Deprecate signature as argument

CHANGELOG.md

@@ -13,6 +13,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - Deprecate vanilla `DataType`
 - Remove `_Encodable` from project
+- Deprecate `signature` as parameter and auto-infer from `.predict`
 
 #### New Features & Functionality
 

superduper/components/graph.py

@@ -197,7 +197,6 @@ class DocumentInput(Model):
 
     spec: t.Union[str, t.List[str]]
     identifier: str = '_input'
-    signature: Signature = 'singleton'
 
     def __post_init__(self, db, example):
         super().__post_init__(db, example)
@@ -232,12 +231,11 @@ class Graph(Model):
     :param edges: Graph edges list.
     :param input: Graph root node.
     :param outputs: Graph output nodes.
-    :param signature: Graph signature.
 
     Example:
     -------
     >> g = Graph(
-    >>   identifier='simple-graph', input=model1, outputs=[model2], signature='*args'
+    >>   identifier='simple-graph', input=model1, outputs=[model2],
     >> )
     >> g.connect(model1, model2)
     >> assert g.predict(1) == [(4, 2)]
@@ -253,15 +251,14 @@ class Graph(Model):
     )
     input: Model
     outputs: t.List[t.Union[str, Model]] = dc.field(default_factory=list)
-    signature: Signature = '*args,**kwargs'
 
     def __post_init__(self, db, example):
         self.G = nx.DiGraph()
         self.nodes = {}
         self.version = 0
         self._db = None
 
-        self.signature = self.input.signature
+        self._signature = self.input.signature
         if isinstance(self.outputs, list):
             self.output_identifiers = [
                 o.identifier if isinstance(o, Model) else o for o in self.outputs

superduper/components/model.py

@@ -45,7 +45,6 @@ def model(
     output_schema: t.Optional[Schema] = None,
     num_workers: int = 0,
     example: t.Any | None = None,
-    signature: Signature = '*args,**kwargs',
 ):
     """Decorator to wrap a function with `ObjectModel`.
 
@@ -59,7 +58,6 @@ def model(
     :param output_schema: Schema for the model outputs.
     :param num_workers: Number of workers to use for parallel processing
     :param example: Example to auto-determine the schema/ datatype.
-    :param signature: Signature for the model.
     """
     if item is not None and (inspect.isclass(item) or callable(item)):
         if inspect.isclass(item):
@@ -74,7 +72,6 @@ def object_model_factory(*args, **kwargs):
                     output_schema=output_schema,
                     num_workers=num_workers,
                     example=example,
-                    signature=signature,
                 )
 
             return object_model_factory
@@ -88,7 +85,6 @@ def object_model_factory(*args, **kwargs):
                 output_schema=output_schema,
                 num_workers=num_workers,
                 example=example,
-                signature=signature,
             )
     else:
 
@@ -105,7 +101,6 @@ def object_model_factory(*args, **kwargs):
                         output_schema=output_schema,
                         num_workers=num_workers,
                         example=example,
-                        signature=signature,
                     )
 
                 return object_model_factory
@@ -119,7 +114,6 @@ def object_model_factory(*args, **kwargs):
                     output_schema=output_schema,
                     num_workers=num_workers,
                     example=example,
-                    signature=signature,
                 )
 
         return decorated_function
@@ -352,13 +346,32 @@ def __new__(mcls, name, bases, dct):
         if 'init' in dct:
             dct['init'] = init_decorator(dct['init'])
         cls = super().__new__(mcls, name, bases, dct)
+
+        signature = inspect.signature(cls.predict)
+        pos = []
+        kw = []
+        for k in signature.parameters:
+            if k == 'self':
+                continue
+            if signature.parameters[k].default == inspect._empty:
+                pos.append(k)
+            else:
+                kw.append(k)
+        if len(pos) == 1 and not kw:
+            cls._signature = 'singleton'
+        elif pos and not kw:
+            cls._signature = '*args'
+        elif pos and kw:
+            cls._signature = '*args,**kwargs'
+        else:
+            assert not pos and kw
+            cls._signature = '**kwargs'
         return cls
 
 
 class Model(Component, metaclass=ModelMeta):
     """Base class for components which can predict.
 
