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 conv node has dynamic weight inputs, fusion does not apply.

Availability

Atlas inference product

Atlas 200I/500 A2 inference product

Atlas training product