AConv2dMulFusion

Description

Fuses the Conv2d+mul or Conv3d+mul operators into one Conv operator.

Before: After:

Restrictions

• The conv node can be Conv2D or Conv3D.
• When the data node has dynamic inputs, fusion is supported.
• When the three inputs of filter, bias, and mul are all of const type, fusion is supported.
• When the conv node is a Conv3D node, the data input must be in NDHWC format. In this case, the const mul input of the Mul operator is a 1D input, and the dimension is equal to the C dimension of the Conv3D operator output.
• When the conv node is a Conv3D node, the data input must be in NDHWC format. In this case, the const mul input of the mul operator must also be in NDHWC format, and the NDHW dimensions must be 1. The fusion is performed only when the C dimension is equal to the C dimension output by the Conv3D operator.

Availability

Atlas 200/300/500 Inference Product

Atlas Training Series Product