Parameter Efficient Training of Deep Convolutional Neural Networks by Dynamic Sparse Reparameterization