Update infer/modules/vc/pipeline.py

infer/modules/vc/pipeline.py

@@ -1,26 +1,24 @@
-import os
-import sys
-import traceback
-import logging
 
-logger = logging.getLogger(__name__)
 
-from functools import lru_cache
-from time import time as ttime
 
-import faiss
-import librosa
-import numpy as np
-import parselmouth
-import pyworld
-import torch
+import numpy as np, parselmouth, torch, pdb, sys, os
+from time import time as ttime
 import torch.nn.functional as F
 import torchcrepe
+from torch import Tensor
+import scipy.signal as signal
+import pyworld, os, faiss, librosa, torchcrepe
 from scipy import signal
+from functools import lru_cache
+import random
+import gc
+import re
 
 now_dir = os.getcwd()
 sys.path.append(now_dir)
 
+from infer.modules.FCPEF0Predictor import FCPEF0Predictor
+
 bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)
 
 input_audio_path2wav = {}
@@ -40,21 +38,22 @@ def cache_harvest_f0(input_audio_path, fs, f0max, f0min, frame_period):
     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):
     # 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
+    rms1 = librosa.feature.rms(y=data1, frame_length=sr1 // 2 * 2, hop_length=sr1 // 2)
     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))
@@ -62,7 +61,7 @@ def change_rms(data1, sr1, data2, sr2, rate):  # 1是输入音频，2是输出
     return data2
 
