| import gradio as gr |
| from torchvision.models import resnet50 |
| import torch |
| import torch.nn as nn |
| from torch.nn import functional as F |
| from torchvision import transforms |
| import numpy as np |
| import os |
| import random |
| def seed_torch(seed=1029): |
| random.seed(seed) |
| os.environ['PYTHONHASHSEED'] = str(seed) |
| np.random.seed(seed) |
| torch.manual_seed(seed) |
| torch.cuda.manual_seed(seed) |
| torch.cuda.manual_seed_all(seed) |
| torch.backends.cudnn.benchmark = False |
| torch.backends.cudnn.deterministic = True |
| torch.backends.cudnn.enabled = False |
| seed_torch(100) |
|
|
| NPRmodel = resnet50() |
| NPRmodel.fc1 = nn.Linear(512, 1) |
| NPRmodel.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=2, padding=1, bias=False) |
| del NPRmodel.layer3, NPRmodel.layer4, NPRmodel.fc |
|
|
| from collections import OrderedDict |
| from copy import deepcopy |
|
|
| NPR_url = 'https://raw.githubusercontent.com/chuangchuangtan/NPR-DeepfakeDetection/main/model_epoch_last_3090.pth' |
| NPRmodel.load_state_dict(torch.hub.load_state_dict_from_url(NPR_url, map_location='cpu'), strict=True) |
| trans = transforms.Compose([ transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) |
|
|
| def interpolate(img, factor): |
| return F.interpolate(F.interpolate(img, scale_factor=factor, mode='nearest', recompute_scale_factor=True), scale_factor=1/factor, mode='nearest', recompute_scale_factor=True) |
|
|
| def predict_image(img): |
| img = trans( img.convert('RGB') ).unsqueeze(0) |
| _, c, w, h = img.shape |
| if w%2 == 1: img = img[:, :, :-1,: ] |
| if h%2 == 1: img = img[:, :, : ,:-1] |
| NPR = img - interpolate(img, 0.5) |
| with torch.no_grad(): |
| x = NPRmodel.conv1(NPR*2.0/3.0) |
| x = NPRmodel.bn1(x) |
| x = NPRmodel.relu(x) |
| x = NPRmodel.maxpool(x) |
| x = NPRmodel.layer1(x) |
| x = NPRmodel.layer2(x).mean(dim=(2,3), keepdim=False) |
| x = NPRmodel.fc1(x) |
| pred = x.sigmoid().cpu().numpy() |
| return {'Fake Image': float(pred), 'Real Image': 1.0 - float(pred)} |
|
|
|
|
| |
| title = "<p>Generalizable Deepfake Detection Neighboring Pixel Relationships(NPR)-CVPR2024<p>" |
| description_html = """ |
| <p> <strong>Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection * CVPR2024</strong><p> |
| <p> This model was only trained on ProGAN for detecting deepfake images. </p> |
| <p> The model achieved an accuracy of <strong>92.2%</strong> on the 28 generation models.</p> |
| <p> Upload a image or pick from the samples below to test model accuracy</p> |
| <p><a href='https://github.com/chuangchuangtan/NPR-DeepfakeDetection' target='_blank'>NPR Github Code</a></p> |
| """ |
|
|
| |
| example_data = ['midjourney1.png', 'CelebAHQ_00000110.png'] |
| custom_css = """ |
| div {background-color: whitesmoke;} |
| """ |
|
|
| gr.Interface( |
| fn=predict_image, |
| inputs=gr.Image(type="pil"), |
| outputs=gr.Label(num_top_classes=2), |
| title=title, |
| description=description_html, |
| allow_flagging='never', |
| examples=example_data, |
| css=custom_css |
| ).launch() |
| |
