update

server/restapi/MMVC_Rest_Fileuploader.py

@@ -63,7 +63,7 @@ class MMVC_Rest_Fileuploader:
     def post_update_settings(
         self, key: str = Form(...), val: Union[int, str, float] = Form(...)
     ):
-        print("post_update_settings", key, val)
+        print("[Voice Changer] update configuration:", key, val)
         info = self.voiceChangerManager.update_settings(key, val)
         json_compatible_item_data = jsonable_encoder(info)
         return JSONResponse(content=json_compatible_item_data)

server/voice_changer/RVC/ModelSlotGenerator.py

@@ -70,7 +70,7 @@ def _setInfoByONNX(slot: ModelSlot, file: str):
     try:
         metadata = json.loads(modelmeta.custom_metadata_map["metadata"])
 
-        slot.modelType = metadata["modelType"]
+        # slot.modelType = metadata["modelType"]
         slot.embChannels = metadata["embChannels"]
 
         if "embedder" not in metadata:

server/voice_changer/RVC/RVC.py

@@ -1,5 +1,6 @@
 import sys
 import os
+from Exceptions import NoModeLoadedException
 from voice_changer.RVC.ModelSlot import ModelSlot
 from voice_changer.RVC.deviceManager.DeviceManager import DeviceManager
 
@@ -42,7 +43,7 @@ import torch
 import traceback
 import faiss
 
-from const import TMP_DIR, UPLOAD_DIR, EnumEmbedderTypes
+from const import UPLOAD_DIR, EnumEmbedderTypes
 
 
 from voice_changer.RVC.custom_vc_infer_pipeline import VC
@@ -89,9 +90,6 @@ class RVC:
             f"[Voice Changer] RVC new model is uploaded,{target_slot_idx}",
             asdict(modelSlot),
         )
-        """
-        [Voice Changer] RVC new model is uploaded,0 {'pyTorchModelFile': 'upload_dir/0/kurage.pth', 'onnxModelFile': None, 'featureFile': None, 'indexFile': None, 'defaultTrans': 16, 'isONNX': False, 'modelType': <EnumInferenceTypes.pyTorchWebUI: 'pyTorchWebUI'>, 'samplingRate': 48000, 'f0': True, 'embChannels': 768, 'deprecated': False, 'embedder': 'hubert-base-japanese'}        
-        """
 
         # 初回のみロード
         if self.initialLoad:
@@ -139,7 +137,6 @@ class RVC:
 
         # Embedder 生成
         try:
-            print("AFASFDAFDAFDASDFASDFSADFASDFA", half, self.settings.gpu)
             embedder = EmbedderManager.getEmbedder(
                 modelSlot.embedder,
                 emmbedderFilename,
@@ -179,8 +176,14 @@ class RVC:
     def prepareModel(self, slot: int):
         if slot < 0:
             return self.get_info()
-        print("[Voice Changer] Prepare Model of slot:", slot)
         modelSlot = self.settings.modelSlots[slot]
+        inferencerFilename = (
+            modelSlot.onnxModelFile if modelSlot.isONNX else modelSlot.pyTorchModelFile
+        )
+        if inferencerFilename == "":
+            return self.get_info()
+
+        print("[Voice Changer] Prepare Model of slot:", slot)
 
