No OneTemporary
Actions

Size

19 KB

Subscribers

None

View Options

	diff --git a/AIMeiSheng/docker_demo/svc_online.py b/AIMeiSheng/docker_demo/svc_online.py
	index f952346..b196832 100644
	--- a/AIMeiSheng/docker_demo/svc_online.py
	+++ b/AIMeiSheng/docker_demo/svc_online.py
	@@ -1,162 +1,170 @@
	# -- coding: UTF-8 --
	"""
	SVC的核心处理逻辑
	"""
	import os
	import shutil
	import hashlib
	import time

	-from AIMeiSheng.meisheng_svc_final import get_svc, process_svc
	+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
	from AIMeiSheng.RawNet3.infererence_fang_meisheng import get_embed, get_embed_model
	from AIMeiSheng.myinfer_multi_spk_embed_in_dec_diff_fi_meisheng import svc_main, load_hubert, get_vc, get_rmvpe

	from AIMeiSheng.docker_demo.common import *

	gs_resource_cache_dir = "/tmp/gs_svc_resource_cache"
	gs_tmp_dir = "/tmp/gs_svc_tmp"
	gs_model_dir = "/tmp/models"

	if os.path.exists(gs_tmp_dir):
	shutil.rmtree(gs_tmp_dir)
	os.makedirs(gs_model_dir, exist_ok=True)

	# 预设参数
	gs_gender_models_url = "https://av-audit-sync-in-1256122840.cos.ap-mumbai.myqcloud.com/hub/voice_classification/models.zip"
	gs_svc_emb_url = ""
	gs_svc_model_url = ""
	gs_volume_bin_url = "https://av-audit-sync-in-1256122840.cos.ap-mumbai.myqcloud.com/dataset/AIMeiSheng/ebur128_tool"


	class GSWorkerAttr:
	def __init__(self, input_data):
	vocal_url = input_data["vocal_url"]
	female_svc_source_url = input_data["female_svc_url"]
	male_svc_source_url = input_data["male_svc_url"]
	st_tm = input_data["st_tm"] # 单位是s
	ed_tm = input_data["ed_tm"] # 单位是s

	- self.distinct_id = hashlib.md5(vocal_url.encode()).hexdigest()
	+ self.distinct_id = hashlib.md5(vocal_url.encode()).hexdigest()#对url进行哈希（通过哈希值建立索引，提高数据库的响应速度）
	self.vocal_url = vocal_url
	self.target_url = input_data["target_url"]

	ext = vocal_url.split(".")[-1]
	self.vocal_path = os.path.join(gs_tmp_dir, self.distinct_id + f"_in.{ext}")
	self.target_wav_path = os.path.join(gs_tmp_dir, self.distinct_id + "_out.wav")
	self.target_wav_ad_path = os.path.join(gs_tmp_dir, self.distinct_id + "_out_ad.wav")
	self.target_path = os.path.join(gs_tmp_dir, self.distinct_id + "_out.m4a")

	self.female_svc_source_url = female_svc_source_url
	self.male_svc_source_url = male_svc_source_url

	ext = female_svc_source_url.split(".")[-1]
	self.female_svc_source_path = hashlib.md5(female_svc_source_url.encode()).hexdigest() + "." + ext
	ext = male_svc_source_url.split(".")[-1]
	self.male_svc_source_path = hashlib.md5(male_svc_source_url.encode()).hexdigest() + "." + ext
	self.st_tm = st_tm
	self.ed_tm = ed_tm
	self.target_loudness = input_data["target_loudness"]

