CoDA/generation.py at main · PlusLabNLP/CoDA · 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
import os
import json
import numpy as np
from dotenv import load_dotenv

from filelock import FileLock
from datetime import datetime
from PIL import Image
from filter import filter_one_return_single_path, print_contrastiveness
import random
import torch
from diffusers import StableDiffusion3Pipeline
version = "stabilityai/stable-diffusion-3.5-large-turbo"
sd_ppl = StableDiffusion3Pipeline.from_pretrained(
    version, torch_dtype=torch.bfloat16).to("cuda")

def generate_image(sd_ppl, prompt, negative_prompt, save_path = None, seed=42, randomize_seed=False, width=1024, height=1024, num_inference_steps=4, guidance_scale=0.0):

    if randomize_seed:
        seed = random.randint(0, np.iinfo(np.int32).max)

    generator = torch.Generator().manual_seed(seed)
    image = sd_ppl(
        prompt=prompt,
        prompt_2=prompt,
        prompt_3=prompt,
        max_sequence_length=512,
        negative_prompt=negative_prompt,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        width=width,
        height=height,
        generator=generator
    ).images[0]
    if save_path:
        image.save(save_path, "PNG")

    return image, seed


if __name__ == "__main__":
    load_dotenv()
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--prompts", type=str, default="contrastive_visual_text")
    parser.add_argument("--num_images", type=int, default=50)
    parser.add_argument("--num_test", type=int, default=5)
    parser.add_argument("--working_dir", type=str, default="YOUR_PATH")

    args = parser.parse_args()
    prompts = args.prompts.split(",")
    working_dir = args.working_dir
    num_images = args.num_images
    num_test = args.num_test
    output_path = working_dir + "/synthetic_improved/34b"
    os.makedirs(output_path, exist_ok=True)

    configs_path = working_dir + "/config.json"
    with open(configs_path, "r") as f:
        configs = json.load(f)

    pairs = configs["pairs"]
    failed_pairs = []
    prompts = ["contrastive_visual_text"]
    for prompt_ in prompts:
        if os.path.exists(output_path + "/failed_pairs.json"):
            with open(output_path + "/failed_pairs.json", "r") as f:
                failed_pairs = json.load(f)

        for j, pair in enumerate(pairs):
            if j in failed_pairs:
                continue
            for mvc in ['mvc', 'cvm']:
                if mvc == 'mvc':
                    main_class = pair["ground_truth"]
                    main_class_full_name = pair["ground_truth_full_name"]
                    confusing_class = pair["confusing_class"]
                    confusing_class_full_name = pair["confusing_class_full_name"]
                else:
                    main_class = pair["confusing_class"]
                    main_class_full_name = pair["confusing_class_full_name"]
                    confusing_class = pair["ground_truth"]
                    confusing_class_full_name = pair["ground_truth_full_name"]

                print("=====================================")
                print(f"{main_class} vs {confusing_class}")
                print("=====================================")

                img_parent_dir = os.path.join(working_dir , "train/" + main_class_full_name)
                image_files = os.listdir(img_parent_dir)
                image_paths = filter_one_return_single_path(img_parent_dir, main_class, top_n = 5)# os.path.join(img_parent_dir, image_files[0])

                confusing_img_parent_dir = os.path.join(working_dir,"train/" + confusing_class_full_name)
                confusing_image_files = os.listdir(confusing_img_parent_dir)
                confusing_image_paths = filter_one_return_single_path(confusing_img_parent_dir, confusing_class, top_n = 5)

                individual_attributes= []
                with open(output_path + f"/{prompt_}/{main_class_full_name}/attributes.json", "r") as f:
                    attributes = json.load(f)
                negative_prompt = "Do not include any human-made objects or structures. Avoid showing other unnatural elements. Ensure the depiction is realistic and not cartoonish"
                if prompt_ == "contrastive_visual" or prompt_ == "contrastive_visual_text":
                    summary = print_contrastiveness(image_paths, confusing_image_paths, attributes)

                    attributes = [attribute.strip().strip('"').strip('\'') for attribute in attributes]
                    attributes = [attr for attr in attributes if summary[attr] > 0.6]
                    generated_images = [output_path + f"/{prompt_}/{main_class_full_name}/{i}.png" for i in range(num_test)]

                    generable_attributes = []
                    all_attributes = []
                    for start in range(0, len(attributes), 4):
                        subgroup = attributes[start:min(start+4, len(attributes))]
                        prompt = f"Generate a 4K realistic image of {main_class} that contains the following attributes: {', '.join(subgroup)}"
                        print(prompt)
                        for it in range(num_test):
                            img, seed = generate_image(sd_ppl, prompt, negative_prompt, randomize_seed=True)
                            img.save(output_path + f"/{prompt_}/{main_class_full_name}/{it}.png")

                        summary_gen = print_contrastiveness(generated_images, confusing_image_paths, subgroup)
                        all_for_this_group = [(attr, summary_gen[attr]) for attr in summary_gen]
                        all_attributes.extend(all_for_this_group)
                    generable_attributes = sorted(all_attributes, key=lambda x: x[1], reverse=True)[:5]

                    st = {}
                    for key, acc in generable_attributes:
                        st[key] = {
                            "real": summary[key],
                            "synthetic": acc
                        }
                    save_json = {
                        "combined": st,
                    }
                    with open(output_path + f"/{prompt_}/{main_class_full_name}/attributes_contrastiveness_statistics.json", "w") as f:
                        json.dump(save_json, f)
                    generable_attributes = [x[0] for x in generable_attributes]
                else:
                    generable_attributes = attributes
                for k in range(num_images):
                    prompt = f"Generate a 4K realistic image of {main_class} that contains the following attributes: {', '.join(generable_attributes)}"
                    img, seed = generate_image(sd_ppl, prompt, negative_prompt, randomize_seed=True)
                    img.save(output_path + f"/{prompt_}/{main_class_full_name}/{k}.png")