Proper out_shape for NumPy resize?

[Kirill Klimov]

Hi! First of all - thank you for awesome library.

I succeed with resizing PyTorch Tensors using channels first representation. But i cannot figure out which out shape should i use for NumPy ndarrays. I tried to directly use channels last image using following code:

import resize_right
import cv2


img = cv2.imread("path_to_my_image")

target_height_width = (img.shape[1] // 8, img.shape[0] // 8)
print(img.shape)
res = resize_right.resize(img, out_shape=target_height_width)
print(res.shape)

Input image has shape (2048, 2048, 3), but output is (256, 256, 2048) which is quite strange.

If i convert my image to channels first via np.rollaxis(2, 0) it returns (256, 256, 3) shape, which is correct but not consistent with input. But result looks broken (basically vertical color lines).

Can you help me to figure out how to use resize_right properly?

[Kirill Klimov]

Just found, that if i specify full out_shape for NumPy it works. Still looks for me as a problem because behavior is not consistent with Torch.

[Ina]

This problem seems come from this.
The correct operation should be：

out_shape = (list(out_shape) + list(in_shape[len(out_shape):])
                 if fw is numpy
                 else list(in_shape[:-len(out_shape)]) + list(out_shape))

[Assaf Shocher]

Thanks, new PR by @LuoXin-s resolves this issue