library_name: transformers tags:

• 深度伪造 datasets:
• elsaEU/ELSA_D3 license: mit

基于对比学习与全局-局部相似性的深度伪造扩散对比方法（ECCV 2024）

项目主页

源代码

模型卡片

import transformers
from huggingface_hub import hf_hub_download
from PIL import Image
import faiss
import timm
import torch
import torch.nn as nn
import joblib
from torchvision import transforms

if torch.cuda.is_available():
    device = torch.device('cuda')
else: device = torch.device('cpu')

'''
线性分类器 - knn --> 0真实 1伪造
svm分类器 --> -1真实 1伪造
'''
class VITContrastiveHF(nn.Module):
    def __init__(self, repo_name, classificator_type):
        super(VITContrastiveHF, self).__init__()
        self.model = transformers.AutoModel.from_pretrained(repo_name)
        self.model.pooler= nn.Identity()

        self.processor = transformers.AutoProcessor.from_pretrained(repo_name)
        self.processor.do_resize= False
        # 定义分类器 /// 考虑使用`cache_dir`参数
        if classificator_type == 'svm':
            file_path = hf_hub_download(repo_id=repo_name, filename='sklearn/ocsvm_kernel_poly_gamma_auto_nu_0_1_crop.joblib')
            self.classifier = joblib.load(file_path)

        elif classificator_type == 'linear':
            file_path = hf_hub_download(repo_id=repo_name, filename='sklearn/linear_tot_classifier_epoch-32.sav')
            self.classifier = joblib.load(file_path)

        elif classificator_type == 'knn':
            file_path = hf_hub_download(repo_id=repo_name, filename='sklearn/knn_tot_classifier_epoch-32.sav')
            self.classifier = joblib.load(file_path)
            
        else:
            raise ValueError('选择了无效的分类器')

    def forward(self, x, return_feature=False):
        features = self.model(x)
        if return_feature:
            return features
        features = features.last_hidden_state[:,0,:].cpu().detach().numpy()
        # features.last_hidden_state[:,0,:].shape
        predictions = self.classifier.predict(features)
        return torch.from_numpy(predictions)
    
# HuggingFace推理代码
classificator_type = 'linear'
model = VITContrastiveHF(repo_name='aimagelab/CoDE', classificator_type=classificator_type)

transform = transforms.Compose([
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

model.eval()
model.model.to(device)
y_pred= []
# 放入待测试图像
img = Image.open('206496010652.png').convert('RGB')

with torch.no_grad():
    in_tens = transform(img).unsqueeze(0)

    in_tens= in_tens.to(device)
    y_pred.extend(model(in_tens).flatten().tolist())

# 检查预测结果的正确标签
for el in y_pred:
    if el == 1:
        print('伪造')
    elif el == 0:
        print('真实')
    elif el == -1:
        print('真实')
    else:
        print('错误')