         # Inferencer, embedderのロード
         inferencer, embedder = self.createPipeline(modelSlot)
@@ -240,13 +243,13 @@ class RVC:
                     and self.embedder.isHalf == half
                 ):
                     print(
-                        "NOT NEED CHAGE TO NEW PIPELINE!!!!!!!!!!!!!!!!!!!!!!!!!!!",
+                        "USE EXSISTING PIPELINE",
                         half,
                     )
                     self.embedder.setDevice(dev)
                     self.inferencer.setDevice(dev)
                 else:
-                    print("CHAGE TO NEW PIPELINE!!!!!!!!!!!!!!!!!!!!!!!!!!!", half)
+                    print("CHAGE TO NEW PIPELINE", half)
                     self.prepareModel(self.settings.modelSlotIndex)
         elif key in self.settings.floatData:
             setattr(self.settings, key, float(val))
@@ -307,7 +310,9 @@ class RVC:
             )
             raise NoModeLoadedException("model_common")
         if self.needSwitch:
-            print(f"Switch model {self.currentSlot} -> {self.settings.modelSlotIndex}")
+            print(
+                f"[Voice Changer] Switch model {self.currentSlot} -> {self.settings.modelSlotIndex}"
+            )
             self.currentSlot = self.settings.modelSlotIndex
             self.switchModel()
             self.needSwitch = False
@@ -363,8 +368,8 @@ class RVC:
         return result
 
     def __del__(self):
-        del self.net_g
+        del self.inferencer
-        del self.onnx_session
+        del self.embedder
 
         print("---------- REMOVING ---------------")
 
@@ -383,57 +388,17 @@ class RVC:
                 pass
 
     def export2onnx(self):
-        if hasattr(self, "net_g") is False or self.net_g is None:
+        modelSlot = self.settings.modelSlots[self.settings.modelSlotIndex]
-            print("[Voice Changer] export2onnx, No pyTorch session.")
+        pyTorchModelFile = modelSlot.pyTorchModelFile
-            return {"status": "ng", "path": ""}
 
-        pyTorchModelFile = self.settings.modelSlots[
+        # PyTorchのファイルが存在しない場合はエラーを返す
-            self.settings.modelSlotIndex
+        if pyTorchModelFile is None or pyTorchModelFile == "":
-        ].pyTorchModelFile  # inference前にexportできるようにcurrentSlotではなくslot
-
-        if pyTorchModelFile is None:
             print("[Voice Changer] export2onnx, No pyTorch filepath.")
             return {"status": "ng", "path": ""}
+
         import voice_changer.RVC.export2onnx as onnxExporter
 
-        output_file = os.path.splitext(os.path.basename(pyTorchModelFile))[0] + ".onnx"
+        output_file_simple = onnxExporter.export2onnx(modelSlot)
-        output_file_simple = (
-            os.path.splitext(os.path.basename(pyTorchModelFile))[0] + "_simple.onnx"
-        )
-        output_path = os.path.join(TMP_DIR, output_file)
-        output_path_simple = os.path.join(TMP_DIR, output_file_simple)
-        print(
-            "embChannels",
-            self.settings.modelSlots[self.settings.modelSlotIndex].embChannels,
-        )
-        metadata = {
-            "application": "VC_CLIENT",
-            "version": "1",
-            "modelType": self.settings.modelSlots[
-                self.settings.modelSlotIndex
-            ].modelType,
-            "samplingRate": self.settings.modelSlots[
-                self.settings.modelSlotIndex
-            ].samplingRate,
-            "f0": self.settings.modelSlots[self.settings.modelSlotIndex].f0,
-            "embChannels": self.settings.modelSlots[
-                self.settings.modelSlotIndex
-            ].embChannels,
-            "embedder": self.settings.modelSlots[self.settings.modelSlotIndex].embedder,
-        }
-
-        if torch.cuda.device_count() > 0:
-            onnxExporter.export2onnx(
-                pyTorchModelFile, output_path, output_path_simple, True, metadata
-            )
-        else:
-            print(
-                "[Voice Changer] Warning!!! onnx export with float32. maybe size is doubled."
-            )
-            onnxExporter.export2onnx(
-                pyTorchModelFile, output_path, output_path_simple, False, metadata
-            )
-
         return {
             "status": "ok",
             "path": f"/tmp/{output_file_simple}",

server/voice_changer/RVC/custom_vc_infer_pipeline.py

@@ -55,7 +55,7 @@ class VC(object):
                 pitchf, device=self.device, dtype=torch.float
             ).unsqueeze(0)
 
