model test logs : "Attempting to upgrade batch norm layers using deprecated params". Issue?

[Jorge L]

When testing a modified model I found this log: 

upgrade_proto.cpp:77] Attempting to upgrade batch norm layers using deprecated params:  model.caffemodel

upgrade_proto.cpp:80] Successfully upgraded batch norm layers using deprecated params.

What would you say are the implications of this?

I may have caused this since I've been experimenting with the filter pruning algorithm  Pruning Filters  for Efficient ConvNets.(He et.al., 2017) . As part of the algorithm, I performed surgery in pretrained model; cut out a portion of the conv weights and the corresponding wights if the BatchNorm and Scale layers that follow.

 I dont trust my code is correct so Im worried the pruned model is corrupted somehow. Should I worry about this log?

Thanks!!

[Jonathan R. Williford]

Hi Jorge,

This is nothing to worry about. 

Caffe used to require lr_mult to be set to 0, 3 times, but no longer requires this. If there are three parameters set in your BatchNorm layer, Caffe just manually sets the lr_mult to be 0. It is for the mean, variance, and moving average parameters, which are never changed by back propagation, hence, they always need to be set to 0.

https://github.com/BVLC/caffe/issues/5120 (there's probably more relevant Github Issues on the topic, but this gives some background)

https://github.com/BVLC/caffe/blob/master/src/caffe/util/upgrade_proto.cpp#L1003

Best,

Jonathan

--
You received this message because you are subscribed to the Google Groups "Caffe Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to caffe-users+unsubscribe@googlegroups.com.
To post to this group, send email to caffe...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/2d4bc7fd-5d06-4f51-b93c-9f054c743f7d%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Jorge L]

Thanks for your help Jonathan ^^