feat!: Update Model Architecture with User-interacting Logic

benchmarks/speed_benchmarks/speed_helper.py

@@ -18,7 +18,6 @@
 
 from mithril.models import (
     MLP,
-    BaseModel,
     Convolution2D,
     Flatten,
     MaxPool2D,
@@ -84,7 +83,7 @@ def create_compl_conv(
 def create_compl_mlp(
     input_size: int,
     dimensions: Sequence[int | None],
-    activations: list[type[BaseModel]],
+    activations: list[type[Model]],
 ):
     """Mithril's MLP wrapper with input size
 

mithril/__init__.py

@@ -35,9 +35,10 @@
     short,
 )
 from .framework.codegen import code_gen_map
-from .framework.common import TBD, Connection, IOKey
+from .framework.common import TBD
+from .framework.logical import Connection, IOKey
 from .framework.physical.model import PhysicalConstantType, PhysicalShapeType
-from .models import BaseModel, PhysicalModel
+from .models import Model, PhysicalModel
 from .models.train_model import TrainModel
 
 __all__ = [
@@ -97,7 +98,7 @@
 
 
 def compile(
-    model: BaseModel,
+    model: Model,
     backend: Backend[DataType],
     *,
     constant_keys: PhysicalConstantType[DataType] | None = None,

mithril/framework/codegen/numpy_gen.py

@@ -35,7 +35,7 @@
     find_intersection_type,
     is_type_adjustment_required,
 )
-from ..logical import PrimitiveModel
+from ..logical import Operator
 from .python_gen import PythonCodeGen, RawGradientType
 from .utils import check_repr_inequality
 
@@ -212,7 +212,7 @@
 
     def get_primitive_details(
         self, output_key: str
-    ) -> tuple[PrimitiveModel, list[str], list[str]]:
+    ) -> tuple[Operator, list[str], list[str]]:
         model = self.pm.flat_graph.get_model(output_key)
 
         global_input_keys = self.pm.flat_graph.get_source_keys(output_key)
@@ -229,7 +229,7 @@
 
     def call_primitive(
         self,
-        model: PrimitiveModel,
+        model: Operator,
         fn: Callable[..., Any],
         l_input_keys: list[str],
         g_input_keys: list[str],
@@ -259,7 +259,7 @@
         return ast.Assign(targets, generated_fn), used_keys | _used_keys
 
     def create_primitive_call_targets(
-        self, output_key: str, model: PrimitiveModel, inference: bool
+        self, output_key: str, model: Operator, inference: bool
     ) -> tuple[list[ast.expr | ast.Name], set[str]]:
         targets: list[ast.expr | ast.Name] = []
 
@@ -271,27 +271,27 @@
 
         if not self.pm.inference:
             # TODO: Change this with cache refactor
-            cache_name = output_key + f"_{model.cache_name}"
+            cache_name = output_key + f"_{Operator.cache_name}"
             used_keys.add(cache_name)
             targets.append(
                 ast.Subscript(
                     value=ast.Name(id=cache_name, ctx=ast.Load()),
-                    slice=ast.Constant(value=PrimitiveModel.output_key),
+                    slice=ast.Constant(value=Operator.output_key),
                     ctx=ast.Store(),
                 )
             )
 
         return targets, used_keys
 
-    def get_cache_name(self, output_key: str, model: PrimitiveModel) -> str:
-        cache_name = "_".join([output_key, model.cache_name])
+    def get_cache_name(self, output_key: str, model: Operator) -> str:
+        cache_name = "_".join([output_key, Operator.cache_name])
         if cache_name not in self.pm.flat_graph.all_data:
-            self.add_cache(model, output_key)
+            self.add_cache(output_key)
 
         return cache_name
 
-    def add_cache(self, model: PrimitiveModel, output_key: str) -> None:
-        cache_name = "_".join([output_key, model.cache_name])
+    def add_cache(self, output_key: str) -> None:
+        cache_name = "_".join([output_key, Operator.cache_name])
         cache_value: dict[str, Any] | None = None if self.pm.inference else {}
         # Create a scalar for caches in manualgrad backend.
         self.pm.flat_graph.update_data(

mithril/framework/codegen/python_gen.py

@@ -31,7 +31,7 @@
     EvaluateType,
     ParamsEvalType,
 )
-from ..logical import PrimitiveModel
+from ..logical import Operator
 from ..physical.model import PhysicalModel
 from ..utils import GeneratedFunction
 from .code_gen import CodeGen
@@ -307,7 +307,7 @@ def is_static_scalar(self, key: str) -> bool:
 
     def get_primitive_details(
         self, output_key: str
-    ) -> tuple[PrimitiveModel, list[str], list[str]]:
+    ) -> tuple[Operator, list[str], list[str]]:
         model = self.pm.flat_graph.get_model(output_key)
 
         global_input_keys = self.pm.flat_graph.get_source_keys(output_key)
@@ -317,7 +317,7 @@ def get_primitive_details(
 
     def call_primitive(
         self,
-        model: PrimitiveModel,
+        model: Operator,
         fn: Callable[..., Any],
         l_input_keys: list[str],
         g_input_keys: list[str],
@@ -354,7 +354,6 @@ def generate_evaluate(self) -> ast.FunctionDef:
         for output_key in self.pm.flat_graph.topological_order:
             model, g_input_keys, l_input_keys = self.get_primitive_details(output_key)
             formula_key = model.formula_key
-
             primitive_function = (
                 self.pm.backend.primitive_function_dict[formula_key]
                 if formula_key in self.pm.backend.primitive_function_dict
@@ -568,7 +567,7 @@ def create_primitive_call(
         return generated_fn, used_keys
 
     def create_primitive_call_targets(
-        self, output_key: str, model: PrimitiveModel, inference: bool
+        self, output_key: str, model: Operator, inference: bool
     ) -> tuple[list[ast.expr], set[str]]:
         if (
             keyword.iskeyword(output_key)

mithril/framework/codegen/torch_gen.py

@@ -19,7 +19,7 @@
 import torch
 
 from ...backends.with_autograd.torch_backend import TorchBackend
-from ..logical import PrimitiveModel
+from ..logical import Operator
 from ..physical.model import PhysicalModel
 from .python_gen import PythonCodeGen
 
@@ -34,7 +34,7 @@ def __init__(self, pm: PhysicalModel[torch.Tensor]) -> None:
 
     def call_primitive(
         self,
-        model: PrimitiveModel,
+        model: Operator,
         fn: Callable[..., Any],
         l_input_keys: list[str],
         g_input_keys: list[str],