Implementing ML into your Elixir/Phoenix app

[evolutionengine]

Hi guys, sorry if this sounds like a silly question, I am a beginner in general. I would like to know how are you implementing machine learning into your apps. A couple of scenarios I could think of are -

1. Interact with a ML service in python over REST or similar.
2. Calling python scripts using erlport or thrift.

Are there any suggestions or other approaches ?

I am trying to build a Phoenix app and planning to use ML to build custom user feed, recommendations and spam filter.

Thanks.

[ityonemo]

here’s one guide. I’m going to look for the other one that I saw:
https://www.erlang-solutions.com/blog/how-to-build-a-machine-learning-project-in-elixir.html

[ityonemo]

above one only works if you are using tensorflow. this one is a more general python interop story:

Real-time Object Detection with Phoenix and Python

[evolutionengine]

Thanks, checking it out!

[patrickdm]

Hello and welcome, there are a few posts in the forum that you could find interesting:

Pelemay (formerly Hastega) - Challenge for GPGPU on Elixir!

Do you want to learn and use machine learning in Elixir / Phoenix / Nerves apps? Of course, you can, but the machine learning is written in Python and invoked from Elixir via ErlPort… Of course, you can implement some machine learning algorithms in Elixir. However, they have no computational power enough to execute real applications. We wish Elixir could have more computational power! This may change by Hastega! Hastega will make a series of code fragments of pipelined Enum.map to be trans…

Deep Learning with Elixir

I am creating a framework for Deep Learning. The network is described by macros as follows. defnetwork init_network2(_x) do _x |> f(5,5) |> flatten |> w(576,100) |> b(100) |> sigmoid |> w(100,10) |> b(10) |> sigmoid end Process backpropagation in parallel. In the future, we will introduce Hastega as a GPU-based engine. GitHub https://github.com/sasagawa888/DeepLearning

Tensorflex: Tensorflow bindings for Elixir

Hello all, I have been working on my proposed project called Tensorflex as part of Google Summer of Code 2018.. Tensorflex can be used for making predictions from input data on pre-trained Tensorflow models. Training as of now is only supported in the Python API (also C++ at a low-level) but Tensorflex fully supports Inference and can be used for making predictions. Moreover, Tensorflex is being worked on and improved everyday so if there are any particular features anyone has in mind, please f…

ElixirConf 2019: Annex: Introducing an Easy-to-Use Composable Deep Learning Framework in Elixir - Jason Goldberger

by @elbow-jason Deep learning has recently exploded as the go-to solution for some very computationally complex problems such as image recognition, text classification, and generalized task learning in robotics. Annex aims to bring practical, extensible, productive, and performant deep learning to Elixir. In the Annex framework, deep neural networks are composed as sequences of Annex. Layer behavior implementing modules/structs that simply transform data; underlying details (pa…

A topology and parameter evolving universal learning network

https://github.com/Rober-t/apxr_run Based on Handbook of Neuroevolution Through Erlang by Gene Sher. Don’t plan on doing too much more with this project but thought I would share it with the community in case others find it useful. Hopefully, it should be easier to work with than the original. Preliminary work was done to integrate with Elixir as well. Cheers

Hth

[ityonemo]

just a caution, unless I’m mistaken I would say except for tensorflex, the others are still experimental and not ready for prod deployment.

[sym_num]

Thank you for introducing my DeepPipe. Currently I am developing DeepPipe2. DP1 was slow and wrong. DP2 utilizes GPU for practical speed. I have finished implementing CNN. I’m also working on RNN for natural language processing.

Deep Learning library with GPU(CUDA/cuBLAS)

I am making a Deep Learning library using GPU. It’s name is DeepPipe2. I improved my previous DeepPipe. It can run simple SGD for MNIST.You can use GPU from Elixir.

[IRLeif]

Another approach to incorporate machine learning components into Elixir applications is to interoperate with other languages such as Python, e.g. through ports, NIFs, etc.

Erlang/OTP gives us other means to interface with- or incorporate external code, each with varying degrees of ease and robustness guarantees. See the Erlang Interoperability Tutorial for details.

Chapter 15 (Interfacing Techniques) of Programming Erlang by Joe Armstrong also contains some valuable information on this topic. Pages 233 through 242.

Here is an excellent article by @alvises which shows how to do this through Elixir. Elixir Mix ep. #91 goes into this article and related topics in more depth together with its author.

Erlang/OTP has excellent support for interoperability, which is one of its strengths. It might make more sense to leverage the existing Python libraries for machine learning in this way.

You can also develop your machine learning model as a separate application using Python and expose it through and API. TensorFlow comes with its own mechanisms for model serving. Then you can make use said API within your Elixir application and create an Elixir wrapper if needed. One benefit of using an Elixir port instead is that you can put your machine learning model behind a GenServer and supervise it as you would any other process in Elixir.

Another alternative is to develop and train your machine learning model using Python, and then bring the resulting model artefact into Elixir. You can serve and expose the model through an API written in Elixir. Have a look at out Tensorflex by @anshuman23 for inspiration on how this can be accomplished for neural network models through TensorFlow bindings.

Yet another approach which I have been thinking about lately is to embed the Python interpreter into Erlang/OTP or Elixir, and then use Elixir’s excellent meta-programming facilities (e.g. macros) to create a domain-specific language (DSL) specifically for machine learning. Supervision trees and GenServer could be used to spawn new processes representing instances of the Python interpreter to “cross-compile” the Elixir-based DSL to Python code and execute it. I found an old thread in the Erlang mailing list on this topic and an abandoned project on GitHub. This would likely require a considerable amount of effort, but could be an interesting way to vastly expand the libraries available in Elixir by essentially opening the door to all Python packages out there. Using ports is simpler.

Lastly, here are two related forum threads which contain some relevant information:

Using Elixir for Data Science and Machine Learning

Will future versions of Elixir be appropriate for data science/machine learning projects, like Scala or F#?

Machine Learning & Elixir

Hello folks, is anyone working on a machine learning library of sorts for Elixir/Erlang? I think that with the upcoming changes to the BEAM, specifically the work on a JIT, and the easy to use of GenStage and Flow, maybe it would become a viable solution to do number crunching (and I really mean maybe). Yes, I’m aware that the BEAM at the current state is not good for such things as number crunching, but, well as this recent talk from Lukas Larsson on EEF17 has given me hope of a faster numeri…

[evolutionengine]

Thanks for the detailed response.

[IRLeif]

You’re very welcome, my friend!

Did you try some of those approaches?