Fully connected multi layer perceptron using PyCaffe

Question

I'm new to Caffe, and its workflow is very different from what I've previously encountered. I've used keras, sklearn, fann (C++) for neural networks before, and I want to use Caffe because of some additional things it offers. But the workflow seems hard to adjust to.

I want to start with a simple, fully connected MLP using PyCaffe. I want to feed it an N-dimensional input vectors and do multi label classification on those. I have the training data. All the Caffe examples seems to be written for images (square matrix inputs).
I also prefer to configure the network programmatically, as opposed to using a lot of configuration files. For example, Keras had a method to sequentially stack layers using add() .

Is it possible to construct a simple network in Caffe using only Python?

Answer 1

You should look into caffe.NetSpec() interface: this allows you to construct a net programatically. For example:

from caffe import layers as L, params as P, to_proto
import caffe

ns = cafe.NetSpec()

ns.fc1 = L.InnerProduct(name='fc1', inner_product_param={'num_output':100,
                                                         'weight_filler':{'type':'xavier','std':0.1},
                                                         'bias_filler':{'type':'constant','value':0}},
                                    param=[{'lr_mult':1,'decay_mult':2},
                                           {'lr_mult':2,'decay_mult':0}])
ns.relu1 = L.ReLU(ns.fc1, inplace=True)