word embeddings should be fixed during training?
wirehack opened this issue · 1 comments
wirehack commented
In your paper Learned in Translation: Contextualized Word Vectors (McCann et. al. 2017), it says
We used the CommonCrawl-840B GloVe model for English word vectors, which were completely fixed during training, so that the MT-LSTM had to learn how to use the pretrained vectors for translation.
However, in your code
Lines 29 to 33 in 67b690b
the Glove embeddings are not fixed.
bmccann commented
Right, in my paper, I found that the weights should be fixed for better performance. I just wanted to leave this a little more flexible in case others wanted to try fine-tuning on different tasks.