[WISH] Bring popular networks into DLS

Hi everyone!

First of all let me start by saying a big thank you for the great work you have been doing with DLS. It is an amazing piece of software that really helps prototype DL models and I am certain it will push the DL (and other) community to the right direction.

In the first webinar I attended I alluded to the possibility of bringing popular deep learning networks into DLS. This would be very helpful for users that are not as comfortable with model development as the expert part of the community (myself included). I think it can be done in two ways, from what I can see. First, the community of users, as it grows, develops, prototypes, and shares their work with each other. I have no doubt this will happen in time. Second, is the DLS team itself provides a helping hand by developing a few of the popular models (or sometimes layers) within DLS.

I thought I would make this thread in order to have a place where users can add their items on the wishlist of models they would like to see implemented in DLS. I’ll start with a few of mine below (as a new user I can only add two links for the items below):

  1. Tensorflow research models has a lot of interesting, and usually bleeding edge, implementations
  2. Deep Markov Model for multiple time-series prediction (https://github.com/clinicalml/dmm)
  3. Any of the implementations of CapsuleNets found on GitHub
  4. Pix2PixHD Nvidia Model
  5. The many flavors of GAN models.
  6. Variational models and bayesian approaches, e.g. variation autoencoder
  7. Attention transfer (https://github.com/szagoruyko/attention-transfer) and attention layer (with context and without)

I hope this is not an unreasonable, or infeasible, matter to discuss. I realize that in some cases there might be some issues with intellectual property. However, the AI community has been developing at incredible speeds mostly due to the fact that the researchers around the world, and big companies, are more than willing to make their research and insights freely accessible. It’s an incredible property for a scientific field and one that could be used in this case.

Thank you very much for reading this far.

Kind regards,

Hi Theodore,

That is a great idea and certainly matches our vision of making latest and greatest models available in DLS.

To make this happen, we can host a repository for DLS models where community can keep adding new models and we keep adding them in DLS.

We will need to agree on a model format. Currently DLS uses Keras 1.2.2. So any models which are compatible with Keras 1.2.2 can be easily be added in DLS.

In fact, we have an option for an user to first load a Keras model as a layer and self-test it with the DLS first. If it works fine, then it can be added to DLS models repo and we can verify and include it in our next release for everyone to benefit from it.

To kickstart this effort, I would like few volunteers who are willing to work on adding model to the DLS models repo.


Hi @rajendra

Thanks for the quick reply! Glad this sounds interesting to you.

I did hear, but did not yet test, the option to load a Keras model. I will give it a try once I install version 2.0.

I think the repository would be a great start and a common format will help people contribute. On my end, as a relatively beginner in this field, I might have a few difficulties bringing in some of these models, but I can sure give it a try once the repo is up and running.

Kind regards,