-    :param signature: Model signature.
     :param datatype: DataType instance.
     :param output_schema: Output schema (mapping of encoders).
     :param model_update_kwargs: The kwargs to use for model update.
@@ -378,7 +391,6 @@ class Model(Component, metaclass=ModelMeta):
 
     breaks: t.ClassVar[t.Sequence] = ('trainer',)
     type_id: t.ClassVar[str] = 'model'
-    signature: Signature = '*args,**kwargs'
     datatype: EncoderArg = None
     output_schema: t.Optional[Schema] = None
     model_update_kwargs: t.Dict = dc.field(default_factory=dict)
@@ -400,6 +412,10 @@ def __post_init__(self, db, example):
         if not self.identifier:
             raise Exception('_Predictor identifier must be non-empty')
 
+    @property
+    def signature(self):
+        return self._signature
+
     @property
     def inputs(self) -> Inputs:
         """Instance of `Inputs` to represent model params."""
@@ -1039,6 +1055,10 @@ class ImportedModel(Model):
     object: Leaf
     method: t.Optional[str] = None
 
+    def __post_init__(self, db, example):
+        super().__post_init__(db, example)
+        self._inferred_signature = None
+
     @staticmethod
     def _infer_signature(object):
         # find positional and key-word parameters from the object
@@ -1059,6 +1079,13 @@ def _infer_signature(object):
             return '**kwargs'
         return '*args,**kwargs'
 
+    @property
+    @ensure_initialized
+    def signature(self):
+        if self._inferred_signature is None:
+            self._inferred_signature = self._infer_signature(self.object)
+        return self._inferred_signature
+
     @property
     def outputs(self):
         """Get an instance of ``IndexableNode`` to index outputs."""
@@ -1197,6 +1224,7 @@ class QueryModel(Model):
     :param preprocess: Preprocess callable
     :param postprocess: Postprocess callable
     :param select: query used to find data (can include `like`)
+    :param signature: signature to use
     """
 
     preprocess: t.Optional[t.Callable] = None
@@ -1258,10 +1286,13 @@ class SequentialModel(Model):
     models: t.List[Model]
 
     def __post_init__(self, db, example):
-        self.signature = self.models[0].signature
         self.datatype = self.models[-1].datatype
         return super().__post_init__(db, example)
 
+    @property
+    def signature(self):
+        return self.models[0].signature
+
     @property
     def inputs(self) -> Inputs:
         """Instance of `Inputs` to represent model params."""
@@ -1291,7 +1322,22 @@ def predict(self, *args, **kwargs):
         :param args: Positional arguments to predict on.
         :param kwargs: Keyword arguments to predict on.
         """
-        return self.predict_batches([(args, kwargs)])[0]
+        for i, p in enumerate(self.models):
+            assert isinstance(p, Model), f'Expected `Model`, got {type(p)}'
+            if i == 0:
+                out = p.predict(*args, **kwargs)
+            else:
+                if p.signature == 'singleton':
+                    out = p.predict(out)
+                elif p.signature == '*args':
+                    out = p.predict(*out)
+                elif p.signature == '**kwargs':
+                    out = p.predict(**out)
+                else:
+                    msg = 'Model defines a predict with no free parameters'
+                    assert p.signature == '*args,**kwargs', msg
+                    out = p.predict(*out[0], **out[1])
+        return out
 
     def predict_batches(self, dataset: t.Union[t.List, QueryDataset]) -> t.List:
         """Execute on series of data point defined in dataset.
@@ -1358,7 +1404,6 @@ class RAGModel(Model):
     breaks: t.ClassVar[t.Sequence] = ('llm', 'prompt_template')
 
     prompt_template: str
-    signature: str = 'singleton'
     select: Query
     key: str
     llm: Model

superduper/rest/build.py

@@ -235,10 +235,9 @@ def db_remove(
     @app.add('/db/show_template', method='get')
     def db_show_template(
         identifier: str,
-        type_id: str = 'template',
         db: 'Datalayer' = DatalayerDependency(),
     ):
-        template: Template = db.load(type_id=type_id, identifier=identifier)
+        template: Template = db.load(type_id='template', identifier=identifier)
         return template.form_template
 
     @app.add('/db/edit', method='get')
@@ -248,10 +247,9 @@ def db_edit(
         db: 'Datalayer' = DatalayerDependency(),
     ):
         component = db.load(type_id, identifier)
-        template = db.load('template', component['build_template'])
-        template: Template = db.load(type_id=type_id, identifier=identifier)
+        template: Template = db.load('template', component.build_template)
         form = template.form_template
-        form['_variables'] = component['build_variables']
+        form['_variables'] = component.build_variables
         return form
 
