Keras: How to train stacked auto-encoder with greedy layer-wise way

[neouyghur]

is there any function available for building stacked auto-encoder in keras library? If no, does some offer some ideas for that. Actually I also have an idea, but I think it is a very naive idea. My idea is that each time train two layer (encode and decode) then freeze them. And add new layers (both decode and encoder) train them. And repeating above processes until reaching the full model.

[jf003320018]

You should check the previous issues firstly. This question has been discussed.
Additionally, you can see the bolg from Francois Chollet if you want to build antoencoder with keras. https://blog.keras.io/building-autoencoders-in-keras.html

[neouyghur]

@jf003320018 You may misunderstand my meaning. I dont mean stack several lays for auto-encoder. I actually did that. However, the result is not good. Here, I want to train each layer separately, then stack them together.

[jf003320018]

I think that this idea is commonly used and you can realize it with keras.

[darthdeus]

@jf003320018 I'm confused. The linked blog post doesn't explain how to train the layers separately. I also don't see any other issue addressing this.

[jgeofil]

A bit late.. but here's an example where each pair of layers are trained independently from @MadhumitaSushil
SDAE