residual basics in lasagne.py

from lasagne.layers import *
import theano
import numpy as np
def build_resmodule3(inp,filter_size=(3,3)):
    assert(len(inp.output_shape))==4
    n_filters = inp.output_shape[1]
    nn = Conv2DLayer(inp,n_filters,filter_size,pad='same')
    nn = batch_norm(nn) #inserts batchnorm before relu
    nn = ElemwiseSumLayer([nn,inp])
    return nn    

nn = l0 = InputLayer((None,3,32,32))
nn = Conv2DLayer(nn,32,(3,3)) #first convolution
nn = build_resmodule3(nn,(3,3)) #resnet block
nn = build_resmodule3(nn,(3,3)) #resnet block
nn = Conv2DLayer(nn,64,(1,1)) #1x1 convolution to increase num units
nn = Pool2DLayer(nn,(2,2))
nn = build_resmodule3(nn,(3,3)) #resnet block
nn = build_resmodule3(nn,(3,3)) #resnet block
nn = DenseLayer(nn,1,nonlinearity=lambda a:1./(1+2.71**a))

predict = theano.function([l0.input_var],get_output(nn))

predict(np.zeros((10,3,32,32),dtype=theano.config.floatX))