	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)

	def __del__(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):
	logging.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):
	logging.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():
	- emb_model_path = os.path.join(gs_model_dir, "RawNet3_weights.pt")
	- if not os.path.exists(emb_model_path):
	- if not download2disk(gs_svc_emb_url, emb_model_path):
	- logging.fatal(f"download svc_emb_model err={gs_svc_emb_url}")
	- embed_model = get_embed_model(emb_model_path)
	- hubert_model = load_hubert()
	+# def init_svc_model():
	+# emb_model_path = os.path.join(gs_model_dir, "RawNet3_weights.pt")
	+# if not os.path.exists(emb_model_path):
	+# if not download2disk(gs_svc_emb_url, emb_model_path):
	+# logging.fatal(f"download svc_emb_model err={gs_svc_emb_url}")
	+# embed_model = get_embed_model(emb_model_path)
	+# hubert_model = load_hubert()
	+#
	+# svc_filename = gs_svc_model_url.split("/")[-1]
	+# svc_model_path = os.path.join(gs_model_dir, svc_filename)
	+# if not os.path.exists(svc_model_path):
	+# if not download2disk(gs_svc_model_url, svc_model_path):
	+# logging.fatal(f"download svc_model err={gs_svc_model_url}")
	+#
	+# # 此处内部会生成全局模型
	+# get_vc(svc_model_path)
	+# return embed_model, hubert_model

	- svc_filename = gs_svc_model_url.split("/")[-1]
	- svc_model_path = os.path.join(gs_model_dir, svc_filename)
	- if not os.path.exists(svc_model_path):
	- if not download2disk(gs_svc_model_url, svc_model_path):
	- logging.fatal(f"download svc_model err={gs_svc_model_url}")

	- # 此处内部会生成全局模型
	- get_vc(svc_model_path)
	- return embed_model, hubert_model
	+def init_svc_model():
	+ meisheng_env_prepare(logging)
	+ embed_model, hubert_model = load_model()
	+ return embed_model, hubert_model


	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")
	if not os.path.exists(volume_bin_path):
	if not download2disk(gs_volume_bin_url, volume_bin_path):
	logging.fatal(f"download volume_bin err={gs_volume_bin_url}")
	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()
	logging.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)
	logging.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'
	logging.info(f"{worker_attr.vocal_url}, modified gender={gender}")
	return gender

	def process(self, worker_attr):
	gender = self.gender_process(worker_attr)
	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_ms, "tnd": worker_attr.ed_tm, 'delay': 0, 'song_path': None}
	st = time.time()
	- similar = process_svc(song_path, worker_attr.vocal_path, worker_attr.target_wav_path, params)
	+ similar = process_svc_online(song_path, worker_attr.vocal_path, worker_attr.target_wav_path, self.embed_model,
	+ self.hubert_model, params)
	logging.info(f"{worker_attr.vocal_url}, similar={similar} process svc sp = {time.time() - st}")
	diff --git a/AIMeiSheng/meisheng_env_preparex.py b/AIMeiSheng/meisheng_env_preparex.py
	new file mode 100644
	index 0000000..bf6aa2e
	--- /dev/null
	+++ b/AIMeiSheng/meisheng_env_preparex.py
	@@ -0,0 +1,38 @@
	+import os
	+from AIMeiSheng.docker_demo.common import *
	+
	+def meisheng_env_prepare(logging,AIMeiSheng_Path='./'):
	+
	+ cos_path = "https://av-audit-sync-sg-1256122840.cos.ap-singapore.myqcloud.com/dataset/AIMeiSheng/"
	+
	+
	+ rmvpe_model_path = os.path.join(AIMeiSheng_Path, 'rmvpe.pt')
	+ gs_rmvpe_model_url = cos_path + "rmvpe.pt"
	+ if not os.path.exists(rmvpe_model_path):
	+ if not download2disk(gs_rmvpe_model_url, rmvpe_model_path):
	+ logging.fatal(f"download rmvpe_model err={gs_rmvpe_model_url}")
	+
	+ hubert_model_path = os.path.join(AIMeiSheng_Path, 'hubert_base.pt')
	+ gs_hubert_model_url = cos_path + "hubert_base.pt"
	+ if not os.path.exists(hubert_model_path):
	+ if not download2disk(gs_hubert_model_url, hubert_model_path):
	+ logging.fatal(f"download hubert_model err={gs_hubert_model_url}")
	+
	+ model_svc = "xusong_v2_org_version_alldata_embed1_enzx_diff_fi_e15_s244110.pth"
	+ svc_model_path = os.path.join(AIMeiSheng_Path, f'weights/{model_svc}')
	+ gs_svc_model_url = cos_path + model_svc
	+ if not os.path.exists(svc_model_path):
	+ if not download2disk(gs_svc_model_url, svc_model_path):
	+ logging.fatal(f"download svc_model err={gs_svc_model_url}")
	+
	+
	+ model_embed = "model.pt"
	+ embed_model_path = os.path.join(AIMeiSheng_Path, f'RawNet3/models/weights/{model_embed}')
	+ gs_embed_model_url = cos_path + model_embed
	+ if not os.path.exists(embed_model_path):
	+ if not download2disk(gs_embed_model_url, embed_model_path):
	+ logging.fatal(f"download embed_model err={gs_embed_model_url}")
	+
	+
	+if __name__ == "__main__":
	+ meisheng_env_prepare()
	diff --git a/AIMeiSheng/meisheng_svc_final.py b/AIMeiSheng/meisheng_svc_final.py
	index e5a6b3f..1ecaaf7 100644
	--- a/AIMeiSheng/meisheng_svc_final.py
	+++ b/AIMeiSheng/meisheng_svc_final.py
	@@ -1,215 +1,227 @@