 
-class Pipeline(object):
+class VC(object):
     def __init__(self, tgt_sr, config):
         self.x_pad, self.x_query, self.x_center, self.x_max, self.is_half = (
             config.x_pad,
@@ -71,15 +70,181 @@ class Pipeline(object):
             config.x_max,
             config.is_half,
         )
-        self.sr = 16000  # hubert输入采样率
-        self.window = 160  # 每帧点数
-        self.t_pad = self.sr * self.x_pad  # 每条前后pad时间
+        self.sr = 16000
+        self.window = 160
+        self.t_pad = self.sr * self.x_pad
         self.t_pad_tgt = tgt_sr * self.x_pad
         self.t_pad2 = self.t_pad * 2
-        self.t_query = self.sr * self.x_query  # 查询切点前后查询时间
-        self.t_center = self.sr * self.x_center  # 查询切点位置
-        self.t_max = self.sr * self.x_max  # 免查询时长阈值
+        self.t_query = self.sr * self.x_query
+        self.t_center = self.sr * self.x_center
+        self.t_max = self.sr * self.x_max
         self.device = config.device
+        self.ref_freqs = [
+            65.41,
+            82.41,
+            110.00,
+            146.83,
+            196.00,
+            246.94,
+            329.63,
+            440.00,
+            587.33,
+            783.99,
+            1046.50,
+        ]
+        # Generate interpolated frequencies
+        self.note_dict = self.generate_interpolated_frequencies()
+
+    def generate_interpolated_frequencies(self):
+        # Generate interpolated frequencies based on the reference frequencies.
+        note_dict = []
+        for i in range(len(self.ref_freqs) - 1):
+            freq_low = self.ref_freqs[i]
+            freq_high = self.ref_freqs[i + 1]
+            # Interpolate between adjacent reference frequencies
+            interpolated_freqs = np.linspace(
+                freq_low, freq_high, num=10, endpoint=False
+            )
+            note_dict.extend(interpolated_freqs)
+        # Add the last reference frequency
+        note_dict.append(self.ref_freqs[-1])
+        return note_dict
+
+    def autotune_f0(self, f0):
+        # Autotunes the given fundamental frequency (f0) to the nearest musical note.
+        autotuned_f0 = np.zeros_like(f0)
+        for i, freq in enumerate(f0):
+            # Find the closest note
+            closest_note = min(self.note_dict, key=lambda x: abs(x - freq))
+            autotuned_f0[i] = closest_note
+        return autotuned_f0
+
+    def get_optimal_torch_device(self, index: int = 0) -> torch.device:
+        if torch.cuda.is_available():
+            return torch.device(f"cuda:{index % torch.cuda.device_count()}")
+        elif torch.backends.mps.is_available():
+            return torch.device("mps")
+        return torch.device("cpu")
+
+    def get_f0_crepe_computation(
+        self,
+        x,
+        f0_min,
+        f0_max,
+        p_len,
+        hop_length,
+        model="full",
+    ):
+        x = x.astype(np.float32)
+        x /= np.quantile(np.abs(x), 0.999)
+        torch_device = self.get_optimal_torch_device()
+        audio = torch.from_numpy(x).to(torch_device, copy=True)
+        audio = torch.unsqueeze(audio, dim=0)
+        if audio.ndim == 2 and audio.shape[0] > 1:
+            audio = torch.mean(audio, dim=0, keepdim=True).detach()
+        audio = audio.detach()
+        pitch: Tensor = torchcrepe.predict(
+            audio,
+            self.sr,
+            hop_length,
+            f0_min,
+            f0_max,
+            model,
+            batch_size=hop_length * 2,
+            device=torch_device,
+            pad=True,
+        )
+        p_len = p_len or x.shape[0] // hop_length
+        source = np.array(pitch.squeeze(0).cpu().float().numpy())
+        source[source < 0.001] = np.nan
+        target = np.interp(
+            np.arange(0, len(source) * p_len, len(source)) / p_len,
+            np.arange(0, len(source)),
+            source,
+        )
+        f0 = np.nan_to_num(target)
+        return f0
+
+    def get_f0_official_crepe_computation(
+        self,
+        x,
+        f0_min,
+        f0_max,
+        model="full",
+    ):
+        batch_size = 512
+        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()
+        return f0
+
+    def get_f0_hybrid_computation(
+        self,
+        methods_str,
+        x,
+        f0_min,
+        f0_max,
+        p_len,
+        hop_length,
+    ):
+        methods_str = re.search("hybrid\[(.+)\]", methods_str)
+        if methods_str:
+            methods = [method.strip() for method in methods_str.group(1).split("+")]
+        f0_computation_stack = []
+        print(f"Calculating f0 pitch estimations for methods {str(methods)}")
+        x = x.astype(np.float32)
+        x /= np.quantile(np.abs(x), 0.999)
+        for method in methods:
+            f0 = None
+            if method == "crepe":
+                f0 = self.get_f0_crepe_computation(
+                    x, f0_min, f0_max, p_len, int(hop_length)
+                )
+            elif method == "rmvpe":
+                if hasattr(self, "model_rmvpe") == False:
+                    from rvc.lib.rmvpe import RMVPE
+
+                    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 = f0[1:]
+            elif method == "fcpe":
+                self.model_fcpe = FCPEF0Predictor(
+                    "fcpe.pt",
+                    f0_min=int(f0_min),
+                    f0_max=int(f0_max),
+                    dtype=torch.float32,
+                    device=self.device,
+                    sampling_rate=self.sr,
+                    threshold=0.03,
+                )
+                f0 = self.model_fcpe.compute_f0(x, p_len=p_len)
+                del self.model_fcpe
+                gc.collect()
+            f0_computation_stack.append(f0)
+
+        print(f"Calculating hybrid median f0 from the stack of {str(methods)}")
+        f0_computation_stack = [fc for fc in f0_computation_stack if fc is not None]
+        f0_median_hybrid = None
+        if len(f0_computation_stack) == 1:
+            f0_median_hybrid = f0_computation_stack[0]
+        else:
+            f0_median_hybrid = np.nanmedian(f0_computation_stack, axis=0)
+        return f0_median_hybrid
 
