import numpy as np
from PIL import Image
import pytest
import torch
import torchvision
from torchvision import transforms as trans
from test_cv2_utils import image_similarity_vectors_via_cos
import torchvision_npu
@pytest.mark.parametrize(
["img_path", "distortion_scale", "p", "interpolation", "fill"],
[
("./test/Data/fish/fish_11.jpg", 0.6, 0.5, 2, 0),
("./test/Data/fish/fish_22.jpg", 0.1, 0.3, 2, 0),
("./test/Data/fish/fish_33.jpg", 1, 1, 0, (255, 0, 0)),
("./test/Data/fish/fish_44.jpg", 0, 0.3, 1, (255, 0, 0)),
("./test/Data/fish/fish_55.jpg", 1, 0.5, 2, (255, 0, 0)),
],
)
def test_perspective(img_path, distortion_scale, p, interpolation, fill):
pil_img = Image.open(img_path)
torchvision.set_image_backend("PIL")
torch.manual_seed(10)
pil_perspective = trans.RandomPerspective(distortion_scale=distortion_scale, p=p, interpolation=interpolation,
fill=fill)(pil_img)
torchvision.set_image_backend("cv2")
torch.manual_seed(10)
cv2_img = np.asarray(pil_img)
cv2_perspective = trans.RandomPerspective(distortion_scale=distortion_scale, p=p, interpolation=interpolation,
fill=fill)(cv2_img)
assert isinstance(pil_perspective, Image.Image) and isinstance(cv2_perspective, np.ndarray)
assert pil_perspective.size == cv2_perspective.shape[:2][::-1]
assert image_similarity_vectors_via_cos(pil_perspective, Image.fromarray(cv2_perspective))
pil_img.close()