-        # tensor
+        # tensor型調整
         feats = torch.from_numpy(audio_pad)
         if self.is_half is True:
             feats = feats.half()
@@ -94,11 +94,9 @@ class VC(object):
                 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)
 
-        # ピッチ抽出
+        # 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]
@@ -108,47 +106,23 @@ class VC(object):
         p_len = torch.tensor([p_len], device=self.device).long()
 
         # 推論実行
-        with torch.no_grad():
+        try:
-            audio1 = (
+            with torch.no_grad():
-                (inferencer.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0] * 32768)
+                audio1 = (
-                .data.cpu()
+                    (
-                .float()
+                        inferencer.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0]
-                .numpy()
+                        * 32768
-                .astype(np.int16)
+                    )
-            )
+                    .data.cpu()
-
+                    .float()
-            # if pitch is not None:
+                    .numpy()
-            #     print("INFERENCE 1 ")
+                    .astype(np.int16)
-            #     audio1 = (
+                )
-            #         (
+        except RuntimeError as e:
-            #             inferencer.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0]
+            if "HALF" in e.__str__().upper():
-            #             * 32768
+                raise HalfPrecisionChangingException()
-            #         )
+            else:
-            #         .data.cpu()
+                raise e
-            #         .float()
-            #         .numpy()
-            #         .astype(np.int16)
-            #     )
-            # else:
-            #     if hasattr(inferencer, "infer_pitchless"):
-            #         print("INFERENCE 2 ")
-
-            #         audio1 = (
-            #             (inferencer.infer_pitchless(feats, p_len, sid)[0][0, 0] * 32768)
-            #             .data.cpu()
-            #             .float()
-            #             .numpy()
-            #             .astype(np.int16)
-            #         )
-            #     else:
-            #         print("INFERENCE 3 ")
-            #         audio1 = (
-            #             (inferencer.infer(feats, p_len, sid)[0][0, 0] * 32768)
-            #             .data.cpu()
-            #             .float()
-            #             .numpy()
-            #             .astype(np.int16)
-            #         )
 
         del feats, p_len, padding_mask
         torch.cuda.empty_cache()

server/voice_changer/RVC/export2onnx.py

@@ -1,7 +1,10 @@
+import os
 import json
 import torch
 from onnxsim import simplify
 import onnx
+from const import TMP_DIR, EnumInferenceTypes
+from voice_changer.RVC.ModelSlot import ModelSlot
 
 from voice_changer.RVC.onnx.SynthesizerTrnMs256NSFsid_ONNX import (
     SynthesizerTrnMs256NSFsid_ONNX,
@@ -15,24 +18,60 @@ from voice_changer.RVC.onnx.SynthesizerTrnMsNSFsidNono_webui_ONNX import (
 from voice_changer.RVC.onnx.SynthesizerTrnMsNSFsid_webui_ONNX import (
     SynthesizerTrnMsNSFsid_webui_ONNX,
 )
-from .const import RVC_MODEL_TYPE_RVC, RVC_MODEL_TYPE_WEBUI
 
 
-def export2onnx(input_model, output_model, output_model_simple, is_half, metadata):
+def export2onnx(modelSlot: ModelSlot):
+    pyTorchModelFile = modelSlot.pyTorchModelFile
+
+    output_file = os.path.splitext(os.path.basename(pyTorchModelFile))[0] + ".onnx"
+    output_file_simple = (
+        os.path.splitext(os.path.basename(pyTorchModelFile))[0] + "_simple.onnx"
+    )
+    output_path = os.path.join(TMP_DIR, output_file)
+    output_path_simple = os.path.join(TMP_DIR, output_file_simple)
+    metadata = {
+        "application": "VC_CLIENT",
+        "version": "2",
+        # ↓EnumInferenceTypesのままだとシリアライズできないのでテキスト化
+        "modelType": modelSlot.modelType.value,
+        "samplingRate": modelSlot.samplingRate,
+        "f0": modelSlot.f0,
+        "embChannels": modelSlot.embChannels,
+        "embedder": modelSlot.embedder.value,
+    }
+
+    if torch.cuda.device_count() > 0:
+        _export2onnx(pyTorchModelFile, output_path, output_path_simple, True, metadata)
+    else:
+        print(
+            "[Voice Changer] Warning!!! onnx export with float32. maybe size is doubled."
+        )
+        _export2onnx(pyTorchModelFile, output_path, output_path_simple, False, metadata)
+    return output_file_simple
+
+
+def _export2onnx(input_model, output_model, output_model_simple, is_half, metadata):
     cpt = torch.load(input_model, map_location="cpu")
     if is_half:
         dev = torch.device("cuda", index=0)
     else:
         dev = torch.device("cpu")
 
