[Project] Search Engine For Pretrained Models


(Brendan Fortuner) #1

I’m thinking of prototyping a search engine for pretrained models. Pretrained models are super useful for transfer learning but they’re hard to find and often buried in academic websites or github repos.

Does anyone know if something like this exists? It’s basically a model zoo with search capabilities to filter by framework, architecture, and topics (medical, satellite). To start we could just manually create json documents for models compiled in wikis like this one, setup a basic React frontend, and use ElasticSearch as our search engine.

Eventually we can host models ourselves and port everything across frameworks so users can download models in any framework.

Anyone interested?

Brendan


(Yash Katariya) #2

This would be awesome. I am interested!!


#3

I’m very much interested!


(Brendan Fortuner) #4

Awesome! I’ll put together a simple prototype this weekend and we can discuss next week where to take it!


(Brendan Fortuner) #5

Okay here’s the prototype. It’s React with Fuse.js for fuzzy search. The corpus is stored in JSON/JS object format. I added ~50 models/datasets, then I got lazy :wink:

http://zoosearch.surge.sh


(Brendan Fortuner) #6

If you’re interested in helping out, probably the best next step is to add more datasets + models. I bet we can collect 90% of what’s out there in a few days if we work together.

To add an entry submit a pull request updating these files with your new content:


Once we settle on the format and see how things are working, we can create a “submit” form on the website. Right now it’s just a static website, no backend.