Support batch normalization

[daylen]

Batch Normalization was added to Caffe in October 2015. (See BVLC/caffe#3229 and BVLC/caffe#3299). The caffe.proto file in this repo is from April 2015, so it doesn't know about Batch Normalization and can't load newer model definition files.

[szagoruyko]

thanks, are there models in caffe format with batchnorm available? should be quick to add, just need a test. resnets won't work.

[ghost]

@szagoruyko You're saying that you cannot use resnets for this test, or you're saying resnets won't work till you get this test working?

[szagoruyko]

@JoshPowell resnets have non-sequential structure which loadcaffe doesn't support now

[nhynes]

@JoshPowell you might want to give [caffegraph](https://github.com/nhynes/caffegraph) a try. It's not yet as fully featured as loadcaffe (it's missing a few modules), but it will load your resnet just fine.

[wuhuo714]

Hi, could loadcaffe load models in caffe format with batch normalization,now?

[szagoruyko]

@wuhuo714 no

[wuhuo714]

@szagoruyko Will batchnorm be available soon?

[szagoruyko]

@wuhuo714 don't have time unfortunately

[wuhuo714]

@szagoruyko Thanks a lot,anyway!

[wuhuo714]

Hi,@szagoruyko,if my caffemodel has batchnorm layers which loadcaffe can not recognize, can the following layers like Relu get its correponding weights?It has confused me a lot.

[szagoruyko]

@wuhuo714 no, it will ignore layers it can't recognize

[wuhuo714]

Although the layers it can't recognize are ignored, will the weights of these unrecognizable layers be used in the following layers it recoginze?
Is that the reason why the following layers like conv can not get their corresponding weights?

[szagoruyko]

@wuhuo714 weights of unrecognized layers will not be loaded at all

[wuhuo714]

@szagoruyko ok, it's very kind of you to answer my doubts!

[Fourure]

Hi @szagoruyko,
First thank you for your work, your package is very useful for me (and probably for lot of peoples).
I'm also interested by loading a caffe model with batch normalization. Is it still not available ?