Flowvex

An app that facilitates comparing different neural network architectures and finding the optimal set of hyperparameters

Comment

Inspiration

This project was inspired by a need to have a convenient and easy way to tune model hyperparameters. While working on an earlier project I found that having to constantly change and recompile my model to tune its parameters , and comparing each iterations, was a big time sink. So I thought I would make an app to help with that.

What it does

Flowwvex lets you build models and run them on a small dataset. The user creates a "Container" which contain multiple models which will be ran on the same data. The user can add and remove layers from their models, add, edit and remove parameters for each of these layers. The user can also upload a dataset to run those models on. Once ran, the models will show their output, based on the metrics sets by the user. The user can then change parameters and run the models again until they are satistifed

For convenience, new users a provided with a "test container" which contains a model and a dataset

--- I want to note that is intended exclusively to be a way to tune model parameters, not to train models. As such the datasets are limited in size

Challenges we ran into

-Started hackathon a bit late, and sickness and travels, means we were short on time

  • Spent a lot of time trying to figure out how to make tfjs-node work
  • Team issues, members joined and did nothing or left

Accomplishments that we're proud of

  • It actually works
  • Managed to host ## What we learned

What's next for Flowvex

  • Selecting from existing datasets -Share datasets
  • Support more data formats for datasets (notably CSV and NPY)
  • Support tfjs-node, which is the faster version of tensorflowJs, but doesn't work with convex just yet
  • More error-checking and sanitizing for models
  • Support using multiple metrics at a time
  • Support multiple optimizer functions
  • Community posts and comments about models
  • Exporting models to a file (needs tfjs-node)
  • Keep compiled models alive while the user is working on them instead of recompiling even if there was no change

Built With

Share this project:

Updates