Experimental: eliminate torchaudio from rvc

server/MMVCServerSIO.py

@@ -239,12 +239,12 @@ if __name__ == "__main__":
             if sys.platform.startswith("win"):
                 process = subprocess.Popen([NATIVE_CLIENT_FILE_WIN, "--disable-gpu", "-u", f"http://localhost:{PORT}/"])
                 return_code = process.wait()
-                logger.info("client closed.")                
+                logger.info("client closed.")
                 p.terminate()
             elif sys.platform.startswith("darwin"):
                 process = subprocess.Popen([NATIVE_CLIENT_FILE_MAC, "--disable-gpu", "-u", f"http://localhost:{PORT}/"])
                 return_code = process.wait()
-                logger.info("client closed.")                
+                logger.info("client closed.")
                 p.terminate()
 
         except Exception as e:

server/voice_changer/DiffusionSVC/DiffusionSVC.py

@@ -14,7 +14,7 @@ from voice_changer.RVC.embedder.EmbedderManager import EmbedderManager
 # from voice_changer.RVC.onnxExporter.export2onnx import export2onnx
 from voice_changer.RVC.deviceManager.DeviceManager import DeviceManager
 
-from Exceptions import DeviceCannotSupportHalfPrecisionException, PipelineCreateException
+from Exceptions import DeviceCannotSupportHalfPrecisionException, PipelineCreateException, PipelineNotInitializedException
 
 logger = VoiceChangaerLogger.get_instance().getLogger()
 
@@ -28,7 +28,6 @@ class DiffusionSVC(VoiceChangerModel):
         InferencerManager.initialize(params)
         self.settings = DiffusionSVCSettings()
         self.params = params
-        self.pitchExtractor = PitchExtractorManager.getPitchExtractor(self.settings.f0Detector, self.settings.gpu)
 
         self.pipeline: Pipeline | None = None
 
@@ -84,6 +83,8 @@ class DiffusionSVC(VoiceChangerModel):
         if self.pipeline is not None:
             pipelineInfo = self.pipeline.getPipelineInfo()
             data["pipelineInfo"] = pipelineInfo
+        else:
+            data["pipelineInfo"] = "None"
         return data
 
     def get_processing_sampling_rate(self):
@@ -137,6 +138,10 @@ class DiffusionSVC(VoiceChangerModel):
         return (self.audio_buffer, self.pitchf_buffer, self.feature_buffer, convertSize, vol)
 
     def inference(self, receivedData: AudioInOut, crossfade_frame: int, sola_search_frame: int):
+        if self.pipeline is None:
+            logger.info("[Voice Changer] Pipeline is not initialized.")
+            raise PipelineNotInitializedException()
+        
         data = self.generate_input(receivedData, crossfade_frame, sola_search_frame)
         audio: AudioInOut = data[0]
         pitchf: PitchfInOut = data[1]

