semo / dataset /data_preprocess.py

Upload folder using huggingface_hub

bd546bf verified over 1 year ago

16.3 kB

import glob
import os
import pickle
import random
from tqdm import tqdm
import imageio.v3 as iio
import os
from multiprocessing import Pool
import os
from moviepy.editor import VideoFileClip
import argparse
import shutil
import cv2


def split_video_whisper():
    def process_video_whisper(videodir,audiodir):
        data_list = []
        video_files = glob.glob(os.path.join(videodir, '**', '*.mp4'), recursive=True)
        
        for f in tqdm(video_files):
            name = os.path.basename(f).split('.')[0]
            cur_whisper_path = os.path.join(audiodir,f'{name}.pt')
            if os.path.exists(cur_whisper_path):
                d = {"video_path":f,"whisper_emb_path":cur_whisper_path}
                data_list.append(d)

        print(len(data_list))
        return data_list

    data = []
    data += process_video_whisper('/mnt/pfs-gv8sxa/tts/dhg/zqy/data/celeb/videos','/mnt/pfs-gv8sxa/tts/dhg/zqy/data/celeb/whisper_embs')
    data += process_video_whisper('/mnt/pfs-mc0p4k/tts/team/digital_avatar_group/liuhuaize/dataset/FaceVid_240h/videos','/mnt/pfs-mc0p4k/tts/team/digital_avatar_group/liuhuaize/dataset/FaceVid_240h/whisper_embs')
    print(f'Total num of data:{len(data)}')
    save_dir = '/mnt/pfs-mc0p4k/tts/team/digital_avatar_group/sunwenzhang/qiyuan/code/AMD2/dataset/path/video_whisper_pose'
    eval_data = data[:100]
    train_data = data[100:]
    with open('./path/video_whisper_pose/train.pkl', 'wb') as file:
        # 使用pickle模块的dump方法保存数据
        pickle.dump(train_data, file)

    with open('./path/video_whisper_pose/eval.pkl', 'wb') as file:
        # 使用pickle模块的dump方法保存数据
        pickle.dump(eval_data, file)

def AMD_process_video(videodir = None):
    if videodir is None :
        video_dir = './train_data_amd.txt'
    else:
        video_dir = videodir

    with open(video_dir, 'r') as file:
        lines = file.readlines()
    video_dirs = [line.strip() for line in lines]

    train_files = []
    eval_files = []
    for dir in video_dirs:
        cur_files = glob.glob(os.path.join(dir, '**', '*.mp4'), recursive=True)
        random.shuffle(cur_files)
        eval_files += cur_files[:20]
        train_files += cur_files[20:]

    print(f'Total {len(train_files)} !!!')
    print(f'Total {len(eval_files)} !!!')

    with open('./path/amd2/train.pkl', 'wb') as file:
        pickle.dump(train_files, file)

    with open('./path/amd2/eval.pkl', 'wb') as file:
        pickle.dump(eval_files, file)

def A2M_process_video(videodirs = None,output_dir = None):
    # if videodir is None :
    #     video_dir = './train_data_a2m.txt'
    # else:
    #     video_dir = videodir

    # with open(video_dir, 'r') as file:
    #     lines = file.readlines()
    # video_dirs = [line.strip() for line in lines]

    train_files = []
    eval_files = []
    max_num= 1000000
    if videodirs is None or output_dir is None:
        return
    if not isinstance(videodirs,list):
        videodirs = [videodirs]
    for dir in videodirs:
        cur_datalist = []
        cur_files = glob.glob(os.path.join(dir, '**', '*.mp4'), recursive=True)
        for f in tqdm(cur_files):
            whisper_path = f.replace('videos','whisper_embs').replace('.mp4','.pt')
            if os.path.exists(whisper_path) :
                cur_datalist.append({'video_path':f,'whisper_emb_path':whisper_path})
            if len(cur_datalist) > max_num:
                break
        random.shuffle(cur_datalist)
        eval_files += cur_datalist[-20:]
        train_files += cur_datalist[:-20]


    print(f'Total {len(train_files)} !!!')
    print(f'Total {len(eval_files)} !!!')

