QuaternionDense layer only accept input of dimension 2

[healthyertech]

Hi, I'm recentrly approaching to quaternion neural networks and I found this fantastic repository which I am exploiting to implement an academic project. I'm developing a convolutional neural network for a project about speech recognition.
After having defined a real-valued neural model, I was trying to convert it into a quaternion model. However, I noticed that the QuaternionDense layer accept ony inputs having dimension 2, while in the fully connected block of my model I would need a dense layer accepting input at least of dimension 3. I tried to slightly modify in my local folder the definition of the quaternion layer, but it slill retuns me a ValueError stressing that the input shape does not mathc the required shape.
Would it actually be possible to make the quaternion dense layer working with input of size 3? If so, how could I proceed?
Kind regards

[TParcollet]

Hi and thanks for you interest.

This is a pretty old code :( If you are interested in E2E ASR with quaternions, please have a look at https://speechbrain.github.io instead :)

[healthyertech]

Ok, thanks a lot for the advice. I also looked at that repository, even though my interest in this one was also related to the fact that I wanted to integrate it into a tensorflow project of mine, extending it by using quaternion neural models
For the tensorflow library, I also found the implementation of a quaternion batch normalization layer, and I would like to ask if something similar could be possibly available also for the quaternion layers library implemented in pytorch.
Kind regards.

[TParcollet]

Yes, here: https://github.com/speechbrain/speechbrain/blob/develop/speechbrain/nnet/quaternion_networks/q_normalization.py

Note: This is not done by using the whitening approach that is WAY too much memory demanding. The reference is given in the doctoring of the class.

[healthyertech]

Thank you very much for the helpful and quick reply and for the useful advices.
Kind regards