Dynamic OV model builder #3137
Dynamic OV model builder #3137
Conversation
github-actions
bot
added
NNCF OpenVINO
This reverts commit abf07e9.
| # We have to have args as the inputs since all of them are nodes and are required to be as input. | ||
| args = [clone_nodes.popleft() for _ in graph_node.inputs()] | ||
|
|
||
| clone_node = self._node_factory.create(node_type, args, attrs) |
There was a problem hiding this comment.
Currently, NodeFactory will create nodes without names since naming is done using a decorator (see https://github.com/openvinotoolkit/openvino/blob/master/src/bindings/python/src/openvino/utils/decorators.py#L25-L34).
It shouldn't be a problem, since the FastBC algorithm utilizes Parameter and Result nodes for statistics. But for BiasCorrection algorithm utilizes the names.
The current approach may be changed by adding the corresponding constructor calls into metatypes, for example, for each layer.
| if node_mapping is None: | ||
| node_mapping = {op.get_friendly_name(): op for op in model.get_ops()} |
There was a problem hiding this comment.
This needs to avoid failures for the algorithms that utilize get_bias_value, but without mapping for now.
| def extract_submodel(self, model_transformer, input_id, output_id): | ||
|
|
||
| return self._model_builder.build( | ||
| input_ids=[input_id], | ||
| output_ids=[output_id], | ||
| node_mapping=self._node_mapping, | ||
| ) |
There was a problem hiding this comment.
To keep the original behaviour of the other backends and reduce code changing, the overloading approach was used here.
|
@alexsu52, @andrey-churkin, could you please review not PR (need to discuss a number of questions), but the proposal of the new graph-building approach (for FastBiasCorrection and as the possible extension of the BiasCorrection)? Thanks. |
Changes
ModelBuilderclass.Reason for changes
Related tickets
Tests