diff --git a/AIMeiSheng/docker_demo/Dockerfile b/AIMeiSheng/docker_demo/Dockerfile index 3ba096e..dd9f0dd 100644 --- a/AIMeiSheng/docker_demo/Dockerfile +++ b/AIMeiSheng/docker_demo/Dockerfile @@ -1,25 +1,25 @@ # 系统版本 CUDA Version 11.8.0 # NAME="CentOS Linux" VERSION="7 (Core)" # FROM starmaker.tencentcloudcr.com/starmaker/av/av:1.1 # 基础镜像, python3.9,cuda118,centos7，外加ffmpeg #FROM starmaker.tencentcloudcr.com/starmaker/av/av_base:1.0 FROM av_base_test:1.0 RUN source /etc/profile && sed -i 's|mirrorlist=|#mirrorlist=|g' /etc/yum.repos.d/CentOS-Base.repo && sed -i 's|#baseurl=http://mirror.centos.org|baseurl=http://vault.centos.org|g' /etc/yum.repos.d/CentOS-Base.repo && yum clean all && yum install -y unzip && yum install -y libsndfile && yum install -y libsamplerate libsamplerate-devel RUN source /etc/profile && pip3 install librosa && pip3 install gradio && pip3 install torch==2.1.2 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 RUN source /etc/profile && pip3 install urllib3==1.26.15 && pip3 install coscmd && coscmd config -a AKIDoQmshFWXGitnQmrfCTYNwEExPaU6RVHm -s F9n9E2ZonWy93f04qMaYFfogHadPt62h -b log-sg-1256122840 -r ap-singapore RUN source /etc/profile && pip3 install asteroid-filterbanks RUN source /etc/profile && pip3 install praat-parselmouth==0.4.3 RUN source /etc/profile && pip3 install pyworld RUN source /etc/profile && pip3 install faiss-cpu RUN source /etc/profile && pip3 install torchcrepe RUN source /etc/profile && pip3 install thop RUN source /etc/profile && pip3 install ffmpeg-python RUN source /etc/profile && pip3 install fairseq -RUN source /etc/profile && pip3 install redis +RUN source /etc/profile && pip3 install redis==4.5.0 WORKDIR /data/code CMD ["/bin/bash", "-c", "source /etc/profile; export PYTHONPATH=/data/code; cd /data/code/AIMeiSheng/docker_demo; python3 offline_server.py"] \ No newline at end of file diff --git a/AIMeiSheng/docker_demo/readme.txt b/AIMeiSheng/docker_demo/readme.txt index beaf3e6..ac14561 100644 --- a/AIMeiSheng/docker_demo/readme.txt +++ b/AIMeiSheng/docker_demo/readme.txt @@ -1,24 +1,24 @@ 简介: ai美声功能，其核心是输入一段15-30s的人声作为音色信息，再给定输入音源，将音源转换为指定音色的声音的效果。例如，孙燕姿演唱的东风破 架构方案: http_server.py (1个) 作为服务端，接收外部传来的数据，塞入到redis中，由offline_server.py (多个服务) 进行承接 # 部署要求: 1. http_server.py 部署在sg-prod-songrefresh-gpu-7 上 2. offline_server.py 使用docker 部署在超级节点上，由运维进行控制 # http_server.py 环境要求: pip install redis pip install flask # offline_server.py 环境要求(docker) cd docker_demo目录下(例子如下): 1. docker build -f Dockerfile -t av_ai_meisheng . (通过docker images 获取av_ai_meisheng的image_id) 2. docker run --gpus all -it -v /data/rsync/jianli.yang/av_svc:/data/code image_id # 即可启动服务 # 测试代码: docker 环境下, offline_server.py 即可验证 # http测试命令: -curl http://127.0.0.1:5004/ai_meisheng -H "Content-Type: application/json" -d '{ "record_song_url": "https://av-audit-sync-sg-1256122840.cos.ap-singapore.myqcloud.com/dataset/AIMeiSheng/vocal_test/yinse.m4a", "target_url": "https://av-audit-sync-sg-1256122840.cos.ap-singapore.myqcloud.com/dataset/AIMeiSheng/vocal_test/out.m4a","start": 0,"end": 15000,"vocal_loudness": -14.57,"female_recording_url": "https://av-audit-sync-sg-1256122840.cos.ap-singapore.myqcloud.com/dataset/AIMeiSheng/vocal_test/female.m4a", "male_recording_url": "https://av-audit-sync-sg-1256122840.cos.ap-singapore.myqcloud.com/dataset/AIMeiSheng/vocal_test/male.m4a"}' +curl http://sg-prod-songrefresh-gpu-7:5000/ai_meisheng -H "Content-Type: application/json" -d '{ "record_song_url": "http://starmaker-in-1256122840.cos.ap-mumbai.myqcloud.com/production/ai_voice/7036874317772028/78d3148926c222585b8079d02fd6aa3e-source.mp4", "target_url": "https://av-audit-sync-sg-1256122840.cos.ap-singapore.myqcloud.com/dataset/AIMeiSheng/vocal_test/out.m4a","start": 