Ncnn: pytorch->onnx->ncnn调用，运算到中间某层后输出为空

[Wilbur529]

Hi，我已经通过直接调用onnx模型验证onnx模型没有问题。
onnx模型转ncnn过程中没有报任何异常，
在实际进行推理时，发现输出层的结果始终为空，但没有报错。
我尝试了一下for循环从1还是按blob_index访问输出，到中间某节点会异常退出：
Process finished with exit code 11
然后，如果单独输出该节点以后的任意节点，输出的blob mat都是空的。
想问一下，这有可能是什么原因？

[ckf1111]

遇到同样的问题，求大佬解答一下，谢谢

[soham24]

+1

[soham24]

it was found that the output layer results were always empty, :
I am facing the exact same issue. output layer always returns -100

[soham24]

@nihui

[soham24]

If this bug gets solved then pytorch-->onnx--->ncnn pipeline will work like charm for all modern arch.
looking forward for this issue to get resolve.

[Wilbur529]

@ckf1111 问题解决了：

1. ONNX转到ncnn的模型，param文件中第二列仍然是blob_name，所以调用的时候需要extractor.extract("123", outputMat);
2. 我的问题是因为UpSampling，ONNX中相应操作的参数是output scale，而NCNN中默认转成了output factor。而由于精度的问题，导致feature map尺寸到后面就对不上了，我把NCNN param中Interp的参数相应修改成output scale以后就好了。

[soham24]

Can you explain lil more @Wilbur529

[Wilbur529]

@soham24
Hi, I think you may fix the problem follow these steps:

1. Check your ONNX model by onnx_runtime package, make sure your format conversion from Pytorch to ONNX is well;
2. Try to extract the output blob of each node in ncnn.param, and compare with the related result from onnx_runtime. Maybe you can see what the problem node is. ^_^

[soham24]

Thank you @Wilbur529 :)

[zhoujinhai]

@soham24
Hi, I think you may fix the problem follow these steps:

1. Check your ONNX model by onnx_runtime package, make sure your format conversion from Pytorch to ONNX is well;
2. Try to extract the output blob of each node in ncnn.param, and compare with the related result from onnx_runtime. Maybe you can see what the problem node is. ^_^

你好 请问一下，用onnx_runtime检测ONNX模型有好的教程推荐的不？