Mxnet模型使用mxnet2ncnn的时候遇到问题

[Ackesnal]

在进行模型转换的时候报错：

~/ncnn/build/tools/mxnet$ ./mxnet2ncnn model-mobilenet-20180202-symbol.json model-mobilenet-20180202-0159.params
_minus_scalar not supported yet!
#scalar=127.5
_mul_scalar not supported yet!
#scalar=0.0078125
BlockGrad not supported yet!
L2Normalization not supported yet!
#mode=instance
_mul_scalar not supported yet!
#scalar=64.0
L2Normalization not supported yet!
#mode=instance
one_hot not supported yet!
#depth=85164
#off_value=0.0
#on_value=1.0
pick not supported yet!
#axis=1
_div_scalar not supported yet!
#scalar=64.0
_minus_scalar not supported yet!
#scalar=-0.87758256189
_mul_scalar not supported yet!
#scalar=0.87758256189
_rminus_scalar not supported yet!
#scalar=1.0
sqrt not supported yet!
_mul_scalar not supported yet!
#scalar=0.479425538604
elemwise_sub not supported yet!
_mul_scalar not supported yet!
#scalar=64.0
_minus_scalar not supported yet!
#scalar=15.3416172353
where not supported yet!
elemwise_sub not supported yet!
expand_dims not supported yet!
#axis=1
broadcast_mul not supported yet!

有很多mxnet模型里的操作似乎ncnn不支持。但是如果把这些操作去掉，改用in.substract_mean_normalize(mean_vals, norm_vals)等函数处理，那在mxnet2ncnn转换模型的时候又会提示Segmentation fault (core dumped) 「可能是因为模型文件和参数文件是一一对应的，无法只修改模型文件。但是参数是经过训练已经封装好了的，也不能修改。」

想请教一下大神这时候可以怎么办？

还有find_blob_index_by_name xxxx failed这个问题该如何解决？我在xxxx中填的和模型里的名字应该已经一样了，但还是说找不到，可能是什么引起？

[jeremycheong]

我也遇到这个问题，请问你那边解决了吗？

[pribadihcr]

+1

[Jancy1072]

+1

[Jancy1072]

ncnn工具有一直在更新，所以重新下载编译能解决

[datafarmers]

_minus_scalar not supported yet!

scalar=127.5

_mul_scalar not supported yet!

scalar=0.0078125

BlockGrad not supported yet!
BlockGrad not supported yet!
slice_axis not supported yet!

axis=0

begin=0

end=20

slice_axis not supported yet!

axis=0

begin=20

end=40

slice_axis not supported yet!

axis=0

begin=40

end=60

_plus_scalar not supported yet!

scalar=0.3

MakeLoss not supported yet!

我也遇到类似的问题

[TonyTangYu]

@Ackesnal 请问你解决了吗？

[nihui]

BlockGrad slice_axis MakeLoss 等训练用的 op 可以自定义 noop 注册解决
https://github.com/Tencent/ncnn/wiki/FAQ-ncnn-throw-error#layer-xyz-not-exists-or-registered