torch.cuda.FloatTensor(im) is really slow compared to torch.FloatTensor(im).cuda(), there should probably be either guards against that or some special case that doesn't do a for loop.