quantized.out int8量化后的模型在CPU后端运行resnet50出错（已更新到最新commit版本）

[YingkunZhou]

#2614

初步测试mobilenet看起来是正常了许多，感谢开发者的努力！

但和上述issue同样的复现方法，我又用resnet50试了试，得到的结果如下：

模型下载地址：https://github.com/onnx/models/raw/main/Computer_Vision/resnet50_Opset16_timm/resnet50_Opset16.onnx

./pictureRecognition.out resnet50.mnn daisy.jpg
Load Cache file error.
The device support i8sdot:1, support fp16:1, support i8mm: 0
Session Info: memory use 29.272980 MB, flops is 4096.363281 M, backendType is 0, batch size = 1
input: w:224 , h:224, bpp: 3
origin size: 2100, 1500
For Image: daisy.jpg
672, -28.451855
657, -28.451855
658, -28.451855
659, -28.451855
660, -28.451855
661, -28.451855
662, -28.451855
663, -28.451855
664, -28.451855
665, -28.451855

有点好奇MNN里的量化框架有没有CI自动测试，为什么连最基本的reset50量化完都跑错

而未量化前的模型，被我命名为tmp.mnn （参考#2614 的操作流程），得到的分类结果就正常很多了

./pictureRecognition.out tmp.mnn  daisy.jpg                  
Load Cache file error.
The device support i8sdot:1, support fp16:1, support i8mm: 0
Session Info: memory use 169.978363 MB, flops is 4096.171875 M, backendType is 0, batch size = 1
input: w:224 , h:224, bpp: 3
origin size: 2100, 1500
For Image: daisy.jpg
985, 9.131153
113, -6.110043
506, -6.371115
308, -6.521829
987, -6.540517
714, -6.673302
307, -6.787087
698, -6.899653
783, -6.938828
652, -6.959727

[YingkunZhou]

继续用经典的模型efficientnetv2_b0试了试
模型的下载地址为：https://github.com/onnx/models/raw/main/Computer_Vision/tf_efficientnetv2_b0_Opset16_timm/tf_efficientnetv2_b0_Opset16.onnx

得到的量化完的运行结果：

./pictureRecognition.out tf_efficientnetv2_b0_Opset16.mnn daisy.jpg
Load Cache file error.
The device support i8sdot:1, support fp16:1, support i8mm: 0
Create execution error : 101
Create execution error : 101
Session Info: memory use 0.005398 MB, flops is 463.154877 M, backendType is 0, batch size = 1
input: w:192 , h:192, bpp: 3
origin size: 2100, 1500
Can't run session because not resized
For Image: daisy.jpg
21, 250908922840517956672101818055251722240.000000
971, 155706892288681663873593671155795886080.000000
558, 144564517325205176731988373268433207296.000000
952, 112714106392544388862589291026817482752.000000
485, 72253792300723863246554176667036680192.000000
280, 70477894939442528461738991265310048256.000000
255, 42867207357076265693574510722641035264.000000
930, 40476864538057105379101055316830191616.000000
234, 39606622401000425894973568566339567616.000000
39, 33638683099101550629631678460033236992.000000

这结果多少是有些抽象了

而未量化前的模型结果

./pictureRecognition.out tmp.mnn daisy.jpg  
Load Cache file error.
The device support i8sdot:1, support fp16:1, support i8mm: 0
Session Info: memory use 34.192852 MB, flops is 537.623291 M, backendType is 0, batch size = 1
input: w:192 , h:192, bpp: 3
origin size: 2100, 1500
For Image: daisy.jpg
985, 9.512913
89, 2.403510
322, 2.085096
108, 2.013274
883, 1.951369
309, 1.885692
113, 1.817128
968, 1.690546
770, 1.643703
738, 1.622023

看着就正常很多

[v0jiuqi]

我测试了结果没有不对啊，你用pictureRecognition_module.out 测试看看

[YingkunZhou]

@v0jiuqi 您确实是测试了量化后的模型了吗，能否贴一下您运行的分类结果让我看看，感谢！

模型就是上面提到的两个onnx官方仓库的，然后图片为

然后我是在arm64的开发板jetson orin （不是在x86的机器上，这一点也请注意）上进行编译运行的

[YingkunZhou]

具体的量化流程可以参考#2614

[YingkunZhou]

另外我想问一下，官方是打算放弃Session接口，改用Module接口了吗

[YingkunZhou]

我测试了结果没有不对啊，你用pictureRecognition_module.out 测试看看

@v0jiuqi 可否提供一下你用pictureRecognition_module.out 测试的config.json文件，感谢

[v0jiuqi]

这个是我们的量化工具和后端不一致导致的，你等我们更新吧

[YingkunZhou]

好的，感谢

[hebangwen]

这个是我们的量化工具和后端不一致导致的，你等我们更新吧

请问这个问题修复了吗

[github-actions]

Marking as stale. No activity in 60 days.