	import os,sys
	import time
	import shutil
	import glob
	import hashlib
	import librosa
	import soundfile
	import gradio as gr
	import pandas as pd
	import numpy as np
	sys.path.append('./RawNet3/')
	from infererence_fang_meisheng import get_embed, get_embed_model
	from myinfer_multi_spk_embed_in_dec_diff_fi_meisheng import svc_main,load_hubert, get_vc, get_rmvpe
	from gender_classify import load_gender_model

	+
	gs_simple_mixer_path = "/data/gpu_env_common/bin/simple_mixer" ##混音执行文件
	tmp_workspace_name = "batch_test_ocean_fi"#工作空间名
	song_folder = "./data_meisheng/" ##song folder
	gs_work_dir = f"./data_meisheng/{tmp_workspace_name}" #工作空间路径
	pth_model_path = "./weights/xusong_v2_org_version_alldata_embed1_enzx_diff_fi_e15_s244110.pth" ##模型文件


	cur_dir = os.path.abspath(os.path.dirname(__file__))
	abs_path = os.path.join(cur_dir,song_folder,tmp_workspace_name) + '/'

	f0_method = None

	def mix(in_path, acc_path, dst_path):
	# svc转码到442
	svc_442_file = in_path + "_442.wav"
	st = time.time()
	cmd = "ffmpeg -i {} -ar 44100 -ac 2 -y {} -loglevel fatal".format(in_path, svc_442_file)
	os.system(cmd)
	if not os.path.exists(svc_442_file):
	return -1
	print("transcode,{},sp={}".format(in_path, time.time() - st))

	# 混合
	st = time.time()
	cmd = "{} {} {} {} 1".format(gs_simple_mixer_path, svc_442_file, acc_path, dst_path)
	os.system(cmd)
	print("mixer,{},sp={}".format(in_path, time.time() - st))


	def load_model():
	global f0_method
	embed_model = get_embed_model()
	hubert_model = load_hubert()
	get_vc(pth_model_path)
	f0_method = get_rmvpe()
	print("model preload finish!!!")
	return embed_model, hubert_model#,svc_model

	-embed_model, hubert_model = load_model() ##提前加载模型
	-gender_model = load_gender_model()
	+def meisheng_init():
	+ embed_model, hubert_model = load_model() ##提前加载模型
	+ gender_model = load_gender_model()
	+ return embed_model, hubert_model, gender_model

	def pyin_process_single_rmvpe(input_file):
	global f0_method
	if f0_method is None:
	f0_method = get_rmvpe()

	rate = 16000 #44100
	# 读取音频文件
	y, sr = librosa.load(input_file, sr=rate)

	len_s = len(y)/sr
	lim_s = 15 #10
	if(len_s > lim_s):
	y1 = y[:sr*lim_s]
	y2 = y[-sr*lim_s:]
	f0 = f0_method.infer_from_audio(y1, thred=0.03)
	f0 = f0[f0 < 600]
	valid_f0 = f0[f0 > 50]
	mean_pitch1 = np.mean(valid_f0)
	f0 = f0_method.infer_from_audio(y2, thred=0.03)
	f0 = f0[f0 < 600]
	valid_f0 = f0[f0 > 50]
	mean_pitch2 = np.mean(valid_f0)

