Reputation: 987
I want to take out a subset of n
parameters in a given layer from a trained network A
, and transfer this subset into a layer of another network B
.
In the layer to which this subset is transferred, the remaining parameters of said layer (total - transferred subset) should be randomly initialized.
Is this achievable? (I know it can be done with whole layers of parameters)
Thank you!
Upvotes: 1
Views: 325
Reputation: 114796
This can be done in python. See net surgery tutorial.
Basically, you should instantiate two caffe.Net
objects one with parameters of A
and the other with random. Then you simply assign the subset of parameters from A
to B
, saves B
and you are done.
Upvotes: 2