Based on the paper" Very Deep Convolutional Networks for Natural Language Processing "(https://arxiv.org/abs/1606.01781) in tensorflow.
Model parameters are initialized using the Xavier Glorot and Yoshua Bengio (2010) initializer: Understanding the difficulty of training deep feedforward neural networks. International conference on artificial intelligence and statistics.
Model trained and tested on a Tesla K80 gpu: - 32min per epoch - Test accuracy of 0.975 (12 epochs)