tspeterkim/attentive-reader

TensorFlow-based implementation of Stanford's Attentive Reader for the CNN/DailyMail datatset

attentive-reader's Issues

dataset is none

wget http://cs.stanford.edu/~danqi/data/cnn.tar.gz
tar xvf cnn.tar.gz
rm cnn.tar.gz

wget http://cs.stanford.edu/~danqi/data/dailymail.tar.gz
tar xvf dailymail.tar.gz
rm dailymail.tar.gz

When I want to run the code, the dataset isn't here.
Would you give me a copy? Please!

Double-check equation for m(t)

Is this line correct: https://github.com/peterkim95/attentive-reader/blob/master/main.py#L107
If I understand your implementation correctly, you are taking the hadamard product. And I don't see the tanh anywhere (not sure if that makes a big difference though).

M = d_output * tf.expand_dims(tf.matmul(q_output, W_bilinear), axis=1) # [batch, h] -> [batch, 1, h]

Should it be this:

M = d_output + tf.expand_dims(tf.matmul(q_output, W_bilinear), axis=1) # [batch, h] -> [batch, 1, h]
M = tf.nn.tanh(M)

Can you also share the results you were getting?

What final parameters you use in model

Hi, recently, I'm do the same project as this. However, I didn't get good accuracy. I;m wondering what hyper parameters you use for training model like # units, learning rate, epochs. Thank you!