder,encoder2-decoder,encoder1-decoder)

Is there a way to extract the encoder and the decoder parts from an Autoencoder object and then stack them?

Answer 1

[David Willingham]

Hi,

That example is using the older Neural Networks functionalilty. I'd recommend looking at one of the examples using the newer framework like this one here: Train Variational Autoencoder (VAE) to Generate Images.

[ytzhak goussha]

Thank you for your reply,

I did try that one as well,

I mannaged to get the loss as low as 5 but then it stoped.

That is why I wanted to train the network like an onion, train the outer encoder-decoder first, and the move inwards, and finally stack all the encoders, and the decoders.