51 lines
2.0 KiB
Python
51 lines
2.0 KiB
Python
import argparse
|
|
import functools
|
|
import torch
|
|
import sys
|
|
sys.path.append(R"../0_EcapaTDNN")
|
|
from macls.predict import MAClsPredictor
|
|
from macls.utils.utils import add_arguments, print_arguments
|
|
|
|
parser = argparse.ArgumentParser(description=__doc__)
|
|
add_arg = functools.partial(add_arguments, argparser=parser)
|
|
add_arg('configs', str, 'configs/ecapa_tdnn.yml', '配置文件')
|
|
add_arg('audio_path', str, 'dataset/126153-9-0-11.wav', '测试音频路径')
|
|
add_arg('model_path', str, '../weights/EcapaTdnn_Fbank/best_model/', '模型权重文件路径')
|
|
add_arg('dst_path', str, '../2_compile/fmodel/EcapaTDNN_predictor_1x398x80.onnx', '导出的onnx模型文件路径')
|
|
args = parser.parse_args()
|
|
print_arguments(args=args)
|
|
# 备注： 太短的音频(0.3175ms < 0.5ms)会无法预测
|
|
# 获取识别器
|
|
|
|
model = MAClsPredictor(configs=args.configs,
|
|
model_path=args.model_path,
|
|
use_gpu=False)
|
|
# 加载音频文件，并进行预处理
|
|
input_data = model._load_audio(audio_data=args.audio_path, sample_rate=16000)
|
|
input_data = torch.tensor(input_data.samples, dtype=torch.float32).unsqueeze(0)
|
|
audio_feature = model._audio_featurizer(input_data)
|
|
print('audio_feature =',audio_feature.shape)
|
|
# 执行预测
|
|
output = model.predictor(audio_feature)
|
|
print('output =',output.shape)
|
|
|
|
# 导出ONNX模型
|
|
input_names = ["audio_feature"]
|
|
output_names = ["logits"]
|
|
torch.onnx.export(
|
|
model.predictor,
|
|
audio_feature,
|
|
args.dst_path,
|
|
verbose=False,
|
|
keep_initializers_as_inputs=False,
|
|
opset_version=17,
|
|
input_names=input_names,
|
|
output_names=output_names,)
|
|
print(f'ONNX export success, saved in: {args.dst_path} ')
|
|
|
|
# 导出pt模型
|
|
# pt_model = model.predictor
|
|
# traced_model = torch.jit.trace(pt_model,audio_feature,strict=False)
|
|
# torch.jit.save(traced_model,args.dst_path)
|
|
# print(f'TorchScript export success, saved in: {args.dst_path} ')
|