Just out of curiosity, How much resources (time, money, etc) does it take to train a model like this?
The vast majority of training for this model was done by Microsoft, so I unfortunately cannot shed much insight there.
From this base model, the amount of additional training and improvement depends on how much you want to optimize the original model. If the research model is sufficient, you may only need a few hours.
That said, we are researching our own model and can share the training time and resources required for this new version once we finish (and assuming the results are superior).
I see sentdex literally burning a $2k Nvidia gpu trying to train some data (on YouTube), so i was just curious on how people did it without burning a massive hole in their pockets.
For which the models are available online. No need to train in this case.
If this was a new project and you were starting from scratch but already had a clean dataset, 0.5 to 3 days of work (excluding training time).
The model is a VAE so I would approximate less than 2 days of training time on a modern GPU.