alan-turing-institute / ARC-SPICE

Sample Level, Pipeline Introduced Cumulative Errors (SPICE). Investigating methods for generalisable measurement of cumulative errors in multi-model (and multi-modal) ML pipelines.
MIT License
0 stars 0 forks source link

Pipeline Class Structure #5

Closed eddableheath closed 3 weeks ago

eddableheath commented 2 months ago

It would be good to have a general class structure for the pipelines, so we can iterate through different versions quickly once we have the models in place. My current idea is to have a config that looks like the following:

{model_a: {
        label: 'model_type',
        model: 'model',
        processor: 'processor',
        input_type: 'input_type',
        output_type: 'output_type',
        **kwargs
        },
 model_b: {...}, ...}

for a linear pipeline, with the class picking up all the details and setting up the functionality.

The output should be the final output, in addition to all the intermediary outputs.

N.B.: for non-linear pipeline structures this is a harder generalisation, should be a separate issue.