server/voice_changer/RVC/RVCr2.py

@@ -0,0 +1,287 @@
+'''
+VoiceChangerV2向け
+'''
+from dataclasses import asdict
+import numpy as np
+import torch
+from data.ModelSlot import RVCModelSlot
+from mods.log_control import VoiceChangaerLogger
+
+from voice_changer.RVC.RVCSettings import RVCSettings
+from voice_changer.RVC.embedder.EmbedderManager import EmbedderManager
+from voice_changer.utils.VoiceChangerModel import AudioInOut, PitchfInOut, FeatureInOut, VoiceChangerModel
+from voice_changer.utils.VoiceChangerParams import VoiceChangerParams
+from voice_changer.RVC.onnxExporter.export2onnx import export2onnx
+from voice_changer.RVC.pitchExtractor.PitchExtractorManager import PitchExtractorManager
+from voice_changer.RVC.pipeline.PipelineGenerator import createPipeline
+from voice_changer.RVC.deviceManager.DeviceManager import DeviceManager
+from voice_changer.RVC.pipeline.Pipeline import Pipeline
+
+from Exceptions import DeviceCannotSupportHalfPrecisionException, PipelineCreateException, PipelineNotInitializedException
+import resampy
+from typing import cast
+
+logger = VoiceChangaerLogger.get_instance().getLogger()
+
+
+class RVCr2(VoiceChangerModel):
+    def __init__(self, params: VoiceChangerParams, slotInfo: RVCModelSlot):
+        logger.info("[Voice Changer] [RVCr2] Creating instance ")
+        self.deviceManager = DeviceManager.get_instance()
+        EmbedderManager.initialize(params)
+        PitchExtractorManager.initialize(params)
+        self.settings = RVCSettings()
+        self.params = params
+        # self.pitchExtractor = PitchExtractorManager.getPitchExtractor(self.settings.f0Detector, self.settings.gpu)
+
+        self.pipeline: Pipeline | None = None
+
+        self.audio_buffer: AudioInOut | None = None
+        self.pitchf_buffer: PitchfInOut | None = None
+        self.feature_buffer: FeatureInOut | None = None
+        self.prevVol = 0.0
+        self.slotInfo = slotInfo
+        # self.initialize()
+
+    def initialize(self):
+        logger.info("[Voice Changer][RVCr2] Initializing... ")
+
+        # pipelineの生成
+        try:
+            self.pipeline = createPipeline(self.slotInfo, self.settings.gpu, self.settings.f0Detector)
+        except PipelineCreateException as e:  # NOQA
+            logger.error("[Voice Changer] pipeline create failed. check your model is valid.")
+            return
+
+        # その他の設定
+        self.settings.tran = self.slotInfo.defaultTune
+        self.settings.indexRatio = self.slotInfo.defaultIndexRatio
+        self.settings.protect = self.slotInfo.defaultProtect
+        logger.info("[Voice Changer] [RVC] Initializing... done")
+
+    def setSamplingRate(self, inputSampleRate, outputSampleRate):
+        self.inputSampleRate = inputSampleRate
+        self.outputSampleRate = outputSampleRate
+        self.initialize()
+
+    def update_settings(self, key: str, val: int | float | str):
+        logger.info(f"[Voice Changer][RVC]: update_settings {key}:{val}")
+        if key in self.settings.intData:
+            setattr(self.settings, key, int(val))
+            if key == "gpu":
+                self.deviceManager.setForceTensor(False)
+                self.initialize()
+        elif key in self.settings.floatData:
+            setattr(self.settings, key, float(val))
+        elif key in self.settings.strData:
+            setattr(self.settings, key, str(val))
+            if key == "f0Detector" and self.pipeline is not None:
+                pitchExtractor = PitchExtractorManager.getPitchExtractor(self.settings.f0Detector, self.settings.gpu)
+                self.pipeline.setPitchExtractor(pitchExtractor)
+        else:
+            return False
+        return True
+
+    def get_info(self):
+        data = asdict(self.settings)
+        if self.pipeline is not None:
+            pipelineInfo = self.pipeline.getPipelineInfo()
+            data["pipelineInfo"] = pipelineInfo
+        else:
+            data["pipelineInfo"] = "None"
+        return data
+
+    def get_processing_sampling_rate(self):
+        return self.slotInfo.samplingRate
+
+    def generate_input(
+        self,
+        newData: AudioInOut,
+        crossfadeSize: int,
+        solaSearchFrame: int,
+        extra_frame: int
+    ):
+        # 16k で入ってくる。
+        inputSize = newData.shape[0]
+        newData = newData.astype(np.float32) / 32768.0
+        newFeatureLength = inputSize // 160  # hopsize:=160
+
+        if self.audio_buffer is not None:
+            # 過去のデータに連結
+            self.audio_buffer = np.concatenate([self.audio_buffer, newData], 0)
+            if self.slotInfo.f0:
+                self.pitchf_buffer = np.concatenate([self.pitchf_buffer, np.zeros(newFeatureLength)], 0)
+            self.feature_buffer = np.concatenate([self.feature_buffer, np.zeros([newFeatureLength, self.slotInfo.embChannels])], 0)
+        else:
+            self.audio_buffer = newData
+            if self.slotInfo.f0:
+                self.pitchf_buffer = np.zeros(newFeatureLength)
+            self.feature_buffer = np.zeros([newFeatureLength, self.slotInfo.embChannels])
+
+        convertSize = inputSize + crossfadeSize + solaSearchFrame + extra_frame
+
+        if convertSize % 160 != 0:  # モデルの出力のホップサイズで切り捨てが発生するので補う。
+            convertSize = convertSize + (160 - (convertSize % 160))
+        outSize = convertSize - extra_frame
+
+        # バッファがたまっていない場合はzeroで補う
+        if self.audio_buffer.shape[0] < convertSize:
+            self.audio_buffer = np.concatenate([np.zeros([convertSize]), self.audio_buffer])
+            if self.slotInfo.f0:
+                self.pitchf_buffer = np.concatenate([np.zeros([convertSize // 160]), self.pitchf_buffer])
