From f1730d42d4b0a1b866d6285cdc629acde4b33f94 Mon Sep 17 00:00:00 2001 From: RVC-Boss <129054828+RVC-Boss@users.noreply.github.com> Date: Sun, 28 May 2023 22:58:33 +0800 Subject: [PATCH] Add files via upload --- MDXNet.py | 198 +++++++++++++++++++++++++++++++++++++++++++ infer-web.py | 99 ++++++++++++++++------ infer_uvr5.py | 172 +++++++++++++++++++++++++++++++------ vc_infer_pipeline.py | 42 ++++++++- 4 files changed, 455 insertions(+), 56 deletions(-) create mode 100644 MDXNet.py diff --git a/MDXNet.py b/MDXNet.py new file mode 100644 index 0000000..02b37f7 --- /dev/null +++ b/MDXNet.py @@ -0,0 +1,198 @@ +import soundfile as sf +import torch,pdb,time,argparse,os,warnings,sys,librosa +import numpy as np +import onnxruntime as ort +from scipy.io.wavfile import write +from tqdm import tqdm +import torch +import torch.nn as nn + +dim_c = 4 +class Conv_TDF_net_trim(): + def __init__(self, device, model_name, target_name, + L, dim_f, dim_t, n_fft, hop=1024): + super(Conv_TDF_net_trim, self).__init__() + + self.dim_f = dim_f + self.dim_t = 2 ** dim_t + self.n_fft = n_fft + self.hop = hop + self.n_bins = self.n_fft // 2 + 1 + self.chunk_size = hop * (self.dim_t - 1) + self.window = torch.hann_window(window_length=self.n_fft, periodic=True).to(device) + self.target_name = target_name + self.blender = 'blender' in model_name + + out_c = dim_c * 4 if target_name == '*' else dim_c + self.freq_pad = torch.zeros([1, out_c, self.n_bins - self.dim_f, self.dim_t]).to(device) + + self.n = L // 2 + + def stft(self, x): + x = x.reshape([-1, self.chunk_size]) + x = torch.stft(x, n_fft=self.n_fft, hop_length=self.hop, window=self.window, center=True, return_complex=True) + x = torch.view_as_real(x) + x = x.permute([0, 3, 1, 2]) + x = x.reshape([-1, 2, 2, self.n_bins, self.dim_t]).reshape([-1, dim_c, self.n_bins, self.dim_t]) + return x[:, :, :self.dim_f] + + def istft(self, x, freq_pad=None): + freq_pad = self.freq_pad.repeat([x.shape[0], 1, 1, 1]) if freq_pad is None else freq_pad + x = torch.cat([x, freq_pad], -2) + c = 4 * 2 if self.target_name == '*' else 2 + x = x.reshape([-1, c, 2, self.n_bins, self.dim_t]).reshape([-1, 2, self.n_bins, self.dim_t]) + x = x.permute([0, 2, 3, 1]) + x = x.contiguous() + x = torch.view_as_complex(x) + x = torch.istft(x, n_fft=self.n_fft, hop_length=self.hop, window=self.window, center=True) + return x.reshape([-1, c, self.chunk_size]) +def get_models(device, dim_f, dim_t, n_fft): + return Conv_TDF_net_trim( + device=device, + model_name='Conv-TDF', target_name='vocals', + L=11, + dim_f=dim_f, dim_t=dim_t, + n_fft=n_fft + ) + +warnings.filterwarnings("ignore") +cpu = torch.device('cpu') +device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') + +class Predictor: + def __init__(self,args): + self.args=args + self.model_ = get_models(device=cpu, dim_f=args.dim_f, dim_t=args.dim_t, n_fft=args.n_fft) + self.model = ort.InferenceSession(os.path.join(args.onnx,self.model_.target_name+'.onnx'), providers=['CUDAExecutionProvider', 'CPUExecutionProvider']) + print('onnx load done') + def demix(self, mix): + samples = mix.shape[-1] + margin = self.args.margin + chunk_size = self.args.chunks*44100 + assert not margin == 0, 'margin cannot be zero!' + if margin > chunk_size: + margin = chunk_size + + segmented_mix = {} + + if self.args.chunks == 0 or samples < chunk_size: + chunk_size = samples + + counter = -1 + for skip in