does anyone successfully run mobilefacenet with ncnn on android armv7?

[yc-huang]

there's no errors when converting the mxnet model, but at runtime it will crash with SIGSEGV when get output from blob 171 or 170 (layer fc1 or pre_fc1); but if changed to get output of blob 169(layer conv_6dw7_7_batchnorm) it seems to work withot errors.

you might get the mxnet checkpoint from:
https://pan.baidu.com/s/1If28BkHde4fiuweJrbicVA

[simon-rob]

I haven't looked at the code... but I suspect it is a bug with the converter as pre_fc1 and fc1 share the same weights. I am not sure if the converter supports shared weights.

[yc-huang]

@simon-rob Thanks for your information. I have a rough go thru of the mxnet2ncnn.cpp, it seems to have weights output to 'ncnn.bin' for each node seperately, so I guess weights are not shared in 'ncnn.bin' anymore?

[yc-huang]

looks like any mxnet network contains FullyConnected layer might have similar issue.
I tried with a simple lenet, which could reproduce this: you can't get output from any layer after the flatten layer...
lenet_mnist.tar.gz

Haven't tried similar network trained on other framework yet...
@nihui do you have any suggestion on this? Thanks...

[yc-huang]

This was caused by over-sized input. Mobilefacenet expect input size of 112x112x3, so crop input image to this size will work, otherwise it might encounter errors.