WIP temporary insertion command

nncf/torch/dynamic_graph/context.py

@@ -11,6 +11,7 @@
 
 import threading
 import weakref
+from collections import OrderedDict
 from collections import defaultdict
 from collections import deque
 from contextlib import contextmanager
@@ -92,9 +93,9 @@ class TracingContext:
     def __init__(self):
         self.graph = DynamicGraph()
 
-        self._post_hooks = defaultdict(list)
-        self._pre_hooks: Dict[PreHookId, List[Callable]] = defaultdict(list)
-        self._num_nested_hooks = 0
+        self._post_hooks = defaultdict(OrderedDict)
+        self._pre_hooks: Dict[PreHookId, List[Callable]] = defaultdict(OrderedDict)
+        self._hooks_counter = 0
 
         self._threading = CopySafeThreadingVars()
 
@@ -260,9 +261,16 @@ def pop_scope(self):
         self.relative_scopes_stack.pop()
         self.module_call_stack.pop()
 
-    def register_pre_hooks(self, fn_list: List[Callable], op_address: OperationAddress, input_port_id: int):
+    def register_pre_hooks(
+        self, fn_list: List[Callable], op_address: OperationAddress, input_port_id: int
+    ) -> List[int]:
         pre_hook_id = PreHookId(op_address, input_port_id)
-        self._pre_hooks[pre_hook_id].extend(fn_list)
+        hooks_ids = []
+        for fn in fn_list:
+            self._hooks_counter += 1
+            self._pre_hooks[pre_hook_id][self._hooks_counter] = fn
+            hooks_ids.append(self._hooks_counter)
+        return pre_hook_id, hooks_ids
 
     def execute_pre_hooks(self, op_address: OperationAddress, op_inputs: OperatorInput) -> OperatorInput:
         in_op = getattr(self, "in_operator", False)
@@ -274,21 +282,26 @@ def execute_pre_hooks(self, op_address: OperationAddress, op_inputs: OperatorInp
         for pre_hook_id in pre_hook_ids_for_curr_op:
             hook_list_for_current_input_port = self._pre_hooks[pre_hook_id]
             input_arg_to_process = pre_hook_id.input_port_id
-            for hook in hook_list_for_current_input_port:
+            for hook in hook_list_for_current_input_port.values():
                 op_inputs[input_arg_to_process] = hook(op_inputs[input_arg_to_process])
         self._threading.thread_local.num_nested_hooks -= 1
         self.in_operator = in_op
         return op_inputs
 
-    def register_post_hooks(self, fn_list: List[Callable], op_address: OperationAddress):
-        self._post_hooks[op_address].extend(fn_list)
+    def register_post_hooks(self, fn_list: List[Callable], op_address: OperationAddress) -> List[int]:
+        hooks_ids = []
+        for fn in fn_list:
+            self._hooks_counter += 1
+            self._post_hooks[op_address][self._hooks_counter] = fn
+            hooks_ids.append(self._hooks_counter)
+        return op_address, hooks_ids
 
     def execute_post_hooks(self, op_address: OperationAddress, outputs):
         in_op = getattr(self, "in_operator", False)
         self.in_operator = False
         self._threading.thread_local.num_nested_hooks += 1
         if op_address in self._post_hooks:
-            for hook in self._post_hooks[op_address]:
+            for hook in self._post_hooks[op_address].values():
                 outputs = hook(outputs)
         self._threading.thread_local.num_nested_hooks -= 1
         self.in_operator = in_op

nncf/torch/graph/transformations/commands.py

@@ -153,6 +153,35 @@ def requires_graph_rebuild(self):
         return self.priority == TransformationPriority.QUANTIZATION_PRIORITY
 
