Tune Kim CNN for SST-2 and Improve SST-1 Results with Dataset/Initialization Changes

castorini / castor

PyTorch deep learning models for text processing

http://castor.ai/

Apache License 2.0

178 stars 58 forks source link

Closed tuzhucheng closed 6 years ago

tuzhucheng commented 6 years ago

Update Kim CNN Dataset to use default min_freq=1 and use Uniform[-0.25, 0.25] distribution for unknown tokens (following original implementation by Yoon Kim)
Tune results for SST-2
Re-tune SST-1

Test Accuracy on SST-1	rand	static	non-static	multichannel
Paper	45.0	45.5	48.0	47.4
PyTorch using above configs	44.3	47.9	48.6	49.2

Test Accuracy on SST-2	rand	static	non-static	multichannel
Paper	82.7	86.8	87.2	88.1
PyTorch using above configs	83.0	86.4	87.3	87.4

tuzhucheng commented 6 years ago

Thanks to @EmilyKeer for some help tuning SST-2.

Related to #58