    with open(os.path.join(output_dir,'train.pkl'), 'wb') as file:
        pickle.dump(train_files, file)

    with open(os.path.join(output_dir,'eval.pkl'), 'wb') as file:
        pickle.dump(eval_files, file)



def A2V_process_video(videodir = None):
    # if videodir is None :
    #     video_dir = './a2v_data.txt'
    # else:
    #     video_dir = videodir

    # with open(video_dir, 'r') as file:
    #     lines = file.readlines()
    # video_dirs = [line.strip() for line in lines]

    video_dirs = ["/mnt/pfs-gv8sxa/tts/dhg/zqy/data/recon_data/videos"]

    files = []

    for dir in video_dirs:
        cur_datalist = []
        cur_files = glob.glob(os.path.join(dir, '**', '*.mp4'), recursive=True)
        random.shuffle(cur_files)
        cur_files = cur_files[:50]

        for f in tqdm(cur_files):
            whisper_path = f.replace('videos','whisper_embs').replace('.mp4','.pt')
            audio_path =  f.replace('videos','audios').replace('.mp4','.wav')
            if os.path.exists(whisper_path)  and os.path.exists(audio_path):
                cur_datalist.append({'video_path':f,
                                     'whisper_emb_path':whisper_path,
                                     'audio_path':audio_path})

        files += cur_datalist

    print(f'Total {len(files)} !!!')

    with open('./path/a2v_recon_demo/data.pkl', 'wb') as file:
        pickle.dump(files, file)




def create_video_from_frames(frame_folder, output_video_path, fps=30):
    # 获取所有图片路径，并按名称排序以保持顺序
    frame_files = sorted([os.path.join(frame_folder, f) for f in os.listdir(frame_folder) if f.endswith('.png')])
    
    # 读取并存储帧
    frames = [iio.imread(f) for f in frame_files]
    
    # 将帧写入视频文件
    iio.imwrite(output_video_path, frames, fps=fps)


def fullbody_imgs2video_multiprocess(datadir,output_dir):
    pool = Pool(64)
    images_folders = []
    
    data_dirs = [os.path.join(datadir,d) for d in os.listdir(datadir)]
    for data in tqdm(data_dirs):
        name = os.path.basename(data)
        frame_folder = os.path.join(data,'content','images')
        video_out_path = os.path.join(output_dir,f'{name}.mp4')

        pool.apply_async(create_video_from_frames, args=(frame_folder,video_out_path))
    
    
    pool.close()

    pool.join()

    print("结束主进程……")


def fullbody_imgs2video(datadir,output_dir):
    pool = Pool(64)
    source_dir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/dataset_0703/datasets_tiktok_0703_clean'
    target_dir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/full_body/dwpose'

    source_name = [n for n in os.listdir(source_dir)]
    target_name = [n.split('.')[0] for n in os.listdir(target_dir)]

    print(f'source {len(source_name)}')
    print(f'target {len(target_name)}')

    set_s = set(source_name)
    set_t = set(target_name)
    need_name = list(set_s-set_t)
    
    data_dirs = [os.path.join(datadir,n) for n in need_name]
    for data in tqdm(data_dirs):
        name = os.path.basename(data)
        frame_folder = os.path.join(data,'content','dwpose')
        video_out_path = os.path.join(output_dir,f'{name}.mp4')

        pool.apply_async(create_video_from_frames, args=(frame_folder,video_out_path))
    
    
    pool.close()

    pool.join()

    print("结束主进程……")

def fullbody_video():
    datadir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/full_body/video'
    video_files = glob.glob(os.path.join(datadir, '**', '*.mp4'), recursive=True)
    cur_files = [f for f in video_files if 'TIKTOK' not in f]
    print(len(cur_files))
    train_files = cur_files
    eval_files = cur_files[:100]

    with open('./path/amd_fullbody/train.pkl', 'wb') as file:
        pickle.dump(train_files, file)

    with open('./path/amd_fullbody/eval.pkl', 'wb') as file:
        pickle.dump(eval_files, file)

def test_data_time():
    import time
    from tqdm import tqdm
    from decord import VideoReader
    from decord import cpu, gpu
    import glob
    import os
    datalist = []


