Enhance ScriptModelConfig to validate and infer missing configuration…

[rhoadesScholar]

… attributes

[mzouink]

i didn't get the point of the pyramid test if key not found check different key. i think the user should give a specific one key value altribute, if not evailable then he gets an error. we can't use this logic as we can never handle all the possible keys

[rhoadesScholar]

@mzouink If we want a strict schema, can we use the one already employed by the cellmap-segmentation-challenge, which groups the input and output array info (shape/scale)?:

input_array_info = {
    "shape": <input shape>,
    "scale": <input voxel size>
}
target_array_info = {
    "shape": <model output shape>,
    "scale": <model output voxel size>
}

[davidackerman]

that seems ok to me, since we haven't really decided on an enforced schema yet.

[rhoadesScholar]

or I can make it work for either your old way or the CSC schema?

In the initial PR I tried to make it as universal an adapter as reasonable, combining the conventions you had been using with CSC conventions. So I allowed split definitions or unified (input_array_info["shape"] vs. just read_shape), as well as "scale" vs. "voxel_size" keywords in *_array_info dicts. seems user friendly to me, but too loose for @mzouink it seems 🙃

[rhoadesScholar]

I forced the ScriptModelConfig parsing to require 1 of 2 schemas or a mix of the two that yields the complete necessary description:

read_shape = ...
input_voxel_size = ...
output_voxel_size = ...

# below can optionally be inferred if a PyTorch model is supplied:
write_shape = ...
output_channels = ...
block_shape = ...

input_array_info = {
    "shape": <input shape>,
    "scale": <input voxel size>
}
target_array_info = {
    "shape": <model output shape>,
    "scale": <model output voxel size>
}

What do you say @mzouink / @davidackerman ?

[mzouink]

good for me. feel free to go through config loader and the examples existing and change them to the new schema
Will be nice if you can add the schema to the docs

[rhoadesScholar]

most of the examples require fly_organelles, but that's not an obviously installable requirement... I'm only going to change the model_setup04.py to the new schema since I think the others should be retired from main imo.