     def get_f0(
         self,
@@ -89,6 +254,8 @@ class Pipeline(object):
         f0_up_key,
         f0_method,
         filter_radius,
+        hop_length,
+        f0autotune,
         inp_f0=None,
     ):
         global input_audio_path2wav
@@ -116,69 +283,63 @@ class Pipeline(object):
         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:
+            if int(filter_radius) > 2:
                 f0 = signal.medfilt(f0, 3)
+        elif f0_method == "dio":
+            f0, t = pyworld.dio(
+                x.astype(np.double),
+                fs=self.sr,
+                f0_ceil=f0_max,
+                f0_floor=f0_min,
+                frame_period=10,
+            )
+            f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.sr)
+            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,
+            f0 = self.get_f0_crepe_computation(
+                x, f0_min, f0_max, p_len, int(hop_length)
+            )
+        elif f0_method == "crepe-tiny":
+            f0 = self.get_f0_crepe_computation(
+                x, f0_min, f0_max, p_len, int(hop_length), "tiny"
             )
-            pd = torchcrepe.filter.median(pd, 3)
-            f0 = torchcrepe.filter.mean(f0, 3)
-            f0[pd < 0.1] = 0
-            f0 = f0[0].cpu().numpy()
         elif f0_method == "rmvpe":
-            if not hasattr(self, "model_rmvpe"):
-                from infer.lib.rmvpe import RMVPE
+            if hasattr(self, "model_rmvpe") == False:
+                from rvc.lib.rmvpe import RMVPE
 
-                logger.info(
-                    "Loading rmvpe model,%s" % "%s/rmvpe.pt" % os.environ["rmvpe_root"]
-                )
                 self.model_rmvpe = RMVPE(
-                    "%s/rmvpe.pt" % os.environ["rmvpe_root"],
-                    is_half=self.is_half,
-                    device=self.device,
-                    # use_jit=self.config.use_jit,
+                    "rmvpe.pt", is_half=self.is_half, device=self.device
                 )
             f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
-
-            if "privateuseone" in str(self.device):  # clean ortruntime memory
-                del self.model_rmvpe.model
-                del self.model_rmvpe
-                logger.info("Cleaning ortruntime memory")
         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,
-                )
-                .squeeze()
-                .cpu()
-                .numpy()
+            self.model_fcpe = FCPEF0Predictor(
+                "fcpe.pt",
+                f0_min=int(f0_min),
+                f0_max=int(f0_max),
+                dtype=torch.float32,
+                device=self.device,
+                sampling_rate=self.sr,
+                threshold=0.03,
+            )
+            f0 = self.model_fcpe.compute_f0(x, p_len=p_len)
+            del self.model_fcpe
+            gc.collect()
+        elif "hybrid" in f0_method:
+            input_audio_path2wav[input_audio_path] = x.astype(np.double)
+            f0 = self.get_f0_hybrid_computation(
+                f0_method,
+                x,
+                f0_min,
+                f0_max,
+                p_len,
+                hop_length,
             )
 
+        if f0autotune == "True":
+            f0 = self.autotune_f0(f0)
+
         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
+        tf0 = self.sr // self.window
         if inp_f0 is not None:
             delta_t = np.round(
                 (inp_f0[:, 0].max() - inp_f0[:, 0].min()) * tf0 + 1
@@ -190,7 +351,6 @@ class Pipeline(object):
             f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)] = replace_f0[
                 :shape
             ]
-        # with open("test_opt.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
         f0bak = f0.copy()
         f0_mel = 1127 * np.log(1 + f0 / 700)
         f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
@@ -198,8 +358,9 @@ class Pipeline(object):
         ) + 1
         f0_mel[f0_mel <= 1] = 1
         f0_mel[f0_mel > 255] = 255
-        f0_coarse = np.rint(f0_mel).astype(np.int32)
-        return f0_coarse, f0bak  # 1-0
+        f0_coarse = np.rint(f0_mel).astype(np.int)
+
+        return f0_coarse, f0bak
 