range(0, samples, chunk_size): + counter+=1 + + s_margin = 0 if counter == 0 else margin + end = min(skip+chunk_size+margin, samples) + + start = skip-s_margin + + segmented_mix[skip] = mix[:,start:end].copy() + if end == samples: + break + + sources = self.demix_base(segmented_mix, margin_size=margin) + ''' + mix:(2,big_sample) + segmented_mix:offset->(2,small_sample) + sources:(1,2,big_sample) + ''' + return sources + def demix_base(self, mixes, margin_size): + chunked_sources = [] + progress_bar = tqdm(total=len(mixes)) + progress_bar.set_description("Processing") + for mix in mixes: + cmix = mixes[mix] + sources = [] + n_sample = cmix.shape[1] + model=self.model_ + trim = model.n_fft//2 + gen_size = model.chunk_size-2*trim + pad = gen_size - n_sample%gen_size + mix_p = np.concatenate((np.zeros((2,trim)), cmix, np.zeros((2,pad)), np.zeros((2,trim))), 1) + mix_waves = [] + i = 0 + while i < n_sample + pad: + waves = np.array(mix_p[:, i:i+model.chunk_size]) + mix_waves.append(waves) + i += gen_size + mix_waves = torch.tensor(mix_waves, dtype=torch.float32).to(cpu) + with torch.no_grad(): + _ort = self.model + spek = model.stft(mix_waves) + if self.args.denoise: + spec_pred = -_ort.run(None, {'input': -spek.cpu().numpy()})[0]*0.5+_ort.run(None, {'input': spek.cpu().numpy()})[0]*0.5 + tar_waves = model.istft(torch.tensor(spec_pred)) + else: + tar_waves = model.istft(torch.tensor(_ort.run(None, {'input': spek.cpu().numpy()})[0])) + tar_signal = tar_waves[:,:,trim:-trim].transpose(0,1).reshape(2, -1).numpy()[:, :-pad] + + start = 0 if mix == 0 else margin_size + end = None if mix == list(mixes.keys())[::-1][0] else -margin_size + if margin_size == 0: + end = None + sources.append(tar_signal[:,start:end]) + + progress_bar.update(1) + + chunked_sources.append(sources) + _sources = np.concatenate(chunked_sources, axis=-1) + # del self.model + progress_bar.close() + return _sources + def prediction(self, m,vocal_root,others_root): + os.makedirs(vocal_root,exist_ok=True) + os.makedirs(others_root,exist_ok=True) + basename = os.path.basename(m) + mix, rate = librosa.load(m, mono=False, sr=44100) + if mix.ndim == 1: + mix = np.asfortranarray([mix,mix]) + mix = mix.T + sources = self.demix(mix.T) + opt=sources[0].T + sf.write("%s/%s_main_vocal.wav"%(vocal_root,basename), mix-opt, rate) + sf.write("%s/%s_others.wav"%(others_root,basename), opt , rate) + +class MDXNetDereverb(): + def __init__(self,chunks): + self.onnx="uvr5_weights/onnx_dereverb_By_FoxJoy" + self.shifts=10#'Predict with randomised equivariant stabilisation' + self.mixing="min_mag"#['default','min_mag','max_mag'] + self.chunks=chunks + self.margin=44100 + self.dim_t=9 + self.dim_f=3072 + self.n_fft=6144 + self.denoise=True + self.pred=Predictor(self) + + def _path_audio_(self,input,vocal_root,others_root): + self.pred.prediction(input,vocal_root,others_root) + +if __name__ == '__main__': + dereverb=MDXNetDereverb(15) + from time import time as ttime + t0=ttime() + dereverb._path_audio_( + "雪雪伴奏对消HP5.wav", + "vocal", + "others", + ) + t1=ttime() + print(t1-t0) + + +''' + +runtime\python.exe MDXNet.py + +6G: +15/9:0.8G->6.8G +14:0.8G->6.5G +25:炸 + +half15:0.7G->6.6G,22.69s +fp32-15:0.7G->6.6G,20.85s + +''' \ No newline at end of file diff --git a/infer-web.py b/infer-web.py index 2d5739b..8596da6 100644 --- a/infer-web.py +++ b/infer-web.py @@ -21,6 +21,7 @@ warnings.filterwarnings("ignore") torch.manual_seed(114514) from