Bilinear Upsampling in Lasagne

bilinear_upsampling.py

import theano
import lasagne
import numpy as np

def upsample(layer, nb_kernels):

    def build_bilinear_kernel(ratio):
        half_kern = np.arange(1, ratio + 1)
        kern = np.concatenate([half_kern, half_kern[-2::-1]])

        kern = kern / ratio
        kern = kern[:,np.newaxis] * kern[np.newaxis, :]
        return kern

    kernel = build_bilinear_kernel(ratio=2) 
    kernel_shape = (nb_kernels, nb_kernels) + kernel.shape
    W_init = np.zeros(kernel_shape)
    W_init[range(nb_kernels), range(nb_kernels), :, :] = kernel
    W_init = np.float32(W_init)

    return lasagne.layers.TransposedConv2DLayer(layer, nb_kernels, 3, stride=(2, 2), W=W_init, b=None, nonlinearity=None) 

if __name__ == '__main__':
    from scipy.misc import face
    from pylab import *

    inp = theano.tensor.tensor4()
    l = lasagne.layers.InputLayer((1,3,256,342),inp)
    l = upsample(l, 3)
    print("Output shape:", lasagne.layers.get_output_shape(l))
    fn = theano.function([inp], lasagne.layers.get_output(l), allow_input_downcast=True)


    X = face()[np.newaxis,::3,::3].transpose((0,3,1,2))
    img = fn(X)[0,...].transpose((1,2,0))
    imshow(img/255.)
    show()