This repo contains a Tensorflow implementation of Dynamic Coattention Network Plus.
Run pip3 install -r requirements.txt.
See the preprocessing directory.
In order to train a DCN+ model, simply run python train.py.
All of the relevant parameters for training are defined in params.json as follows:
model_directory: The directory where all of the graph and session information is saved
model_repository: The directory where copies of the model_directory are stored at the end of every epoch
train_directory: The directory that contains the training data
test_directory: The directory that contains the test data
n_epochs: The number of epochs that you want to run
start_epoch: The starting epoch (usually 1, but could be higher if you're resuming training for some reason)
train_batch_size: The batch size used for training
test_batch_size: The batch size used when computing predictions in testing
state_size: The state size for all RNNs in the model
max_sequence_length: The maximum sequence length for all RNNs in the model
use_mixed_loss: Flag to indicate whether to use the mixed loss. If false, just the cross-entropy loss is used.
keep_prob: The keep probability for dropout for all RNNs in the model
word_keep_prob: The keep probability for dropout as applied to words in the context document
embedding_length: The embedding length of the input data
lr: The learning rate of the optimizer
Using the training params in params.json, I was able to achieve 61.79% exact match score and 72.56% F1 score on the dev dataset. The model directory corresponding to this model is stored in this repo as pretrained. Make sure to clone the repo using git lfs clone if you want to use the pretrained model.