def forward(self, img):
factors = []
# noise & blur
blur_factor = 0
if self.blur is not None:
blur_factor = random.uniform(*self.blur)
img = imfilter(
img,
torch.tensor(gaussian_kernel(15, blur_factor),
device=img.device), self.blur_padding)
awgn_factor = (0, 0, 0)
if self.awgn is not None:
_r = random.uniform(*self.awgn)
_g = random.uniform(*self.awgn)
_b = random.uniform(*self.awgn)
img += gaussian_noise(img, stddev=(_r, _g, _b))
awgn_factor = (_r, _g, _b)
poisson_factor = (_r, _g, _b)
if self.poisson is not None:
_r = random.uniform(*self.poisson)
_g = random.uniform(*self.poisson)
_b = random.uniform(*self.poisson)
img += poisson_noise(img, stddev=(_r, _g, _b))
poisson_factor = (_r, _g, _b)
fac = [blur_factor, *awgn_factor, *poisson_factor]
factors.append(torch.tensor(fac))
img = img.clamp(0, 1)
return img, torch.stack(factors).to(img.device)
def gen_kernel(self, ktype, ksize, l1, l2=None, theta=0):
if ktype == 'isotropic':
kernel = gaussian_kernel(ksize, l1)
elif ktype == 'anisotropic':
kernel = anisotropic_gaussian_kernel(ksize, theta, l1, l2 or l1)
else:
# TODO(wenyi) this is "directKernel"
raise NotImplementedError("DirectKernel not implemented.")
return kernel
# Copyright (c) 2017-2020 Wenyi Tang. # Author: Wenyi Tang # Email: [email protected] # Update: 2020 - 2 - 13 import unittest import numpy as np from VSR.Util.Math import (anisotropic_gaussian_kernel, gaussian_kernel) from VSR.Backend import BACKEND _K1 = gaussian_kernel(15, 2) _K2 = anisotropic_gaussian_kernel(15, 1, 5, 3) class ImFilter(unittest.TestCase): # For ones([4, 4]) y_gold = np.array([[0.3151, 0.3776, 0.3776, 0.3151], [0.3776, 0.4524, 0.4524, 0.3776], [0.3776, 0.4524, 0.4524, 0.3776], [0.3151, 0.3776, 0.3776, 0.3151]]) z_gold = np.array([[0.3391, 0.3950, 0.3774, 0.2950], [0.3850, 0.4627, 0.4557, 0.3677], [0.3677, 0.4557, 0.4627, 0.3850], [0.2950, 0.3774, 0.3950, 0.3391]]) def test_torch(self): if BACKEND != 'pytorch': return import torch