export_onnx.py
2.9 KB
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
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
#!/usr/bin/env python3
# Copyright 2025 Xiaomi Corp. (authors: Fangjun Kuang)
from pathlib import Path
from typing import Dict
import os
import nemo.collections.asr as nemo_asr
import onnx
import torch
from onnxruntime.quantization import QuantType, quantize_dynamic
def add_meta_data(filename: str, meta_data: Dict[str, str]):
"""Add meta data to an ONNX model. It is changed in-place.
Args:
filename:
Filename of the ONNX model to be changed.
meta_data:
Key-value pairs.
"""
model = onnx.load(filename)
while len(model.metadata_props):
model.metadata_props.pop()
for key, value in meta_data.items():
meta = model.metadata_props.add()
meta.key = key
meta.value = str(value)
if filename == "encoder.onnx":
external_filename = "encoder"
onnx.save(
model,
filename,
save_as_external_data=True,
all_tensors_to_one_file=True,
location=external_filename + ".weights",
)
else:
onnx.save(model, filename)
@torch.no_grad()
def main():
if Path("./parakeet-tdt-0.6b-v3.nemo").is_file():
asr_model = nemo_asr.models.ASRModel.restore_from(
restore_path="./parakeet-tdt-0.6b-v3.nemo"
)
else:
asr_model = nemo_asr.models.ASRModel.from_pretrained(
model_name="nvidia/parakeet-tdt-0.6b-v3"
)
asr_model.eval()
with open("./tokens.txt", "w", encoding="utf-8") as f:
for i, s in enumerate(asr_model.joint.vocabulary):
f.write(f"{s} {i}\n")
f.write(f"<blk> {i+1}\n")
print("Saved to tokens.txt")
asr_model.encoder.export("encoder.onnx")
asr_model.decoder.export("decoder.onnx")
asr_model.joint.export("joiner.onnx")
os.system("ls -lh *.onnx")
normalize_type = asr_model.cfg.preprocessor.normalize
if normalize_type == "NA":
normalize_type = ""
meta_data = {
"vocab_size": asr_model.decoder.vocab_size,
"normalize_type": normalize_type,
"pred_rnn_layers": asr_model.decoder.pred_rnn_layers,
"pred_hidden": asr_model.decoder.pred_hidden,
"subsampling_factor": 8,
"model_type": "EncDecRNNTBPEModel",
"version": "2",
"model_author": "NeMo",
"url": "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3",
"comment": "Only the transducer branch is exported",
"feat_dim": 128,
}
for m in ["encoder", "decoder", "joiner"]:
quantize_dynamic(
model_input=f"./{m}.onnx",
model_output=f"./{m}.int8.onnx",
weight_type=QuantType.QUInt8 if m == "encoder" else QuantType.QInt8,
)
os.system("ls -lh *.onnx")
add_meta_data("encoder.int8.onnx", meta_data)
add_meta_data("encoder.onnx", meta_data)
print("meta_data", meta_data)
if __name__ == "__main__":
main()