i18n import I18nAuto import ffmpeg +from MDXNet import MDXNetDereverb i18n = I18nAuto() i18n.print() @@ -82,7 +83,7 @@ import gradio as gr import logging from vc_infer_pipeline import VC from config import Config -from infer_uvr5 import _audio_pre_ +from infer_uvr5 import _audio_pre_,_audio_pre_new from my_utils import load_audio from train.process_ckpt import show_info, change_info, merge, extract_small_model @@ -133,7 +134,7 @@ for root, dirs, files in os.walk(index_root, topdown=False): index_paths.append("%s/%s" % (root, name)) uvr5_names = [] for name in os.listdir(weight_uvr5_root): - if name.endswith(".pth"): + if name.endswith(".pth")or "onnx"in name: uvr5_names.append(name.replace(".pth", "")) @@ -150,6 +151,7 @@ def vc_single( filter_radius, resample_sr, rms_mix_rate, + protect ): # spk_item, input_audio0, vc_transform0,f0_file,f0method0 global tgt_sr, net_g, vc, hubert_model, version if input_audio_path is None: @@ -197,6 +199,7 @@ def vc_single( resample_sr, rms_mix_rate, version, + protect, f0_file=f0_file, ) if resample_sr >= 16000 and tgt_sr != resample_sr: @@ -232,6 +235,7 @@ def vc_multi( filter_radius, resample_sr, rms_mix_rate, + protect ): try: dir_path = ( @@ -262,6 +266,7 @@ def vc_multi( filter_radius, resample_sr, rms_mix_rate, + protect ) if "Success" in info: try: @@ -288,12 +293,16 @@ def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg): save_root_ins = ( save_root_ins.strip(" ").strip('"').strip("\n").strip('"').strip(" ") ) - pre_fun = _audio_pre_( - agg=int(agg), - model_path=os.path.join(weight_uvr5_root, model_name + ".pth"), - device=config.device, - is_half=config.is_half, - ) + if(model_name=="onnx_dereverb_By_FoxJoy"): + pre_fun=MDXNetDereverb(15) + else: + func=_audio_pre_ if "DeEcho"not in model_name else _audio_pre_new + pre_fun = func( + agg=int(agg), + model_path=os.path.join(weight_uvr5_root, model_name + ".pth"), + device=config.device, + is_half=config.is_half, + ) if inp_root != "": paths = [os.path.join(inp_root, name) for name in os.listdir(inp_root)] else: @@ -336,8 +345,12 @@ def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg): yield "\n".join(infos) finally: try: - del pre_fun.model - del pre_fun + if (model_name == "onnx_dereverb_By_FoxJoy"): + del pre_fun.pred.model + del pre_fun.pred.model_ + else: + del pre_fun.model + del pre_fun except: traceback.print_exc() print("clean_empty_cache") @@ -790,7 +803,7 @@ def train_index(exp_dir1, version19): faiss.write_index( index, "%s/trained_IVF%s_Flat_nprobe_%s_%s_%s.index" - % (exp_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19), + % (exp_dir, n_ivf, index_ivf.nprobe,exp_dir1, version19), ) # faiss.write_index(index, '%s/trained_IVF%s_Flat_FastScan_%s.index'%(exp_dir,n_ivf,version19)) infos.append("adding") @@ -801,11 +814,11 @@ def train_index(exp_dir1, version19): faiss.write_index( index, "%s/added_IVF%s_Flat_nprobe_%s_%s_%s.index" - % (exp_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19), + % (exp_dir, n_ivf, index_ivf.nprobe,exp_dir1, version19), ) infos.append( "成功构建索引，added_IVF%s_Flat_nprobe_%s_%s_%s.index" - % (n_ivf, index_ivf.nprobe, exp_dir1, version19) + % (n_ivf, index_ivf.nprobe,exp_dir1, version19) ) # faiss.write_index(index, '%s/added_IVF%s_Flat_FastScan_%s.index'%(exp_dir,n_ivf,version19)) # infos.append("成功构建索引，added_IVF%s_Flat_FastScan_%s.index"%(n_ivf,version19)) @@ -1030,7 +1043,7 @@ def