VerboseImages [ICLR24]¶

2025-08-02 01:45:22,698 - OPT - INFO - PARAMETER ...
2025-08-02 01:45:22,699 - OPT - INFO - Namespace(epsilon=0.032, step_size=0.0039, iter=100, gpu=0, seed=256, root_path='/root/autodl-tmp/Capstone/Verbose_Images', dataset='/root/autodl-tmp/Capstone/Verbose_Images/dataset')
2025-08-02 01:48:57,167 - OPT - INFO - Original sequences: layers of tall buildings
2025-08-02 01:48:57,168 - OPT - INFO - Verbose sequences: the last essex community bus service has left the network
2025-08-02 01:48:57,168 - OPT - INFO - ------------------------
2025-08-02 01:48:57,168 - OPT - INFO - Original sequences: a city bus moving down a road with another bus behind it
2025-08-02 01:48:57,168 - OPT - INFO - Verbose sequences: ’2149qb bfti grnt 45 hp twin barrier tri dup plasma plus interurbia 82 leadless all colour with rainbow rainbow stock clips from the blue and rainbow sofa where can i buy a sofa in india, search blue sofa in india, person, indian sofa, couch, city bus, transportation, in the city, the city bus, orlando bus, electric bus, metro bus, metro system, regional bus, ftl buses, metros, btm, btdl
2025-08-02 01:48:57,168 - OPT - INFO - ------------------------
2025-08-02 01:48:57,168 - OPT - INFO - Original sequences: a purple bus sitting next to some buildings
2025-08-02 01:48:57,168 - OPT - INFO - Verbose sequences: the purple bus is driving on the street
2025-08-02 01:48:57,168 - OPT - INFO - ------------------------
2025-08-02 01:48:57,168 - OPT - INFO - Original images, Length: 8.00, Energy: 14.31, Latency: 0.34
2025-08-02 01:48:57,168 - OPT - INFO - Verbose images, Length: 28.67, Energy: 90.84, Latency: 0.91

source /etc/network_turbo

import os
import random
import shutilh

def random_select_images(source_folder, target_folder, num_images=1000):
    """
    从源文件夹随机选择指定数量的图片到目标文件夹
    
    参数:
        source_folder: 源图片文件夹路径
        target_folder: 目标文件夹路径
        num_images: 要选择的图片数量(默认1000)
    """
    # 确保目标文件夹存在
    os.makedirs(target_folder, exist_ok=True)
    
    # 获取所有图片文件
    all_images = [f for f in os.listdir(source_folder) 
                 if f.lower().endswith(('.png', '.jpg', '.jpeg', '.bmp', '.gif'))]
    
    # 检查是否有足够图片
    if len(all_images) < num_images:
        raise ValueError(f"文件夹中只有 {len(all_images)} 张图片，不足 {num_images} 张")
    
    # 随机选择
    selected_images = random.sample(all_images, num_images)
    
    # 复制到目标文件夹
    for img in selected_images:
        src_path = os.path.join(source_folder, img)
        dst_path = os.path.join(target_folder, img)
        shutil.copy2(src_path, dst_path)
    
    print(f"成功复制 {len(selected_images)} 张图片到 {target_folder}")

# 使用示例
source = "/root/autodl-tmp/train2017"  # 替换为你的图片文件夹路径
target = "/Capstone/Verbose_Images/dataset"       # 替换为输出文件夹路径
random_select_images(source, target, 1000)

uv env --python=3.9.2

uv pip install -r requirements.txt

conda create -n VI python=3.9.2

conda activate VI

pip install -e .

from scipy import log
log(x)

import numpy as np
ratio1 = 10.0 * np.log(tdx + 1) - 20.0
ratio2 = 0.5 * np.log(tdx + 1) + 1.0

from moviepy.editor import VideoFileClip

#提取mp4视频里的音频保存为mp3.py
from moviepy.video.io.VideoFileClip import VideoFileClip

def init_tokenizer(cls, truncation_side="right"):
        tokenizer = BertTokenizer.from_pretrained("models/bert-base-uncased", truncation_side=truncation_side)
        tokenizer.add_special_tokens({"bos_token": "[DEC]"})
        return tokenizer

def parse_args():
    '''PARAMETERS'''
    parser = argparse.ArgumentParser('generate verbose images')
    parser.add_argument('--epsilon', type=float, default=0.032, help='the perturbation magnitude')
    parser.add_argument('--step_size', type=float, default=0.0039, help='the step size')
    parser.add_argument('--iter', type=int, default=1000, help='the iteration')
    parser.add_argument('--gpu', type=int, default=0, help='GPU index')
    parser.add_argument('--seed', type=int, default=256, help='random seed')
    # 增加两行
    parser.add_argument('--root_path', type=str, default='.', 
                      help='Root directory')
    parser.add_argument('--dataset', type=str, required=True,
                      help='Dataset name')

pip install open3d

pip install -U spacy==3.6.0
pip install numpy==1.26.4
pip install thinc==8.1.8

pip install peft==0.10.0
pip install transformers==4.31.0

pip install huggingface_hub==0.25.2 -i https://pypi.tuna.tsinghua.edu.cn/simple

vi lavis/configs/models/blip2/blip2_instruct_vicuna7b.yaml