PublicMirrors/voice-changer
1
0
Fork 0
mirror of https://github.com/w-okada/voice-changer.git synced 2025-01-23 21:45:00 +03:00
Code Issues Actions Packages Projects Releases Wiki Activity

WIP: support hubert onnx 1

Browse Source
This commit is contained in:
wataru 2023-03-29 23:11:35 +09:00
parent 06770746d9
commit 4bbcbd66bc
1 changed files with 45 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

58
server/voice_changer/SoVitsSvc40/SoVitsSvc40.py
View File

@ -122,7 +122,12 @@ class SoVitsSvc40:
onnx_model_file, onnx_model_file,
providers=providers providers=providers
) )
input_info = self.onnx_session.get_inputs() # input_info = self.onnx_session.get_inputs()
# for i in input_info:
# print("input", i)
# output_info = self.onnx_session.get_outputs()
# for i in output_info:
# print("output", i)
return self.get_info() return self.get_info()
def update_setteings(self, key: str, val: any): def update_setteings(self, key: str, val: any):
@ -191,8 +196,8 @@ class SoVitsSvc40:
uv = uv.unsqueeze(0) uv = uv.unsqueeze(0)
# wav16k = librosa.resample(audio_buffer, orig_sr=24000, target_sr=16000) # wav16k = librosa.resample(audio_buffer, orig_sr=24000, target_sr=16000)
wav16k = librosa.resample(audio_buffer, orig_sr=self.hps.data.sampling_rate, target_sr=16000) wav16k_numpy = librosa.resample(audio_buffer, orig_sr=self.hps.data.sampling_rate, target_sr=16000)
wav16k = torch.from_numpy(wav16k) wav16k_tensor = torch.from_numpy(wav16k_numpy)
if (self.settings.gpu < 0 or self.gpu_num == 0) or self.settings.framework == "ONNX": if (self.settings.gpu < 0 or self.gpu_num == 0) or self.settings.framework == "ONNX":
dev = torch.device("cpu") dev = torch.device("cpu")
@ -200,11 +205,38 @@ class SoVitsSvc40:
dev = torch.device("cuda", index=self.settings.gpu) dev = torch.device("cuda", index=self.settings.gpu)
self.hubert_model = self.hubert_model.to(dev) self.hubert_model = self.hubert_model.to(dev)
wav16k = wav16k.to(dev) wav16k_tensor = wav16k_tensor.to(dev)
uv = uv.to(dev) uv = uv.to(dev)
f0 = f0.to(dev) f0 = f0.to(dev)
c = utils.get_hubert_content(self.hubert_model, wav_16k_tensor=wav16k) import time
start = time.time()
for i in range(10):
c = utils.get_hubert_content(self.hubert_model, wav_16k_tensor=wav16k_tensor)
end = time.time()
elapse = end - start
print("torch time", elapse, elapse / 10)
import onnxruntime
ort_options = onnxruntime.SessionOptions()
ort_options.intra_op_num_threads = 8
if not hasattr(self, "hubert_onnx"):
self.hubert_onnx = onnxruntime.InferenceSession(
"model_hubert/hubert_simple.onnx",
# providers=['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider']
providers=['CPUExecutionProvider']
)
start = time.time()
for i in range(10):
c_onnx = utils.get_hubert_content2(self.hubert_onnx, wav16k_numpy)
end = time.time()
elapse = end - start
print("onnx time", elapse, elapse / 10)
print("torch units:", c)
print("onnx units:", c_onnx)
c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[1]) c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[1])
if self.settings.clusterInferRatio != 0 and hasattr(self, "cluster_model") and self.cluster_model != None: if self.settings.clusterInferRatio != 0 and hasattr(self, "cluster_model") and self.cluster_model != None:
@ -256,23 +288,22 @@ class SoVitsSvc40:
return np.zeros(convertSize).astype(np.int16) return np.zeros(convertSize).astype(np.int16)
c, f0, uv = [x.numpy() for x in data] c, f0, uv = [x.numpy() for x in data]
sid_target = torch.LongTensor([self.settings.dstId]).unsqueeze(0).numpy()
audio1 = self.onnx_session.run( audio1 = self.onnx_session.run(
["audio"], ["audio"],
{ {
"c": c, "c": c.astype(np.float32),
"f0": f0, "f0": f0.astype(np.float32),
"g": np.array([self.settings.dstId]).astype(np.int64), "uv": uv.astype(np.float32),
"uv": uv, "g": sid_target.astype(np.int64),
"predict_f0": np.array([self.settings.dstId]).astype(np.int64), "noice_scale": np.array([self.settings.noiceScale]).astype(np.float32),
"noice_scale": np.array([self.settings.dstId]).astype(np.int64), # "predict_f0": np.array([self.settings.dstId]).astype(np.int64),
})[0][0, 0] * self.hps.data.max_wav_value })[0][0, 0] * self.hps.data.max_wav_value
audio1 = audio1 * vol audio1 = audio1 * vol
result = audio1 result = audio1
return result return result
pass pass
@ -296,6 +327,7 @@ class SoVitsSvc40:
with torch.no_grad(): with torch.no_grad():
c, f0, uv = [x.to(dev)for x in data] c, f0, uv = [x.to(dev)for x in data]
print("Sahape", c.shape, f0.shape, uv.shape)
sid_target = torch.LongTensor([self.settings.dstId]).to(dev).unsqueeze(0) sid_target = torch.LongTensor([self.settings.dstId]).to(dev).unsqueeze(0)
self.net_g.to(dev) self.net_g.to(dev)
# audio1 = self.net_g.infer(c, f0=f0, g=sid_target, uv=uv, predict_f0=True, noice_scale=0.1)[0][0, 0].data.float() # audio1 = self.net_g.infer(c, f0=f0, g=sid_target, uv=uv, predict_f0=True, noice_scale=0.1)[0][0, 0].data.float()