     @app.add('/db/metadata/show_jobs', method='get')

test/unittest/base/test_datalayer.py

@@ -473,7 +473,6 @@ def test_replace(db: Datalayer):
         object=lambda x: x + 1,
         identifier='m',
         datatype=DataType(identifier='base'),
-        signature='singleton',
     )
     model.version = 0
     db.apply(model)
@@ -482,7 +481,8 @@ def test_replace(db: Datalayer):
     assert db.load('model', 'm').predict(1) == 2
 
     new_model = ObjectModel(
-        object=lambda x: x + 2, identifier='m', signature='singleton'
+        object=lambda x: x + 2,
+        identifier='m',
     )
     new_model.version = 0
     db.replace(new_model)
@@ -494,7 +494,8 @@ def test_replace(db: Datalayer):
 
     # replace the last version of the model
     new_model = ObjectModel(
-        object=lambda x: x + 3, identifier='m', signature='singleton'
+        object=lambda x: x + 3,
+        identifier='m',
     )
     new_model.version = 0
     db.replace(new_model)

test/unittest/component/test_graph.py

@@ -11,7 +11,7 @@ def model1(db):
     def model_object(x):
         return x + 1
 
-    model = ObjectModel(identifier='m1', object=model_object, signature='singleton')
+    model = ObjectModel(identifier='m1', object=model_object)
     yield model
 
 
@@ -53,13 +53,17 @@ def model_object(x, y):
 
 def test_simple_graph(model1, model2):
     g = Graph(
-        identifier='simple-graph', input=model1, outputs=model2, signature='*args'
+        identifier='simple-graph',
+        input=model1,
+        outputs=model2,
     )
     g.connect(model1, model2)
     assert g.predict(1) == (4, 2)
 
     g = Graph(
-        identifier='simple-graph', input=model1, outputs=model2, signature='*args'
+        identifier='simple-graph',
+        input=model1,
+        outputs=model2,
     )
     g.connect(model1, model2)
     assert g.predict_batches([1, 2, 3]) == [(4, 2), (5, 3), (6, 4)]
@@ -70,7 +74,6 @@ def test_graph_output_indexing(model2_multi_dict, model2, model1):
         identifier='simple-graph',
         input=model1,
         outputs=[model2],
-        signature='**kwargs',
     )
     g.connect(model1, model2_multi_dict, on=(None, 'x'))
     g.connect(model2_multi_dict, model2, on=('x', 'x'))

test/unittest/component/test_model.py

@@ -220,7 +220,8 @@ def fit(self, *args, **kwargs):
 @patch.object(Mapping, '__call__')
 def test_model_validate(mock_call):
     # Check the metadadata recieves the correct values
-    model = Validator('test', object=object())
+    model = Validator('test', object=lambda x: x)
+    model._signature = 'singleton'
     my_metric = MagicMock(spec=Metric)
     my_metric.identifier = 'acc'
     my_metric.return_value = 0.5
@@ -382,13 +383,12 @@ def test_sequential_model():
             ObjectModel(
                 identifier='test-predictor-2',
                 object=lambda x: x + 1,
-                signature='singleton',
             ),
         ],
     )
 
-    assert m.predict(x=1) == 4
-    assert m.predict_batches([((1,), {}) for _ in range(4)]) == [4, 4, 4, 4]
+    assert m.predict(1) == 4
+    assert m.predict_batches([1 for _ in range(4)]) == [4, 4, 4, 4]
 
 
 def test_pm_predict_with_select_ids_multikey(monkeypatch, predict_mixin_multikey):
@@ -439,7 +439,8 @@ def _test(X, docs):
 @pytest.fixture
 def object_model():
     return ObjectModel(
-        'test', object=lambda x: numpy.array(x) + 1, signature='singleton'
+        'test',
+        object=lambda x: numpy.array(x) + 1,
     )
 
 
@@ -468,3 +469,34 @@ def test_object_model_as_a_listener(db, object_model):
     r = results[0].unpack()
     key = next(k for k in r.keys() if k.startswith('_outputs__test'))
     assert all(np.allclose(r.unpack()[key], sample_data + 1) for r in results)
+
+
+class MyModelSingleton(Model):
+    def predict(self, X):
+        return X
+
+
+class MyModelArgs(Model):
+    def predict(self, X, Y):
+        return X
+
+
+class MyModelArgsKwargs(Model):
+    def predict(self, X, Y=None):
+        return X
+
+
+class MyModelKwargs(Model):
+    def predict(self, Y=None):
+        return Y
+
+
+def test_signature_inference():
+    assert MyModelSingleton._signature == 'singleton'
+    assert MyModelArgs._signature == '*args'
+    assert MyModelArgsKwargs._signature == '*args,**kwargs'
+    assert MyModelKwargs._signature == '**kwargs'
+
+    m = ObjectModel('test', object=lambda x, y: x + y)
+
+    assert m.signature == '*args'