 
+class PTInsertionTemporaryCommand(PTTransformationCommand):
+    """
+    Insertion operation to the models.
+    """
+
+    def __init__(
+        self,
+        point: PTTargetPoint,
+        fn: Callable,
+        priority: TransformationPriority = TransformationPriority.DEFAULT_PRIORITY,
+    ):
+        super().__init__(TransformationType.INSERT, point)
+        self.fn: Callable = fn
+        self.priority: TransformationPriority = priority
+
+    def union(self, other: "PTTransformationCommand") -> "PTTransformationCommand":
+        # TODO: keep all TransformationCommands atomic, refactor TransformationLayout instead
+        raise NotImplementedError()
+
+    def requires_graph_rebuild(self):
+        """
+        Return boolean flag to rebuild graph of model.
+
+        :return: Boolean flag.
+        """
+        # Rebuild graph when adding quantization nodes or an op.
+        return False
+
+
 class PTSharedFnInsertionCommand(PTTransformationCommand):
     def __init__(
         self,

nncf/torch/model_transformer.py

@@ -11,7 +11,7 @@
 
 import copy
 from collections import defaultdict
-from typing import Callable, Dict, List, Tuple
+from typing import Callable, Dict, Iterator, List, Tuple
 
 from torch import Tensor
 from torch import nn
@@ -24,6 +24,7 @@
 from nncf.torch.external_hook import EXTERNAL_OP_STORAGE_NAME
 from nncf.torch.graph.transformations.commands import PTBiasCorrectionCommand
 from nncf.torch.graph.transformations.commands import PTInsertionCommand
+from nncf.torch.graph.transformations.commands import PTInsertionTemporaryCommand
 from nncf.torch.graph.transformations.commands import PTModelExtractionWithFusedBiasCommand
 from nncf.torch.graph.transformations.commands import PTQuantizerInsertionCommand
 from nncf.torch.graph.transformations.commands import PTSharedFnInsertionCommand
@@ -52,6 +53,7 @@ def __init__(self, model: NNCFNetwork):
         self._command_transformation_ordered_pairs = [
             (PTModelExtractionWithFusedBiasCommand, self._apply_extraction_with_fused_bias_transformations),
             (PTInsertionCommand, self._apply_insertion_transformations),
+            (PTInsertionTemporaryCommand, self._apply_temporary_insertion_transformation),
             (PTQuantizerInsertionCommand, self._apply_quantizer_insertion_transformations),
             (PTBiasCorrectionCommand, self._apply_bias_correction_transformations),
             (PTSharedFnInsertionCommand, self._apply_shared_nodes_insertion),
@@ -82,6 +84,33 @@ def _apply_insertion_transformations(model: NNCFNetwork, transformations: List[P
         """
         Applies insertion transformations to the model.
 
+        :param model: Model to apply transformations.
+        :param transformations: List of the bias correction transformations.
+        """
+        for insert_args in PTModelTransformer._get_nncf_network_insert_arguments(model, transformations):
+            model.nncf.insert_at_point(*insert_args)
+        return model
+
+    @staticmethod
+    def _apply_temporary_insertion_transformation(
+        model: NNCFNetwork, transformations: List[PTInsertionCommand]
+    ) -> NNCFNetwork:
+        """
+        Applies temporary insertion transformations to the model.
+
+        :param model: Model to apply transformations.
+        :param transformations: List of the bias correction transformations.
+        """
+        for insert_args in PTModelTransformer._get_nncf_network_insert_arguments(model, transformations):
+            model.nncf.temporary_insert_at_point(*insert_args)
+        return model
+
+    def _get_nncf_network_insert_arguments(
+        model: NNCFNetwork, transformations: List[PTInsertionCommand]
+    ) -> Iterator[Tuple[PTInsertionPoint, List[Callable]]]:
+        """
+        Applies insertion transformations to the model.
+
         :param model: Model to apply transformations.
         :param transformations: List of the bias correction transformations.
         """
@@ -107,8 +136,7 @@ def _apply_insertion_transformations(model: NNCFNetwork, transformations: List[P
 
         for pt_ip, fn_list_with_priority in fns_grouped_by_points.items():
             fn_list_with_priority = sorted(fn_list_with_priority, key=lambda x: x[1])
-            model.nncf.insert_at_point(pt_ip, [x[0] for x in fn_list_with_priority])
-        return model
+            yield (pt_ip, [x[0] for x in fn_list_with_priority])
 