    video_dir = '/mnt/pfs-mc0p4k/tts/team/digital_avatar_group/sunwenzhang/qiyuan/code/AMD2/dataset/train_data_amd.txt'

    with open(video_dir, 'r') as file:
        lines = file.readlines()
    video_dirs = [line.strip() for line in lines]

    files = []
    for dir in video_dirs:
        cur_files = glob.glob(os.path.join(dir, '**', '*.mp4'), recursive=True)
        files += cur_files

    for f in tqdm(files):
        try:
            start_time = time.time()  # 记录开始时间

            test_idx = [1,2,3,4,5,6,7]
            video_reader = VideoReader(f, ctx=cpu(0))
            video_length = len(video_reader)
            if video_length < 10 :
                continue
            video = video_reader.get_batch(test_idx)

            end_time = time.time()   # 记录结束时间

            execution_time = end_time - start_time
            if execution_time <= 3.0:
                datalist.append(f)
            else:
                print(f'{f} # Time:{execution_time} # Frame:{video_length}')
        except Exception as e:
            continue

    with open('./amd_available_data.pkl', 'wb') as file:
        pickle.dump(datalist, file)


def get_split_video_motion_path_pkl(video_path_list:list,output_dir:str):

    def split_list_into_n(lst, n=8):
        if n > len(lst):
            n = len(lst)  # 如果 n 大于列表长度，则设置 n 为列表长度
        k, m = divmod(len(lst), n)

        result = []
        for i in range(n):
            start_index = i * k + min(i, m)
            end_index = (i + 1) * k + min(i + 1, m)
            result.append(lst[start_index:end_index])
        
        return result

    datas = [] 
    for f in tqdm(video_path_list):
        name = os.path.basename(f).split('.')[0]
        motion_dir = os.path.join(os.path.dirname(os.path.dirname(f)),'motion')
        motion_path = os.path.join(motion_dir,f'{name}.pt')

        datas.append({'video_path':f,'motion_path':motion_path})

    print(f'total avail data:{len(datas)}')

    # split
    split_data_list = split_list_into_n(datas,8)

    for i,f in enumerate(split_data_list):
        output_path = os.path.join(output_dir,f'{i}.pkl')
        print(f'send data to {output_path},len:{len(f)}')
        with open(output_path, 'wb') as file:
            pickle.dump(f, file)

def get_exp_a2m_data(video_dir,
                     save_dir,
                     num=200,
                     duration=5):
    path = os.path.join("test", "exp", "sample")

    # 递归创建目录（exist_ok=True确保目录存在时不报错）
    video_save_dir = os.path.join(save_dir, "videos")
    wav_save_dir = os.path.join(save_dir, "wav")
    mp3_save_dir = os.path.join(save_dir, "mp3")
    whisper_save_dir = os.path.join(save_dir, "whisper_embs")
    refimg_save_dir = os.path.join(save_dir, "refimg")
    os.makedirs(video_save_dir, exist_ok=True)
    os.makedirs(wav_save_dir, exist_ok=True)
    os.makedirs(mp3_save_dir, exist_ok=True)
    os.makedirs(whisper_save_dir, exist_ok=True)
    os.makedirs(refimg_save_dir, exist_ok=True)


    # get all video path
    video_files = glob.glob(os.path.join(video_dir, '**', '*.mp4'), recursive=True)
    random.shuffle(video_files)

    save_num = 0
    for f in tqdm(video_files):
        if save_num >= num :
            break

        name = f.split('/')[-1].split('.')[0]
        whisper_path = f.replace('videos','whisper_embs').replace('.mp4','.pt')
        audio_path = f.replace('videos','audios').replace('.mp4','.wav')

        try:
            assert os.path.exists(whisper_path)

            with VideoFileClip(f) as clip:
                # save path
                video_output_path = os.path.join(video_save_dir, f"{name}.mp4")
                wav_output_path = os.path.join(wav_save_dir, f"{name}.wav")
                mp3_output_path = os.path.join(mp3_save_dir, f"{name}.mp3")
                whisper_output_path = os.path.join(whisper_save_dir, f"{name}.pt")
                refimg_output_path = os.path.join(refimg_save_dir, f"{name}.png")

