tuna2134/sbv2-v2
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

Add: Support for ONNX inference and ONNX conversion of Non-JP-Extra models

Browse Source
This commit is contained in:
tsukumi
2024-09-23 05:59:12 +09:00
parent cc1d6120fe
commit db0acd7b5f
3 changed files with 140 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
style_bert_vits2/models/hyper_parameters.py
View File

@@ -38,7 +38,8 @@ class HyperParametersTrain(BaseModel):
class HyperParametersData(BaseModel):
use_jp_extra: bool = False # このフィールドが存在しない旧モデルとの互換性のために False をデフォルト値とする
# use_jp_extra フィールドが存在しない旧モデルとの互換性のために False をデフォルト値とする
use_jp_extra: bool = False
training_files: str = "Data/Dummy/train.list"
validation_files: str = "Data/Dummy/val.list"
max_wav_value: float = 32768.0

66
style_bert_vits2/models/infer_onnx.py
View File

@@ -75,15 +75,15 @@ def get_text_onnx(
if language_str == Languages.ZH:
bert = bert_ori
ja_bert = np.zeros((1024, len(phone)))
en_bert = np.zeros((1024, len(phone)))
ja_bert = np.zeros((1024, len(phone)), dtype=np.float32)
en_bert = np.zeros((1024, len(phone)), dtype=np.float32)
elif language_str == Languages.JP:
bert = np.zeros((1024, len(phone)))
bert = np.zeros((1024, len(phone)), dtype=np.float32)
ja_bert = bert_ori
en_bert = np.zeros((1024, len(phone)))
en_bert = np.zeros((1024, len(phone)), dtype=np.float32)
elif language_str == Languages.EN:
bert = np.zeros((1024, len(phone)))
ja_bert = np.zeros((1024, len(phone)))
bert = np.zeros((1024, len(phone)), dtype=np.float32)
ja_bert = np.zeros((1024, len(phone)), dtype=np.float32)
en_bert = bert_ori
else:
raise ValueError("language_str should be ZH, JP or EN")
@@ -170,27 +170,41 @@ def infer_onnx(
np.array([noise_scale], dtype=np.float32),
np.array([noise_scale_w], dtype=np.float32),
]
first_provider = onnx_session.get_providers()[0]
if first_provider == "CUDAExecutionProvider":
device_type = "cuda"
elif first_provider == "DmlExecutionProvider":
device_type = "dml"
else:
device_type = "cpu"
# GPU メモリに入力テンソルを割り当て
io_binding = onnx_session.io_binding()
for name, value in zip(input_names, input_tensor):
gpu_tensor = onnxruntime.OrtValue.ortvalue_from_numpy(value, device_type)
io_binding.bind_ortvalue_input(name, gpu_tensor)
# 推論の実行
io_binding.bind_output(output_name, device_type)
onnx_session.run_with_iobinding(io_binding)
output = io_binding.get_outputs()
else:
raise NotImplementedError("Not implemented yet")
input_tensor = [
x_tst,
x_tst_lengths,
sid_tensor,
tones,
lang_ids,
bert,
ja_bert,
en_bert,
style_vec_tensor,
np.array([length_scale], dtype=np.float32),
np.array([sdp_ratio], dtype=np.float32),
np.array([noise_scale], dtype=np.float32),
np.array([noise_scale_w], dtype=np.float32),
]
first_provider = onnx_session.get_providers()[0]
if first_provider == "CUDAExecutionProvider":
device_type = "cuda"
elif first_provider == "DmlExecutionProvider":
device_type = "dml"
else:
device_type = "cpu"
# GPU メモリに入力テンソルを割り当て
io_binding = onnx_session.io_binding()
for name, value in zip(input_names, input_tensor):
gpu_tensor = onnxruntime.OrtValue.ortvalue_from_numpy(value, device_type)
io_binding.bind_ortvalue_input(name, gpu_tensor)
# 推論の実行
io_binding.bind_output(output_name, device_type)
onnx_session.run_with_iobinding(io_binding)
output = io_binding.get_outputs()
audio = output[0].numpy()[0, 0]