ASplitConv2dConcatPass

Description

Fuses the Split/SplitV, Conv2D*N, and Concat/ConcatV2 operators into a group convolution to simplify the graph structure.

Restrictions

• The number of inputs of each Conv2D must be the same and greater than or equal to 2.
• The shape and format of each Conv2D filter must be the same (only HWCN or NCHW is supported).
• Each Conv2D supports only one output.
• The filter and bias of Conv2D must be of any of the following types: {"Const", "Constant", "QuantBiasOptimization", "QuantWeightRollBack", "QuantBiasRollBack", "AscendWeightQuant"}.
• The Split/SplitV node and the other input of the Concat/ConcatV2 node must be of const type, and the number of outputs of the Split/SplitV node must be the same as the number of inputs of the Concat/ConcatV2 node.
• The split axis of the Split/SplitV node and the combination axis of the Concat/ConcatV2 node must both be the channel dimension.
• Conv2D supports only float, float16, int8, and int32 inputs, and does not allow dynamic inputs. This restriction applies to the following models:
  • Atlas inference product
  • Atlas 200I/500 A2 inference product
  • Atlas training product
• Conv2D supports only float, float16, and int8 inputs. This restriction applies to the following models:
  • Atlas 350 Accelerator Card

Availability

Atlas inference product

Atlas 200I/500 A2 inference product

Atlas training product

Atlas 350 Accelerator Card