0,"end": 15000,"vocal_loudness": -14.57,"female_recording_url": "https://av-audit-sync-sg-1256122840.cos.ap-singapore.myqcloud.com/dataset/AIMeiSheng/vocal_test/female.m4a", "male_recording_url": "https://av-audit-sync-sg-1256122840.cos.ap-singapore.myqcloud.com/dataset/AIMeiSheng/vocal_test/male.m4a"}' {"gender":"male","schedule":100,"status":0,"target_song_url":"https://av-audit-sync-sg-1256122840.cos.ap-singapore.myqcloud.com/dataset/AIMeiSheng/vocal_test/out.m4a"} # 资源消耗: 显存占用约2G,但是最高能到9G, 所以，一台机器部署一个即可 注意: 通过common.py 的prod可以控制是否是线上环境 \ No newline at end of file diff --git a/AIMeiSheng/docker_demo/svc_online.py b/AIMeiSheng/docker_demo/svc_online.py index db0f255..e910f5f 100644 --- a/AIMeiSheng/docker_demo/svc_online.py +++ b/AIMeiSheng/docker_demo/svc_online.py @@ -1,188 +1,188 @@ # -*- coding: UTF-8 -*- """ SVC的核心处理逻辑 """ import os import time import socket import shutil import hashlib from AIMeiSheng.meisheng_svc_final import load_model, process_svc_online from AIMeiSheng.meisheng_env_preparex import meisheng_env_prepare from AIMeiSheng.voice_classification.online.voice_class_online_fang import VoiceClass, download_volume_balanced from AIMeiSheng.docker_demo.common import * import logging hostname = socket.gethostname() log_file_name = f"{os.path.dirname(os.path.abspath(__file__))}/av_meisheng_{hostname}.log" # 设置logger svc_offline_logger = logging.getLogger("svc_offline") file_handler = logging.FileHandler(log_file_name) file_handler.setLevel(logging.INFO) formatter = logging.Formatter('%(asctime)s %(levelname)s %(message)s', datefmt='%Y-%m-%d %I:%M:%S') file_handler.setFormatter(formatter) if gs_prod: svc_offline_logger.addHandler(file_handler) if os.path.exists(gs_tmp_dir): shutil.rmtree(gs_tmp_dir) os.makedirs(gs_model_dir, exist_ok=True) os.makedirs(gs_resource_cache_dir, exist_ok=True) # 预设参数 gs_gender_models_url = "https://av-audit-sync-sg-1256122840.cos.ap-singapore.myqcloud.com/hub/voice_classification/models.zip" gs_volume_bin_url = "https://av-audit-sync-sg-1256122840.cos.ap-singapore.myqcloud.com/dataset/AIMeiSheng/ebur128_tool" class GSWorkerAttr: def __init__(self, input_data): # 取出输入资源 vocal_url = input_data["record_song_url"] target_url = input_data["target_url"] start = input_data["start"] # 单位是ms end = input_data["end"] # 单位是ms vocal_loudness = input_data["vocal_loudness"] female_recording_url = input_data["female_recording_url"] male_recording_url = input_data["male_recording_url"] self.distinct_id = hashlib.md5(vocal_url.encode()).hexdigest() self.tmp_dir = os.path.join(gs_tmp_dir, self.distinct_id) if os.path.exists(self.tmp_dir): shutil.rmtree(self.tmp_dir) os.makedirs(self.tmp_dir) self.vocal_url = vocal_url self.target_url = target_url ext = vocal_url.split(".")[-1] self.vocal_path = os.path.join(self.tmp_dir, self.distinct_id + f"_in.{ext}") self.target_wav_path = os.path.join(self.tmp_dir, self.distinct_id + "_out.wav") self.target_wav_ad_path = os.path.join(self.tmp_dir, self.distinct_id + "_out_ad.wav") self.target_path = os.path.join(self.tmp_dir, self.distinct_id + "_out.m4a") self.female_svc_source_url = female_recording_url self.male_svc_source_url = male_recording_url ext = female_recording_url.split(".")[-1] self.female_svc_source_path = os.path.join(gs_resource_cache_dir, hashlib.md5(female_recording_url.encode()).hexdigest() + "." + ext) ext = male_recording_url.split(".")