                # 计算实际截取时长（处理短于5秒的视频）
                assert clip.duration >= 5
                
                # 截取视频片段
                subclip = clip.subclip(0, duration)
                
                # 保存视频片段（包含音频）
                subclip.write_videofile(video_output_path, 
                                      codec='libx264', 
                                      audio_codec='aac',
                                      logger=None)  # 禁用进度输出
                
                # 提取音频并保存为两种格式
                audio = subclip.audio
                audio.write_audiofile(wav_output_path, 
                                    codec='pcm_s16le',
                                    logger=None)
                audio.write_audiofile(mp3_output_path, 
                                    codec='libmp3lame',
                                    bitrate='192k',
                                    logger=None)
            
                # copy whisper emb   
                shutil.copy(whisper_path, whisper_output_path)     

                # extract refimg
                clip.save_frame(refimg_output_path, t=0)

                # 显式关闭资源
                subclip.close()
                audio.close()

                save_num += 1
        except Exception as e:
            print(e)
            continue

def extract_first_frame(video_dir, save_dir):
    video_files = glob.glob(os.path.join(video_dir, '**', '*.mp4'), recursive=True)
    refimg_save_dir = os.path.join(save_dir, "refimg")
    os.makedirs(refimg_save_dir, exist_ok=True)

    idx = 0
    for f in tqdm(video_files):
        name = os.path.basename(f).split('.')[0]
        output_path = os.path.join(refimg_save_dir, f"{name}.png")
        # 读取视频文件
        vidcap = cv2.VideoCapture(f)
        
        # 获取第一帧
        success, image = vidcap.read()
        
        if success:
            # 转换 BGR 到 RGB（可选，取决于是否需要颜色校正）
            # image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
            
            # 保存为 PNG（无损格式）
            cv2.imwrite(output_path, image)
            idx+=1
            print(f"保存成功: {output_path},保存数量{idx}")
        else:
            print("无法读取视频帧")



if __name__ == '__main__': 
    # data_dir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/dataset_0703/datasets_tiktok_0703_clean'
    # output_dir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/full_body/dwpose'
    # fullbody_imgs2video(data_dir,output_dir)

    # # AMD Video
    # AMD_process_video()


    # video_dir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/celeb/whisper_embs'
    # video_files = os.listdir(video_dir)
    # print(len(video_files)) 

    # split_video_whisper()
       

    # # ------------------ test data time ---------
    # test_data_time()

    # # ------------------ split video motion data ---------
    # video_audio_dir = ['/mnt/pfs-gv8sxa/tts/dhg/zqy/data/celeb/videos',
    #                    '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/Asian_emo/videos',
    #                    '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/FaceVid_240h/videos',
    #                    '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/mead/videos',
    #                    '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/hdtf_dit/videos']
    # video_path_list = []
    # for directory in video_audio_dir:
    #     cur_video_files = glob.glob(os.path.join(directory, '**', '*.mp4'), recursive=True)
    #     video_path_list += cur_video_files
    # print('一共有这么多视频',len(video_path_list))
    # output_dir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/code/AMD2/dataset/path/a2m/videos_pkl'
    # get_split_video_motion_path_pkl(video_path_list,output_dir)

    # # -------------------- A2M video ------------------
    # A2M_process_video(["/mnt/pfs-gv8sxa/tts/dhg/zqy/data/FaceVid_240h/videos",
    #                    "/mnt/pfs-gv8sxa/tts/dhg/zqy/data/hdtf_dit/videos"],"/mnt/pfs-gv8sxa/tts/dhg/zqy/code/AMD2/dataset/path/a2m_hdtf_facevid")

    # -------------------- A2V video -------------------
    A2V_process_video()

    # # --------------------- A2M exp --------------
    # get_exp_a2m_data(video_dir='/mnt/pfs-gv8sxa/tts/dhg/zqy/data/hdtf_dit/videos',
    #                  save_dir='/mnt/pfs-gv8sxa/tts/dhg/zqy/code/evaluate/A2M/data/a2m/hdtf_200')