     def vc(
         self,
@@ -209,19 +370,18 @@ class Pipeline(object):
         audio0,
         pitch,
         pitchf,
-        times,
         index,
         big_npy,
         index_rate,
         version,
         protect,
-    ):  # ,file_index,file_big_npy
+    ):
         feats = torch.from_numpy(audio0)
         if self.is_half:
             feats = feats.half()
         else:
             feats = feats.float()
-        if feats.dim() == 2:  # double channels
+        if feats.dim() == 2:
             feats = feats.mean(-1)
         assert feats.dim() == 1, feats.dim()
         feats = feats.view(1, -1)
@@ -236,20 +396,17 @@ class Pipeline(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 and pitch is not None and pitchf is not None:
+        if protect < 0.5 and pitch != None and pitchf != None:
             feats0 = feats.clone()
         if (
-            not isinstance(index, type(None))
-            and not isinstance(big_npy, type(None))
+            isinstance(index, type(None)) == False
+            and isinstance(big_npy, type(None)) == False
             and index_rate != 0
         ):
             npy = feats[0].cpu().numpy()
             if self.is_half:
                 npy = npy.astype("float32")
 
-            # _, I = index.search(npy, 1)
-            # npy = big_npy[I.squeeze()]
-
             score, ix = index.search(npy, k=8)
             weight = np.square(1 / score)
             weight /= weight.sum(axis=1, keepdims=True)
@@ -263,7 +420,7 @@ class Pipeline(object):
             )
 
         feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
-        if protect < 0.5 and pitch is not None and pitchf is not None:
+        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
             )
@@ -271,11 +428,11 @@ class Pipeline(object):
         p_len = audio0.shape[0] // self.window
         if feats.shape[1] < p_len:
             p_len = feats.shape[1]
-            if pitch is not None and pitchf is not None:
+            if pitch != None and pitchf != None:
                 pitch = pitch[:, :p_len]
                 pitchf = pitchf[:, :p_len]
 
-        if protect < 0.5 and pitch is not None and pitchf is not None:
+        if protect < 0.5 and pitch != None and pitchf != None:
             pitchff = pitchf.clone()
             pitchff[pitchf > 0] = 1
             pitchff[pitchf < 1] = protect
@@ -284,16 +441,21 @@ class Pipeline(object):
             feats = feats.to(feats0.dtype)
         p_len = torch.tensor([p_len], device=self.device).long()
         with torch.no_grad():
-            hasp = pitch is not None and pitchf is not None
-            arg = (feats, p_len, pitch, pitchf, sid) if hasp else (feats, p_len, sid)
-            audio1 = (net_g.infer(*arg)[0][0, 0]).data.cpu().float().numpy()
-            del hasp, arg
+            if pitch != None and pitchf != None:
+                audio1 = (
+                    (net_g.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0])
+                    .data.cpu()
+                    .float()
+                    .numpy()
+                )
+            else:
+                audio1 = (
+                    (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():
             torch.cuda.empty_cache()
         t2 = ttime()
-        times[0] += t1 - t0
-        times[2] += t2 - t1
         return audio1
 
