1berry/wokada-voice-changer
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

WIP: DML onnx inferencer

Browse Source
This commit is contained in:
wataru 2023-05-29 17:34:35 +09:00
parent f1683cb4a8
commit d45467a3c4
19 changed files with 113 additions and 194 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
server/voice_changer/MMVCv13/MMVCv13.py
View File

@ -102,10 +102,11 @@ class MMVCv13:
return self.get_info()
def getOnnxExecutionProvider(self):
if self.settings.gpu >= 0:
availableProviders = onnxruntime.get_available_providers()
if self.settings.gpu >= 0 and "CUDAExecutionProvider" in availableProviders:
return ["CUDAExecutionProvider"], [{"device_id": self.settings.gpu}]
elif "DmlExecutionProvider" in onnxruntime.get_available_providers():
return ["DmlExecutionProvider"], []
elif self.settings.gpu >= 0 and "DmlExecutionProvider" in availableProviders:
return ["DmlExecutionProvider"], [{}]
else:
return ["CPUExecutionProvider"], [
{

7
server/voice_changer/MMVCv15/MMVCv15.py
View File

@ -125,10 +125,11 @@ class MMVCv15:
return self.get_info()
def getOnnxExecutionProvider(self):
if self.settings.gpu >= 0:
availableProviders = onnxruntime.get_available_providers()
if self.settings.gpu >= 0 and "CUDAExecutionProvider" in availableProviders:
return ["CUDAExecutionProvider"], [{"device_id": self.settings.gpu}]
elif "DmlExecutionProvider" in onnxruntime.get_available_providers():
return ["DmlExecutionProvider"], []
elif self.settings.gpu >= 0 and "DmlExecutionProvider" in availableProviders:
return ["DmlExecutionProvider"], [{}]
else:
return ["CPUExecutionProvider"], [
{

20
server/voice_changer/RVC/RVC.py
View File

@ -41,13 +41,6 @@ from const import RVC_MAX_SLOT_NUM, RVC_MODEL_DIRNAME, SAMPLES_JSONS, UPLOAD_DIR
import shutil
import json
providers = [
"OpenVINOExecutionProvider",
"CUDAExecutionProvider",
"DmlExecutionProvider",
"CPUExecutionProvider",
]
class RVC:
initialLoad: bool = True
@ -193,19 +186,8 @@ class RVC:
setattr(self.settings, key, val)
if key == "gpu":
dev = self.deviceManager.getDevice(val)
half = self.deviceManager.halfPrecisionAvailable(val)
self.prepareModel(self.settings.modelSlotIndex)
# half-precisionの使用可否が変わるときは作り直し
if self.pipeline is not None and self.pipeline.isHalf == half:
print(
"USE EXISTING PIPELINE",
half,
)
self.pipeline.setDevice(dev)
else:
print("CHAGE TO NEW PIPELINE", half)
self.prepareModel(self.settings.modelSlotIndex)
if key == "enableDirectML":
if self.pipeline is not None and val == 0:
self.pipeline.setDirectMLEnable(False)

16
server/voice_changer/RVC/deviceManager/DeviceManager.py
View File

@ -1,4 +1,5 @@
import torch
import onnxruntime
class DeviceManager(object):
@ -26,6 +27,21 @@ class DeviceManager(object):
dev = torch.device("cuda", index=id)
return dev
def getOnnxExecutionProvider(self, gpu: int):
availableProviders = onnxruntime.get_available_providers()
if gpu >= 0 and "CUDAExecutionProvider" in availableProviders:
return ["CUDAExecutionProvider"], [{"device_id": gpu}]
elif gpu >= 0 and "DmlExecutionProvider" in availableProviders:
return ["DmlExecutionProvider"], [{}]
else:
return ["CPUExecutionProvider"], [
{
"intra_op_num_threads": 8,
"execution_mode": onnxruntime.ExecutionMode.ORT_PARALLEL,
"inter_op_num_threads": 8,
}
]
def halfPrecisionAvailable(self, id: int):
if self.gpu_num == 0:
return False

