This repository has been archived by the owner on Nov 11, 2023. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 4.7k
/
onnx_export.py
56 lines (50 loc) · 2.2 KB
/
onnx_export.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
import torch
from onnxexport.model_onnx import SynthesizerTrn
import utils
def main(NetExport):
path = "SoVits4.0"
if NetExport:
device = torch.device("cpu")
hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json")
SVCVITS = SynthesizerTrn(
hps.data.filter_length // 2 + 1,
hps.train.segment_size // hps.data.hop_length,
**hps.model)
_ = utils.load_checkpoint(f"checkpoints/{path}/model.pth", SVCVITS, None)
_ = SVCVITS.eval().to(device)
for i in SVCVITS.parameters():
i.requires_grad = False
n_frame = 10
hidden_channels = 256 #(Hubert's shape[2])
test_hidden_unit = torch.rand(1, n_frame, hidden_channels)
test_pitch = torch.rand(1, n_frame)
test_mel2ph = torch.arange(0, n_frame, dtype=torch.int64)[None] # torch.LongTensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]).unsqueeze(0)
test_uv = torch.ones(1, n_frame, dtype=torch.float32)
test_noise = torch.randn(1, 192, n_frame)
test_sid = torch.LongTensor([0])
input_names = ["c", "f0", "mel2ph", "uv", "noise", "sid"]
output_names = ["audio", ]
torch.onnx.export(SVCVITS,
(
test_hidden_unit.to(device),
test_pitch.to(device),
test_mel2ph.to(device),
test_uv.to(device),
test_noise.to(device),
test_sid.to(device)
),
f"checkpoints/{path}/model.onnx",
dynamic_axes={
"c": [0, 1],
"f0": [1],
"mel2ph": [1],
"uv": [1],
"noise": [2],
},
do_constant_folding=False,
opset_version=16,
verbose=False,
input_names=input_names,
output_names=output_names)
if __name__ == '__main__':
main(True)