     @staticmethod
     def _apply_shared_nodes_insertion(

nncf/torch/nncf_network.py

@@ -17,7 +17,7 @@
 from copy import deepcopy
 from enum import Enum
 from enum import IntEnum
-from typing import Callable, Dict, Iterator, List, Optional, Tuple, TypeVar
+from typing import Any, Callable, Dict, Iterator, List, Optional, Tuple, TypeVar
 
 import torch
 from torch import nn
@@ -224,6 +224,7 @@ def __init__(
         self._target_scopes = target_scopes
         self._user_dummy_forward_fn = dummy_forward_fn
         self._kd_loss_handler = None
+        self._temprorary_hooks_adresses = []
 
         if wrap_inputs_fn is not None:
             self._wrap_inputs_fn = wrap_inputs_fn
@@ -397,11 +398,36 @@ def get_modules_in_nncf_modules_by_type(self, class_names: List[str]) -> Dict[Sc
     def update_model_ref(self, model: torch.nn.Module) -> None:
         object.__setattr__(self, "__model_ref", model)
 
+    def temporary_insert_at_point(self, point: PTInsertionPoint, fn_list: List[Callable]):
+        hook_addresses = self.insert_at_point(point, fn_list)
+        self._temprorary_hooks_adresses.append(hook_addresses)
+        return hook_addresses
+
+    def remove_temporary_ops(self):
+        for point, hook_address, hook_ids in self._temprorary_hooks_adresses:
+            for hook_idx in hook_ids:
+                if point.insertion_type == PTInsertionType.OPERATOR_PRE_HOOK:
+                    hooks = self._compressed_context._pre_hooks[hook_address]
+                elif point.insertion_type == PTInsertionType.OPERATOR_POST_HOOK:
+                    hooks = self._compressed_context._post_hooks[hook_address]
+                else:
+                    nncf_module = self.get_module_by_scope(point.module_scope)
+                    if point.insertion_type == PTInsertionType.NNCF_MODULE_PRE_OP:
+                        hooks = nncf_module.pre_ops
+                    else:
+                        hooks = nncf_module.post_ops
+
+                del hooks[hook_idx]
+        self._temprorary_hooks_adresses.clear()
+
     def insert_at_point(self, point: PTInsertionPoint, fn_list: List[Callable]):
+        hooks_ids = None
         if point.insertion_type == PTInsertionType.OPERATOR_PRE_HOOK:
-            self._compressed_context.register_pre_hooks(fn_list, point.op_address, point.input_port_id)
+            hook_address, hooks_ids = self._compressed_context.register_pre_hooks(
+                fn_list, point.op_address, point.input_port_id
+            )
         elif point.insertion_type == PTInsertionType.OPERATOR_POST_HOOK:
-            self._compressed_context.register_post_hooks(fn_list, point.op_address)
+            hook_address, hooks_ids = self._compressed_context.register_post_hooks(fn_list, point.op_address)
         elif point.insertion_type in [PTInsertionType.NNCF_MODULE_PRE_OP, PTInsertionType.NNCF_MODULE_POST_OP]:
             nncf_module = self.get_module_by_scope(point.module_scope)
             if not isinstance(nncf_module, _NNCFModuleMixin):
@@ -419,14 +445,18 @@ def insert_at_point(self, point: PTInsertionPoint, fn_list: List[Callable]):
             for scope_list_for_module in self.get_nncf_module_scopes():
                 norm_nncf_scopes.extend([self._normalize_variable_recurrent_scope(x) for x in scope_list_for_module])
             assert norm_target_scope in norm_nncf_scopes  # Required for proper Recurrent/VariableRecurrent addressing
+            hooks_ids = []
             if point.insertion_type == PTInsertionType.NNCF_MODULE_PRE_OP:
                 for fn in fn_list:
-                    nncf_module.register_pre_forward_operation(fn)
+                    hooks_ids.append(nncf_module.register_pre_forward_operation(fn))
             elif point.insertion_type == PTInsertionType.NNCF_MODULE_POST_OP:
                 for fn in fn_list:
-                    nncf_module.register_post_forward_operation(fn)
+                    hooks_ids.append(nncf_module.register_post_forward_operation(fn))
+            hook_address = None
         else:
             raise RuntimeError("Unsupported insertion type: {}".format(point.insertion_type))
+        hook_addresses = (point, hook_address, hooks_ids)
+        return hook_addresses
 
