The ArmNN Converter and Quantizer

Table of Contents

• The ArmNN Converter
• The ArmNN Quantizer

The ArmNN Converter

The ArmnnConverter is a program for converting neural networks from other formats to Arm NN format. Currently the program supports models in Caffe, Onnx, Tensorflow Protocol Buffers and Tensorflow Lite FlatBuffers formats. Run the program with no arguments to see command-line help.

For more information about the layers that are supported, see parsers.

The ArmNN Quantizer

The ArmnnQuantizer is a program for loading a 32-bit float network into ArmNN and converting it into a quantized asymmetric 8-bit or quantized symmetric 16-bit network. It supports static quantization by default, dynamic quantization is enabled if CSV file of raw input tensors is provided. Run the program with no arguments to see command-line help.

Cmd:
-h	–help	Display help messages
-f	–infile	Input file containing float 32 ArmNN Input Graph
-s	–scheme	Quantization scheme, "QAsymm8" or "QSymm16". Default value: QAsymm8
-c	–csvfile	CSV file containing paths for raw input tensors for dynamic quantization. If unset, static quantization is used
-p	–preserve-data-type	Preserve the input and output data types. If unset, input and output data types are not preserved
-d	–outdir	Directory that output file will be written to
-o	–outfile	ArmNN output file name

./ArmnnQuantizer -f /path/to/armnn/input/graph/ -s "QSymm16" -c /path/to/csv/file -p 1 -d /path/to/output -o outputFileName