Practical recommendations for gradient-based training of deep architectures