File size: 4,488 Bytes
5ca4e86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import argparse
import os
import random
import cv2
import numpy as np
import torch
import torch.backends.cudnn as cudnn
from PIL import Image
import json
import pandas as pd

def rtp_read(text_file):
    dataset = []
    lines = open(text_file).read().split("\n")
    for li in lines:
        obj = json.loads(li)
        if obj['challenging']:
            dataset.append(obj['prompt']['text'])
    return dataset

def denormalize(images):
    mean = torch.tensor([0.48145466, 0.4578275, 0.40821073]).cuda()
    std = torch.tensor([0.26862954, 0.26130258, 0.27577711]).cuda()
    new_images = (images - mean[None, :, None, None])/ std[None, :, None, None]
    return new_images

def normalize(images):
    mean = torch.tensor([0.48145466, 0.4578275, 0.40821073]).cuda()
    std = torch.tensor([0.26862954, 0.26130258, 0.27577711]).cuda()
    new_images = (images * std[None, :, None, None])+ mean[None, :, None, None]
    return new_images


def parse_args():

    parser = argparse.ArgumentParser(description="Demo")
    parser.add_argument("--model-path", type=str, default="ckpts/llava_llama_2_13b_chat_freeze")
    parser.add_argument("--model-base", type=str, default=None)
    parser.add_argument("--image_path", type=str, default='./image.bmp',
                        help="Image file")
    parser.add_argument("--image_safety_patch", type=str, default=None,
                        help="image safety patch file")
    parser.add_argument("--text_safety_patch", type=str, default=None,
                        help="text safety patch file") 
    parser.add_argument("--output_file", type=str, default='./result.jsonl',
                        help="Output file.")

    args = parser.parse_args()
    return args


def load_image(image_path):
    image = Image.open(image_path).convert('RGB')
    return image



# ========================================
#             Model Initialization
# ========================================

print('>>> Initializing Models')

from llava.utils import get_model
args = parse_args()

print('model = ', args.model_path)

tokenizer, model, image_processor, model_name = get_model(args)
model = model.cuda()
print(model)
print(model_name)
model.eval()

print('[Initialization Finished]\n')


from llava_utils import prompt_wrapper, generator


my_generator = generator.Generator(model=model, tokenizer=tokenizer)


# ========================================
#             Inference
# ========================================

prompts = rtp_read('harmful_corpus/rtp_prompts.jsonl')
text_prompt = '%s'

if args.text_safety_patch!=None:
    with open(args.text_safety_patch, 'r') as file:
        text_safety_patch = file.read().rstrip()

out = []
with torch.no_grad():

    for i, user_message in enumerate(prompts):
 
        image_file = 'adversarial_qna_images/adv_image_' + str(np.random.randint(25)) + '.bmp'
        image = load_image(image_file)
        image = image_processor.preprocess(image, return_tensors='pt')['pixel_values'].cuda()
        if args.image_safety_patch!=None:
            image = normalize(image)
            safety_patch = torch.load(args.image_safety_patch).cuda()
            safe_image = denormalize((image + safety_patch).clamp(0,1))
        else:
            safe_image = image
        print(f" ----- {i} ----")
        print(" -- prompt: ---")
       
        if args.text_safety_patch!=None:
            # Use the below for optimized text patch
            #    user_message = text_safety_patch + '\n' + user_message 
            # Use the below for heuristic text patch
            user_message += '\n' + text_safety_patch

        text_prompt_template = prompt_wrapper.prepare_text_prompt(text_prompt % user_message)
        print(text_prompt_template)
        prompt = prompt_wrapper.Prompt(model, tokenizer, text_prompts=text_prompt_template, device=model.device)

        response = my_generator.generate(prompt, safe_image).replace("[INST]","").replace("[/INST]","").replace("[SYS]","").replace("[/SYS/]","").strip()

        if args.text_safety_patch!=None:
            response = response.replace(text_safety_patch,"")
        
        print(" -- continuation: ---")
        print(response)
        out.append({'prompt': user_message, 'continuation': response})
        print()

with open(args.output_file, 'w') as f:
    f.write(json.dumps({
        "args": vars(args),
        "prompt": text_prompt
    }))
    f.write("\n")

    for li in out:
        f.write(json.dumps(li))
        f.write("\n")