	if abs(mean_pitch1 - mean_pitch2) > 55:
	mean_pitch_cur = min(mean_pitch1, mean_pitch2)
	else:
	mean_pitch_cur = (mean_pitch1 + mean_pitch2) / 2

	else:
	f0 = f0_method.infer_from_audio(y, thred=0.03)
	f0 = f0[f0 < 600]
	valid_f0 = f0[f0 > 50]
	mean_pitch_cur = np.mean(valid_f0)


	return mean_pitch_cur

	-def meisheng_svc(song_wav, target_wav, svc_out_path, embed_npy, paras):

	+def meisheng_svc(song_wav, target_wav, svc_out_path, embed_npy, embed_md, hubert_md, paras):
	##计算pitch
	f0up_key = pyin_process_single_rmvpe(target_wav)
	## get embed
	- get_embed(target_wav, embed_npy, embed_model)
	-
	- print("svc main start...")
	- svc_main(song_wav,svc_out_path,pth_model_path,embed_npy,f0up_key,hubert_model,paras)
	+ get_embed(target_wav, embed_npy, embed_md)
	+
	+ print("svc main start...")
	+ svc_main(song_wav, svc_out_path, pth_model_path, embed_npy, f0up_key, hubert_md, paras)
	print("svc main finished!!")

	return 0
	-def process_svc(song_wav, target_wav, svc_out_path,paras):
	+
	+def process_svc_online(song_wav, target_wav, svc_out_path, embed_md, hubert_md, paras):
	+
	+ embed_npy = target_wav[:-4] + '.npy' ##embd npy存储位置
	+ similar = meisheng_svc(song_wav, target_wav, svc_out_path, embed_npy, embed_md, hubert_md, paras)
	+
	+ return similar
	+
	+def process_svc(song_wav, target_wav, svc_out_path, embed_md, hubert_md, paras):

	song_wav1, target_wav, svc_out_path = os.path.basename(song_wav), os.path.basename(
	target_wav), os.path.basename(svc_out_path) #绝对路径
	song_wav, target_wav, svc_out_path = song_wav, abs_path + target_wav, abs_path + svc_out_path
	embed_npy = target_wav[:-4] + '.npy' ##embd npy存储位置

	- similar = meisheng_svc(song_wav,target_wav,svc_out_path,embed_npy,paras)
	-
	+ # similar = meisheng_svc(song_wav,target_wav,svc_out_path,embed_npy,paras)
	+ similar = meisheng_svc(song_wav, target_wav, svc_out_path, embed_npy, embed_md, hubert_md, paras)

	return similar


	-def get_svc(target_yinse_wav, song_name, paras):
	+def get_svc(target_yinse_wav, song_name, embed_model, hubert_model, paras):
	'''
	:param target_yinse_wav: 目标音色
	:param song_name: 歌曲名字
	;param paras: 其他参数
	:return: svc路径名
	'''

	##清空工作空间临时路径
	if os.path.exists(gs_work_dir):
	#shutil.rmtree(gs_work_dir)
	cmd = f"rm -rf {gs_work_dir}/*"
	os.system(cmd)
	else:
	os.makedirs(gs_work_dir)

	gender = paras['gender']##为了确定歌曲

	##目标音色读取
	f_dst = os.path.join(gs_work_dir, os.path.basename(target_yinse_wav))
	#print("dir :", f_dst,"target_yinse_wav:",target_yinse_wav)
	#shutil.move(target_yinse_wav, f_dst) ##放在工作目录
	shutil.copy(target_yinse_wav, f_dst)
	target_yinse_wav = f_dst

	##歌曲/伴奏读取（路径需要修改）
	song_wav = os.path.join("{}{}/{}/vocal321.wav".format(song_folder, gender, song_name)) # 歌曲vocal
	inf_acc_path = os.path.join("{}{}/{}/acc.wav".format(song_folder, gender, song_name))
	#song_wav = './xusong_long.wav'
	svc_out_path = os.path.join(gs_work_dir, "svc.wav") ###svc结果名字
	print("inputMsg:", song_wav, target_yinse_wav, svc_out_path)

