search

galsang / BIMPM-pytorch

Re-implementation of BIMPM (Bilateral Multi-Perspective Matching for Natural Language Sentences, Zhiguo Wang et al.) on Pytorch.
106 stars 31 forks source link

Less accuracy in reimplementation #5

Closed amirkargar closed 6 years ago

amirkargar commented 6 years ago

Hi,

I noticed you are only limiting the number of words in the train phase, not in test/dev phase. Could it be the reason the reimplementation accuracies are less than the reported ones from the papers? or they do the same?

Thanks, Amir

galsang commented 6 years ago

I don't understand which part you are mentioning. In case of word embedding (utils.py, line 15 for SNLI), self.TEXT.build_vocab(self.train, self.dev, self.test, vectors=GloVe(name='840B', dim=300)) makes embedding matrix contains all of the words in train/dev/test datasets but only words in each dataset are used for each phase because of the nature of the embedding itself.