Don't support type [BatchNorm] onnx转化MNN失败

[gofugoo]

平台(如果交叉编译请再附上交叉编译目标平台):

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

MAC 上

Github版本:

Github Version:

https://github.com/alibaba/MNN/commit/3ea9dd1866e1c502f3632e1564b2fbc546559863

编译方式:

Compiling Method

cmake -GXcode -DMNN_BUILD_CONVERTER=ON ..

按照这个修复resize算子，还有如下下面错误。 https://github.com/alibaba/MNN/issues/1133

转化失败日志： onnxConverter.cpp:29: ONNX Model ir version: 6 Start to Optimize the MNN Net... Don't support type [BatchNorm], 222 Don't support type [BatchNorm], 222 Check failed: shape->dimensions() == 1 ==> "shape rank should be one" Check failed: indices->length(indicesDimension - 1) == 1 ==> "indices.shape[-1] = shape.rank" Check failed: updates->dimensions() == dimension ==> "updates dimension should be equal to given shape" Don't support type [BatchNorm], 222 Don't support type [BatchNorm], 222 Check failed: shape->dimensions() == 1 ==> "shape rank should be one" Check failed: indices->length(indicesDimension - 1) == 1 ==> "indices.shape[-1] = shape.rank" Check failed: updates->dimensions() == dimension ==> "updates dimension should be equal to given shape"

模型 yolov_simp.onnx.zip

[hjchen2]

@gofugoo 看你的模型中存在多个convolution输出concat之后计算BatchNorm的结构，这种网络结构在现在的MNN中不支持。由于MNN本身不支持BatchNorm算子，因此BatchNorm都需要fuse到前一个convolution中，但由于中间存在concat的情况，BatchNorm无法被fuse，转换就会失败。解决方法： 1、修改模型结构，把batchnorm挪到concat之前。 2、修改MNN的转换工具，把BatchNorm转换成scale + bias_add，（必要时需要修改一下convolution和batchnorm的fusion逻辑）。 3、等MNN更新，时间可能会比较长～

[gofugoo]

@hjchen2 谢谢回复，修改这个类把OnnxBatchNormTransform默认的转换修改成scale + bias_add ?

[hjchen2]

OnnxBatchNormTransform

是的

[gofugoo]

OnnxBatchNormTransform

是的 2、修改MNN的转换工具，把BatchNorm转换成scale + bias_add，（必要时需要修改一下convolution和batchnorm的fusion逻辑）。 大神能不能简单写一点代码粘贴上来，我对项目不是很熟？

[jxt1234]

Fixed