damir_smoothing.py

import pyiqa
import torch
import random
import json
import numpy as np

from tqdm import tqdm
from pathlib import Path
from PIL import Image
from torchvision import transforms

def generate_image_noise(shape):
    return torch.randn(*shape)

def normalize_in_sphere(image_noise, radius):
    norm_per_pixel = torch.linalg.norm(image_noise, axis=-1, keepdims=True)
    normalized_noise = image_noise / norm_per_pixel

    scaled_noise = normalized_noise * radius

    return scaled_noise

device = torch.device("cuda")

iqa_metric = pyiqa.create_metric('maniqa', device=device)

current_directory = Path.cwd()
data_dir = Path('./data/images/')

image_files = list(data_dir.iterdir())

scores_true=[]
scores_new=[]
for i in range(1):
    image_path = image_files[i]
    score = iqa_metric(image_path.as_posix())
    numeric_value = score.item()
    scores_true.append(numeric_value)

    image = Image.open(image_path.as_posix())
    original_array = transforms.PILToTensor()(image)

    dimensions = original_array.shape

    scores_fr = []

    for j in tqdm(range(50000)):

        image_noise = generate_image_noise(dimensions)
        normalized_noise = normalize_in_sphere(image_noise, radius = 0.05)

        adversarial_array = torch.clamp(original_array + normalized_noise, 0, 255)#astype(np.uint8)
        #adversarial_image = transforms.fromarray(adversarial_array)

        score_fr = iqa_metric(adversarial_array/255)
        scores_fr.append(score_fr.item())
    total_sum = sum(scores_fr)
    average = total_sum / len(scores_fr)
    scores_new.append(average)
    print(i, 'score orig=', numeric_value, 'scores_fr_avg=', average)

result = max([abs(x - y) for x, y in zip(scores_true, scores_new)])
print('max difference=', result)
with open('results.json','w') as f:
    json.dump({"scores_true": scores_true,
               "scores_new": scores_new,
               "scores_fr": scores_fr}, 
               f)