train1key( faiss.write_index( index, "%s/trained_IVF%s_Flat_nprobe_%s_%s_%s.index" - % (model_log_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19), + % (model_log_dir, n_ivf, index_ivf.nprobe,exp_dir1, version19), ) yield get_info_str("adding index") batch_size_add = 8192 @@ -1039,11 +1052,11 @@ def train1key( faiss.write_index( index, "%s/added_IVF%s_Flat_nprobe_%s_%s_%s.index" - % (model_log_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19), + % (model_log_dir, n_ivf, index_ivf.nprobe,exp_dir1, version19), ) yield get_info_str( "成功构建索引, added_IVF%s_Flat_nprobe_%s_%s_%s.index" - % (n_ivf, index_ivf.nprobe, exp_dir1, version19) + % (n_ivf, index_ivf.nprobe, exp_dir1,version19) ) yield get_info_str(i18n("全流程结束！")) @@ -1161,8 +1174,8 @@ with gr.Blocks() as app: value="E:\\codes\\py39\\test-20230416b\\todo-songs\\冬之花clip1.wav", ) f0method0 = gr.Radio( - label=i18n("选择音高提取算法,输入歌声可用pm提速,harvest低音好但巨慢无比"), - choices=["pm", "harvest"], + label=i18n("选择音高提取算法,输入歌声可用pm提速,harvest低音好但巨慢无比,crepe效果好但吃GPU"), + choices=["pm", "harvest","crepe"], value="pm", interactive=True, ) @@ -1197,9 +1210,10 @@ with gr.Blocks() as app: minimum=0, maximum=1, label=i18n("检索特征占比"), - value=0.76, + value=0.88, interactive=True, ) + with gr.Column(): resample_sr0 = gr.Slider( minimum=0, maximum=48000, @@ -1215,9 +1229,17 @@ with gr.Blocks() as app: value=1, interactive=True, ) + protect0 = gr.Slider( + minimum=0, + maximum=0.5, + label=i18n("保护清辅音和呼吸声，防止电音撕裂等artifact，拉满0.5不开启，调低加大保护力度但可能降低索引效果"), + value=0.33, + step=0.01, + interactive=True, + ) f0_file = gr.File(label=i18n("F0曲线文件, 可选, 一行一个音高, 代替默认F0及升降调")) but0 = gr.Button(i18n("转换"), variant="primary") - with gr.Column(): + with gr.Row(): vc_output1 = gr.Textbox(label=i18n("输出信息")) vc_output2 = gr.Audio(label=i18n("输出音频(右下角三个点,点了可以下载)")) but0.click( @@ -1235,6 +1257,7 @@ with gr.Blocks() as app: filter_radius0, resample_sr0, rms_mix_rate0, + protect0 ], [vc_output1, vc_output2], ) @@ -1249,8 +1272,8 @@ with gr.Blocks() as app: ) opt_input = gr.Textbox(label=i18n("指定输出文件夹"), value="opt") f0method1 = gr.Radio( - label=i18n("选择音高提取算法,输入歌声可用pm提速,harvest低音好但巨慢无比"), - choices=["pm", "harvest"], + label=i18n("选择音高提取算法,输入歌声可用pm提速,harvest低音好但巨慢无比,crepe效果好但吃GPU"), + choices=["pm", "harvest","crepe"], value="pm", interactive=True, ) @@ -1285,6 +1308,7 @@ with gr.Blocks() as app: value=1, interactive=True, ) + with gr.Column(): resample_sr1 = gr.Slider( minimum=0, maximum=48000, @@ -1300,6 +1324,14 @@ with gr.Blocks() as app: value=1, interactive=True, ) + protect1 = gr.Slider( + minimum=0, + maximum=0.5, + label=i18n("保护清辅音和呼吸声，防止电音撕裂等artifact，拉满0.5不开启，调低加大保护力度但可能降低索引效果"), + value=0.33, + step=0.01, + interactive=True, + ) with gr.Column(): dir_input = gr.Textbox( label=i18n("输入待处理音频文件夹路径(去文件管理器地址栏拷就行了)"), @@ -1308,8 +1340,9 @@ with gr.Blocks() as app: inputs = gr.File( file_count="multiple", label=i18n("也可批量输入音频文件, 二选一, 优先读文件夹") ) - but1 = gr.Button(i18n("转换"), variant="primary") - vc_output3 = gr.Textbox(label=i18n("输出信息")) + with gr.Row(): + but1 = gr.Button(i18n("转换"), variant="primary") + vc_output3 = gr.Textbox(label=i18n("输出信息")) but1.click( vc_multi, [ @@ -1326,14 +1359,26 @@ with gr.Blocks() as app: filter_radius1, resample_sr1, rms_mix_rate1, + protect1 ], [vc_output3], ) - with gr.TabItem(i18n("伴奏人声分离")): + with gr.TabItem(i18n("伴奏人声分离&去混响&去回声")): with gr.Group(): gr.Markdown( value=i18n( - "人声伴奏分离批量处理, 使用UVR5模型.
