span

Swift Parameter-free Attention Network for Efficient Super-Resolution

Input

(Image from https://github.com/JingyunLiang/SwinIR/tree/main/testsets)

Ailia input shape : (1, 3, 256, 256)

Output

Ailia output shape : (1, 3, 256 * scale, 256 * scale)

default : scale=2

Usage

Automatically downloads the onnx and prototxt files when running. It is necessary to be connected to the Internet while downloading.

$ python3 span.py

If you want to specify the input image, put the image path after the --input option.
You can use --savepath option to cspange the name of the output file to save.

$ python3 span.py --input IMAGE_PATH --savepath SAVE_IMAGE_PATH

By adding the --arch option, you can specify model type which is selected from "span_ch48","span_ch52". (default is span_ch48)

$ python3 span.py --arch ARCH

If you want to specify the scale for the resolution, put the scale after the --scale option.
Choose the scale in [2, 4].

$ python3 span.py --scale SCALE 

By adding the --video option, you can input the video.
If you pass 0 as an argument to VIDEO_PATH, you can use the webcam input instead of the video file.

$ python3 span.py --video VIDEO_PATH

Reference

SPAN

Framework

Pytorch 1.30.0

Model Format

ONNX opset = 17

Netron

spanx2_ch48.onnx.prototxt

spanx2_ch52.onnx.prototxt

spanx4_ch48.onnx.prototxt

spanx4_ch52.onnx.prototxt