Florian Müller-Fouarge

I can add the API I already wrote and we can work from there. I'll have to check for write permissions.

I'll give this a shot in the next days. We can easily have a cpu and a gpu version. I think for the beginning user rights might be a bit...

1. That's a valid argument. In this case we could also simply move the HTTPClient to another file or let the user provide their own. 2. The most elegant way...

I created my own branch from your 'inference' branch. I added what I had already impelemnted + docstrings. I'm not sure whether to return the response or its jsonified version.

I don't think we should implement a logic depending on the input type inside of the InferenceApi. It should work no matter if the `input_` is a buffer, a byte...

I do agree. I just had the problem of loosing annotations when working with a dataset loaded from local storage. The button _Save_, visible on right side of the screenshot...

I used v0.14.16 on Ubuntu 20.10

This is also the case in your online demo version on https://udt.dev/

FYI, using a nested structure isn't a problem I you directly modify fields 'imageUrl' the udt.json file. But that's quite more complicated than importing the images in the UDT