56
server/voice_changer/RVC/inferencer/Inferencer.py
View File

@ -1,21 +1,18 @@
from typing import Any, Protocol
import torch
from torch import device
from const import EnumInferenceTypes
import onnxruntime
class Inferencer(Protocol):
inferencerType: EnumInferenceTypes = EnumInferenceTypes.pyTorchRVC
file: str
isHalf: bool = True
dev: device
# inferencerType: EnumInferenceTypes = EnumInferenceTypes.pyTorchRVC
# file: str
# isHalf: bool = True
# dev: device | None
# onnxProviders: list[str] | None
# onnxProviderOptions: Any | None
model: onnxruntime.InferenceSession | Any | None = None
def loadModel(self, file: str, dev: device, isHalf: bool = True):
def loadModel(self, file: str, gpu: int):
...
def infer(
@ -28,27 +25,18 @@ class Inferencer(Protocol):
) -> torch.Tensor:
...
def setProps(
self,
inferencerType: EnumInferenceTypes,
file: str,
dev: device,
isHalf: bool = True,
):
self.inferencerType = inferencerType
self.file = file
self.isHalf = isHalf
self.dev = dev
def setHalf(self, isHalf: bool):
self.isHalf = isHalf
if self.model is not None and isHalf:
self.model = self.model.half()
elif self.model is not None and isHalf is False:
self.model = self.model.float()
def setDevice(self, dev: device):
self.dev = dev
if self.model is not None:
self.model = self.model.to(self.dev)
return self
# def setProps(
# self,
# inferencerType: EnumInferenceTypes,
# file: str,
# dev: device | None,
# onnxProviders: list[str] | None,
# onnxProviderOptions: Any | None,
# isHalf: bool = True,
# ):
# self.inferencerType = inferencerType
# self.file = file
# self.isHalf = isHalf
# self.dev = dev
# self.onnxProviders = onnxProviders
# self.onnxProviderOptions = onnxProviderOptions

30
server/voice_changer/RVC/inferencer/InferencerManager.py
View File

@ -1,5 +1,3 @@
from torch import device
from const import EnumInferenceTypes
from voice_changer.RVC.inferencer.Inferencer import Inferencer
from voice_changer.RVC.inferencer.OnnxRVCInferencer import OnnxRVCInferencer
@ -17,54 +15,60 @@ class InferencerManager:
@classmethod
def getInferencer(
cls, inferencerType: EnumInferenceTypes, file: str, isHalf: bool, dev: device
cls,
inferencerType: EnumInferenceTypes,
file: str,
gpu: int,
) -> Inferencer:
cls.currentInferencer = cls.loadInferencer(inferencerType, file, isHalf, dev)
cls.currentInferencer = cls.loadInferencer(inferencerType, file, gpu)
return cls.currentInferencer
@classmethod
def loadInferencer(
cls, inferencerType: EnumInferenceTypes, file: str, isHalf: bool, dev: device
cls,
inferencerType: EnumInferenceTypes,
file: str,
gpu: int,
) -> Inferencer:
if (
inferencerType == EnumInferenceTypes.pyTorchRVC
or inferencerType == EnumInferenceTypes.pyTorchRVC.value
):
return RVCInferencer().loadModel(file, dev, isHalf)
return RVCInferencer().loadModel(file, gpu)
elif (
inferencerType == EnumInferenceTypes.pyTorchRVCNono
or inferencerType == EnumInferenceTypes.pyTorchRVCNono.value
):
return RVCInferencerNono().loadModel(file, dev, isHalf)
return RVCInferencerNono().loadModel(file, gpu)
elif (
inferencerType == EnumInferenceTypes.pyTorchRVCv2
or inferencerType == EnumInferenceTypes.pyTorchRVCv2.value
):
return RVCInferencerv2().loadModel(file, dev, isHalf)
return RVCInferencerv2().loadModel(file, gpu)
elif (
inferencerType == EnumInferenceTypes.pyTorchRVCv2Nono
or inferencerType == EnumInferenceTypes.pyTorchRVCv2Nono.value
):
return RVCInferencerv2Nono().loadModel(file, dev, isHalf)
return RVCInferencerv2Nono().loadModel(file, gpu)
elif (
inferencerType == EnumInferenceTypes.pyTorchWebUI
or inferencerType == EnumInferenceTypes.pyTorchWebUI.value
):
return WebUIInferencer().loadModel(file, dev, isHalf)
return WebUIInferencer().loadModel(file, gpu)
elif (
inferencerType == EnumInferenceTypes.pyTorchWebUINono
or inferencerType == EnumInferenceTypes.pyTorchWebUINono.value
):
return WebUIInferencerNono().loadModel(file, dev, isHalf)
return WebUIInferencerNono().loadModel(file, gpu)
elif (
inferencerType == EnumInferenceTypes.onnxRVC
or inferencerType == EnumInferenceTypes.onnxRVC.value
):
return OnnxRVCInferencer().loadModel(file, dev, isHalf)
return OnnxRVCInferencer().loadModel(file, gpu)
elif (
inferencerType == EnumInferenceTypes.onnxRVCNono
or inferencerType == EnumInferenceTypes.onnxRVCNono.value
):
return OnnxRVCInferencerNono().loadModel(file, dev, isHalf)
return OnnxRVCInferencerNono().loadModel(file, gpu)
else:
raise RuntimeError("[Voice Changer] Inferencer not found", inferencerType)

