QANet

Replaces RNNs with depthwise separable convolution and self-attention layers for faster training and inference, aligning with the paper's efficiency goals as stated in the philosophy.

Includes scripts for data preprocessing, training, testing, and an interactive demo server, adapted from R-Net for workflow efficiency, with modes configurable via config.py.

Provides a results table comparing performance with the original paper and explains adaptations like reduced hidden size and single-head attention due to GPU constraints.

Employs dropout, stochastic depth dropout, and exponential moving average for stable training, as detailed in the implementation section to prevent overfitting.