Hey guys,
Does the model which gets 84.5 on SNLI test set use dropout(in the encoder and fc layers)? If so, what's the dropout rate ?
Also, did you guys experiment with fc layers with more neurons than 512 ? Does it help ?
Hi,
dropout didn't help much in our case no. The default setting uses a number of hidden units that was found to work best on validation data.
Best,
Alexis
Hey guys, Does the model which gets 84.5 on SNLI test set use dropout(in the encoder and fc layers)? If so, what's the dropout rate ? Also, did you guys experiment with fc layers with more neurons than 512 ? Does it help ?
And thanks so much for sharing the code !