How to generate customer Pooling layer prototxt in pycaffe? - machine-learning

Question

My goal is to achieve a pooling layer prototxt from code as follows:
layer {
name: "my_pooling"
type: "Pooling_Custom"
bottom: "conv1"
top: "my_pooling"
pooling_custom_param {
pool: MAX
kernel_size: 2
stride: 2
engine : CAFFE
}
}
where Pooking_Custom and pooling_custom_param are my modify Pooling. I want to use python to generate above output. I am using NetSpec to do it:
from caffe import params as P
n = caffe.NetSpec()
my_pooling = L.Pooling(conv1, type="Pooling_Custom", pool=P.Pooling.MAX, kernel_size=2, stride=2, engine=1)
However, I cannot generate pooling_custom_param. How can I do it? My current result is
layer {
name: "my_pooling"
type: "Pooling_Custom"
bottom: "conv1"
top: "my_pooling"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
engine : CAFFE
}
}

Answer 1

If you're doing this as a subtype of Pooling, you can't: those attribute names are defined by the parent template. All you can change are the values (right side of each item, after the colon).
The functional reason is that you're incorporating methods from the parent class in which those field names are hard-coded. For instance, there will be something that refers to pooling_param["kernel_size"]. If you have changed the name to pooling_custom_param, you need to also reconfigure that software ... and that's outside the language's software implementation.
If you really need to change that attribute name, you're welcome to clone the entire Pooling class, make the change, and instantiate layers of that new class.