52
server/voice_changer/RVC/inferencer/OnnxRVCInferencer.py
View File

@ -1,20 +1,20 @@
import torch
from torch import device
import onnxruntime
from const import EnumInferenceTypes
from voice_changer.RVC.deviceManager.DeviceManager import DeviceManager
from voice_changer.RVC.inferencer.Inferencer import Inferencer
import numpy as np
providers = ["CPUExecutionProvider"]
class OnnxRVCInferencer(Inferencer):
def loadModel(self, file: str, dev: device, isHalf: bool = True):
super().setProps(EnumInferenceTypes.onnxRVC, file, dev, isHalf)
# ort_options = onnxruntime.SessionOptions()
# ort_options.intra_op_num_threads = 8
def loadModel(self, file: str, gpu: int):
(
onnxProviders,
onnxProviderOptions,
) = DeviceManager.get_instance().getOnnxExecutionProvider(gpu)
onnx_session = onnxruntime.InferenceSession(file, providers=providers)
onnx_session = onnxruntime.InferenceSession(
file, providers=onnxProviders, provider_options=onnxProviderOptions
)
# check half-precision
first_input_type = onnx_session.get_inputs()[0].type
@ -24,7 +24,6 @@ class OnnxRVCInferencer(Inferencer):
self.isHalf = True
self.model = onnx_session
self.setDevice(dev)
return self
def infer(
@ -66,36 +65,3 @@ class OnnxRVCInferencer(Inferencer):
)
return torch.tensor(np.array(audio1))
def setHalf(self, isHalf: bool):
self.isHalf = isHalf
pass
# raise RuntimeError("half-precision is not changable.", self.isHalf)
def setDevice(self, dev: device):
index = dev.index
type = dev.type
if type == "cpu":
self.model.set_providers(providers=["CPUExecutionProvider"])
elif type == "cuda":
provider_options = [{"device_id": index}]
self.model.set_providers(
providers=["CUDAExecutionProvider"],
provider_options=provider_options,
)
else:
self.model.set_providers(providers=["CPUExecutionProvider"])
return self
def setDirectMLEnable(self, enable: bool):
if "DmlExecutionProvider" not in onnxruntime.get_available_providers():
print("[Voice Changer] DML is not available.")
return
if enable:
self.model.set_providers(
providers=["DmlExecutionProvider", "CPUExecutionProvider"]
)
else:
self.model.set_providers(providers=["CPUExecutionProvider"])

22
server/voice_changer/RVC/inferencer/OnnxRVCInferencerNono.py
View File

@ -1,32 +1,10 @@
import torch
from torch import device
import onnxruntime
from const import EnumInferenceTypes
import numpy as np
from voice_changer.RVC.inferencer.OnnxRVCInferencer import OnnxRVCInferencer
providers = ["CPUExecutionProvider"]
class OnnxRVCInferencerNono(OnnxRVCInferencer):
def loadModel(self, file: str, dev: device, isHalf: bool = True):
super().setProps(EnumInferenceTypes.onnxRVC, file, dev, isHalf)
# ort_options = onnxruntime.SessionOptions()
# ort_options.intra_op_num_threads = 8
onnx_session = onnxruntime.InferenceSession(file, providers=providers)
# check half-precision
first_input_type = onnx_session.get_inputs()[0].type
if first_input_type == "tensor(float)":
self.isHalf = False
else:
self.isHalf = True
self.model = onnx_session
return self
def infer(
self,
feats: torch.Tensor,

10
server/voice_changer/RVC/inferencer/RVCInferencer.py
View File

@ -1,7 +1,6 @@
import torch
from torch import device
from const import EnumInferenceTypes
from voice_changer.RVC.deviceManager.DeviceManager import DeviceManager
from voice_changer.RVC.inferencer.Inferencer import Inferencer
from infer_pack.models import ( # type:ignore
SynthesizerTrnMs256NSFsid,
@ -9,9 +8,10 @@ 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)
def loadModel(self, file: str, gpu: int):
dev = DeviceManager.get_instance().getDevice(gpu)
isHalf = DeviceManager.get_instance().halfPrecisionAvailable(gpu)
cpt = torch.load(file, map_location="cpu")
model = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=isHalf)

