tutorials

Tutorials

This directory contains end-to-end tutorials that show how to generate a binary file for a specific model using the VectorBlox SDK. Tutorials are organized by model source. For a detailed breakdown of three of our tutorials, please refer to the Tutorial Walkthrough Guide.

Tutorials Overview

Each tutorial includes a shell script that demonstrates the complete pipeline:

1. Download the source model.
2. Convert the model to TensorFlow Lite (if needed) using one of the following: openvino2tensorflow, onnx2tf, or tflite_quantize.
3. Run tflite_preprocess to add a preprocessing layer before generating the binary.
4. Generate the binary with VectorBlox's graph generation tool: vnnx_compile.
5. Simulate and run inference using the Python script in example/python with the provided test image.

These scripts illustrate the complete pipeline for generating a model's binary file. Users may use or modify these scripts to generate a binary file for their own custom model, tailored to their use cases.

To run a tutorial script, change into the model's tutorial directory and run:

(vbx_env) ~/SDK/VectorBlox-SDK/tutorials/SOURCE_NAME/MODEL_NAME$ bash MODEL_NAME.sh

Tutorial Metrics

• Runtime in milliseconds (ms) measured on SoC Video Kit. Accuracy is measured over 1000 samples.
• For a complete list of metrics for every VectorBlox tutorial, please visit the Tutorial Metrics Appendix Markdown Page in the docs folder.

No Compression (subset)

Source	Tutorial	Input (H,W,C)	Runtime (ms)	Task	Metric	TFLITE	VNNX
PINTO	081_MiDaS_v2	[256, 256, 3]	121.921	depth estimation	depthdelta1 (nyuv2)	60.89	59.93
kaggle	efficientnet-lite0	[224, 224, 3]	16.481	classification	Top1	70.8	70.4
onnx	onnx_resnet18-v1	[224, 224, 3]	25.741	classification	Top1	69.1	68.9
openvino	mobilenet-v1-1.0-224	[224, 224, 3]	11.811	classification	Top1	70.1	70.1
qualcomm	FFNet-122NS-LowRes_512x288	[288, 512, 3]	76.168	segmentation	meanIoU (cityscapes)	44.82	44.45
qualcomm	MobileNet-v3-Large-Quantized	[224, 224, 3]	22.895	classification	Top1	69.6	68.8
qualcomm	QuickSRNetMedium-Quantized	[128, 128, 3]	9.353	image enhancement	PSNR (bsd300)	26.79	26.81
tensorflow	mobilenet_v2	[224, 224, 3]	12.875	classification	Top1	70.2	70.1
ultralytics	yolov5n	[640, 640, 3]	38.817	object detection	mAP⁵⁰⁻⁹⁵	22.88	22.95
ultralytics	yolov8n	[640, 640, 3]	54.129	object detection	mAP⁵⁰⁻⁹⁵	37.4	37.38
ultralytics	yolov8n-cls	[224, 224, 3]	4.104	classification	Top1	67.2	67.3
ultralytics	yolov8n-obb	[1024, 1024, 3]	142.268	obb detection	mAP⁵⁰⁻⁹⁵
ultralytics	yolov8n-pose_512x288_split	[288, 512, 3]	21.998	pose detection	Pose Detection
ultralytics	yolov8n-seg	[640, 640, 3]	70.769	instance segmentation
ultralytics	yolov9t	[640, 640, 3]	65.068	object detection	mAP⁵⁰⁻⁹⁵	37.92	38.18

Compression

Source	Tutorial	Input (H,W,C)	Runtime (ms)	Task	Metric	TFLITE	VNNX
compressed	yolov8n_comp66	[640, 640, 3]	35.85	object detection	mAP⁵⁰⁻⁹⁵	37.18	37.3
compressed	yolov8s_comp68	[640, 640, 3]	90.812	object detection	mAP⁵⁰⁻⁹⁵	46.11	46.15

Unstructured Compression

Source	Tutorial	Input (H,W,C)	Runtime (ms)	Task	Metric	TFLITE
unstructure_compressed	resnet18_86s_07p	[224, 224, 3]	14.63	classification	Top1	65.6
unstructure_compressed	yolov5n_70s_512x512	[512, 512, 3]	19.779	object detection	mAP⁵⁰⁻⁹⁵	20.2
unstructure_compressed	yolov8n_50s_25p_512x288	[288, 512, 3]	15.428	object detection	mAP⁵⁰⁻⁹⁵	34.6
unstructure_compressed	yolov8n_50s_25p_512x512	[512, 512, 3]	25.265	object detection	mAP⁵⁰⁻⁹⁵
unstructure_compressed	yolov8n_pose_50s_25p_512x288_split	[288, 512, 3]	15.507	pose detection
unstructure_compressed	yolov9s_70s_15p_512x288	[288, 512, 3]	32.95	object detection		43.2
unstructure_compressed	yolov9s_70s_15p_512x512	[512, 512, 3]	55.576	object detection