onnx推理结果和mnn推理结果不一致，使用正确性校验出现TESTERROR

[QHenry1990]

平台(如果交叉编译请再附上交叉编译目标平台): PCx86

Platform(Include target platform as well if cross-compiling): ubuntu-20.04

Github版本:2.4.0

Github Version:

直接下载ZIP包请提供下载日期以及压缩包注释里的git版本(可通过7z l zip包路径命令并在输出信息中搜索Comment 获得，形如Comment = bc80b11110cd440aacdabbf59658d630527a7f2b)。 git clone请提供 git commit 第一行的commit id

Provide date (or better yet, git revision from the comment section of the zip. Obtainable using 7z l PATH/TO/ZIP and search for Comment in the output) if downloading source as zip,otherwise provide the first commit id from the output of git commit

编译方式: cmake

Compiling Method

请在这里粘贴cmake参数或使用的cmake脚本路径以及完整输出
Paste cmake arguments or path of the build script used here as well as the full log of the cmake proess here or pastebin

mkdir build; cd build; cmake ..; make -j 4

编译日志:

Build Log:

粘贴在这里
Paste log here or pastebin

python ../tools/script/testMNNFromOnnx.py ../hhquan_models/vadmodel2.onnx
onnx/test.onnx
tensor(float)
tensor(float)
tensor(float)
tensor(float)
tensor(float)
tensor(float)
tensor(float)
tensor(float)
['yn', 'cnn1_c', 'cnn2_c', 'cnn3_c', 'ht']
inputs:
xn
onnx/
cnn1_a
onnx/
cnn2_a
onnx/
cnn3_a
onnx/
cnn1_b
onnx/
cnn2_b
onnx/
cnn3_b
onnx/
h
onnx/
outputs:
onnx/yn.txt (1, 1, 1)
onnx/
onnx/cnn1_c.txt (1, 1, 40, 16)
onnx/
onnx/cnn2_c.txt (1, 1, 20, 16)
onnx/
onnx/cnn3_c.txt (1, 1, 20, 32)
onnx/
onnx/ht.txt (2, 1, 32)
onnx/
The device support dot:0, support fp16:0, support i8mm: 0
Start to Convert Other Model Format To MNN Model...
[14:00:09] /home/hhquan/Projects/MNN_test/MNN-2.4.0/tools/converter/source/onnx/onnxConverter.cpp:40: ONNX Model ir version: 7
[14:00:09] /home/hhquan/Projects/MNN_test/MNN-2.4.0/tools/converter/source/onnx/onnxConverter.cpp:108: Check it out ==> /rnn/GRU_output_0 has empty input, the index is 4
[14:00:09] /home/hhquan/Projects/MNN_test/MNN-2.4.0/tools/converter/source/onnx/onnxConverter.cpp:108: Check it out ==> /rnn/GRU_1_output_0 has empty input, the index is 4
Start to Optimize the MNN Net...
[14:00:09] /home/hhquan/Projects/MNN_test/MNN-2.4.0/tools/converter/source/optimizer/onnxextra/OnnxSequenceGRUMerge.cpp:68: OnnxSequenceGRUMerge: W shape:{1, 96, 320}; R shape: {1, 96, 32}, inputs num:6, outputs num:2
[14:00:09] /home/hhquan/Projects/MNN_test/MNN-2.4.0/tools/converter/source/optimizer/onnxextra/OnnxSequenceGRUMerge.cpp:68: OnnxSequenceGRUMerge: W shape:{1, 96, 32}; R shape: {1, 96, 32}, inputs num:6, outputs num:2
268 op name is empty or dup, set to Unsqueeze268
inputTensors : [ xn, cnn1_a, cnn1_b, cnn2_a, cnn2_b, cnn3_a, cnn3_b, h, ]
outputTensors: [ cnn1_c, cnn2_c, cnn3_c, ht, yn, ]
Converted Success!
Check convert result by onnx, thredhold is 0.01
xn
cnn1_a
cnn2_a
cnn3_a
cnn1_b
cnn2_b
cnn3_b
h
output: yn
output: cnn1_c
output: cnn2_c
output: cnn3_c
output: ht
yn: (1, 1, 1, )
TESTERROR yn value error : absMaxV:0.855153 - DiffMax 0.144592
Error for output yn
cnn1_c: (1, 1, 40, 16, )
cnn2_c: (1, 1, 20, 16, )
TESTERROR cnn2_c value error : absMaxV:3.355157 - DiffMax 3.234371
Error for output cnn2_c
cnn3_c: (1, 1, 20, 32, )
TESTERROR cnn3_c value error : absMaxV:2.566657 - DiffMax 1.898886
Error for output cnn3_c
ht: (2, 1, 32, )
TESTERROR ht value error : absMaxV:1.087269 - DiffMax 2.051899
Error for output ht
Save mnn result to .error director

[QHenry1990]

commit dcb080c

[jxt1234]

onnx 模型可以发一下?

[QHenry1990]

testmodel.zip
正确性自查，在Error is between /Transpose and /pool/MaxPool

[QHenry1990]

onnx 模型可以发一下?

模型已上传。

[jxt1234]

ok ，复现了，定位中

[jxt1234]

CPURaster singleConvert 优化的 bug ，修复中

[jxt1234]

2.4.2 修正

[QHenry1990]

辛苦了，非常感谢