rename a layer?

[G Voysey]

I have an existing *.prototext and *.caffemodel. 

Unfortunately, the final layer's name is "class", which causes all sorts of trouble with pycaffe (because "class" is a reserved word).

I'd like to rename this to "class_".

Here's how i'm trying it now:

  def convert(prototext, new_prototext, model, newmodel):
      net_old = caffe.Net(prototext, model, caffe.TEST)
      net_new = caffe.Net(new_prototext, model, caffe.TEST)
      net_new = update_key(net_new)

      for i in range(0,2):
          net_new.params['class_'][i].data[...] = net_old.params['class'][i].data[...]
      net_new.save(newmodel)


  def update_key(net):
      net.params['class_'] = net.params['class']
      net.params.pop('class')
      return net

This executes cleanly.  But, the newly saved prototext has not changed the name of "class" to "class_".  I can inspect the state of net_new at the end of convert() and determine that the key in net_new.params is really different... what's going on?

[Przemek D]

I don't think you can do this with your net instantiated. Editing the params list does not modify the layers - that information is held in layers and _layer_names, but I'm not sure right now if you can modify them.

The simplest approach here would be to manually change "class" to "class_" in the prototext, and then continue with the weight transplant like you already do. If you want to automate this though, I suppose you should work with the NetSpec. Unfortunately the documentation is scarce, maybe this snippet will guide you onto the right direction.