experiments/overlap/augmentations/additive_noise.py [243:286]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
                for i in range(5)]) * self.im_size
        #radii = np.array([0.1 for i in range(5)]) * self.im_size
        #amounts = np.array([float_parameter(sample_level(self.severity, self.max_intensity), 50)\
        #        for i in range(5)])
        amounts = np.array([50 for i in range(5)])
        color = np.array([255, 255, 255])
        randomness = 25
        seed = np.random.randint(low=0, high=2**32)
        nrays = np.random.randint(low=50, high=200, size=5)

        return {'centers' : centers, 'radii' : radii, 'color' : color, 'randomness' : randomness,
                'seed' : seed, 'nrays' : nrays, 'amounts' : amounts
                }

    def transform(self, image, centers, radii, nrays, amounts, color, randomness, seed):

        def kernel(point, value, center, radius, ray_lengths, amount, color):
            rays = len(ray_lengths)
            dp = point - center
            dist = np.linalg.norm(dp)
            angle = np.arctan2(dp[1], dp[0])
            d = (angle + np.pi) / (2 * np.pi) * rays
            i = int(d)
            f = d - i 

            if radius != 0:
                length = ray_lengths[i % rays] + f * (ray_lengths[(i+1) % rays] - ray_lengths[i % rays])
                g = length**2 / (dist**2 + 1e-4)
                g = g ** ((100 - amount) / 50.0)
                f -= 0.5
                f = 1 - f**2
                f *= g
            f = np.clip(f, 0, 1)
            return value + f * (color - value)

        random_state = np.random.RandomState(seed=seed)
        for center, rays, amount, radius in zip(centers, nrays, amounts, radii):
            ray_lengths = [max(1,radius + randomness / 100.0 * radius * random_state.randn())\
                for i in range(rays)]

            image = np.array([[kernel(np.array([y,x]), image[y,x,:].astype(np.float32), center, radius, ray_lengths, amount, color)\
                for x in range(self.im_size)] for y in range(self.im_size)])

        return np.clip(image, 0, 255).astype(np.uint8)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



imagenet_c_bar/corrupt.py [227:267]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
                for i in range(5)]) * self.im_size
        amounts = np.array([50 for i in range(5)])
        color = np.array([255, 255, 255])
        randomness = 25
        seed = np.random.randint(low=0, high=2**32)
        nrays = np.random.randint(low=50, high=200, size=5)

        return {'centers' : centers, 'radii' : radii, 'color' : color, 'randomness' : randomness,
                'seed' : seed, 'nrays' : nrays, 'amounts' : amounts
                }

    def transform(self, image, centers, radii, nrays, amounts, color, randomness, seed):

        def kernel(point, value, center, radius, ray_lengths, amount, color):
            rays = len(ray_lengths)
            dp = point - center
            dist = np.linalg.norm(dp)
            angle = np.arctan2(dp[1], dp[0])
            d = (angle + np.pi) / (2 * np.pi) * rays
            i = int(d)
            f = d - i 

            if radius != 0:
                length = ray_lengths[i % rays] + f * (ray_lengths[(i+1) % rays] - ray_lengths[i % rays])
                g = length**2 / (dist**2 + 1e-4)
                g = g ** ((100 - amount) / 50.0)
                f -= 0.5
                f = 1 - f**2
                f *= g
            f = np.clip(f, 0, 1)
            return value + f * (color - value)

        random_state = np.random.RandomState(seed=seed)
        for center, rays, amount, radius in zip(centers, nrays, amounts, radii):
            ray_lengths = [max(1,radius + randomness / 100.0 * radius * random_state.randn())\
                for i in range(rays)]

            image = np.array([[kernel(np.array([y,x]), image[y,x,:].astype(np.float32), center, radius, ray_lengths, amount, color)\
                for x in range(self.im_size)] for y in range(self.im_size)])

        return np.clip(image, 0, 255).astype(np.uint8)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -