Other Tools

Table of Contents

• The ImageCSVFileGenerator
• The ImageTensorGenerator
• The ModelAccuracyTool-ArmNN

The ImageCSVFileGenerator

The ImageCSVFileGenerator is a program for creating a CSV file that contains a list of .raw tensor files. These .raw tensor files can be generated using theImageTensorGenerator.

Cmd:
-h	–help	Display help messages
-i	–indir	Directory that .raw files are stored in
-o	–outfile	Output CSV file path

Example usage:

./ImageCSVFileGenerator -i /path/to/directory/ -o /output/path/csvfile.csv

The ImageTensorGenerator

The ImageTensorGenerator is a program for pre-processing a .jpg image before generating a .raw tensor file from it.

Build option: To build ModelAccuracyTool, pass the following options to Cmake: -DBUILD_ARMNN_QUANTIZER=1

Cmd:
-h	–help	Display help messages
-f	–model-format	Format of the intended model file that uses the images.Different formats have different image normalization styles.Accepted values (caffe, tensorflow, tflite)
-i	–infile	Input image file to generate tensor from
-o	–outfile	Output raw tensor file path
-z	–output-type	The data type of the output tensors.If unset, defaults to "float" for all defined inputs. Accepted values (float, int or qasymm8)
	–new-width	Resize image to new width. Keep original width if unspecified
	–new-height	Resize image to new height. Keep original height if unspecified
-l	–layout	Output data layout, "NHWC" or "NCHW". Default value: NHWC

Example usage:

.sh ./ImageTensorGenerator -i /path/to/image/dog.jpg -o /output/path/dog.raw --new-width 224 --new-height 224

The ModelAccuracyTool-ArmNN

The ModelAccuracyTool-Armnn is a program for measuring the Top 5 accuracy results of a model against an image dataset.

Prerequisites:

1. The model is in .armnn format model file. The ArmnnConverter can be used to convert a model to this format.

Build option: To build ModelAccuracyTool, pass the following options to Cmake: -DFLATC_DIR=/path/to/flatbuffers/x86build/ -DBUILD_ACCURACY_TOOL=1 -DBUILD_ARMNN_SERIALIZER=1

Cmd:
-h	–help	Display help messages
-m	–model-path	Path to armnn format model file
-f	–model-format	The model format. Supported values: caffe, tensorflow, tflite
-i	–input-name	Identifier of the input tensors in the network separated by comma
-o	–output-name	Identifier of the output tensors in the network separated by comma
-d	–data-dir	Path to directory containing the ImageNet test data
-p	–model-output-labels	Path to model output labels file.
-v	–validation-labels-path	Path to ImageNet Validation Label file
-l	–data-layout ]	Data layout. Supported value: NHWC, NCHW. Default: NHWC
-c	–compute	Which device to run layers on by default. Possible choices: CpuRef, CpuAcc, GpuAcc. Default: CpuAcc, CpuRef
-r	–validation-range	The range of the images to be evaluated. Specified in the form <begin index>="">:<end index>="">. The index starts at 1 and the range is inclusive. By default the evaluation will be performed on all images.
-b	–blacklist-path	Path to a blacklist file where each line denotes the index of an image to be excluded from evaluation.

Example usage:

./ModelAccuracyTool -m /path/to/model/model.armnn -f tflite -i input -o output -d /path/to/test/directory/ -p /path/to/model-output-labels -v /path/to/file/val.txt -c CpuRef -r 1:100