import sys sys.path.append("MMVC_Client/python") import os from dataclasses import dataclass, asdict import numpy as np import torch import onnxruntime import pyworld as pw from voice_changer.client_modules import convert_continuos_f0, spectrogram_torch, TextAudioSpeakerCollate, get_hparams_from_file, load_checkpoint from models import SynthesizerTrn from const import ERROR_NO_ONNX_SESSION, TMP_DIR providers = ['OpenVINOExecutionProvider', "CUDAExecutionProvider", "DmlExecutionProvider", "CPUExecutionProvider"] @dataclass class MMVCv15Settings(): gpu: int = 0 srcId: int = 0 dstId: int = 101 inputSampleRate: int = 24000 # 48000 or 24000 crossFadeOffsetRate: float = 0.1 crossFadeEndRate: float = 0.9 crossFadeOverlapSize: int = 4096 f0Factor: float = 1.0 f0Detector: str = "dio" # dio or harvest recordIO: int = 0 # 0:off, 1:on framework: str = "PyTorch" # PyTorch or ONNX pyTorchModelFile: str = "" onnxModelFile: str = "" configFile: str = "" # ↓mutableな物だけ列挙 intData = ["gpu", "srcId", "dstId", "inputSampleRate", "crossFadeOverlapSize", "recordIO"] floatData = ["crossFadeOffsetRate", "crossFadeEndRate", "f0Factor"] strData = ["framework", "f0Detector"] class MMVCv15: def __init__(self): # 初期化 self.settings = MMVCv15Settings() self.net_g = None self.onnx_session = None self.gpu_num = torch.cuda.device_count() self.text_norm = torch.LongTensor([0, 6, 0]) self.audio_buffer = torch.zeros(1, 0) self.mps_enabled = getattr(torch.backends, "mps", None) is not None and torch.backends.mps.is_available() print(f"VoiceChanger Initialized (GPU_NUM:{self.gpu_num}, mps_enabled:{self.mps_enabled})") def loadModel(self, config: str, pyTorch_model_file: str = None, onnx_model_file: str = None): self.settings.configFile = config self.hps = get_hparams_from_file(config) if pyTorch_model_file != None: self.settings.pyTorchModelFile = pyTorch_model_file if onnx_model_file: self.settings.onnxModelFile = onnx_model_file # PyTorchモデル生成 if pyTorch_model_file != None: self.net_g = SynthesizerTrn( spec_channels=self.hps.data.filter_length // 2 + 1, segment_size=self.hps.train.segment_size // self.hps.data.hop_length, inter_channels=self.hps.model.inter_channels, hidden_channels=self.hps.model.hidden_channels, upsample_rates=self.hps.model.upsample_rates, upsample_initial_channel=self.hps.model.upsample_initial_channel, upsample_kernel_sizes=self.hps.model.upsample_kernel_sizes, n_flow=self.hps.model.n_flow, dec_out_channels=1, dec_kernel_size=7, n_speakers=self.hps.data.n_speakers, gin_channels=self.hps.model.gin_channels, requires_grad_pe=self.hps.requires_grad.pe, requires_grad_flow=self.hps.requires_grad.flow, requires_grad_text_enc=self.hps.requires_grad.text_enc, requires_grad_dec=self.hps.requires_grad.dec ) self.net_g.eval() load_checkpoint(pyTorch_model_file, self.net_g, None) # ONNXモデル生成 if onnx_model_file != None: ort_options = onnxruntime.SessionOptions() ort_options.intra_op_num_threads = 8 self.onnx_session = onnxruntime.InferenceSession( onnx_model_file, providers=providers ) return self.get_info() def destroy(self): del self.net_g del self.onnx_session def get_info(self): data = asdict(self.settings) data["onnxExecutionProviders"] = self.onnx_session.get_providers() if self.onnx_session != None else [] files = ["configFile", "pyTorchModelFile", "onnxModelFile"] for f in files: if data[f] != None and os.path.exists(data[f]): data[f] = os.path.basename(data[f]) else: data[f] = "" return data def update_setteings(self, key: str, val: any): if key == "onnxExecutionProvider" and self.onnx_session != None: if val == "CUDAExecutionProvider": if self.settings.gpu < 0 or self.settings.gpu >= self.gpu_num: self.settings.gpu = 0 provider_options = [{'device_id': self.settings.gpu}] self.onnx_session.set_providers(providers=[val], provider_options=provider_options) else: self.onnx_session.set_providers(providers=[val]) elif