+            self.feature_buffer = np.concatenate([np.zeros([convertSize // 160, self.slotInfo.embChannels]), self.feature_buffer])
+
+        # 不要部分をトリミング
+        convertOffset = -1 * convertSize
+        featureOffset = convertOffset // 160
+        self.audio_buffer = self.audio_buffer[convertOffset:]  # 変換対象の部分だけ抽出
+        if self.slotInfo.f0:
+            self.pitchf_buffer = self.pitchf_buffer[featureOffset:]
+        self.feature_buffer = self.feature_buffer[featureOffset:]
+
+        # 出力部分だけ切り出して音量を確認。(TODO:段階的消音にする)
+        cropOffset = -1 * (inputSize + crossfadeSize)
+        cropEnd = -1 * (crossfadeSize)
+        crop = self.audio_buffer[cropOffset:cropEnd]
+        vol = np.sqrt(np.square(crop).mean())
+        vol = max(vol, self.prevVol * 0.0)
+        self.prevVol = vol
+
+        return (self.audio_buffer, self.pitchf_buffer, self.feature_buffer, convertSize, vol, outSize)
+
+    def inference(self, receivedData: AudioInOut, crossfade_frame: int, sola_search_frame: int):
+        if self.pipeline is None:
+            logger.info("[Voice Changer] Pipeline is not initialized.")
+            raise PipelineNotInitializedException()
+
+        # 処理は16Kで実施(Pitch, embed, (infer))
+        receivedData = cast(
+            AudioInOut,
+            resampy.resample(
+                receivedData,
+                self.inputSampleRate,
+                16000,
+            ),
+        )
+        crossfade_frame = int((crossfade_frame / self.inputSampleRate) * 16000)
+        sola_search_frame = int((sola_search_frame / self.inputSampleRate) * 16000)
+        extra_frame = int((self.settings.extraConvertSize / self.inputSampleRate) * 16000)
+
+        # 入力データ生成
+        data = self.generate_input(receivedData, crossfade_frame, sola_search_frame, extra_frame)
+
+        audio = data[0]
+        pitchf = data[1]
+        feature = data[2]
+        convertSize = data[3]
+        vol = data[4]
+        outSize = data[5]
+
+        if vol < self.settings.silentThreshold:
+            return np.zeros(convertSize).astype(np.int16) * np.sqrt(vol)
+
+        device = self.pipeline.device
+        
+        audio = torch.from_numpy(audio).to(device=device, dtype=torch.float32)
+        repeat = 1 if self.settings.rvcQuality else 0
+        sid = self.settings.dstId
+        f0_up_key = self.settings.tran
+        index_rate = self.settings.indexRatio
+        protect = self.settings.protect
+
+        if_f0 = 1 if self.slotInfo.f0 else 0
+        embOutputLayer = self.slotInfo.embOutputLayer
+        useFinalProj = self.slotInfo.useFinalProj
+
+        try:
+            audio_out, self.pitchf_buffer, self.feature_buffer = self.pipeline.exec(
+                sid,
+                audio,
+                pitchf,
+                feature,
+                f0_up_key,
+                index_rate,
+                if_f0,
+                self.settings.extraConvertSize / self.slotInfo.samplingRate if self.settings.silenceFront else 0.,  # extaraDataSizeの秒数。RVCのモデルのサンプリングレートで処理(★１)。
+                embOutputLayer,
+                useFinalProj,
+                repeat,
+                protect
+            )
+            outSize = outSize // 16000 * self.slotInfo.samplingRate
+            result = audio_out[-outSize:].detach().cpu().numpy() * np.sqrt(vol)
+
+            result = cast(
+                AudioInOut,
+                resampy.resample(
+                    result,
+                    self.slotInfo.samplingRate,
+                    self.outputSampleRate,
+                ),
+            )
+
+            return result
+        except DeviceCannotSupportHalfPrecisionException as e:  # NOQA
+            logger.warn("[Device Manager] Device cannot support half precision. Fallback to float....")
+            self.deviceManager.setForceTensor(True)
+            self.initialize()
+            # raise e
+
+        return
+
+    def __del__(self):
+        del self.pipeline
+
+        # print("---------- REMOVING ---------------")
+
+        # remove_path = os.path.join("RVC")
+        # sys.path = [x for x in sys.path if x.endswith(remove_path) is False]
+
+        # for key in list(sys.modules):
+        #     val = sys.modules.get(key)
+        #     try:
+        #         file_path = val.__file__
+        #         if file_path.find("RVC" + os.path.sep) >= 0:
+        #             # print("remove", key, file_path)
+        #             sys.modules.pop(key)
+        #     except Exception:  # type:ignore
+        #         # print(e)
+        #         pass
+
+    def export2onnx(self):
+        modelSlot = self.slotInfo
+
+        if modelSlot.isONNX:
+            logger.warn("[Voice Changer] export2onnx, No pyTorch filepath.")
+            return {"status": "ng", "path": ""}
+
+        if self.pipeline is not None:
+            del self.pipeline
+            self.pipeline = None
+
+        torch.cuda.empty_cache()
+        self.initialize()
+
+        output_file_simple = export2onnx(self.settings.gpu, modelSlot)
+
+        return {
+            "status": "ok",
+            "path": f"/tmp/{output_file_simple}",
+            "filename": output_file_simple,
+        }
+
+    def get_model_current(self):
+        return [
+            {
+                "key": "defaultTune",
+                "val": self.settings.tran,
+            },
+            {
+                "key": "defaultIndexRatio",
+                "val": self.settings.indexRatio,
+            },
+            {
+                "key": "defaultProtect",
+                "val": self.settings.protect,
+            },
+        ]

