1
0
mirror of synced 2024-11-27 17:00:54 +01:00
Retrieval-based-Voice-Conve.../tools/export_onnx.py

55 lines
2.0 KiB
Python
Raw Permalink Normal View History

2023-04-08 17:36:25 +02:00
import torch
2023-08-29 09:14:11 +02:00
from infer.lib.infer_pack.models_onnx import SynthesizerTrnMsNSFsidM
2023-04-08 17:36:25 +02:00
if __name__ == "__main__":
MoeVS = True # 模型是否为MoeVoiceStudio原MoeSS使用
2023-04-08 17:36:25 +02:00
ModelPath = "Shiroha/shiroha.pth" # 模型路径
ExportedPath = "model.onnx" # 输出路径
hidden_channels = 256 # hidden_channels为768Vec做准备
cpt = torch.load(ModelPath, map_location="cpu")
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0] # n_spk
print(*cpt["config"])
2023-04-08 17:36:25 +02:00
test_phone = torch.rand(1, 200, hidden_channels) # hidden unit
test_phone_lengths = torch.tensor([200]).long() # hidden unit 长度(貌似没啥用)
test_pitch = torch.randint(size=(1, 200), low=5, high=255) # 基频(单位赫兹)
test_pitchf = torch.rand(1, 200) # nsf基频
test_ds = torch.LongTensor([0]) # 说话人ID
test_rnd = torch.rand(1, 192, 200) # 噪声(加入随机因子)
2023-04-08 17:36:25 +02:00
device = "cpu" # 导出时设备(不影响使用模型)
net_g = SynthesizerTrnMsNSFsidM(
*cpt["config"], is_half=False
) # fp32导出C++要支持fp16必须手动将内存重新排列所以暂时不用fp16
net_g.load_state_dict(cpt["weight"], strict=False)
input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
output_names = [
"audio",
]
# net_g.construct_spkmixmap(n_speaker) 多角色混合轨道导出
torch.onnx.export(
net_g,
(
test_phone.to(device),
test_phone_lengths.to(device),
test_pitch.to(device),
test_pitchf.to(device),
test_ds.to(device),
test_rnd.to(device),
),
ExportedPath,
dynamic_axes={
"phone": [1],
"pitch": [1],
"pitchf": [1],
"rnd": [2],
},
do_constant_folding=False,
opset_version=16,
verbose=False,
input_names=input_names,
output_names=output_names,
)