     def get_graph(self) -> PTNNCFGraph:
         if self._compressed_context.graph.get_nodes_count() == 0 or self._compressed_graph is None:
@@ -772,6 +802,14 @@ def strip(self, do_copy: bool = True) -> "NNCFNetwork":
         return self.compression_controller.strip(do_copy)
 
 
+class TemporaryOp:
+    def __init__(self, op: Callable) -> None:
+        self._op = op
+
+    def __call__(self, *args: Any, **kwargs: Any) -> Any:
+        return self._op(*args, **kwargs)
+
+
 class NNCFNetworkMeta(type):
     """
     Metaclass for the NNCFNetwork mixin. Has magic methods defined so that the original model object could be

nncf/torch/statistics/aggregator.py

@@ -21,7 +21,7 @@
 from nncf.common.graph.transformations.layout import TransformationLayout
 from nncf.common.tensor_statistics.aggregator import StatisticPointsContainer
 from nncf.common.tensor_statistics.aggregator import StatisticsAggregator
-from nncf.torch.graph.transformations.commands import PTInsertionCommand
+from nncf.torch.graph.transformations.commands import PTInsertionTemporaryCommand
 from nncf.torch.nncf_network import NNCFNetwork
 from nncf.torch.tensor import PTNNCFTensor
 from nncf.torch.tensor_statistics.algo import create_register_input_hook
@@ -60,8 +60,10 @@ def __exit__(self, exc_type, exc_val, exc_tb):
 class PTStatisticsAggregator(StatisticsAggregator):
     def collect_statistics(self, model: NNCFNetwork, graph: NNCFGraph) -> None:
         with torch.no_grad():
-            with ModelView(model) as intermediate_model:
-                super().collect_statistics(intermediate_model, graph)
+            # with ModelView(model) as intermediate_model:
+            #    super().collect_statistics(intermediate_model, graph)
+            super().collect_statistics(model, graph)
+            model.nncf.remove_temporary_ops()
 
     def _register_statistics(
         self, outputs: Dict[str, PTNNCFTensor], statistic_points: StatisticPointsContainer
@@ -79,7 +81,7 @@ def _get_transformation_layout_extra_outputs(
                 for collectors in _statistic_point.algorithm_to_tensor_collectors.values():
                     for collector in collectors:
                         transformation_commands.append(
-                            PTInsertionCommand(
+                            PTInsertionTemporaryCommand(
                                 _statistic_point.target_point,
                                 create_register_input_hook(collector=collector),
                                 TransformationPriority.FP32_TENSOR_STATISTICS_OBSERVATION,

tests/torch/helpers.py

@@ -561,7 +561,7 @@ def _check_weight_update_hooks(ref_hooks):
     def _check_pre_post_hooks(hooks, ref_hooks):
         assert len(hooks) == len(ref_hooks)
         for op_address, ref_hooks in ref_hooks.items():
-            actual_hooks = hooks[op_address]
+            actual_hooks = hooks[op_address].values()
             assert len(actual_hooks) == len(ref_hooks)
             for actual_hook, ref_hook in zip(actual_hooks, ref_hooks):
                 assert actual_hook is ref_hook