Differentiable Quantization of Deep Neural Networks