DCN+

This is a PyTorch Implenetation of Dynamic Coattention Networks + described in the paper DCN+: MIXED OBJECTIVE AND DEEP RESIDUAL COATTENTION FOR QUESTION ANSWERING

A TensorFlow implementation can be found here

Dataset

The SQuAD dataset can be found here

Preprocessing data and training the Model

Preprocess the SQuAD data by running

python utils.py

Optional Arguments

   --train_file           path to training data for SQuAD
   --dev_file             path to development data for SQuAD
   --preprocessed_train   path to store the processed training data
   --preprocessed_dev     path to store the processed development data

To train the Model

python train.py

Optional Arguments

     --learning_rate              learning rate
     --hidden_size                hidden dimension for LSTM's and for Experts and Maxouts layers
     --pooling_size               pooling size for Maxout and Experts layers
     --dropout                    dropout for LSTM's
     --tokenized_train_data_path  path to the preprocessed training data
     --tokenized_dev_data_path    path to the preprocessed development data
 ```