DRAFT: split dlc schema into train, model, pose

[CBroz1]

WIP. Does this split of schema roles work for you, @ttngu207? If so, I'll update documentation to reflect the split.

See corresponding workflow PR, specifically the pipeline which now includes VideoRecording.

forthcoming updates:

• Attempt full workflow usage, resolve issues with pieces that connect schemas.
• documentation for usage, including upstream dependencies

[ttngu207]

I think the split should be into two modules

• model and pose-estimation
• model training (which must import the other module)

I think model and pose-estimation should really go together, it wouldn't be much of a deeplabcut element with just model

[CBroz1]

My motivation for splitting model and pose was the imagined use-case where there is only one model. A user could supply their own model table with only the project_path field and pose estimation would still function the same.

I expect...

1. Every user will want to do pose estimation and therefore use the pose schema
2. A subset will want to also do model evaluation and therefore need the full model schema
3. A further subset will also want to launch training and therefore use all 3

In the ultimate version of this schema, will we still support the current 'run from on-disk files' process? Or will we want to push users away from that standard? If the former, it seems to me like 'only use the schema you need' is within the a-la-carte design philosophy of the elements.

[ttngu207]

The way the pose and model is implemented here is quite interconnected. I.e. to use pose, users must use our model, I'm not sure if they can just import and use pose by supplying their own model table and the pose estimation step can work.

[CBroz1]

oh - you're right. I previously though that we only relied on project_path, but I took another look at do_pose_estimation where we require the model table to have shuffle, trainingsetindex, and model_prefix - I'll revert to 2 schemas and request your review when complete