	## svc process
	st = time.time()
	print("start inference...")
	- similar = process_svc(song_wav, target_yinse_wav, svc_out_path,paras)
	+ similar = process_svc(song_wav, target_yinse_wav, svc_out_path, embed_model, hubert_model,paras)
	print("svc finished!!")
	print("time cost = {}".format(time.time() - st))
	print("out path name {} ".format(svc_out_path))

	#'''
	##加混响
	print("add reverbration...")
	svc_out_path_effect = svc_out_path[:-4] + '_effect.wav'
	cmd = f"/data/gpu_env_common/bin/effect_tool {svc_out_path} {svc_out_path_effect}"
	print("cmd :", cmd)
	os.system(cmd)
	# # 人声伴奏合并
	print("add acc...")
	out_path = svc_out_path_effect[:-4] + '_music.wav'
	mix(svc_out_path_effect, inf_acc_path, out_path)

	print("time cost = {}".format(time.time() - st))
	print("out path name {} ".format(out_path))
	#'''


	return svc_out_path


	-if __name__=='__main__':
	+def meisheng_func(target_yinse_wav,song_name, paras):
	+
	+ ##init
	+ embed_model, hubert_model, gender_model = meisheng_init()

	###gender predict
	- target_yinse_wav = "./raw/meisheng_yinse/female/target_yinse_cloris.m4a"
	gender, female_rate, is_pure = gender_model.process(target_yinse_wav)
	print('=====================')
	- print("gender:{}, female_rate:{},is_pure:{}".format(gender,female_rate,is_pure))
	+ print("gender:{}, female_rate:{},is_pure:{}".format(gender, female_rate, is_pure))
	if gender == 0:
	gender = 'female'
	elif gender == 1:
	gender = 'male'
	elif female_rate > 0.5:
	gender = 'female'
	else:
	gender = 'male'
	print("modified gender:{} ".format(gender))
	print('=====================')

	- ###接口函数
	- '''
	- target_yinse_wav = "./raw/meisheng_yinse/female/changying.wav" #需要完整路径
	- song_name = "drivers_license" #"Levitating" ##路径会自动添加(要更改)
	- paras = {'gender': 'female', 'tst': 0, "tnd": None, 'delay': 0, 'song_path': None} ##单位都是ms
	- #paras = {'gender': 'female', 'tst': 0, "tnd": 30, 'delay': 0} ###片段svc测试
	- #'''
	+ ##美声main
	+ paras['gender'] = gender ##单位都是ms
	+ get_svc(target_yinse_wav, song_name, embed_model, hubert_model, paras)
	+
	+
	+if __name__=='__main__':
	+
	+ #target_yinse_wav = "./raw/meisheng_yinse/female/changying.wav" # 需要完整路径
	+ target_yinse_wav = "./raw/meisheng_yinse/female/target_yinse_cloris.m4a"
	+ song_name = "lost_stars" ##歌曲名字
	+ paras = {'gender': None, 'tst': 0, "tnd": None, 'delay': 0, 'song_path': None}
	+ # paras = {'gender': 'female', 'tst': 0, "tnd": 30, 'delay': 0} ###片段svc测试
	+ meisheng_func(target_yinse_wav, song_name, paras)

	- #'''
	- #target_yinse_wav = "./raw/meisheng_yinse/female/target_yinse_cloris.m4a"
	- song_name = "lost_stars"
	- #paras = {'gender': 'female', 'tst': 0, "tnd": None, 'delay': 0, 'song_path': None}
	- paras = {'gender': gender, 'tst': 0, "tnd": None, 'delay': 0, 'song_path': None }
	- get_svc(target_yinse_wav, song_name, paras)
	- #'''

File Metadata

Mime Type: text/x-diff
Expires: Sun, Jan 12, 08:33 (1 d, 15 h)
Storage Engine: blob
Storage Format: Raw Data
Storage Handle: 1347138
Default Alt Text: (19 KB)

No OneTemporaryActions

View Options

File Metadata

Event Timeline

No OneTemporary
Actions