RIW/Injector.py at main · BennyTMT/RIW · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
import torch , os
import torch.nn as nn
import numpy as np
from PIL import Image
import lpips,clip
from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize,InterpolationMode
import torch.optim as optim

class target_Injector():
    def __init__(self, model, eps=16/255, alpha=2/255, steps=1010, random_start=False, clip_loss = False,device=''):
        super(target_Injector, self).__init__()
        self.clip_loss = False
        self.model = model
        self.device= device
        self.eps = eps
        self.alpha = alpha
        self.steps = steps
        self.random_start = random_start
        self.supported_mode = ['default', 'targeted']
        self.auto_bound= True
        self.l2_loss = nn.MSELoss()
        self.lpips = lpips.LPIPS(net='vgg').to(self.device)
        if clip_loss:
            self.clip_loss = clip_loss
            self.clip_model , _ =  clip.load("ViT-L/14", device=self.device)
            # 224
            n_px= self.clip_model.visual.input_resolution
            self.transform_ = Compose([
                    Resize(n_px , interpolation=InterpolationMode.BICUBIC),
                    CenterCrop(n_px),
                    Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
                ])

    def run(self, origin_imgs, x_prime = None , water= None  ,
            noised_degree_latent=0 ,  decoder_loss = False , lambda3 =1 ,  lambda2 = 1.0 ):

        self.decoder_loss = decoder_loss
        self.origin_imgs  = origin_imgs
        self.lambda3  = lambda3
        self.lambda2  = lambda2

        self.decoder_x_prime = None
        self.noised_degree = noised_degree_latent

        print('lambda : '  , self.lambda2 , self.lambda3 )

        if x_prime != None :
            tar_images = x_prime.clone().detach().to(self.device)
        if water!= None :
            e_water = self.model.encode_first_stage(water.clone().detach().to(self.device) ).mode()

        if self.random_start:
            # Starting at a uniformly random point
            adv_images = adv_images + \
                torch.empty_like(adv_images).uniform_(-self.eps, self.eps)
            adv_images = torch.clamp(adv_images, min=-1, max=1).detach()
        else:
            adv_images = origin_imgs.clone().detach().to(self.model.device)

        if self.auto_bound:
            low_bound = float(torch.min( origin_imgs.clone().detach().to(self.device) ))
            upp_bound = float(torch.max( origin_imgs.clone().detach().to(self.device) ))
        else:
            low_bound , upp_bound = -1. , 1.

        lowest_cost = 1000
        return_image = None
        self.tar_image_features = self.model.encode_first_stage(tar_images).mode()
        costs = []
        for i in range(1, self.steps +1):
            adv_images.requires_grad = True
            adv_image_features = self.model.encode_first_stage(adv_images).mode()

            cost = -self.loss_fn(adv_images ,origin_imgs  , adv_image_features, self.tar_image_features)

            grad = torch.autograd.grad(cost, adv_images,
                                        retain_graph=False, create_graph=False)[0]

            adv_images = adv_images.detach() + self.alpha*grad.sign()
            delta      = torch.clamp(adv_images - origin_imgs , min=-self.eps, max= self.eps)
            adv_images = torch.clamp(origin_imgs + delta, min=low_bound, max=upp_bound).detach()

            if i%50 == 0 :
                curr_cost = abs(float( cost.detach().data.cpu().numpy() ) )
                print(i , curr_cost )
                costs.append(curr_cost)
                # self.image_save(adv_images.clone(), i , inject_type , noised_degree = noised_degree )
                # if lowest_cost > curr_cost :
                #     return_image = adv_images.clone().detach()
                #     lowest_cost  = curr_cost
                #     best_step = i
        # return return_image, best_step
        return return_image, costs

    def check_loaded_encoder(self,  adv_image_features, tar_image_features , store =  True ):
        print(adv_image_features.shape)
        print(tar_image_features.shape)
        tar_ecodering = tar_image_features.data.cpu().numpy()

        if store : np.save('tar_encode.npy' , tar_ecodering  ) ; print('saved!') ; exit()
        else:
            tar_ecodering_saved = np.load('tar_encode.npy' )
            a = tar_ecodering - tar_ecodering_saved
            print(np.mean( np.abs(a)))
            print(np.mean( np.abs(tar_ecodering_saved)))
            print(a.shape)
            exit()

    def loss_fn(self, adv_image  ,  ori_image , adv_encode , tar_encode ):
        c1 = 1
        c2 = self.lambda2
        c3 = self.lambda3

        l1 = torch.abs(self.lpips(adv_image,ori_image) )
        l2 = torch.mean(torch.abs(adv_encode - tar_encode ))
        if self.decoder_loss  :
            adv_encode = (adv_encode-adv_encode.mean())/adv_encode.std()
            loss_decoder = self.l2_loss( self.origin_imgs,  self.model.decode_first_stage(adv_encode))
        # if self.clip_loss :
        #     l3 = self.clip_distance(  ori_image ,adv_image )
        #     return  c1* l1 + c2 * l2 + c3* l3
        if self.decoder_loss :
            return  c1* l1 + c2 * l2 + c3 * loss_decoder
        else :
            print('there will be no punishment!')
            return  c1* l1 + c2 * l2

    def clip_distance(self , x_0 , x  ):
        x_0 = self.transform_(x_0) ; x = self.transform_(x)
        orig_image_features  = self.clip_model.encode_image(x_0)
        targ_image_features  = self.clip_model.encode_image(x)
        return  self.l2_loss(orig_image_features ,targ_image_features )

    def image_save(self, image_tensor , step , inject_type , noised_degree = 0 ):
        image_tensor = image_tensor.cpu().detach()

        IMGPATH = '/YOURDIR/instruct-pix2pix/imgs/injection/{}/'.format(inject_type)
        if not os.path.exists(IMGPATH): os.makedirs(IMGPATH)

        # image_tensor = (0,1)
        image_tensor = ( (image_tensor +1) * 125.0 ).clamp_(0, 255)
        image_np = np.squeeze(image_tensor.numpy().astype(np.uint8))
        # Transpose the NumPy array from (C, H, W) to (H, W, C)
        image_np = np.transpose(image_np, (1, 2, 0))
        # Convert the NumPy array to a PIL Image object
        image_pil = Image.fromarray(image_np)
        # Save the PIL Image object
        image_pil.save(IMGPATH+"{}_{}.png".format(noised_degree,step))

    def image_save_from_tensor(self, tensorImg, save_path ):
        image_tensor = tensorImg.cpu().detach()
        # 2 * image / 255 - 1
        image_tensor = ( (image_tensor +1) * 125.0 ).clamp_(0, 255)
        image_np = np.squeeze(image_tensor.numpy().astype(np.uint8))
        image_np = np.transpose(image_np, (1, 2, 0))
        image_pil = Image.fromarray(image_np)
        image_pil.save(save_path)