    # extract_first_frame('/mnt/pfs-gv8sxa/tts/dhg/zqy/code/evaluate/A2M/data/a2m/hdtf_200/videos',
    #                     '/mnt/pfs-gv8sxa/tts/dhg/zqy/code/evaluate/A2M/data/a2m/hdtf_200')

	import glob
	import os
	import pickle
	import random
	from tqdm import tqdm
	import imageio.v3 as iio
	import os
	from multiprocessing import Pool
	import os
	from moviepy.editor import VideoFileClip
	import argparse
	import shutil
	import cv2


	def split_video_whisper():
	def process_video_whisper(videodir,audiodir):
	data_list = []
	video_files = glob.glob(os.path.join(videodir, '*', '.mp4'), recursive=True)

	for f in tqdm(video_files):
	name = os.path.basename(f).split('.')[0]
	cur_whisper_path = os.path.join(audiodir,f'{name}.pt')
	if os.path.exists(cur_whisper_path):
	d = {"video_path":f,"whisper_emb_path":cur_whisper_path}
	data_list.append(d)

	print(len(data_list))
	return data_list

	data = []
	data += process_video_whisper('/mnt/pfs-gv8sxa/tts/dhg/zqy/data/celeb/videos','/mnt/pfs-gv8sxa/tts/dhg/zqy/data/celeb/whisper_embs')
	data += process_video_whisper('/mnt/pfs-mc0p4k/tts/team/digital_avatar_group/liuhuaize/dataset/FaceVid_240h/videos','/mnt/pfs-mc0p4k/tts/team/digital_avatar_group/liuhuaize/dataset/FaceVid_240h/whisper_embs')
	print(f'Total num of data:{len(data)}')
	save_dir = '/mnt/pfs-mc0p4k/tts/team/digital_avatar_group/sunwenzhang/qiyuan/code/AMD2/dataset/path/video_whisper_pose'
	eval_data = data[:100]
	train_data = data[100:]
	with open('./path/video_whisper_pose/train.pkl', 'wb') as file:
	# 使用pickle模块的dump方法保存数据
	pickle.dump(train_data, file)

	with open('./path/video_whisper_pose/eval.pkl', 'wb') as file:
	# 使用pickle模块的dump方法保存数据
	pickle.dump(eval_data, file)

	def AMD_process_video(videodir = None):
	if videodir is None :
	video_dir = './train_data_amd.txt'
	else:
	video_dir = videodir

	with open(video_dir, 'r') as file:
	lines = file.readlines()
	video_dirs = [line.strip() for line in lines]

	train_files = []
	eval_files = []
	for dir in video_dirs:
	cur_files = glob.glob(os.path.join(dir, '*', '.mp4'), recursive=True)
	random.shuffle(cur_files)
	eval_files += cur_files[:20]
	train_files += cur_files[20:]

	print(f'Total {len(train_files)} !!!')
	print(f'Total {len(eval_files)} !!!')

	with open('./path/amd2/train.pkl', 'wb') as file:
	pickle.dump(train_files, file)

	with open('./path/amd2/eval.pkl', 'wb') as file:
	pickle.dump(eval_files, file)

	def A2M_process_video(videodirs = None,output_dir = None):
	# if videodir is None :
	# video_dir = './train_data_a2m.txt'
	# else:
	# video_dir = videodir

	# with open(video_dir, 'r') as file:
	# lines = file.readlines()
	# video_dirs = [line.strip() for line in lines]

	train_files = []
	eval_files = []
	max_num= 1000000
	if videodirs is None or output_dir is None:
	return
	if not isinstance(videodirs,list):
	videodirs = [videodirs]
	for dir in videodirs:
	cur_datalist = []
	cur_files = glob.glob(os.path.join(dir, '*', '.mp4'), recursive=True)
	for f in tqdm(cur_files):
	whisper_path = f.replace('videos','whisper_embs').replace('.mp4','.pt')
	if os.path.exists(whisper_path) :
	cur_datalist.append({'video_path':f,'whisper_emb_path':whisper_path})
	if len(cur_datalist) > max_num:
	break
	random.shuffle(cur_datalist)
	eval_files += cur_datalist[-20:]
	train_files += cur_datalist[:-20]


	print(f'Total {len(train_files)} !!!')
	print(f'Total {len(eval_files)} !!!')

