TensorFlow implementation of Attentive Pooling Networks paper.
Welcome to post issues and pull requests.
- Python 2.7
- TensorFlow >= 1.8
Origin paper use one filter size in AP-CNN, but here we can use multiple filter sizes. CNN and LSTM should share weights, but should attentive pooling matrix U share or not ?
There are many details need to be considered in implementation.
- add evaluation metrics MAP MRR
- add other datasets
- deal with last smaller batch for prediction, now only work for predict_batch_size=1