how to use onnx models correctly?

[frederik-uni]

I tried
(image is C, H, W)

let pad_x = (8 - image.width % 8) % 8;
let pad_y = (8 - image.height % 8) % 8;
image  = add_padding(image, pad_x, pad_y);
image  = image / 255.0;
image  = expand_dims(image, axis=0)
out = onnx.run(x => image);
img = squeeze(out, axis=0)

it resulted in 400x400 => 728 x 728.

the image is a bit zoomed in. I guess I need some padding, but how much or is it something else? It seems to work with a padding of 18, but why

[nagadomi]

Yes, padding is required.

nunif/waifu2x/models/cunet.py

Line 143 in bedce31

super(UpCUNet, self).__init__(locals(), scale=2, offset=36, in_channels=in_channels)

For the cunet 2x model, offset=36. This means 36 pixels are cropped from each edge of the 2x output. To keep the same image area as the original input, all sides of the input should be padded by 18 pixels (18=36px/2x).

The ONNX model is used in unlimited:waifu2x, so it will probably be helpful as a reference.
https://github.com/nagadomi/nunif/tree/master/waifu2x/unlimited_waifu2x