server/voice_changer/RVC/pipeline/Pipeline.py

@@ -18,6 +18,7 @@ from voice_changer.RVC.inferencer.OnnxRVCInferencer import OnnxRVCInferencer
 from voice_changer.RVC.inferencer.OnnxRVCInferencerNono import OnnxRVCInferencerNono
 
 from voice_changer.RVC.pitchExtractor.PitchExtractor import PitchExtractor
+from voice_changer.utils.Timer import Timer
 
 logger = VoiceChangaerLogger.get_instance().getLogger()
 
@@ -89,174 +90,174 @@ class Pipeline(object):
         protect=0.5,
         out_size=None,
     ):
-        # 16000のサンプリングレートで入ってきている。以降この世界は16000で処理。
+        with Timer("main-process") as t:
+            # 16000のサンプリングレートで入ってきている。以降この世界は16000で処理。
+            search_index = self.index is not None and self.big_npy is not None and index_rate != 0
+            # self.t_pad = self.sr * repeat  # 1秒
+            # self.t_pad_tgt = self.targetSR * repeat  # 1秒　出力時のトリミング(モデルのサンプリングで出力される)
+            audio = audio.unsqueeze(0)
 
-        search_index = self.index is not None and self.big_npy is not None and index_rate != 0
+            quality_padding_sec = (repeat * (audio.shape[1] - 1)) / self.sr  # padding(reflect)のサイズは元のサイズより小さい必要がある。
-        # self.t_pad = self.sr * repeat  # 1秒
-        # self.t_pad_tgt = self.targetSR * repeat  # 1秒　出力時のトリミング(モデルのサンプリングで出力される)
-        audio = audio.unsqueeze(0)
 
-        quality_padding_sec = (repeat * (audio.shape[1] - 1)) / self.sr  # padding(reflect)のサイズは元のサイズより小さい必要がある。
+            self.t_pad = round(self.sr * quality_padding_sec)  # 前後に音声を追加
+            self.t_pad_tgt = round(self.targetSR * quality_padding_sec)  # 前後に音声を追加　出力時のトリミング(モデルのサンプリングで出力される)
+            audio_pad = F.pad(audio, (self.t_pad, self.t_pad), mode="reflect").squeeze(0)
+            p_len = audio_pad.shape[0] // self.window
+            sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
 