	with open(os.path.join(output_dir,'train.pkl'), 'wb') as file:
	pickle.dump(train_files, file)

	with open(os.path.join(output_dir,'eval.pkl'), 'wb') as file:
	pickle.dump(eval_files, file)



	def A2V_process_video(videodir = None):
	# if videodir is None :
	# video_dir = './a2v_data.txt'
	# else:
	# video_dir = videodir

	# with open(video_dir, 'r') as file:
	# lines = file.readlines()
	# video_dirs = [line.strip() for line in lines]

	video_dirs = ["/mnt/pfs-gv8sxa/tts/dhg/zqy/data/recon_data/videos"]

	files = []

	for dir in video_dirs:
	cur_datalist = []
	cur_files = glob.glob(os.path.join(dir, '*', '.mp4'), recursive=True)
	random.shuffle(cur_files)
	cur_files = cur_files[:50]

	for f in tqdm(cur_files):
	whisper_path = f.replace('videos','whisper_embs').replace('.mp4','.pt')
	audio_path = f.replace('videos','audios').replace('.mp4','.wav')
	if os.path.exists(whisper_path) and os.path.exists(audio_path):
	cur_datalist.append({'video_path':f,
	'whisper_emb_path':whisper_path,
	'audio_path':audio_path})

	files += cur_datalist

	print(f'Total {len(files)} !!!')

	with open('./path/a2v_recon_demo/data.pkl', 'wb') as file:
	pickle.dump(files, file)




	def create_video_from_frames(frame_folder, output_video_path, fps=30):
	# 获取所有图片路径，并按名称排序以保持顺序
	frame_files = sorted([os.path.join(frame_folder, f) for f in os.listdir(frame_folder) if f.endswith('.png')])

	# 读取并存储帧
	frames = [iio.imread(f) for f in frame_files]

	# 将帧写入视频文件
	iio.imwrite(output_video_path, frames, fps=fps)


	def fullbody_imgs2video_multiprocess(datadir,output_dir):
	pool = Pool(64)
	images_folders = []

	data_dirs = [os.path.join(datadir,d) for d in os.listdir(datadir)]
	for data in tqdm(data_dirs):
	name = os.path.basename(data)
	frame_folder = os.path.join(data,'content','images')
	video_out_path = os.path.join(output_dir,f'{name}.mp4')

	pool.apply_async(create_video_from_frames, args=(frame_folder,video_out_path))


	pool.close()

	pool.join()

	print("结束主进程……")


	def fullbody_imgs2video(datadir,output_dir):
	pool = Pool(64)
	source_dir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/dataset_0703/datasets_tiktok_0703_clean'
	target_dir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/full_body/dwpose'

	source_name = [n for n in os.listdir(source_dir)]
	target_name = [n.split('.')[0] for n in os.listdir(target_dir)]

	print(f'source {len(source_name)}')
	print(f'target {len(target_name)}')

	set_s = set(source_name)
	set_t = set(target_name)
	need_name = list(set_s-set_t)

	data_dirs = [os.path.join(datadir,n) for n in need_name]
	for data in tqdm(data_dirs):
	name = os.path.basename(data)
	frame_folder = os.path.join(data,'content','dwpose')
	video_out_path = os.path.join(output_dir,f'{name}.mp4')

	pool.apply_async(create_video_from_frames, args=(frame_folder,video_out_path))


	pool.close()

	pool.join()

	print("结束主进程……")

	def fullbody_video():
	datadir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/full_body/video'
	video_files = glob.glob(os.path.join(datadir, '*', '.mp4'), recursive=True)
	cur_files = [f for f in video_files if 'TIKTOK' not in f]
	print(len(cur_files))
	train_files = cur_files
	eval_files = cur_files[:100]

	with open('./path/amd_fullbody/train.pkl', 'wb') as file:
	pickle.dump(train_files, file)

	with open('./path/amd_fullbody/eval.pkl', 'wb') as file:
	pickle.dump(eval_files, file)

	def test_data_time():
	import time
	from tqdm import tqdm
	from decord import VideoReader
	from decord import cpu, gpu
	import glob
	import os
	datalist = []


	video_dir = '/mnt/pfs-mc0p4k/tts/team/digital_avatar_group/sunwenzhang/qiyuan/code/AMD2/dataset/train_data_amd.txt'

