optimize(f0): move fcpe into rvc.f0

Author: fumiama

infer/lib/rtrvc.py

@@ -13,7 +13,6 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-import torchcrepe
 from torchaudio.transforms import Resample
 
 from rvc.synthesizer import load_synthesizer
@@ -323,20 +322,17 @@ def get_f0_rmvpe(self, x, f0_up_key):
 
     def get_f0_fcpe(self, x, f0_up_key):
         if hasattr(self, "model_fcpe") == False:
-            from torchfcpe import spawn_bundled_infer_model
+            from rvc.f0 import FCPE
 
             printt("Loading fcpe model")
-            if "privateuseone" in str(self.device):
-                self.device_fcpe = "cpu"
-            else:
-                self.device_fcpe = self.device
-            self.model_fcpe = spawn_bundled_infer_model(self.device_fcpe)
-        f0 = self.model_fcpe.infer(
-            x.to(self.device_fcpe).unsqueeze(0).float(),
-            sr=16000,
-            decoder_mode="local_argmax",
-            threshold=0.006,
-        )
+            self.model_fcpe = FCPE(
+                160,
+                self.f0_min,
+                self.f0_max,
+                16000,
+                self.device,
+            )
+        f0 = self.model_fcpe.compute_f0(x)
         f0 *= pow(2, f0_up_key / 12)
         return self.get_f0_post(f0)
 

infer/modules/vc/pipeline.py

@@ -14,7 +14,7 @@
 import torch.nn.functional as F
 from scipy import signal
 
-from rvc.f0 import PM, Harvest, RMVPE, CRePE, Dio
+from rvc.f0 import PM, Harvest, RMVPE, CRePE, Dio, FCPE
 
 now_dir = os.getcwd()
 sys.path.append(now_dir)
@@ -118,21 +118,15 @@ def get_f0(
 
         elif f0_method == "fcpe":
             if not hasattr(self, "model_fcpe"):
-                from torchfcpe import spawn_bundled_infer_model
-
                 logger.info("Loading fcpe model")
-                self.model_fcpe = spawn_bundled_infer_model(self.device)
-            f0 = (
-                self.model_fcpe.infer(
-                    torch.from_numpy(x).to(self.device).unsqueeze(0).float(),
-                    sr=16000,
-                    decoder_mode="local_argmax",
-                    threshold=0.006,
+                self.model_fcpe = FCPE(
+                    self.window,
+                    f0_min,
+                    f0_max,
+                    self.sr,
+                    self.device,
                 )
-                .squeeze()
-                .cpu()
-                .numpy()
-            )
+            f0 = self.model_fcpe.compute_f0(x, p_len=p_len)
 
         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()]))

rvc/f0/__init__.py

@@ -2,8 +2,9 @@
 
 from .crepe import CRePE
 from .dio import Dio
+from .fcpe import FCPE
 from .harvest import Harvest
 from .pm import PM
 from .rmvpe import RMVPE
 
-__all__ = ["F0Predictor", "CRePE", "Dio", "Harvest", "PM", "RMVPE"]
+__all__ = ["F0Predictor", "CRePE", "Dio", "FCPE", "Harvest", "PM", "RMVPE"]

rvc/f0/crepe.py

@@ -16,6 +16,8 @@ def __init__(
         sampling_rate=44100,
         device="cpu",
     ):
+        if "privateuseone" in str(device):
+            device = "cpu"
         super().__init__(
             hop_length,
             f0_min,
@@ -32,11 +34,13 @@ def compute_f0(
     ):
         if p_len is None:
             p_len = wav.shape[0] // self.hop_length
+        if not torch.is_tensor(wav):
+            wav = torch.from_numpy(wav)
         # Pick a batch size that doesn't cause memory errors on your gpu
         batch_size = 512
         # Compute pitch using device 'device'
         f0, pd = torchcrepe.predict(
-            torch.tensor(np.copy(wav))[None].float().to(self.device),
+            wav.float().to(self.device).unsqueeze(dim=0),
             self.sampling_rate,
             self.hop_length,
             self.f0_min,

rvc/f0/fcpe.py

@@ -0,0 +1,45 @@
+from typing import Any, Optional, Union
+
+import numpy as np
+import torch
+from torchfcpe import spawn_bundled_infer_model
+
+from .f0 import F0Predictor
+
+
+class FCPE(F0Predictor):
+    def __init__(
+        self,
+        hop_length=512,
+        f0_min=50,
+        f0_max=1100,
+        sampling_rate=44100,
+        device="cpu",
+    ):
+        super().__init__(
+            hop_length,
+            f0_min,
+            f0_max,
+            sampling_rate,
+            device,
+        )
+
+        self.model = spawn_bundled_infer_model(self.device)
+
+    def compute_f0(
+        self,
+        wav: np.ndarray[Any, np.dtype],
+        p_len: Optional[int] = None,
+        filter_radius: Optional[Union[int, float]] = 0.006,
+    ):
+        if p_len is None:
+            p_len = wav.shape[0] // self.hop_length
+        if not torch.is_tensor(wav):
+            wav = torch.from_numpy(wav)
+        f0 = self.model.infer(
+            wav.float().to(self.device).unsqueeze(0),
+            sr=self.sampling_rate,
+            decoder_mode="local_argmax",
+            threshold=filter_radius,
+        ).squeeze().cpu().numpy()
+        return self._interpolate_f0(self._resize_f0(f0, p_len))[0]

web.py

@@ -861,9 +861,7 @@ def change_f0_method(f0method8):
                                 "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement"
                             ),
                             choices=(
-                                ["pm", "dio", "harvest", "rmvpe"]
-                                if config.dml
-                                else ["pm", "dio", "harvest", "crepe", "rmvpe"]
+                                ["pm", "dio", "harvest", "crepe", "rmvpe", "fcpe"]
                             ),
                             value="rmvpe",
                             interactive=True,