-        self.t_pad = round(self.sr * quality_padding_sec)  # 前後に音声を追加
+            # RVC QualityがOnのときにはsilence_frontをオフに。
-        self.t_pad_tgt = round(self.targetSR * quality_padding_sec)  # 前後に音声を追加　出力時のトリミング(モデルのサンプリングで出力される)
+            silence_front = silence_front if repeat == 0 else 0
-        audio_pad = F.pad(audio, (self.t_pad, self.t_pad), mode="reflect").squeeze(0)
+            pitchf = pitchf if repeat == 0 else np.zeros(p_len)
-        p_len = audio_pad.shape[0] // self.window
+            out_size = out_size if repeat == 0 else None
-        sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
 
-        # RVC QualityがOnのときにはsilence_frontをオフに。
+            # ピッチ検出
-        silence_front = silence_front if repeat == 0 else 0
-        pitchf = pitchf if repeat == 0 else np.zeros(p_len)
-        out_size = out_size if repeat == 0 else None
-
-        # ピッチ検出
-        try:
-            if if_f0 == 1:
-                pitch, pitchf = self.pitchExtractor.extract(
-                    audio_pad,
-                    pitchf,
-                    f0_up_key,
-                    self.sr,
-                    self.window,
-                    silence_front=silence_front,
-                )
-                # pitch = pitch[:p_len]
-                # pitchf = pitchf[:p_len]
-                pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
-                pitchf = torch.tensor(pitchf, device=self.device, dtype=torch.float).unsqueeze(0)
-            else:
-                pitch = None
-                pitchf = None
-        except IndexError as e:  # NOQA
-            # print(e)
-            # import traceback
-            # traceback.print_exc()
-            raise NotEnoughDataExtimateF0()
-
-        # tensor型調整
-        feats = audio_pad
-        if feats.dim() == 2:  # double channels
-            feats = feats.mean(-1)
-        assert feats.dim() == 1, feats.dim()
-        feats = feats.view(1, -1)
-
-        # embedding
-        with autocast(enabled=self.isHalf):
             try:
-                feats = self.embedder.extractFeatures(feats, embOutputLayer, useFinalProj)
+                if if_f0 == 1:
-                if torch.isnan(feats).all():
+                    pitch, pitchf = self.pitchExtractor.extract(
-                    raise DeviceCannotSupportHalfPrecisionException()
+                        audio_pad,
+                        pitchf,
+                        f0_up_key,
+                        self.sr,
+                        self.window,
+                        silence_front=silence_front,
+                    )
+                    # pitch = pitch[:p_len]
+                    # pitchf = pitchf[:p_len]
+                    pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
+                    pitchf = torch.tensor(pitchf, device=self.device, dtype=torch.float).unsqueeze(0)
+                else:
+                    pitch = None
+                    pitchf = None
+            except IndexError as e:  # NOQA
+                # print(e)
+                # import traceback
+                # traceback.print_exc()
+                raise NotEnoughDataExtimateF0()
+
+            # tensor型調整
+            feats = audio_pad
+            if feats.dim() == 2:  # double channels
+                feats = feats.mean(-1)
+            assert feats.dim() == 1, feats.dim()
+            feats = feats.view(1, -1)
+
+            # embedding
+            with autocast(enabled=self.isHalf):
+                try:
+                    feats = self.embedder.extractFeatures(feats, embOutputLayer, useFinalProj)
+                    if torch.isnan(feats).all():
+                        raise DeviceCannotSupportHalfPrecisionException()
+                except RuntimeError as e:
+                    if "HALF" in e.__str__().upper():
+                        raise HalfPrecisionChangingException()
+                    elif "same device" in e.__str__():
+                        raise DeviceChangingException()
+                    else:
+                        raise e
+
+            # Index - feature抽出
+            # if self.index is not None and self.feature is not None and index_rate != 0:
+            if search_index:
+                npy = feats[0].cpu().numpy()
+                # apply silent front for indexsearch
+                npyOffset = math.floor(silence_front * 16000) // 360
+                npy = npy[npyOffset:]
+
+                if self.isHalf is True:
+                    npy = npy.astype("float32")
+
+                # TODO: kは調整できるようにする
+                k = 1
+                if k == 1:
+                    _, ix = self.index.search(npy, 1)
+                    npy = self.big_npy[ix.squeeze()]
+                else:
+                    score, ix = self.index.search(npy, k=8)
+                    weight = np.square(1 / score)
+                    weight /= weight.sum(axis=1, keepdims=True)
+                    npy = np.sum(self.big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
+
+                # recover silient font
+                npy = np.concatenate([np.zeros([npyOffset, npy.shape[1]], dtype=np.float32), feature[:npyOffset:2].astype("float32"), npy])[-feats.shape[1]:]
+                feats = torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate + (1 - index_rate) * feats
+            feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
+            if protect < 0.5 and search_index:
+                feats0 = feats.clone()
+
+            # ピッチサイズ調整
+            p_len = audio_pad.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:
+                    pitch = pitch[:, :p_len]
+                    pitchf = pitchf[:, :p_len]
+
+            feats_len = feats.shape[1]
+            if pitch is not None and pitchf is not None:
+                pitch = pitch[:, -feats_len:]
+                pitchf = pitchf[:, -feats_len:]
+            p_len = torch.tensor([feats_len], device=self.device).long()
+
+            # pitchの推定が上手くいかない(pitchf=0)場合、検索前の特徴を混ぜる
+            # pitchffの作り方の疑問はあるが、本家通りなので、このまま使うことにする。
+            # https://github.com/w-okada/voice-changer/pull/276#issuecomment-1571336929
+            if protect < 0.5 and search_index:
+                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()
+
+            # apply silent front for inference
+            if type(self.inferencer) in [OnnxRVCInferencer, OnnxRVCInferencerNono]:
+                npyOffset = math.floor(silence_front * 16000) // 360
+                feats = feats[:, npyOffset * 2 :, :]  # NOQA
+
+            feats_len = feats.shape[1]
+            if pitch is not None and pitchf is not None:
+                pitch = pitch[:, -feats_len:]
+                pitchf = pitchf[:, -feats_len:]
+            p_len = torch.tensor([feats_len], device=self.device).long()
+
+            # 推論実行
+            try:
+                with torch.no_grad():
+                    with autocast(enabled=self.isHalf):
+                        audio1 = (
+                            torch.clip(
+                                self.inferencer.infer(feats, p_len, pitch, pitchf, sid, out_size)[0][0, 0].to(dtype=torch.float32),
+                                -1.0,
+                                1.0,
+                            )
+                            * 32767.5
+                        ).data.to(dtype=torch.int16)
             except RuntimeError as e:
                 if "HALF" in e.__str__().upper():
+                    print("11", e)
                     raise HalfPrecisionChangingException()
-                elif "same device" in e.__str__():
-                    raise DeviceChangingException()
                 else:
                     raise e
 