	with open(video_dir, 'r') as file:
	lines = file.readlines()
	video_dirs = [line.strip() for line in lines]

	files = []
	for dir in video_dirs:
	cur_files = glob.glob(os.path.join(dir, '*', '.mp4'), recursive=True)
	files += cur_files

	for f in tqdm(files):
	try:
	start_time = time.time() # 记录开始时间

	test_idx = [1,2,3,4,5,6,7]
	video_reader = VideoReader(f, ctx=cpu(0))
	video_length = len(video_reader)
	if video_length < 10 :
	continue
	video = video_reader.get_batch(test_idx)

	end_time = time.time() # 记录结束时间

	execution_time = end_time - start_time
	if execution_time <= 3.0:
	datalist.append(f)
	else:
	print(f'{f} # Time:{execution_time} # Frame:{video_length}')
	except Exception as e:
	continue

	with open('./amd_available_data.pkl', 'wb') as file:
	pickle.dump(datalist, file)


	def get_split_video_motion_path_pkl(video_path_list:list,output_dir:str):

	def split_list_into_n(lst, n=8):
	if n > len(lst):
	n = len(lst) # 如果 n 大于列表长度，则设置 n 为列表长度
	k, m = divmod(len(lst), n)

	result = []
	for i in range(n):
	start_index = i * k + min(i, m)
	end_index = (i + 1) * k + min(i + 1, m)
	result.append(lst[start_index:end_index])

	return result

	datas = []
	for f in tqdm(video_path_list):
	name = os.path.basename(f).split('.')[0]
	motion_dir = os.path.join(os.path.dirname(os.path.dirname(f)),'motion')
	motion_path = os.path.join(motion_dir,f'{name}.pt')

	datas.append({'video_path':f,'motion_path':motion_path})

	print(f'total avail data:{len(datas)}')

	# split
	split_data_list = split_list_into_n(datas,8)

	for i,f in enumerate(split_data_list):
	output_path = os.path.join(output_dir,f'{i}.pkl')
	print(f'send data to {output_path},len:{len(f)}')
	with open(output_path, 'wb') as file:
	pickle.dump(f, file)

	def get_exp_a2m_data(video_dir,
	save_dir,
	num=200,
	duration=5):
	path = os.path.join("test", "exp", "sample")

	# 递归创建目录（exist_ok=True确保目录存在时不报错）
	video_save_dir = os.path.join(save_dir, "videos")
	wav_save_dir = os.path.join(save_dir, "wav")
	mp3_save_dir = os.path.join(save_dir, "mp3")
	whisper_save_dir = os.path.join(save_dir, "whisper_embs")
	refimg_save_dir = os.path.join(save_dir, "refimg")
	os.makedirs(video_save_dir, exist_ok=True)
	os.makedirs(wav_save_dir, exist_ok=True)
	os.makedirs(mp3_save_dir, exist_ok=True)
	os.makedirs(whisper_save_dir, exist_ok=True)
	os.makedirs(refimg_save_dir, exist_ok=True)


	# get all video path
	video_files = glob.glob(os.path.join(video_dir, '*', '.mp4'), recursive=True)
	random.shuffle(video_files)

	save_num = 0
	for f in tqdm(video_files):
	if save_num >= num :
	break

	name = f.split('/')[-1].split('.')[0]
	whisper_path = f.replace('videos','whisper_embs').replace('.mp4','.pt')
	audio_path = f.replace('videos','audios').replace('.mp4','.wav')

	try:
	assert os.path.exists(whisper_path)

	with VideoFileClip(f) as clip:
	# save path
	video_output_path = os.path.join(video_save_dir, f"{name}.mp4")
	wav_output_path = os.path.join(wav_save_dir, f"{name}.wav")
	mp3_output_path = os.path.join(mp3_save_dir, f"{name}.mp3")
	whisper_output_path = os.path.join(whisper_save_dir, f"{name}.pt")
	refimg_output_path = os.path.join(refimg_save_dir, f"{name}.png")

