89 lines
3.2 KiB
Python
89 lines
3.2 KiB
Python
import torch
|
|
import torchvision
|
|
from torchvision import datasets, models, transforms
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
import os
|
|
import clip
|
|
from utils import normalize
|
|
|
|
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
|
|
|
|
|
def get_prompt(cfg):
|
|
model, preprocess = clip.load("ViT-B/32", device=device)
|
|
prompt=cfg.prompt
|
|
text=clip.tokenize(prompt).to(device)
|
|
with torch.no_grad():
|
|
prompt = model.encode_text(text)
|
|
return prompt
|
|
# class GanAttack(nn.Module):
|
|
|
|
class GanAttack(nn.Module):
|
|
def __init__(self, net,prompt):
|
|
super().__init__()
|
|
|
|
self.net= net
|
|
# self.generator.eval()
|
|
# self.inverter=inverter
|
|
# self.images_resize=images_resize
|
|
self.prompt=prompt
|
|
text_len=self.prompt.shape[0]
|
|
self.mlp=nn.Sequential(
|
|
nn.Linear(text_len+512, 4096),
|
|
nn.ReLU(inplace=True),
|
|
nn.Linear(4096, 512)
|
|
)
|
|
|
|
|
|
def forward(self, img):
|
|
codes = self.net.encoder(img)
|
|
codes = codes + self.net.latent_avg.repeat(codes.shape[0], 1, 1)
|
|
# result_images, result_latent = self.net.decoder([codes], input_is_latent=True, randomize_noise=False, return_latents=False)
|
|
# result_images = self.net.face_pool(result_images)
|
|
# _, _, _, refine_images, latent_codes, _=self.inverter(img,self.images_resize,img_path)
|
|
x=codes
|
|
batch_size=img.shape[0]
|
|
prompt=self.prompt.repeat(batch_size).to(device)
|
|
x_prompt=torch.cat([codes,prompt],dim=1)
|
|
x_prompt=self.mlp(x_prompt)
|
|
x=x_prompt+x
|
|
im,_=self.net.decoder([x], input_is_latent=True, randomize_noise=False, return_latents=False)
|
|
result_images = self.net.face_pool(im)
|
|
|
|
return result_images,x
|
|
|
|
class CLIPLoss(torch.nn.Module):
|
|
def __init__(self):
|
|
super(CLIPLoss, self).__init__()
|
|
self.model, self.preprocess = clip.load("ViT-B/32", device="cuda")
|
|
self.model.eval()
|
|
self.face_pool = torch.nn.AdaptiveAvgPool2d((224, 224))
|
|
# self.mean = torch.tensor([0.48145466, 0.4578275, 0.40821073], device="cuda").view(1,3,1,1)
|
|
# self.std = torch.tensor([0.26862954, 0.26130258, 0.27577711], device="cuda").view(1,3,1,1)
|
|
|
|
def forward(self, image, text):
|
|
image=normalize(image)
|
|
image = self.face_pool(image)
|
|
similarity = 1 - self.model(image, text)[0]/ 100
|
|
return similarity
|
|
|
|
|
|
class VggLoss(torch.nn.Module):
|
|
def __init__(self):
|
|
super(VggLoss, self).__init__()
|
|
self.model=models.vgg11(pretrained=True)
|
|
self.model.features=nn.Sequential()
|
|
|
|
# self.mean = torch.tensor([0.48145466, 0.4578275, 0.40821073], device="cuda").view(1,3,1,1)
|
|
# self.std = torch.tensor([0.26862954, 0.26130258, 0.27577711], device="cuda").view(1,3,1,1)
|
|
|
|
def forward(self, image1, image2):
|
|
# image=normalize(image)
|
|
with torch.no_grad:
|
|
feature1=self.model(image1)
|
|
feature2=self.model(image2)
|
|
feature1=torch.flatten(feature1)
|
|
feature2=torch.flatten(feature2)
|
|
similarity = F.cosine_similarity(feature1,feature2)
|
|
return similarity |