Caffe: how to get the phase of a Python layer?

2020-02-10 05:02发布

I created a "Python" layer "myLayer" in caffe, and use it in the net train_val.prototxt I insert the layer like this:

layer {
  name: "my_py_layer"
  type: "Python"
  bottom: "in"
  top: "out"
  python_param {
    module: "my_module_name"
    layer: "myLayer"
  }
  include { phase: TRAIN } # THIS IS THE TRICKY PART!
}

Now, my layer only participates in the TRAINing phase of the net,
how can I know that in my layer's setup function??

class myLayer(caffe.Layer):
  def setup(self, bottom, top):
     # I want to know here what is the phase?!!
  ...

PS,
I posted this question on "Caffe Users" google group as well. I'll udpdate if anything pops there.

2条回答
唯我独甜
2楼-- · 2020-02-10 05:31

As pointed out by galloguille, caffe is now exposing phase to the python layer class. This new feature makes this answer a bit redundant. Still it is useful to know about the param_str in caffe python layer for passing other parameters to the layer.

Original answer:

AFAIK there is no trivial way of getting the phase. However, one can pass arbitrary parameters from the net prototxt to python. This can be done using the param_str parameters of the python_param.
Here's how it's done:

layer {
  type: "Python"
  ...
  python_param {
    ...
    param_str: '{"phase":"TRAIN","numeric_arg":5}' # passing params as a STRING

In python, you get param_str in the layer's setup function:

import caffe, json
class myLayer(caffe.Layer):
  def setup(self, bottom, top):
    param = json.loads( self.param_str ) # use JSON to convert string to dict
    self.phase = param['phase']
    self.other_param = int( param['numeric_arg'] ) # I might want to use this as well...
查看更多
Melony?
3楼-- · 2020-02-10 05:37

This is a very good workaround, but if you are only interested in passing the phase as a parameter, now you can access the phase as an attribute of the layer. This feature was merged just 6 days ago https://github.com/BVLC/caffe/pull/3995.

Specific commit: https://github.com/BVLC/caffe/commit/de8ac32a02f3e324b0495f1729bff2446d402c2c

With this new feature you just need to use the attribute self.phase. For example you can do the following:

class PhaseLayer(caffe.Layer):
"""A layer for checking attribute `phase`"""

def setup(self, bottom, top):
    pass

def reshape(self, bootom, top):
    top[0].reshape()

def forward(self, bottom, top):
    top[0].data[()] = self.phase
查看更多
登录 后发表回答