w-okada / yolov7-onnx-test

YOLOv7-ONNX-TEST

This repository is intended only to evaluate the performance of the yolov7 onnx model. Model type is only Tiny.

1. python onnx (cpu)

Prerequisite

This repository use docker.

Experiment

python onnx (cpu)

This code is used.

start_time = time.time()
for i in range(10):
    outputs = session.run(outname, inp)[0]
elapsed_time = time.time() - start_time
print(img_size, 'fin. avr time:', (elapsed_time / 10) * 1000, "msec")

Result

Result on "Intel(R) Core(TM) i9-9900KF CPU @ 3.60GHz"

Size	from	python onnx (cpu)
320x320	Official	9.98msec
640x640	Official	39.14msec
1280x1280	Official	183.11msec
1920x1920	Official	411.53msec
256x320	PINTO (*1)	(*2)
256x480	PINTO	11.26msec
256x640	PINTO	14.43msec
348x640	PINTO	23.73msec
480x640	PINTO	28.63msec
640x640	PINTO	44.12msec
736x1280	PINTO	100.65msec

(*1) PINTO model includes postprocess. link (*2) seems not to work correctly.

Operations

You can reproduce the experiment according to the following way.

build docker

$ npm run build:docker

export onnx model

Only yolox_nano.pth is supported.

$ npm run export:yolov7

run python onnx

$ npm run run:yolov7_onnx
$ npm run run:yolov7_onnx_pinto

Wait for a while. Then you can see the output on the termianl.

Misc

web demo only for onnx.

for tfjs is under construction. (stacked. technical problem occured.)

yolov7_onnx3.mp4

Reference

1. https://github.com/WongKinYiu/yolov7
2. https://github.com/PINTO0309/PINTO_model_zoo/tree/main/307_YOLOv7