jasp-stats / jasp-issues

This repository is solely meant for reporting of bugs, feature requests and other issues in JASP.
56 stars 29 forks source link

[Feature Request]: #1913

Open luiscunhacsc opened 1 year ago

luiscunhacsc commented 1 year ago

Description

Deep Learning Module(s)

Purpose

Allow using 80% of Deep Learning Modelling via an intuitive and beautiful GUI, just like the JASP's ML module

Use-case

Someone whats to apply a CNN to train a model to classify images

Is your feature request related to a problem?

Deep learning is simple, but for now hidden behind Python or R code

Is your feature request related to a JASP module?

Machine Learning

Describe the solution you would like

A similar module of the one created for Machine Learning, but with CNN's, LSTM's and Transformers

Describe alternatives that you have considered

Either putting those three main methods in one menu, or (if necessary, or in a secound phase), create separate modules: one for computer vision with CNN's, other for Temporal Series with LSTM's, and another for Applications of Transformers (e.g. Language, but also Temporal Series, and other domains)

Additional context

Today, the Pyhton library Keras, by Fraçois Chollet is the most accessible entry point to use Deep Learning. But, just as JASP's wonderful Machine Learning module suffices for 80% of students / teachers use cases, it would be wonderful if there was a Module in JASP that allowed the use of Deep Learning methods.

It could be possible to develop such a Module based on Keras (I believe there is even a version of Keras that runs in R, or at least can be accessed from R).

Also, there are pre-trained modules available for free online, that could be used to enhance the possibilities of the module.

Even if in a first moment only CNN and LSTM modules were made available, it would already be immensely valuable.

I could only dream with the day JASP provided a Menu with CNN's, LSTM's and Transformers, with their sample data sets.

As a teacher, who loves the concepts, knows that they are simple, but spends 90% of the time teaching stuff about Python or Jupyter/Management, rather than the models themselves, that would certainly revolutionize the accessibility of these methods, and would make them much more widely known and democratic.

koenderks commented 1 year ago

I would love to implement this! Hopefully I can find the time somewhere. Do you have some examples of this being done in R, like some of those pre-trained models? That would be a good starting point.

luiscunhacsc commented 1 year ago

Hi. That would be so great! I think it could have an enormous impact! One thing I know is that Francois Chollet has a book based in R. I noticed that the source code, in R, for the chapters of the book is available online, here.
For instance, chapter 8 has a section named "8.3 Leveraging a pretrained model", and the corresponding code is in the source code publicly available. For instance, that part starts at line 314, of the file ch08.R (leveraging vgg16)

you would be legend is you could manage to to anything resembling the Machine Learning module...

thank you very much for your time, and interest.

all the best,

Luís

luiscunhacsc commented 1 year ago

Please see also here:

I will add any other links / resources that I might find.

tomtomme commented 5 months ago

@luiscunhacsc Could you change the title of your request to [Feature Request] Deep Learning incl. CNN's, LSTM's and Transformers or similar? Thx!