     def pipeline(
@@ -303,7 +465,6 @@ class Pipeline(object):
         sid,
         audio,
         input_audio_path,
-        times,
         f0_up_key,
         f0_method,
         file_index,
@@ -315,21 +476,16 @@ class Pipeline(object):
         rms_mix_rate,
         version,
         protect,
+        hop_length,
+        f0autotune,
         f0_file=None,
     ):
-        if (
-            file_index != ""
-            # and file_big_npy != ""
-            # and os.path.exists(file_big_npy) == True
-            and os.path.exists(file_index)
-            and index_rate != 0
-        ):
+        if file_index != "" and os.path.exists(file_index) == True and index_rate != 0:
             try:
                 index = faiss.read_index(file_index)
-                # big_npy = np.load(file_big_npy)
                 big_npy = index.reconstruct_n(0, index.ntotal)
-            except:
-                traceback.print_exc()
+            except Exception as error:
+                print(error)
                 index = big_npy = None
         else:
             index = big_npy = None
@@ -339,14 +495,14 @@ class Pipeline(object):
         if audio_pad.shape[0] > self.t_max:
             audio_sum = np.zeros_like(audio)
             for i in range(self.window):
-                audio_sum += np.abs(audio_pad[i : i - self.window])
+                audio_sum += audio_pad[i : i - self.window]
             for t in range(self.t_center, audio.shape[0], self.t_center):
                 opt_ts.append(
                     t
                     - self.t_query
                     + np.where(
-                        audio_sum[t - self.t_query : t + self.t_query]
-                        == audio_sum[t - self.t_query : t + self.t_query].min()
+                        np.abs(audio_sum[t - self.t_query : t + self.t_query])
+                        == np.abs(audio_sum[t - self.t_query : t + self.t_query]).min()
                     )[0][0]
                 )
         s = 0
@@ -356,7 +512,7 @@ class Pipeline(object):
         audio_pad = np.pad(audio, (self.t_pad, self.t_pad), mode="reflect")
         p_len = audio_pad.shape[0] // self.window
         inp_f0 = None
-        if hasattr(f0_file, "name"):
+        if hasattr(f0_file, "name") == True:
             try:
                 with open(f0_file.name, "r") as f:
                     lines = f.read().strip("\n").split("\n")
@@ -364,8 +520,8 @@ class Pipeline(object):
                 for line in lines:
                     inp_f0.append([float(i) for i in line.split(",")])
                 inp_f0 = np.array(inp_f0, dtype="float32")
-            except:
-                traceback.print_exc()
+            except Exception as error:
+                print(error)
         sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
         pitch, pitchf = None, None
         if if_f0 == 1:
@@ -376,16 +532,17 @@ class Pipeline(object):
                 f0_up_key,
                 f0_method,
                 filter_radius,
+                hop_length,
+                f0autotune,
                 inp_f0,
             )
             pitch = pitch[:p_len]
             pitchf = pitchf[:p_len]
-            if "mps" not in str(self.device) or "xpu" not in str(self.device):
+            if self.device == "mps":
                 pitchf = pitchf.astype(np.float32)
             pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
             pitchf = torch.tensor(pitchf, device=self.device).unsqueeze(0).float()
         t2 = ttime()
-        times[1] += t2 - t1
         for t in opt_ts:
             t = t // self.window * self.window
             if if_f0 == 1:
@@ -397,7 +554,6 @@ class Pipeline(object):
                         audio_pad[s : t + self.t_pad2 + self.window],
                         pitch[:, s // self.window : (t + self.t_pad2) // self.window],
                         pitchf[:, s // self.window : (t + self.t_pad2) // self.window],
-                        times,
                         index,
                         big_npy,
                         index_rate,
@@ -414,7 +570,6 @@ class Pipeline(object):
                         audio_pad[s : t + self.t_pad2 + self.window],
                         None,
                         None,
-                        times,
                         index,
                         big_npy,
                         index_rate,
@@ -432,7 +587,6 @@ class Pipeline(object):
                     audio_pad[t:],
                     pitch[:, t // self.window :] if t is not None else pitch,
                     pitchf[:, t // self.window :] if t is not None else pitchf,
-                    times,
                     index,
                     big_npy,
                     index_rate,
@@ -449,7 +603,6 @@ class Pipeline(object):
                     audio_pad[t:],
                     None,
                     None,
-                    times,
                     index,
                     big_npy,
                     index_rate,
@@ -460,7 +613,7 @@ class Pipeline(object):
         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 tgt_sr != resample_sr >= 16000:
+        if resample_sr >= 16000 and tgt_sr != resample_sr:
             audio_opt = librosa.resample(
                 audio_opt, orig_sr=tgt_sr, target_sr=resample_sr
             )
@@ -472,4 +625,4 @@ class Pipeline(object):
         del pitch, pitchf, sid
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
-        return audio_opt
+        return audio_opt