9
server/voice_changer/RVC/inferencer/RVCInferencerNono.py
View File

@ -1,7 +1,6 @@
import torch
from torch import device
from const import EnumInferenceTypes
from voice_changer.RVC.deviceManager.DeviceManager import DeviceManager
from voice_changer.RVC.inferencer.Inferencer import Inferencer
from infer_pack.models import ( # type:ignore
SynthesizerTrnMs256NSFsid_nono,
@ -9,8 +8,10 @@ from infer_pack.models import ( # type:ignore
class RVCInferencerNono(Inferencer):
def loadModel(self, file: str, dev: device, isHalf: bool = True):
super().setProps(EnumInferenceTypes.pyTorchRVCNono, file, dev, isHalf)
def loadModel(self, file: str, gpu: int):
dev = DeviceManager.get_instance().getDevice(gpu)
isHalf = DeviceManager.get_instance().halfPrecisionAvailable(gpu)
cpt = torch.load(file, map_location="cpu")
model = SynthesizerTrnMs256NSFsid_nono(*cpt["config"], is_half=isHalf)

11
server/voice_changer/RVC/inferencer/RVCInferencerv2.py
View File

@ -1,7 +1,5 @@
import torch
from torch import device
from const import EnumInferenceTypes
from voice_changer.RVC.deviceManager.DeviceManager import DeviceManager
from voice_changer.RVC.inferencer.Inferencer import Inferencer
from infer_pack.models import ( # type:ignore
SynthesizerTrnMs768NSFsid,
@ -9,9 +7,10 @@ from infer_pack.models import ( # type:ignore
class RVCInferencerv2(Inferencer):
def loadModel(self, file: str, dev: device, isHalf: bool = True):
super().setProps(EnumInferenceTypes.pyTorchRVCv2, file, dev, isHalf)
print("load inf", file)
def loadModel(self, file: str, gpu: int):
dev = DeviceManager.get_instance().getDevice(gpu)
isHalf = DeviceManager.get_instance().halfPrecisionAvailable(gpu)
cpt = torch.load(file, map_location="cpu")
model = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=isHalf)

9
server/voice_changer/RVC/inferencer/RVCInferencerv2Nono.py
View File

@ -1,7 +1,6 @@
import torch
from torch import device
from const import EnumInferenceTypes
from voice_changer.RVC.deviceManager.DeviceManager import DeviceManager
from voice_changer.RVC.inferencer.Inferencer import Inferencer
from infer_pack.models import ( # type:ignore
SynthesizerTrnMs768NSFsid_nono,
@ -9,8 +8,10 @@ from infer_pack.models import ( # type:ignore
class RVCInferencerv2Nono(Inferencer):
def loadModel(self, file: str, dev: device, isHalf: bool = True):
super().setProps(EnumInferenceTypes.pyTorchRVCv2Nono, file, dev, isHalf)
def loadModel(self, file: str, gpu: int):
dev = DeviceManager.get_instance().getDevice(gpu)
isHalf = DeviceManager.get_instance().halfPrecisionAvailable(gpu)
cpt = torch.load(file, map_location="cpu")
model = SynthesizerTrnMs768NSFsid_nono(*cpt["config"], is_half=isHalf)

9
server/voice_changer/RVC/inferencer/WebUIInferencer.py
View File

@ -1,14 +1,15 @@
import torch
from torch import device
from voice_changer.RVC.deviceManager.DeviceManager import DeviceManager
from const import EnumInferenceTypes
from voice_changer.RVC.inferencer.Inferencer import Inferencer
from .models import SynthesizerTrnMsNSFsid
class WebUIInferencer(Inferencer):
def loadModel(self, file: str, dev: device, isHalf: bool = True):
super().setProps(EnumInferenceTypes.pyTorchWebUI, file, dev, isHalf)
def loadModel(self, file: str, gpu: int):
dev = DeviceManager.get_instance().getDevice(gpu)
isHalf = DeviceManager.get_instance().halfPrecisionAvailable(gpu)
cpt = torch.load(file, map_location="cpu")
model = SynthesizerTrnMsNSFsid(**cpt["params"], is_half=isHalf)

