Contributing new models

[thekevinscott]

I'd like to explore making it easier to submit new models to the repo.

To train a new model, a developer needs to be willing to sling some Python. However, I think there's ways we can ease the process.

Google Colab offers free GPUs, and I believe that training idealo's implementation of ESRGAN can fit within the Colab time limits. (Note to self: confirm this.)

Training a brand new architecture from scratch is difficult in that it requires ascertaining that the Python implementation is compatible with Tensorflow.js. However, re-training an existing implementation could be a lot more straight forward - find a dataset, set up the code to split the dataset into appropriate train / test sets (if not already done), run the training on Colab and babysit as needed, and then submit the model to this repo.

Specific datasets (like Flickr's faces dataset) could allow us to develop more domain-specific models that offer better accuracy for specific domains.

Some open questions:

• Can Colab support this process? If not, is there a service we could use to train models that is free? (Introducing a paid tier would I think significantly hinder abilities to recruit users willing to train new models)
• How much can we automate the process of integrating a new dataset into the training script?
• What does the submission process look like? I assume we'd want to record information about how the model was trained - even maybe include a copy of the notebook associated with training. Automatic benchmarking would also be nice. Would a model be auto-deployed, or would we need to cut a new release?