不带和声用HP2, 带和声且提取的人声不需要和声用HP5
合格的文件夹路径格式举例: E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华测试样例(去文件管理器地址栏拷就行了)" + "人声伴奏分离批量处理， 使用UVR5模型。
" + "合格的文件夹路径格式举例： E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华测试样例(去文件管理器地址栏拷就行了)。
" + "模型分为三类：
" + "1、保留人声：不带和声的音频选这个，对主人声保留比HP5更好。内置HP2和HP3两个模型，HP3可能轻微漏伴奏但对主人声保留比HP2稍微好一丁点；
" + "2、仅保留主人声：带和声的音频选这个，对主人声可能有削弱。内置HP5一个模型；
" + "3、去混响、去延迟模型（by FoxJoy）：
" + "  (1)MDX-Net:对于双通道混响是最好的选择，不能去除单通道混响；
" + " (234)DeEcho:去除延迟效果。Aggressive比Normal去除得更彻底，DeReverb额外去除混响，可去除单声道混响，但是对高频重的板式混响去不干净。
" + "去混响/去延迟，附：
" + "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；
" + "2、MDX-Net-Dereverb模型挺慢的；
" + "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。" ) ) with gr.Row(): @@ -1384,7 +1429,7 @@ with gr.Blocks() as app: exp_dir1 = gr.Textbox(label=i18n("输入实验名"), value="mi-test") sr2 = gr.Radio( label=i18n("目标采样率"), - choices=["32k", "40k", "48k"], + choices=["40k", "48k"], value="40k", interactive=True, ) diff --git a/infer_uvr5.py b/infer_uvr5.py index 4aada2d..4948d17 100644 --- a/infer_uvr5.py +++ b/infer_uvr5.py @@ -1,5 +1,7 @@ import os, sys, torch, warnings, pdb - +now_dir = os.getcwd() +sys.path.append(now_dir) +from json import load as ll warnings.filterwarnings("ignore") import librosa import importlib @@ -10,7 +12,8 @@ from uvr5_pack.lib_v5 import spec_utils from uvr5_pack.utils import _get_name_params, inference from uvr5_pack.lib_v5.model_param_init import ModelParameters from scipy.io import wavfile - +from uvr5_pack.lib_v5.nets_new import CascadedNet +from uvr5_pack.lib_v5 import nets_61968KB as nets class _audio_pre_: def __init__(self, agg, model_path, device, is_half): @@ -25,28 +28,7 @@ class _audio_pre_: "agg": agg, "high_end_process": "mirroring", } - nn_arch_sizes = [ - 31191, # default - 33966, - 61968, - 123821, - 123812, - 537238, # custom - ] - self.nn_architecture = list("{}KB".format(s) for s in nn_arch_sizes) - model_size = math.ceil(os.stat(model_path).st_size / 1024) - nn_architecture = "{}KB".format( - min(nn_arch_sizes, key=lambda x: abs(x - model_size)) - ) - nets = importlib.import_module( - "uvr5_pack.lib_v5.nets" - + f"_{nn_architecture}".replace("_{}KB".format(nn_arch_sizes[0]), ""), - package=None, - ) - model_hash = hashlib.md5(open(model_path, "rb").read()).hexdigest() - param_name, model_params_d = _get_name_params(model_path, model_hash) - - mp = ModelParameters(model_params_d) + mp = ModelParameters("uvr5_pack/lib_v5/modelparams/4band_v2.json") model = nets.CascadedASPPNet(mp.param["bins"] * 2) cpk = torch.load(model_path, map_location="cpu") model.load_state_dict(cpk) @@ -164,12 +146,148 @@ class _audio_pre_: (np.array(wav_vocals) * 32768).astype("int16"), ) +class _audio_pre_new: + def __init__(self, agg, model_path, device, is_half): + self.model_path = model_path + self.device = device + self.data = { + # Processing Options + "postprocess": False, + "tta": False, + # Constants + "window_size": 512, + "agg": agg, + "high_end_process": "mirroring", + } + mp=ModelParameters("uvr5_pack/lib_v5/modelparams/4band_v3.json") + nout=64 if "DeReverb"in model_path else 48 + model = CascadedNet(mp.param["bins"] * 2,nout) + cpk = torch.load(model_path, map_location="cpu") + model.load_state_dict(cpk) + model.eval() + if is_half: + model = model.half().to(device) + else: + model = model.to(device) + + self.mp = mp + self.model = model + + def _path_audio_(self, music_file, vocal_root=None, ins_root=None):#3个VR模型vocal和ins是反的 + if