ConvBatchnormFusionPass

Description

Fuses the Conv2d+batchnorm or Conv3d+batchnorm operators into one operator; or fuses the DepthwiseConv2d+batchnorm operators into one operator when batchnorm has two inputs.

Before:

After:

Restrictions

• The conv node can be Conv2D or Conv3D. When the batchnorm node has only two inputs (mean and variance), the conv node can also be DepthwiseConv2D.
• The batchnorm node can be batchnorm or BNInference.
• The filter and bias nodes must be of const type. If the filter node is QuantWeightRollBack, fusion fails.
• The batchnorm node can have two inputs (mean and variance) or four inputs (as shown in the preceding figure). All inputs must be of const type. Otherwise, fusion fails.
• When the data node has dynamic inputs, fusion is supported.
• When the weight input of the convolution node is dynamic, fusion is not performed.

Availability

Atlas 200/300/500 Inference Product

Atlas Training Series Product