	# 计算实际截取时长（处理短于5秒的视频）
	assert clip.duration >= 5

	# 截取视频片段
	subclip = clip.subclip(0, duration)

	# 保存视频片段（包含音频）
	subclip.write_videofile(video_output_path,
	codec='libx264',
	audio_codec='aac',
	logger=None) # 禁用进度输出

	# 提取音频并保存为两种格式
	audio = subclip.audio
	audio.write_audiofile(wav_output_path,
	codec='pcm_s16le',
	logger=None)
	audio.write_audiofile(mp3_output_path,
	codec='libmp3lame',
	bitrate='192k',
	logger=None)

	# copy whisper emb
	shutil.copy(whisper_path, whisper_output_path)

	# extract refimg
	clip.save_frame(refimg_output_path, t=0)

	# 显式关闭资源
	subclip.close()
	audio.close()

	save_num += 1
	except Exception as e:
	print(e)
	continue

	def extract_first_frame(video_dir, save_dir):
	video_files = glob.glob(os.path.join(video_dir, '*', '.mp4'), recursive=True)
	refimg_save_dir = os.path.join(save_dir, "refimg")
	os.makedirs(refimg_save_dir, exist_ok=True)

	idx = 0
	for f in tqdm(video_files):
	name = os.path.basename(f).split('.')[0]
	output_path = os.path.join(refimg_save_dir, f"{name}.png")
	# 读取视频文件
	vidcap = cv2.VideoCapture(f)

	# 获取第一帧
	success, image = vidcap.read()

	if success:
	# 转换 BGR 到 RGB（可选，取决于是否需要颜色校正）
	# image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

	# 保存为 PNG（无损格式）
	cv2.imwrite(output_path, image)
	idx+=1
	print(f"保存成功: {output_path},保存数量{idx}")
	else:
	print("无法读取视频帧")



	if __name__ == '__main__':
	# data_dir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/dataset_0703/datasets_tiktok_0703_clean'
	# output_dir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/full_body/dwpose'
	# fullbody_imgs2video(data_dir,output_dir)

	# # AMD Video
	# AMD_process_video()


	# video_dir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/celeb/whisper_embs'
	# video_files = os.listdir(video_dir)
	# print(len(video_files))

	# split_video_whisper()


	# # ------------------ test data time ---------
	# test_data_time()

	# # ------------------ split video motion data ---------
	# video_audio_dir = ['/mnt/pfs-gv8sxa/tts/dhg/zqy/data/celeb/videos',
	# '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/Asian_emo/videos',
	# '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/FaceVid_240h/videos',
	# '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/mead/videos',
	# '/mnt/pfs-gv8sxa/tts/dhg/zqy/data/hdtf_dit/videos']
	# video_path_list = []
	# for directory in video_audio_dir:
	# cur_video_files = glob.glob(os.path.join(directory, '*', '.mp4'), recursive=True)
	# video_path_list += cur_video_files
	# print('一共有这么多视频',len(video_path_list))
	# output_dir = '/mnt/pfs-gv8sxa/tts/dhg/zqy/code/AMD2/dataset/path/a2m/videos_pkl'
	# get_split_video_motion_path_pkl(video_path_list,output_dir)

	# # -------------------- A2M video ------------------
	# A2M_process_video(["/mnt/pfs-gv8sxa/tts/dhg/zqy/data/FaceVid_240h/videos",
	# "/mnt/pfs-gv8sxa/tts/dhg/zqy/data/hdtf_dit/videos"],"/mnt/pfs-gv8sxa/tts/dhg/zqy/code/AMD2/dataset/path/a2m_hdtf_facevid")

	# -------------------- A2V video -------------------
	A2V_process_video()

	# # --------------------- A2M exp --------------
	# get_exp_a2m_data(video_dir='/mnt/pfs-gv8sxa/tts/dhg/zqy/data/hdtf_dit/videos',
	# save_dir='/mnt/pfs-gv8sxa/tts/dhg/zqy/code/evaluate/A2M/data/a2m/hdtf_200')

	# extract_first_frame('/mnt/pfs-gv8sxa/tts/dhg/zqy/code/evaluate/A2M/data/a2m/hdtf_200/videos',
	# '/mnt/pfs-gv8sxa/tts/dhg/zqy/code/evaluate/A2M/data/a2m/hdtf_200')