File size: 5,086 Bytes
223aff6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
# Copyright (c) 2022, Yongqiang Li ([email protected])
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import argparse
import json
import os
import sys

import torch

from models import SynthesizerTrn
import utils

try:
    import onnxruntime as ort
except ImportError:
    print('Please install onnxruntime!')
    sys.exit(1)


def to_numpy(tensor):
    return tensor.detach().cpu().numpy() if tensor.requires_grad \
        else tensor.detach().numpy()


def get_args():
    parser = argparse.ArgumentParser(description='export onnx model')
    parser.add_argument('--checkpoint', required=True, help='checkpoint')
    parser.add_argument('--cfg', required=True, help='config file')
    parser.add_argument('--onnx_model', required=True, help='onnx model name')
    # parser.add_argument('--phone_table',
    #                     required=True,
    #                     help='input phone dict')
    # parser.add_argument('--speaker_table', default=None, help='speaker table')
    # parser.add_argument("--speaker_num", required=True,
    #                     type=int, help="speaker num")
    parser.add_argument(
        '--providers',
        required=False,
        default='CPUExecutionProvider',
        choices=['CUDAExecutionProvider', 'CPUExecutionProvider'],
        help='the model to send request to')
    args = parser.parse_args()
    return args


def get_data_from_cfg(cfg_path: str):
    assert os.path.isfile(cfg_path)
    with open(cfg_path, 'r') as f:
        data = json.load(f)
        symbols = data["symbols"]
        speaker_num = data["data"]["n_speakers"]
    return len(symbols), speaker_num


def main():
    args = get_args()
    os.environ['CUDA_VISIBLE_DEVICES'] = '0'

    hps = utils.get_hparams_from_file(args.cfg)
    # with open(args.phone_table) as p_f:
    #     phone_num = len(p_f.readlines()) + 1
    # num_speakers = 1
    # if args.speaker_table is not None:
    #     num_speakers = len(open(args.speaker_table).readlines()) + 1
    phone_num, num_speakers = get_data_from_cfg(args.cfg)
    net_g = SynthesizerTrn(phone_num,
                           hps.data.filter_length // 2 + 1,
                           hps.train.segment_size // hps.data.hop_length,
                           n_speakers=num_speakers,
                           **hps.model)
    utils.load_checkpoint(args.checkpoint, net_g, None)
    net_g.forward = net_g.export_forward
    net_g.eval()

    seq = torch.randint(low=0, high=phone_num, size=(1, 10), dtype=torch.long)
    seq_len = torch.IntTensor([seq.size(1)]).long()

    # noise(可用于控制感情等变化程度) lenth(可用于控制整体语速) noisew(控制音素发音长度变化程度)
    # 参考 https://github.com/gbxh/genshinTTS
    scales = torch.FloatTensor([0.667, 1.0, 0.8])
    # make triton dynamic shape happy
    scales = scales.unsqueeze(0)
    sid = torch.IntTensor([0]).long()

    dummy_input = (seq, seq_len, scales, sid)
    torch.onnx.export(model=net_g,
                      args=dummy_input,
                      f=args.onnx_model,
                      input_names=['input', 'input_lengths', 'scales', 'sid'],
                      output_names=['output'],
                      dynamic_axes={
                          'input': {
                              0: 'batch',
                              1: 'phonemes'
                          },
                          'input_lengths': {
                              0: 'batch'
                          },
                          'scales': {
                              0: 'batch'
                          },
                          'sid': {
                              0: 'batch'
                          },
                          'output': {
                              0: 'batch',
                              1: 'audio',
                              2: 'audio_length'
                          }
                      },
                      opset_version=13,
                      verbose=False)

    # Verify onnx precision
    torch_output = net_g(seq, seq_len, scales, sid)
    providers = [args.providers]
    ort_sess = ort.InferenceSession(args.onnx_model, providers=providers)
    ort_inputs = {
        'input': to_numpy(seq),
        'input_lengths': to_numpy(seq_len),
        'scales': to_numpy(scales),
        'sid': to_numpy(sid),
    }
    onnx_output = ort_sess.run(None, ort_inputs)


if __name__ == '__main__':
    main()