Update vc_infer_pipeline.py

vc_infer_pipeline.py

@@ -1,4 +1,4 @@
-import numpy as np, parselmouth, torch, pdb
+import numpy as np, parselmouth, torch, pdb, sys, os
 from time import time as ttime
 import torch.nn.functional as F
 import scipy.signal as signal
@@ -6,13 +6,17 @@ import pyworld, os, traceback, faiss, librosa, torchcrepe
 from scipy import signal
 from functools import lru_cache
 
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
 bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)
 
-input_audio_path2wav={}
+input_audio_path2wav = {}
+
 
 @lru_cache
-def cache_harvest_f0(input_audio_path,fs,f0max,f0min,frame_period):
-    audio=input_audio_path2wav[input_audio_path]
+def cache_harvest_f0(input_audio_path, fs, f0max, f0min, frame_period):
+    audio = input_audio_path2wav[input_audio_path]
     f0, t = pyworld.harvest(
         audio,
         fs=fs,
@@ -23,18 +27,29 @@ def cache_harvest_f0(input_audio_path,fs,f0max,f0min,frame_period):
     f0 = pyworld.stonemask(audio, f0, t, fs)
     return f0
 
-def change_rms(data1,sr1,data2,sr2,rate):#1是输入音频，2是输出音频,rate是2的占比
+
+def change_rms(data1, sr1, data2, sr2, rate):  # 1是输入音频，2是输出音频,rate是2的占比
     # print(data1.max(),data2.max())
-    rms1 = librosa.feature.rms(y=data1, frame_length=sr1//2*2, hop_length=sr1//2)#每半秒一个点
-    rms2 = librosa.feature.rms(y=data2, frame_length=sr2//2*2, hop_length=sr2//2)
-    rms1=torch.from_numpy(rms1)
-    rms1=F.interpolate(rms1.unsqueeze(0), size=data2.shape[0],mode='linear').squeeze()
-    rms2=torch.from_numpy(rms2)
-    rms2=F.interpolate(rms2.unsqueeze(0), size=data2.shape[0],mode='linear').squeeze()
-    rms2=torch.max(rms2,torch.zeros_like(rms2)+1e-6)
-    data2*=(torch.pow(rms1,torch.tensor(1-rate))*torch.pow(rms2,torch.tensor(rate-1))).numpy()
+    rms1 = librosa.feature.rms(
+        y=data1, frame_length=sr1 // 2 * 2, hop_length=sr1 // 2
+    )  # 每半秒一个点
+    rms2 = librosa.feature.rms(y=data2, frame_length=sr2 // 2 * 2, hop_length=sr2 // 2)
+    rms1 = torch.from_numpy(rms1)
+    rms1 = F.interpolate(
+        rms1.unsqueeze(0), size=data2.shape[0], mode="linear"
+    ).squeeze()
+    rms2 = torch.from_numpy(rms2)
+    rms2 = F.interpolate(
+        rms2.unsqueeze(0), size=data2.shape[0], mode="linear"
+    ).squeeze()
+    rms2 = torch.max(rms2, torch.zeros_like(rms2) + 1e-6)
+    data2 *= (
+        torch.pow(rms1, torch.tensor(1 - rate))
+        * torch.pow(rms2, torch.tensor(rate - 1))
+    ).numpy()
     return data2
 
+
 class VC(object):
     def __init__(self, tgt_sr, config):
         self.x_pad, self.x_query, self.x_center, self.x_max, self.is_half = (
@@ -54,7 +69,16 @@ class VC(object):
         self.t_max = self.sr * self.x_max  # 免查询时长阈值
         self.device = config.device
 
-    def get_f0(self, input_audio_path,x, p_len, f0_up_key, f0_method,filter_radius, inp_f0=None):
+    def get_f0(
+        self,
+        input_audio_path,
+        x,
+        p_len,
+        f0_up_key,
+        f0_method,
+        filter_radius,
+        inp_f0=None,
+    ):
         global input_audio_path2wav
         time_step = self.window / self.sr * 1000
         f0_min = 50
@@ -78,9 +102,9 @@ class VC(object):
                     f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant"
                 )
         elif f0_method == "harvest":
-            input_audio_path2wav[input_audio_path]=x.astype(np.double)
-            f0=cache_harvest_f0(input_audio_path,self.sr,f0_max,f0_min,10)
-            if(filter_radius>2):
+            input_audio_path2wav[input_audio_path] = x.astype(np.double)
+            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"
@@ -103,6 +127,15 @@ class VC(object):
             f0 = torchcrepe.filter.mean(f0, 3)
             f0[pd < 0.1] = 0
             f0 = f0[0].cpu().numpy()
+        elif f0_method == "rmvpe":
+            if hasattr(self, "model_rmvpe") == False:
+                from rmvpe import RMVPE
+
+                print("loading rmvpe model")
+                self.model_rmvpe = RMVPE(
+                    "rmvpe.pt", is_half=self.is_half, device=self.device
+                )
+            f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
         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
@@ -125,7 +158,7 @@ class VC(object):
         ) + 1
         f0_mel[f0_mel <= 1] = 1
         f0_mel[f0_mel > 255] = 255
-        f0_coarse = np.rint(f0_mel).astype(int)
+        f0_coarse = np.rint(f0_mel).astype(np.int)
         return f0_coarse, f0bak  # 1-0
 