[-1] self.male_svc_source_path = os.path.join(gs_resource_cache_dir, hashlib.md5(male_recording_url.encode()).hexdigest() + "." + ext) self.st_tm = start self.ed_tm = end self.target_loudness = vocal_loudness def log_info_name(self): return f"d_id={self.distinct_id}, vocal_url={self.vocal_url}" def rm_cache(self): if os.path.exists(self.tmp_dir): shutil.rmtree(self.tmp_dir) def init_gender_model(): """ 下载模型 :return: """ dst_model_dir = os.path.join(gs_model_dir, "voice_classification") if not os.path.exists(dst_model_dir): dst_zip_path = os.path.join(gs_model_dir, "models.zip") if not download2disk(gs_gender_models_url, dst_zip_path): svc_offline_logger.fatal(f"download gender_model err={gs_gender_models_url}") cmd = f"cd {gs_model_dir}; unzip {dst_zip_path}; mv models voice_classification; rm -f {dst_zip_path}" os.system(cmd) if not os.path.exists(dst_model_dir): svc_offline_logger.fatal(f"unzip {dst_zip_path} err") music_voice_pure_model = os.path.join(dst_model_dir, "voice_005_rec_v5.pth") music_voice_no_pure_model = os.path.join(dst_model_dir, "voice_10_v5.pth") gender_pure_model = os.path.join(dst_model_dir, "gender_8k_ratev5_v6_adam.pth") gender_no_pure_model = os.path.join(dst_model_dir, "gender_8k_v6_adam.pth") vc = VoiceClass(music_voice_pure_model, music_voice_no_pure_model, gender_pure_model, gender_no_pure_model) return vc def init_svc_model(): meisheng_env_prepare(logging, gs_model_dir) embed_model, hubert_model = load_model() return embed_model, hubert_model def download_volume_adjustment(): """ 下载音量调整工具 :return: """ volume_bin_path = os.path.join(gs_model_dir, "ebur128_tool") if not os.path.exists(volume_bin_path): if not download2disk(gs_volume_bin_url, volume_bin_path): svc_offline_logger.fatal(f"download volume_bin err={gs_volume_bin_url}") os.system(f"chmod +x {volume_bin_path}") def volume_adjustment(wav_path, target_loudness, out_path): """ 音量调整 :param wav_path: :param target_loudness: :param out_path: :return: """ volume_bin_path = os.path.join(gs_model_dir, "ebur128_tool") cmd = f"{volume_bin_path} {wav_path} {target_loudness} {out_path}" os.system(cmd) class SVCOnline: def __init__(self): st = time.time() self.gender_model = init_gender_model() self.embed_model, self.hubert_model = init_svc_model() download_volume_adjustment() download_volume_balanced() svc_offline_logger.info(f"svc init finished, sp = {time.time() - st}") def gender_process(self, worker_attr): st = time.time() gender, female_rate, is_pure = self.gender_model.process(worker_attr.vocal_path) svc_offline_logger.info( f"{worker_attr.vocal_url}, gender={gender}, female_rate={female_rate}, is_pure={is_pure}, " f"gender_process sp = {time.time() - st}") if gender == 0: gender = 'female' elif gender == 1: gender = 'male' elif female_rate > 0.5: gender = 'female' else: gender = 'male' svc_offline_logger.info(f"{worker_attr.vocal_url}, modified gender={gender}") - err = gs_err_code_success - if female_rate == -1: - err = gs_err_code_target_silence - return gender, err + # err = gs_err_code_success + # if female_rate == -1: + # err = gs_err_code_target_silence + return gender, gs_err_code_success def process(self, worker_attr): gender, err = self.gender_process(worker_attr) if err != gs_err_code_success: return gender, err song_path = worker_attr.female_svc_source_path if gender == "male": song_path = worker_attr.male_svc_source_path params = {'gender': gender, 'tst': worker_attr.st_tm, "tnd": worker_attr.ed_tm, 'delay': 0, 'song_path': None} st = time.time() err_code = process_svc_online(song_path, worker_attr.vocal_path, worker_attr.target_wav_path, self.embed_model, self.hubert_model, params) svc_offline_logger.info(f"{worker_attr.vocal_url}, err_code={err_code} process svc sp = {time.time() - st}") return gender, err_code