optimize(f0): move fcpe into rvc.f0

rvc/f0/__init__.py

@@ -2,8 +2,9 @@ from .f0 import F0Predictor
 
 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 @@ class CRePE(F0Predictor):
         sampling_rate=44100,
         device="cpu",
     ):
+        if "privateuseone" in str(device):
+            device = "cpu"
         super().__init__(
             hop_length,
             f0_min,
@@ -32,11 +34,13 @@ class CRePE(F0Predictor):
     ):
         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]