How to implement a cascaded neural network in Keras where 1st NN's forward output is cloned twice to perform forward on the 2nd NN?

[Parvathy Prasad]

Using Keras, I am trying to reproduce a few basic results from a published paper.

In this task, there are two neural networks - A & B, that are connected in a cascade formation, i.e. the output of model A feeds the input of model B. Please see the figure below.

Here are the specifics steps that need to be implemented:

1. Perform one forward pass on model A with a batch size of 64, and 13155 rows.
2. Clone the output to create a batch of 128 rows i.e. use each training sample twice. (There are auxiliary inputs which distinguish the duplicate rows, but I have omitted that info for the sake of brevity).
3. Perform one forward pass on model B by feeding it with #2 above.
4. Now backward the two networks, calculate a loss for each one, sum the loss & send it to the optimizer with the gradients of both networks.

As I am quite new to the world of neural networks & Keras, I'd really appreciate any help from the community to implement the above steps in Keras. In particular, I am quite confused about implementing step 2 (please see the attached figure).

The details of customization to my specific problem is fuzzy, for example how to implement the forward & backward logic separately etc.

[Krishnakumar Gopalakrishnan]

I am also interested in how to implement the tasks in this question using Keras.  In particular, I am intrigued by the part wherein the output of the 1st network is doubled/cloned and fed as input to the second network. Can someone from the community please point in a suitable direction?

Regards,

Krishnakumar

[Lance Norskog]

In the Keras class hierarchy, a Model is a subclass of Layer.

You can create two inner Models, then use them as Layers in a Model that wraps the two and drives the training loop.

Each training batch will push the input sample through all Layers and compare with its output.

You can then save & load these inner Models however you see fit. 

Does this have anything to do with this problem?

https://www.ee.cuhk.edu.hk/~xgwang/CUHK_identification.html

--
You received this message because you are subscribed to the Google Groups "Keras-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to keras-users...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/2408083a-c79c-424e-8a37-8d0ddcc90b67n%40googlegroups.com.

--

Lance Norskog
lance....@gmail.com
Redwood City, CA

[Lance Norskog]

To implement a network that compares two inputs, you need to create two channels.

Here is an example of a 'Siamese' network, where both input channels use matching structure and weights.

You do not need to do this, the different channels can have different processing.

The 'feature_model' in this example is used twice in the 'similarity_model' in this example:

https://www.kaggle.com/kmader/image-similarity-with-siamese-networks/code#Siamese-Model