wire a model training for images

[eroux]

Prodigy doesn't seem to have a built-in model for images so we need to include a connection to something in our recipe. There are some examples with tensorflow on https://github.com/explosion/prodigy-recipes/tree/master/image/tf_odapi but perhaps we want to try pytorch? This is of course something that we need to make sure we have right before annotators start

may or may not be relevant: https://thinc.ai/docs

something that we also need to determine is where the training should take place, it will require a machine with some GPUs and possibly a very big hard drive

[eroux]

Here's an example of annotations we get from prodigy and that we need to train our model on:

http://eroux.fr/bdrc_crop.jsonl

the steps are:

• [ ] determine the framework that we will train the model on (torch?)
• [ ] determine the hardware
• [ ] get / rent the hardware
• [ ] build a pipeline to train a model on annotations in the format above
• [ ] train a model
• [ ] build a pipeline to run the model on more data, with results in a format that prodigy can use
• [ ] run the model on more data and get a file that we can import on prodigy
• [ ] build a pipeline to reimport the file in prodigy

All of that should be recorded some documented logs, as there are many potentials points of failure