MNN模型转换后对比结果错误。

[cwp-wind]

MNNConvert 版本：2.9.3
使用ResNet34_LM后通过命令：
python wespeaker/bin/export_onnx.py --config model/config.yaml --checkpoint model/model_5.pt --output_model model.onnx
转onnx后再通过命令：
python wespeaker/bin/export_mnn.py --onnx_model model.onnx --output_model model.mnn
提示：
The device supports: i8sdot:0, fp16:0, i8mm: 0, sve2: 0 The device supports: i8sdot:0, fp16:0, i8mm: 0, sve2: 0 Start to Convert Other Model Format To MNN Model..., target version: 2.9 [10:50:32] /root/code/MNN-master/tools/converter/source/onnx/onnxConverter.cpp:46: ONNX Model ir version: 7 [10:50:32] /root/code/MNN-master/tools/converter/source/onnx/onnxConverter.cpp:47: ONNX Model opset version: 14 Start to Optimize the MNN Net... inputTensors : [ feats, ] outputTensors: [ embs, ] Converted Success! Exported MNN model to model.mnn The device supports: i8sdot:0, fp16:0, i8mm: 0, sve2: 0 Model default dimensionFormat is NCHW Model Inputs: [ feats ]: dimensionFormat: NCHW, size: [ 1,-1,80 ], type is float Model Outputs: [ embs ] Model Version: 2.9.3 MNN use high precision

使用生成的模型进行测试，命令为：
”
./asv_main
--enroll_wav warmup.wav
--test_wav warmup.wav
--threshold 0.5
--speaker_model_path model.mnn
--embedding_size 256
I0820 10:51:01.697679 3446718 asv_main.cc:38] model.mnn
I0820 10:51:01.697815 3446718 asv_main.cc:39] Init model ...
I0820 10:51:01.697841 3446718 speaker_engine.cc:35] Reading model model.mnn
I0820 10:51:01.697861 3446718 speaker_engine.cc:37] Embedding size: 256
I0820 10:51:01.697881 3446718 speaker_engine.cc:39] per_chunk_samples: 32000
I0820 10:51:01.697898 3446718 speaker_engine.cc:41] Sample rate: 16000
The device support i8sdot:0, support fp16:0, support i8mm: 0
I0820 10:51:01.765399 3446718 asv_main.cc:44] embedding size: 256
I0820 10:51:01.773716 3446718 asv_main.cc:53] 81750
I0820 10:51:02.185937 3446718 mnn_speaker_model.cc:65] dynamic shape.
I0820 10:51:02.231452 3446718 asv_main.cc:62] 81750
I0820 10:51:02.668128 3446718 asv_main.cc:65] compute score ...
I0820 10:51:02.668175 3446718 asv_main.cc:67] Cosine socre: 0
I0820 10:51:02.668211 3446718 asv_main.cc:71] Warning! It's a different speaker“

同一个文件，比对后显示2个文件不是同一个发音人，warmup.wav里面只有一个人讲话。
同样的音频用onnx模型执行同样的模型：
”./asv_main --enroll_wav warmup.wav
--test_wav warmup.wav
--threshold 0.5
--speaker_model_path model.onnx
--embedding_size 256
I0820 11:02:23.524395 3484515 asv_main.cc:38] model.onnx
I0820 11:02:23.524457 3484515 asv_main.cc:39] Init model ...
I0820 11:02:23.524466 3484515 speaker_engine.cc:35] Reading model model.onnx
I0820 11:02:23.524472 3484515 speaker_engine.cc:37] Embedding size: 256
I0820 11:02:23.524477 3484515 speaker_engine.cc:39] per_chunk_samples: 32000
I0820 11:02:23.524483 3484515 speaker_engine.cc:41] Sample rate: 16000
I0820 11:02:23.591621 3484515 onnx_speaker_model.cc:60] Ouput name: feats
I0820 11:02:23.591661 3484515 onnx_speaker_model.cc:68] Output name: embs
I0820 11:02:23.591672 3484515 asv_main.cc:44] embedding size: 256
I0820 11:02:23.594732 3484515 asv_main.cc:53] 43306
I0820 11:02:26.240854 3484515 asv_main.cc:62] 43306
I0820 11:02:28.911242 3484515 asv_main.cc:65] compute score ...
I0820 11:02:28.911299 3484515 asv_main.cc:67] Cosine socre: 1
I0820 11:02:28.911329 3484515 asv_main.cc:69] It's the same speaker!“

[cdliang11]

I0820 11:02:23.594732 3484515 asv_main.cc:53] 43306
I0820 11:02:26.240854 3484515 asv_main.cc:62] 43306

@cwp-wind asv_main.cc这两行显示的的音频长度。 onnx 和 mnn测试用的是不同的输入文件吗？

[cwp-wind]

是同一个文件。

[cdliang11]

是同一个文件。

日志里显示的音频长度不一样，mnn是81750，onnx是43306

[cwp-wind]

抱歉，图搞错了，我重新跑了一遍，以这个为准：

另外我上传了我转换的模型以及测试的音频

[cwp-wind]

感谢支持，模型网盘链接为：
链接：https://pan.baidu.com/s/1sZJFuCgqIQh0mdCJAR637w
提取码：j2or

[cdliang11]

感谢支持，模型网盘链接为： 链接：https://pan.baidu.com/s/1sZJFuCgqIQh0mdCJAR637w 提取码：j2or

使用你给的模型，我这边没问题。

(base) ➜  mnn git:(master) ✗ ./build/bin/asv_main --enroll_wav ../../1.wav --test_wav ../../1.wav --speaker_model_path ../../model.mnn                     
I1031 18:51:49.387418 3852557824 asv_main.cc:38] ../../model.mnn
I1031 18:51:49.387877 3852557824 asv_main.cc:39] Init model ...
I1031 18:51:49.387885 3852557824 speaker_engine.cc:35] Reading model ../../model.mnn
I1031 18:51:49.387892 3852557824 speaker_engine.cc:37] Embedding size: 256
I1031 18:51:49.387897 3852557824 speaker_engine.cc:39] per_chunk_samples: 32000
I1031 18:51:49.387902 3852557824 speaker_engine.cc:41] Sample rate: 16000
hw.cpufamily: 3660830781 , size = 4
The device support i8sdot:1, support fp16:1, support i8mm: 0
I1031 18:51:49.469084 3852557824 asv_main.cc:44] embedding size: 256
I1031 18:51:49.470297 3852557824 asv_main.cc:53] 4200
I1031 18:51:49.485879 3852557824 mnn_speaker_model.cc:65] dynamic shape.
I1031 18:51:49.810196 3852557824 asv_main.cc:62] 4200
I1031 18:51:49.948745 3852557824 asv_main.cc:65] compute score ...
I1031 18:51:49.948784 3852557824 asv_main.cc:67] Cosine socre: 1
I1031 18:51:49.948794 3852557824 asv_main.cc:69] It's the same speaker!