neural_compressor.adaptor.tf_utils.graph_rewriter.generic.graph_cse_optimizer
CSE Graph Rewriter.
Module Contents
Classes
We introduce the CSE optimizer to optimize the nodes that contains identical op type. |
- class neural_compressor.adaptor.tf_utils.graph_rewriter.generic.graph_cse_optimizer.GraphCseOptimizer(model)[source]
We introduce the CSE optimizer to optimize the nodes that contains identical op type.
- Case 1. Node A has three output nodes(B,C,D) and those child nodes has their own outputs
(B1/C1C2/D1). Node A x x x x x x
NODE B NODE C NODE D x x x x B1 C1 C2 D1 If Node B/C/D have the identical memory-bound op, like relu/relu6. The graph will be converted as below. We removed the Node C & Node D, updated the B as the input of C1/C2/D1.
Node A x Node B x x x x x x x x
x x x x B1 C1 C2 D1
- Case 2. Node A has three output nodes(B,C,D) and those child nodes has their own outputs
(B1/C1C2/D1). Node A x x x x x x
- NODE B NODE C NODE D
x x x x B1 C1 C2 D1
If Node B and C have the identical memory-bound op, like relu/relu6. The graph will be converted as below. We removed the Node C, updated the B as the input of C1/C2.
Node A x x
Node B Node D x x x x x | x x x | x x B1 C1 C2 D1
- Returns:
A optimized graphdef object.
- Return type:
[graphdef]