9
server/voice_changer/RVC/inferencer/WebUIInferencerNono.py
View File

@ -1,14 +1,15 @@
import torch
from torch import device
from const import EnumInferenceTypes
from voice_changer.RVC.deviceManager.DeviceManager import DeviceManager
from voice_changer.RVC.inferencer.Inferencer import Inferencer
from .models import SynthesizerTrnMsNSFsidNono
class WebUIInferencerNono(Inferencer):
def loadModel(self, file: str, dev: device, isHalf: bool = True):
super().setProps(EnumInferenceTypes.pyTorchWebUINono, file, dev, isHalf)
def loadModel(self, file: str, gpu: int):
dev = DeviceManager.get_instance().getDevice(gpu)
isHalf = DeviceManager.get_instance().halfPrecisionAvailable(gpu)
cpt = torch.load(file, map_location="cpu")
model = SynthesizerTrnMsNSFsidNono(**cpt["params"], is_half=isHalf)

8
server/voice_changer/RVC/pipeline/Pipeline.py
View File

@ -59,14 +59,6 @@ class Pipeline(object):
self.sr = 16000
self.window = 160
self.device = device
self.isHalf = isHalf
def setDevice(self, device: torch.device):
self.device = device
self.embedder.setDevice(device)
self.inferencer.setDevice(device)
def setDirectMLEnable(self, enable: bool):
if hasattr(self.inferencer, "setDirectMLEnable"):
self.inferencer.setDirectMLEnable(enable)

9
server/voice_changer/RVC/pipeline/PipelineGenerator.py
View File

@ -13,18 +13,11 @@ from voice_changer.RVC.pitchExtractor.PitchExtractorManager import PitchExtracto
def createPipeline(modelSlot: ModelSlot, gpu: int, f0Detector: str):
dev = DeviceManager.get_instance().getDevice(gpu)
half = DeviceManager.get_instance().halfPrecisionAvailable(gpu)
# # ファイル名特定(Inferencer)
# inferencerFilename = (
# modelSlot.onnxModelFile if modelSlot.isONNX else modelSlot.pyTorchModelFile
# )
# Inferencer 生成
try:
inferencer = InferencerManager.getInferencer(
modelSlot.modelType,
modelSlot.modelFile,
half,
dev,
modelSlot.modelType, modelSlot.modelFile, gpu
)
except Exception as e:
print("[Voice Changer] exception! loading inferencer", e)

7
server/voice_changer/SoVitsSvc40/SoVitsSvc40.py
View File

@ -162,10 +162,11 @@ class SoVitsSvc40:
return self.get_info()
def getOnnxExecutionProvider(self):
if self.settings.gpu >= 0:
availableProviders = onnxruntime.get_available_providers()
if self.settings.gpu >= 0 and "CUDAExecutionProvider" in availableProviders:
return ["CUDAExecutionProvider"], [{"device_id": self.settings.gpu}]
elif "DmlExecutionProvider" in onnxruntime.get_available_providers():
return ["DmlExecutionProvider"], []
elif self.settings.gpu >= 0 and "DmlExecutionProvider" in availableProviders:
return ["DmlExecutionProvider"], [{}]
else:
return ["CPUExecutionProvider"], [
{

7
server/voice_changer/SoVitsSvc40v2/SoVitsSvc40v2.py
View File

@ -139,10 +139,11 @@ class SoVitsSvc40v2:
return self.get_info()
def getOnnxExecutionProvider(self):
if self.settings.gpu >= 0:
availableProviders = onnxruntime.get_available_providers()
if self.settings.gpu >= 0 and "CUDAExecutionProvider" in availableProviders:
return ["CUDAExecutionProvider"], [{"device_id": self.settings.gpu}]
elif "DmlExecutionProvider" in onnxruntime.get_available_providers():
return ["DmlExecutionProvider"], []
elif self.settings.gpu >= 0 and "DmlExecutionProvider" in availableProviders:
return ["DmlExecutionProvider"], [{}]
else:
return ["CPUExecutionProvider"], [
{

9
server/voice_changer/VoiceChanger.py
View File

@ -29,13 +29,6 @@ import time
import sounddevice as sd
import librosa
providers = [
"OpenVINOExecutionProvider",
"CUDAExecutionProvider",
"DmlExecutionProvider",
"CPUExecutionProvider",
]
STREAM_INPUT_FILE = os.path.join(TMP_DIR, "in.wav")
STREAM_OUTPUT_FILE = os.path.join(TMP_DIR, "out.wav")
@ -379,7 +372,7 @@ class VoiceChanger:
else:
ret = self.voiceChanger.update_settings(key, val)
if ret is False:
print(f"{key} is not mutable variable or unknown variable!")
print(f"({key} is not mutable variable or unknown variable)")
return self.get_info()
def _generate_strength(self, crossfadeSize: int):