-    if metadata["f0"] is True and metadata["modelType"] == RVC_MODEL_TYPE_RVC:
+    # EnumInferenceTypesのままだとシリアライズできないのでテキスト化
+    if metadata["modelType"] == EnumInferenceTypes.pyTorchRVC.value:
         net_g_onnx = SynthesizerTrnMs256NSFsid_ONNX(*cpt["config"], is_half=is_half)
-    elif metadata["f0"] is True and metadata["modelType"] == RVC_MODEL_TYPE_WEBUI:
+    elif metadata["modelType"] == EnumInferenceTypes.pyTorchWebUI.value:
         net_g_onnx = SynthesizerTrnMsNSFsid_webui_ONNX(**cpt["params"], is_half=is_half)
-    elif metadata["f0"] is False and metadata["modelType"] == RVC_MODEL_TYPE_RVC:
+    elif metadata["modelType"] == EnumInferenceTypes.pyTorchRVCNono.value:
         net_g_onnx = SynthesizerTrnMs256NSFsid_nono_ONNX(*cpt["config"])
-    elif metadata["f0"] is False and metadata["modelType"] == RVC_MODEL_TYPE_WEBUI:
+    elif metadata["modelType"] == EnumInferenceTypes.pyTorchWebUINono.value:
         net_g_onnx = SynthesizerTrnMsNSFsidNono_webui_ONNX(**cpt["params"])
+    else:
+        print(
+            "unknwon::::: ",
+            metadata["modelType"],
+            EnumInferenceTypes.pyTorchWebUI.value,
+        )
 
     net_g_onnx.eval().to(dev)
     net_g_onnx.load_state_dict(cpt["weight"], strict=False)

server/voice_changer/RVC/inferencer/OnnxRVCInferencer.py

@@ -24,6 +24,7 @@ class OnnxRVCInferencer(Inferencer):
             self.isHalf = True
 
         self.model = onnx_session
+        self.setDevice(dev)
         return self
 
     def infer(
@@ -37,9 +38,9 @@ class OnnxRVCInferencer(Inferencer):
         if pitch is None or pitchf is None:
             raise RuntimeError("[Voice Changer] Pitch or Pitchf is not found.")
 
-        print("INFER1", self.model.get_providers())
+        # print("INFER1", self.model.get_providers())
-        print("INFER2", self.model.get_provider_options())
+        # print("INFER2", self.model.get_provider_options())
-        print("INFER3", self.model.get_session_options())
+        # print("INFER3", self.model.get_session_options())
         if self.isHalf:
             audio1 = self.model.run(
                 ["audio"],

server/voice_changer/RVC/inferencer/RVCInferencer.py

@@ -11,6 +11,7 @@ from infer_pack.models import (  # type:ignore
 class RVCInferencer(Inferencer):
     def loadModel(self, file: str, dev: device, isHalf: bool = True):
         super().setProps(EnumInferenceTypes.pyTorchRVC, file, dev, isHalf)
+        print("load inf", file)
         cpt = torch.load(file, map_location="cpu")
         model = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=isHalf)