Fusion Support
Currently, this tool mainly implements the following forms of BN fusion (single-operators involved in the following fusion forms must meet the restrictions described in Quantization):
- Conv+BN+Scale fusion: Before AMCT quantization, "Conv+BN+Scale" fusion is performed on the "Convolution+BatchNorm+Scale" composite in the model. The BatchNorm and Scale layers are removed.
- Deconv+BN+Scale fusion: Before AMCT quantization, "Deconv+BN+Scale" fusion is performed on the "Deconvolution+BatchNorm+Scale" composite in the model. The BatchNorm and Scale layers are removed.
- BN+Scale+Conv fusion: applicable only to PTQ. Before AMCT quantization, "BN+Scale+Conv" fusion is performed on the "BatchNorm+Scale+Convolution" composite in the model. The BatchNorm and Scale layers are removed.
- FC+BN+Scale fusion: applicable only to PTQ. Before AMCT quantization, "FC+BN+Scale" fusion is performed on the "InnerProduct+BatchNorm+Scale" composite in the model. The BatchNorm and Scale layers are removed.
Parent topic: Reference