key in self.settings.intData: setattr(self.settings, key, int(val)) if key == "gpu" and val >= 0 and val < self.gpu_num and self.onnx_session != None: providers = self.onnx_session.get_providers() print("Providers:", providers) if "CUDAExecutionProvider" in providers: provider_options = [{'device_id': self.settings.gpu}] self.onnx_session.set_providers(providers=["CUDAExecutionProvider"], provider_options=provider_options) if key == "crossFadeOffsetRate" or key == "crossFadeEndRate": self.unpackedData_length = 0 elif key in self.settings.floatData: setattr(self.settings, key, float(val)) elif key in self.settings.strData: setattr(self.settings, key, str(val)) else: print(f"{key} is not mutalbe variable!") return self.get_info() def _generate_input(self, unpackedData: any, convertSize: int): # 今回変換するデータをテンソルとして整形する audio = torch.FloatTensor(unpackedData.astype(np.float32)) # float32でtensorfを作成 audio_norm = audio / self.hps.data.max_wav_value # normalize audio_norm = audio_norm.unsqueeze(0) # unsqueeze self.audio_buffer = torch.cat([self.audio_buffer, audio_norm], axis=1) # 過去のデータに連結 # audio_norm = self.audio_buffer[:, -(convertSize + 1280 * 2):] # 変換対象の部分だけ抽出 audio_norm = self.audio_buffer[:, -(convertSize):] # 変換対象の部分だけ抽出 self.audio_buffer = audio_norm # TBD: numpy <--> pytorch変換が行ったり来たりしているが、まずは動かすことを最優先。 audio_norm_np = audio_norm.squeeze().numpy().astype(np.float64) if self.settings.f0Detector == "dio": _f0, _time = pw.dio(audio_norm_np, self.hps.data.sampling_rate, frame_period=5.5) f0 = pw.stonemask(audio_norm_np, _f0, _time, self.hps.data.sampling_rate) else: f0, t = pw.harvest(audio_norm_np, self.hps.data.sampling_rate, frame_period=5.5, f0_floor=71.0, f0_ceil=1000.0) f0 = convert_continuos_f0(f0, int(audio_norm_np.shape[0] / self.hps.data.hop_length)) f0 = torch.from_numpy(f0.astype(np.float32)) spec = spectrogram_torch(audio_norm, self.hps.data.filter_length, self.hps.data.sampling_rate, self.hps.data.hop_length, self.hps.data.win_length, center=False) # dispose_stft_specs = 2 # spec = spec[:, dispose_stft_specs:-dispose_stft_specs] # f0 = f0[dispose_stft_specs:-dispose_stft_specs] spec = torch.squeeze(spec, 0) sid = torch.LongTensor([int(self.settings.srcId)]) # data = (self.text_norm, spec, audio_norm, sid) # data = TextAudioSpeakerCollate()([data]) data = TextAudioSpeakerCollate( sample_rate=self.hps.data.sampling_rate, hop_size=self.hps.data.hop_length, f0_factor=self.settings.f0Factor )([(spec, sid, f0)]) return data def _onnx_inference(self, data, inputSize): if hasattr(self, "onnx_session") == False or self.onnx_session == None: print("[Voice Changer] No ONNX session.") return np.zeros(1).astype(np.int16) # x, x_lengths, spec, spec_lengths, y, y_lengths, sid_src = [x for x in data] # sid_tgt1 = torch.LongTensor([self.settings.dstId]) spec, spec_lengths, sid_src, sin, d = data sid_tgt1 = torch.LongTensor([self.settings.dstId]) # if spec.size()[2] >= 8: audio1 = self.onnx_session.run( ["audio"], { "specs": spec.numpy(), "lengths": spec_lengths.numpy(), "sin": sin.numpy(), "d0": d[0][:1].numpy(), "d1": d[1][:1].numpy(), "d2": d[2][:1].numpy(), "d3": d[3][:1].numpy(), "sid_src": sid_src.numpy(), "sid_tgt": sid_tgt1.numpy() })[0][0, 0] * self.hps.data.max_wav_value if hasattr(self, 'np_prev_audio1') == True: overlapSize = min(self.settings.crossFadeOverlapSize, inputSize) prev_overlap = self.np_prev_audio1[-1 * overlapSize:] cur_overlap = audio1[-1 * (inputSize + overlapSize):-1 * inputSize] # print(prev_overlap.shape, self.np_prev_strength.shape, cur_overlap.shape, self.np_cur_strength.shape) # print(">>>>>>>>>>>", -1*(inputSize + overlapSize) , -1*inputSize) powered_prev = prev_overlap * self.np_prev_strength powered_cur = cur_overlap * self.np_cur_strength powered_result = powered_prev + powered_cur cur = audio1[-1 * inputSize:-1 * overlapSize] result = np.concatenate([powered_result, cur], axis=0) else: result = np.zeros(1).astype(np.int16) self.np_prev_audio1 = audio1 return result def _pyTorch_inference(self, data, inputSize): if hasattr(self, "net_g") == False or self.net_g == None: print("[Voice Changer] No pyTorch session.") return np.zeros(1).astype(np.int16) if self.settings.gpu < 0 or self.gpu_num == 0: with torch.no_grad(): spec, spec_lengths, sid_src, sin, d = data spec = spec.cpu() spec_lengths = spec_lengths.cpu() sid_src = sid_src.cpu() sin = sin.cpu() d = tuple([d[:1].cpu() for d in d]) sid_target = torch.LongTensor([self.settings.dstId]).cpu() audio1 = self.net_g.cpu().voice_conversion(spec, spec_lengths, sin, d, sid_src, sid_target)[0, 0].data * self.hps.data.max_wav_value if self.prev_strength.device != torch.device('cpu'): print(f"prev_strength move from {self.prev_strength.device} to cpu") self.prev_strength = self.prev_strength.cpu() if self.cur_strength.device != torch.device('cpu'): print(f"cur_strength move from {self.cur_strength.device} to cpu") self.cur_strength = self.cur_strength.cpu() if hasattr(self, 'prev_audio1') == True and self.prev_audio1.device == torch.device('cpu'): # prev_audio1が所望のデバイスに無い場合は一回休み。 overlapSize = min(self.settings.crossFadeOverlapSize, inputSize) prev_overlap = self.prev_audio1[-1 * overlapSize:] cur_overlap = audio1[-1 * (inputSize + overlapSize):-1 * inputSize] powered_prev = prev_overlap * self.prev_strength powered_cur = cur_overlap * self.cur_strength powered_result = powered_prev + powered_cur cur = audio1[-1 * inputSize:-1 * overlapSize] # 今回のインプットの生部分。(インプット - 次回のCrossfade部分)。 result = torch.cat([powered_result, cur], axis=0) # Crossfadeと今回のインプットの生部分を結合 else: cur = audio1[-2 * inputSize:-1 * inputSize] result = cur self.prev_audio1 = audio1 result = result.cpu().float().numpy() else: with torch.no_grad(): spec, spec_lengths, sid_src, sin, d = data spec = spec.cuda(self.settings.gpu) spec_lengths = spec_lengths.cuda(self.settings.gpu) sid_src = sid_src.cuda(self.settings.gpu) sin = sin.cuda(self.settings.gpu) d = tuple([d[:1].cuda(self.settings.gpu) for d in d]) sid_target = torch.LongTensor([self.settings.dstId]).cuda(self.settings.gpu) # audio1 = self.net_g.cuda(self.settings.gpu).voice_conversion(spec, spec_lengths, sid_src=sid_src, # sid_tgt=sid_tgt1)[0, 0].data * self.hps.data.max_wav_value audio1 = self.net_g.cuda(self.settings.gpu).voice_conversion(spec, spec_lengths, sin, d, sid_src, sid_target)[0, 0].data * self.hps.data.max_wav_value if self.prev_strength.device != torch.device('cuda', self.settings.gpu): print(f"prev_strength move from {self.prev_strength.device} to gpu{self.settings.gpu}") self.prev_strength = self.prev_strength.cuda(self.settings.gpu) if self.cur_strength.device != torch.device('cuda', self.settings.gpu): print(f"cur_strength move from {self.cur_strength.device} to gpu{self.settings.gpu}") self.cur_strength = self.cur_strength.cuda(self.settings.gpu) if hasattr(self, 'prev_audio1') == True and self.prev_audio1.device == torch.device('cuda', self.settings.gpu): overlapSize = min(self.settings.crossFadeOverlapSize, inputSize) prev_overlap = self.prev_audio1[-1 * overlapSize:] cur_overlap = audio1[-1 * (inputSize + overlapSize):-1 * inputSize] powered_prev = prev_overlap * self.prev_strength powered_cur = cur_overlap * self.cur_strength powered_result = powered_prev + powered_cur # print(overlapSize, prev_overlap.shape, cur_overlap.shape, self.prev_strength.shape, self.cur_strength.shape) # print(self.prev_audio1.shape, audio1.shape, inputSize, overlapSize) cur = audio1[-1 * inputSize:-1 * overlapSize] # 今回のインプットの生部分。(インプット - 次回のCrossfade部分)。 result = torch.cat([powered_result, cur], axis=0) # Crossfadeと今回のインプットの生部分を結合 else: cur = audio1[-2 * inputSize:-1 * inputSize] result = cur self.prev_audio1 = audio1 result = result.cpu().float().numpy() return result