ins_root is None and vocal_root is None: + return "No save root." + name = os.path.basename(music_file) + if ins_root is not None: + os.makedirs(ins_root, exist_ok=True) + if vocal_root is not None: + os.makedirs(vocal_root, exist_ok=True) + X_wave, y_wave, X_spec_s, y_spec_s = {}, {}, {}, {} + bands_n = len(self.mp.param["band"]) + # print(bands_n) + for d in range(bands_n, 0, -1): + bp = self.mp.param["band"][d] + if d == bands_n: # high-end band + ( + X_wave[d], + _, + ) = librosa.core.load( # 理论上librosa读取可能对某些音频有bug，应该上ffmpeg读取，但是太麻烦了弃坑 + music_file, + bp["sr"], + False, + dtype=np.float32, + res_type=bp["res_type"], + ) + if X_wave[d].ndim == 1: + X_wave[d] = np.asfortranarray([X_wave[d], X_wave[d]]) + else: # lower bands + X_wave[d] = librosa.core.resample( + X_wave[d + 1], + self.mp.param["band"][d + 1]["sr"], + bp["sr"], + res_type=bp["res_type"], + ) + # Stft of wave source + X_spec_s[d] = spec_utils.wave_to_spectrogram_mt( + X_wave[d], + bp["hl"], + bp["n_fft"], + self.mp.param["mid_side"], + self.mp.param["mid_side_b2"], + self.mp.param["reverse"], + ) + # pdb.set_trace() + if d == bands_n and self.data["high_end_process"] != "none": + input_high_end_h = (bp["n_fft"] // 2 - bp["crop_stop"]) + ( + self.mp.param["pre_filter_stop"] - self.mp.param["pre_filter_start"] + ) + input_high_end = X_spec_s[d][ + :, bp["n_fft"] // 2 - input_high_end_h : bp["n_fft"] // 2, : + ] + + X_spec_m = spec_utils.combine_spectrograms(X_spec_s, self.mp) + aggresive_set = float(self.data["agg"] / 100) + aggressiveness = { + "value": aggresive_set, + "split_bin": self.mp.param["band"][1]["crop_stop"], + } + with torch.no_grad(): + pred, X_mag, X_phase = inference( + X_spec_m, self.device, self.model, aggressiveness, self.data + ) + # Postprocess + if self.data["postprocess"]: + pred_inv = np.clip(X_mag - pred, 0, np.inf) + pred = spec_utils.mask_silence(pred, pred_inv) + y_spec_m = pred * X_phase + v_spec_m = X_spec_m - y_spec_m + + if ins_root is not None: + if self.data["high_end_process"].startswith("mirroring"): + input_high_end_ = spec_utils.mirroring( + self.data["high_end_process"], y_spec_m, input_high_end, self.mp + ) + wav_instrument = spec_utils.cmb_spectrogram_to_wave( + y_spec_m, self.mp, input_high_end_h, input_high_end_ + ) + else: + wav_instrument = spec_utils.cmb_spectrogram_to_wave(y_spec_m, self.mp) + print("%s instruments done" % name) + wavfile.write( + os.path.join( + ins_root, "main_vocal_{}_{}.wav".format(name, self.data["agg"]) + ), + self.mp.param["sr"], + (np.array(wav_instrument) * 32768).astype("int16"), + ) # + if vocal_root is not None: + if self.data["high_end_process"].startswith("mirroring"): + input_high_end_ = spec_utils.mirroring( + self.data["high_end_process"], v_spec_m, input_high_end, self.mp + ) + wav_vocals = spec_utils.cmb_spectrogram_to_wave( + v_spec_m, self.mp, input_high_end_h, input_high_end_ + ) + else: + wav_vocals = spec_utils.cmb_spectrogram_to_wave(v_spec_m, self.mp) + print("%s vocals done" % name) + wavfile.write( + os.path.join( + vocal_root, "others_{}_{}.wav".format(name, self.data["agg"]) + ), + self.mp.param["sr"], + (np.array(wav_vocals) * 32768).astype("int16"), + ) + if __name__ == "__main__": device = "cuda" is_half = True - model_path = "uvr5_weights/2_HP-UVR.pth" - pre_fun = _audio_pre_(model_path=model_path, device=device, is_half=True) - audio_path = "神女劈观.aac" + # model_path = "uvr5_weights/2_HP-UVR.pth" + # model_path = "uvr5_weights/VR-DeEchoDeReverb.pth" + # model_path = "uvr5_weights/VR-DeEchoNormal.pth" + model_path = "uvr5_weights/DeEchoNormal.pth" + # pre_fun = _audio_pre_(model_path=model_path, device=device, is_half=True,agg=10) + pre_fun = _audio_pre_new(model_path=model_path, device=device, is_half=True,agg=10) + audio_path = "雪雪伴奏对消HP5.wav" save_path = "opt" pre_fun._path_audio_(audio_path, save_path, save_path) diff --git a/vc_infer_pipeline.py b/vc_infer_pipeline.py index a8f0540..f45d4c8 100644 --- a/vc_infer_pipeline.py +++ b/vc_infer_pipeline.py @@ -2,7 +2,7 @@ import numpy as np, parselmouth, torch, pdb from time import time as ttime import torch.nn.functional as F import scipy.signal as signal -import pyworld, os, traceback, faiss, librosa +import pyworld, os, traceback, faiss, librosa,torchcrepe from scipy import signal from functools import lru_cache @@ -103,6 +103,27 @@ class VC(object): f0 = cache_harvest_f0(input_audio_path, self.sr, f0_max, f0_min, 10) if filter_radius > 2: f0 = signal.medfilt(f0, 3) + elif f0_method == "crepe": + model = "full" + # Pick a batch size that doesn't cause memory errors on your gpu + batch_size = 512 + # Compute pitch using first gpu + audio = torch.tensor(np.copy(x))[None].float() + f0, pd = torchcrepe.predict( + audio, + self.sr, + self.window, + f0_min, + f0_max, + model, + batch_size=batch_size, + device=self.device, + return_periodicity=True, + ) + pd = torchcrepe.filter.median(pd, 3) + f0 = torchcrepe.filter.mean(f0, 3) + f0[pd < 0.1] = 0 + f0 = f0[0].cpu().numpy() f0 *= pow(2, f0_up_key / 12) # with open("test.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()])) tf0 = self.sr // self.window # 每秒f0点数 @@ -141,6 +162,7 @@ class VC(object): big_npy, index_rate, version, + protect ): # ,file_index,file_big_npy feats = torch.from_numpy(audio0) if self.is_half: @@ -162,7 +184,8 @@ class VC(object): with torch.no_grad(): logits = model.extract_features(**inputs) feats = model.final_proj(logits[0]) if version == "v1" else logits[0] - + if(protect<0.5): + feats0=feats.clone() if ( isinstance(index, type(None)) == False and isinstance(big_npy, type(None)) == False @@ -188,6 +211,8 @@ class VC(object): ) feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1) + if(protect<0.5): + feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1) t1 = ttime() p_len = audio0.shape[0] // self.window if feats.shape[1] < p_len: @@ -195,6 +220,14 @@ class VC(object): if pitch != None and pitchf != None: pitch = pitch[:, :p_len] pitchf = pitchf[:, :p_len] + + if(protect<0.5): + pitchff = pitchf.clone() + pitchff[pitchf > 0] = 1 + pitchff[pitchf < 1] = protect + pitchff = pitchff.unsqueeze(-1) + feats = feats * pitchff + feats0 * (1 - pitchff) + feats=feats.to(feats0.dtype) p_len = torch.tensor([p_len], device=self.device).long() with torch.no_grad(): if pitch != None and pitchf != None: @@ -235,6 +268,7 @@ class VC(object): resample_sr, rms_mix_rate, version, + protect, f0_file=None, ): if ( @@ -322,6 +356,7 @@ class VC(object): big_npy, index_rate, version, + protect )[self.t_pad_tgt : -self.t_pad_tgt] ) else: @@ -338,6 +373,7 @@ class VC(object): big_npy, index_rate, version, + protect )[self.t_pad_tgt : -self.t_pad_tgt] ) s = t @@ -355,6 +391,7 @@ class VC(object): big_npy, index_rate, version, + protect )[self.t_pad_tgt : -self.t_pad_tgt] ) else: @@ -371,6 +408,7 @@ class VC(object): big_npy, index_rate, version, + protect )[self.t_pad_tgt : -self.t_pad_tgt] ) audio_opt = np.concatenate(audio_opt)