-        # Index - feature抽出
+            feats_buffer = feats.squeeze(0).detach().cpu()
-        # if self.index is not None and self.feature is not None and index_rate != 0:
+            if pitchf is not None:
-        if search_index:
+                pitchf_buffer = pitchf.squeeze(0).detach().cpu()
-            npy = feats[0].cpu().numpy()
-            # apply silent front for indexsearch
-            npyOffset = math.floor(silence_front * 16000) // 360
-            npy = npy[npyOffset:]
-
-            if self.isHalf is True:
-                npy = npy.astype("float32")
-
-            # TODO: kは調整できるようにする
-            k = 1
-            if k == 1:
-                _, ix = self.index.search(npy, 1)
-                npy = self.big_npy[ix.squeeze()]
             else:
-                score, ix = self.index.search(npy, k=8)
+                pitchf_buffer = None
-                weight = np.square(1 / score)
-                weight /= weight.sum(axis=1, keepdims=True)
-                npy = np.sum(self.big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
 
-            # recover silient font
+            del p_len, pitch, pitchf, feats
-            npy = np.concatenate([np.zeros([npyOffset, npy.shape[1]], dtype=np.float32), feature[:npyOffset:2].astype("float32"), npy])[-feats.shape[1]:]
+            # torch.cuda.empty_cache()
-            feats = torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate + (1 - index_rate) * feats
-        feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
-        if protect < 0.5 and search_index:
-            feats0 = feats.clone()
 
-        # ピッチサイズ調整
+            # inferで出力されるサンプリングレートはモデルのサンプリングレートになる。
-        p_len = audio_pad.shape[0] // self.window
+            # pipelineに（入力されるときはhubertように16k）
-        if feats.shape[1] < p_len:
+            if self.t_pad_tgt != 0:
-            p_len = feats.shape[1]
+                offset = self.t_pad_tgt
-            if pitch is not None and pitchf is not None:
+                end = -1 * self.t_pad_tgt
-                pitch = pitch[:, :p_len]
+                audio1 = audio1[offset:end]
-                pitchf = pitchf[:, :p_len]
 
