darts
darts copied to clipboard
architecture parameters initialization
Thanks for your great work! I find some inconsistency about the architecture parameters initialization between your paper and your code. In your paper, you said using zero initialization for architecture variables, which implies equal amount attention over all possible ops. However, in your code, you initialize the architecture variables randomly: ' Variable(1e-3*torch.randn(k, num_ops).cuda(), requires_grad=True)' in your code. Could you please explain why you use this random initialization?