[FCT] Functional approach to graph models and listeners

[blythed]

Idea of @jieguangzhou to create models inside a context manager.

# Create model instances
model1 = Model()
model2_multi = Model()
model3 = Model()

# Using the Graph context manager
with Graph.creating() as G:
    o_m1 = model1(G.input)
    o_m2 = model2(y=model1)
    o_m3 = model3(x=o_m1, y=o_m2)

G.predict_one()     # is ok
G(model4)             # raise NoBuildContextError

Idea, use same formal approach for creating Listener instances:

with db.listener(coll.find()) as list:
    # this is a `Listener` equivalent to 
    # `Listener(key={'a': 'txt', 'b': 'brand'}, select=coll.find(), model=model1)
    o_m1 = model1(a='txt', b='brand')         
    # This is a downstream listener
    # `Listener(key={'x': 'brand', 'y': '_outputs....{model1.identifier}.{model1.version}'}, select=coll.find(), model=model1)`
    o_m2 = model2(x='brand', y=o_m1)      
    ...

Suggestion to make easier to debug:

with db.listener(coll.find(), eager=True):
    # Same as `model1.predict(list(db.execute(coll.find())))`
    o_m1 = model1(a='txt', b='brand')         
    # This is a downstream prediction
    # Same as `model2.predict(...)`
    o_m2 = model2(x='brand', y=o_m1)      

[blythed]

Solution for multiple inputs and models - create a routing model node:

import dataclasses as dc
import typing as t
from superduperdb.components.model import _Predictor, objectmodel

db = ...
coll = ...


@objectmodel
def f(a, b, c=2):
    ...


# Data in database looks like this:
{'txt': 'sdsds', 'brand': '12324hd'}
{'txt': 'sdsd1s', 'brand': '212324hd'}



f.predict_in_db(X={'txt': 'b', 'brand': 'a', 'bla': 'c'}, db=db, select=coll.find())
f.predict_in_db(X=(
        ['brand', 'txt'],
        {'other': 'c'}
    ),
    db=db, select=coll.find()
)


# Question, if we have a GraphModel what are the input parameters?
# If we have 1 input node, it's easy, same as input node
# If multiple input nodes ... ?

@dc.dataclass
class InputNode(_Predictor):
    inputs: t.List[str]
    signature: ...

    @property
    def inputs(self):
        ...

    def predict_one(self, *args, **kwargs) -> int:
        return ...


with Graph.create() as g:
    ...

[blythed]

Suggestion: create a GraphModel and then do G.listen(db=db, select=coll.find())