-        feats_len = feats.shape[1]
+            del sid
-        if pitch is not None and pitchf is not None:
+            # torch.cuda.empty_cache()
-            pitch = pitch[:, -feats_len:]
-            pitchf = pitchf[:, -feats_len:]
-        p_len = torch.tensor([feats_len], device=self.device).long()
-
-        # pitchの推定が上手くいかない(pitchf=0)場合、検索前の特徴を混ぜる
-        # pitchffの作り方の疑問はあるが、本家通りなので、このまま使うことにする。
-        # https://github.com/w-okada/voice-changer/pull/276#issuecomment-1571336929
-        if protect < 0.5 and search_index:
-            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()
-
-        # apply silent front for inference
-        if type(self.inferencer) in [OnnxRVCInferencer, OnnxRVCInferencerNono]:
-            npyOffset = math.floor(silence_front * 16000) // 360
-            feats = feats[:, npyOffset * 2 :, :]  # NOQA
-
-        feats_len = feats.shape[1]
-        if pitch is not None and pitchf is not None:
-            pitch = pitch[:, -feats_len:]
-            pitchf = pitchf[:, -feats_len:]
-        p_len = torch.tensor([feats_len], device=self.device).long()
-
-        # 推論実行
-        try:
-            with torch.no_grad():
-                with autocast(enabled=self.isHalf):
-                    audio1 = (
-                        torch.clip(
-                            self.inferencer.infer(feats, p_len, pitch, pitchf, sid, out_size)[0][0, 0].to(dtype=torch.float32),
-                            -1.0,
-                            1.0,
-                        )
-                        * 32767.5
-                    ).data.to(dtype=torch.int16)
-        except RuntimeError as e:
-            if "HALF" in e.__str__().upper():
-                print("11", e)
-                raise HalfPrecisionChangingException()
-            else:
-                raise e
-
-        feats_buffer = feats.squeeze(0).detach().cpu()
-        if pitchf is not None:
-            pitchf_buffer = pitchf.squeeze(0).detach().cpu()
-        else:
-            pitchf_buffer = None
-
-        del p_len, pitch, pitchf, feats
-        # torch.cuda.empty_cache()
-
-        # inferで出力されるサンプリングレートはモデルのサンプリングレートになる。
-        # pipelineに（入力されるときはhubertように16k）
-        if self.t_pad_tgt != 0:
-            offset = self.t_pad_tgt
-            end = -1 * self.t_pad_tgt
-            audio1 = audio1[offset:end]
-
-        del sid
-        # torch.cuda.empty_cache()
         return audio1, pitchf_buffer, feats_buffer
 
     def __del__(self):

server/voice_changer/VoiceChangerManager.py

@@ -214,11 +214,18 @@ class VoiceChangerManager(ServerDeviceCallbacks):
             return
         elif slotInfo.voiceChangerType == "RVC":
             logger.info("................RVC")
-            from voice_changer.RVC.RVC import RVC
+            # from voice_changer.RVC.RVC import RVC
 
-            self.voiceChangerModel = RVC(self.params, slotInfo)
+            # self.voiceChangerModel = RVC(self.params, slotInfo)
-            self.voiceChanger = VoiceChanger(self.params)
+            # self.voiceChanger = VoiceChanger(self.params)
+            # self.voiceChanger.setModel(self.voiceChangerModel)
+
+            from voice_changer.RVC.RVCr2 import RVCr2
+
+            self.voiceChangerModel = RVCr2(self.params, slotInfo)
+            self.voiceChanger = VoiceChangerV2(self.params)
             self.voiceChanger.setModel(self.voiceChangerModel)
+
         elif slotInfo.voiceChangerType == "MMVCv13":
             logger.info("................MMVCv13")
             from voice_changer.MMVCv13.MMVCv13 import MMVCv13

server/voice_changer/VoiceChangerV2.py

@@ -6,7 +6,7 @@
 
 - 適用VoiceChangerModel
 ・DiffusionSVC
-
+・RVC
 '''
 
 from typing import Any, Union