     def vc(
@@ -141,6 +174,7 @@ class VC(object):
         big_npy,
         index_rate,
         version,
+        protect,
     ):  # ,file_index,file_big_npy
         feats = torch.from_numpy(audio0)
         if self.is_half:
@@ -161,8 +195,9 @@ class VC(object):
         t0 = ttime()
         with torch.no_grad():
             logits = model.extract_features(**inputs)
-            feats = model.final_proj(logits[0])if version=="v1"else logits[0]
-
+            feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
+        if protect < 0.5 and pitch != None and pitchf != None:
+            feats0 = feats.clone()
         if (
             isinstance(index, type(None)) == False
             and isinstance(big_npy, type(None)) == False
@@ -188,6 +223,10 @@ class VC(object):
             )
 
         feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
+        if protect < 0.5 and pitch != None and pitchf != None:
+            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 +234,14 @@ class VC(object):
             if pitch != None and pitchf != None:
                 pitch = pitch[:, :p_len]
                 pitchf = pitchf[:, :p_len]
+
+        if protect < 0.5 and pitch != None and pitchf != None:
+            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:
@@ -206,10 +253,7 @@ class VC(object):
                 )
             else:
                 audio1 = (
-                    (net_g.infer(feats, p_len, sid)[0][0, 0])
-                    .data.cpu()
-                    .float()
-                    .numpy()
+                    (net_g.infer(feats, p_len, sid)[0][0, 0]).data.cpu().float().numpy()
                 )
         del feats, p_len, padding_mask
         if torch.cuda.is_available():
@@ -238,6 +282,7 @@ class VC(object):
         resample_sr,
         rms_mix_rate,
         version,
+        protect,
         f0_file=None,
     ):
         if (
@@ -292,7 +337,15 @@ class VC(object):
         sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
         pitch, pitchf = None, None
         if if_f0 == 1:
-            pitch, pitchf = self.get_f0(input_audio_path,audio_pad, p_len, f0_up_key, f0_method,filter_radius, inp_f0)
+            pitch, pitchf = self.get_f0(
+                input_audio_path,
+                audio_pad,
+                p_len,
+                f0_up_key,
+                f0_method,
+                filter_radius,
+                inp_f0,
+            )
             pitch = pitch[:p_len]
             pitchf = pitchf[:p_len]
             if self.device == "mps":
@@ -317,6 +370,7 @@ class VC(object):
                         big_npy,
                         index_rate,
                         version,
+                        protect,
                     )[self.t_pad_tgt : -self.t_pad_tgt]
                 )
             else:
@@ -333,6 +387,7 @@ class VC(object):
                         big_npy,
                         index_rate,
                         version,
+                        protect,
                     )[self.t_pad_tgt : -self.t_pad_tgt]
                 )
             s = t
@@ -350,6 +405,7 @@ class VC(object):
                     big_npy,
                     index_rate,
                     version,
+                    protect,
                 )[self.t_pad_tgt : -self.t_pad_tgt]
             )
         else:
@@ -366,20 +422,22 @@ class VC(object):
                     big_npy,
                     index_rate,
                     version,
+                    protect,
                 )[self.t_pad_tgt : -self.t_pad_tgt]
             )
         audio_opt = np.concatenate(audio_opt)
-        if(rms_mix_rate!=1):
-            audio_opt=change_rms(audio,16000,audio_opt,tgt_sr,rms_mix_rate)
-        if(resample_sr>=16000 and tgt_sr!=resample_sr):
+        if rms_mix_rate != 1:
+            audio_opt = change_rms(audio, 16000, audio_opt, tgt_sr, rms_mix_rate)
+        if resample_sr >= 16000 and tgt_sr != resample_sr:
             audio_opt = librosa.resample(
                 audio_opt, orig_sr=tgt_sr, target_sr=resample_sr
             )
-        audio_max=np.abs(audio_opt).max()/0.99
-        max_int16=32768
-        if(audio_max>1):max_int16/=audio_max
-        audio_opt=(audio_opt * max_int16).astype(np.int16)
+        audio_max = np.abs(audio_opt).max() / 0.99
+        max_int16 = 32768
+        if audio_max > 1:
+            max_int16 /= audio_max
+        audio_opt = (audio_opt * max_int16).astype(np.int16)
         del pitch, pitchf, sid
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
-        return audio_opt
+        return audio_opt