I think GraphFormer should be a pip installable thing.

[tensor-ninja]

This library should should have a wrapper around the GraphFormer model and provide a high level API for users. I imagine the flow to be something like this pseudo code:

from GraphFormer.data import GraphDataLoaders       # This can probably be a patch for fastai's DataLoaders base class 
from GraphFormer.models import GraphFormerBase
from GraphFormer.learner import GraphLearnerBase

dataset = fetch_graph_dataset()

dls = GraphDataLoaders(dataset)                                 # This should validate the schema needed for GraphFormer

model = GraphFromerBase(
      # config
)

learner = GraphLearner(dls, model, metrics=metrics_func)

learner.train() # train model, extra kudos if this can spin up infra in RLEF

learner.export() # this would export the model to be used as an endpoint, extra kudos for a leaner.serve() type functionality which could serve the model on RLEF infra