PyTorch 0.4, torchtext 0.2.3, All Features Implemented

• Fixed bug in horizontal comparison algorithm - implementation used to be different from paper
• Fixed use of double NLLLoss (since CrossEntropyLoss already include it)
• Uses latest versions of PyTorch and torchtext (PyTorch 0.4.0 and torchtext 0.2.3)

This is the version of the code used for my thesis.

Full Implementation of MP-CNN with Sparse Features and Attention

This release does not use torchtext.

For the SICK dataset, running python main.py mpcnn.sick.model.castor --dataset sick --epochs 19 --epsilon 1e-7 --dropout 0, one can get Pearson's p to be 0.8744 and Spearman's r to be 0.8183, slightly better than the results obtained in the paper (0.8686 and 0.8047).

For the MSRVID dataset, running python main.py mpcnn.msrvid.model.castor --dataset msrvid --batch-size 16 --epsilon 1e-7 --epochs 32 --dropout 0 --regularization 0.0025, one can get Pearson's p to be 0.9072, very close to the performance in the paper (0.9090).