Preserve the constant subgraph

[andrey-churkin]

Reason for changes

Removing the ShapeOf subgraph may lead to the removal of some input edges of operations, and QPS doesn't place FQ on these edges. In this case, the operations are quantized incorrectly because not all required inputs are quantized. These operations don't change their runtime precision in the execution graph, which means we do not have a fully quantized model.

Changes

To resolve the described problem, this PR introduces the following algorithm:

• Step 1: Find all nodes belonging to the ShapeOf subgraphs in the graph. Let's denote the found nodes as $S$. At this step, we don't remove the found subgraphs; we only store the nodes they consist of. See find_shapeof_subgraphs() method.

• Step 2: Check the producers for the specified nodes (currently, we are considering only nodes of the LSTMSequence and Convolution types). If any of these producers are in set $S$, it means that we will lose this input edge when the ShapeOf subgraph is removed. Therefore, we try to exclude float subgraphs from this ShapeOf subgraph. Let's denote the nodes that are part of such float subgraphs as $P$. We can see that $P \subset S$. See find_preserved_nodes() method.

• Step 3: Find all nodes belonging to the Constant subgraphs in the graph. Let's denote the found nodes as $C$. At this step, we don't remove the found subgraphs; we only store the nodes they consist of. See find_constant_subgraphs() method.

• Step 4: Remove $S \cup C - P$ nodes from the graph.

Related tickets

• 146088

[andrey-churkin]

post_training_quantization Build # 456

[alexsu52]

LGTM