Drag-and-Drop GUI for deep learning experimentation thread

[rumbleFTW]

Greetings!

I just wanted to clarify some doubts and present my rough idea and plan for idea [12] "Develop a drag-and-drop GUI for deep learning experimentation".

First of all, the queries I'd like to be resolved are:

• Will the proposed application be used just for working with digital imagery, utilizing torchvision, or for all-round neural network development/experimentation, supporting base torch, and torchaudio?
• In the idea description, it is mentioned

Predefined network architectures should be embedded as blocks

What would these predefined network architectures actually be? Do they refer to pre-trained architectures, like VGG or ResNet? Or basic Hidden-layer or CNN architectures ?

• In the idea description, PyQT5 is mentioned. Is this absolutely necessary for the project? A browser-based front-end an be easily used on low-end devices, where the actual torch model would run in a remote server.

• As we know, torch is a decently large library, hence it is tough to wrap all the features that the python library provides, by a GUI. So, there must be a level of abstraction in the whole project. We have to sacrifice some tweaking and specific functionalities for the sake of clutter-free and comfortable experience. I'd like to know what level of abstraction should be followed in this case.

Rough Idea:

I have decided to follow a Modular approach with low coupling and better modularity. This way, a browser based or any other GUI can be easily connected with the back-end of the application. The back-end and wrapper would provide some APIs, that can be used by any other front-end implementation/libraries(if necessary in the future). The GUI built would define a config.json file, which will be forwarded to a wrapper code, that will decode the JSON file and communicate with the torch back-end to make and interact with the neural networks. The JSON file would be mutex locked to synchronize the file between the ends and avoid overwrites.

Current state:

Currently, I am working on a demo, and it in its current state supports data loading, creating basic regression and classification and regression models. It also supports creating neural networks layer-by-layer and using the Sequential() API of PyTorch. I have tested the application by creating neural networks on the classic Titanic dataset and California House prices dataset.

I currently plan to improve the prototype/demo and create a basic documentation. After the prototype reaches a presentable stage, I'll be sharing it along with my final proposal.

I would also appreciate reviews on the current approach and possible improvements that can be done or alternate paths that can be followed.

Thank you!