If you want to create a deconvolution layer for bilinear upsampling, there's already a way to initialize weights appropriately, with no need to train anything. Take a look at
this code, just bear in mind that the proper weight filler type is
not bilinear_upsampling, simply just:
bilinear.