Freezing glove word embeddings for few epochs about pytorch-sentiment-analysis

[bentrevett]

Great idea. I think I'll do something like freezing the embedding weights until the validation loss stops decreasing and then unfreeze the embedding weights. Sound good?

[404akhan]

Sounds reasonable, yeah try it. Although most papers i've seen did different things. But this also looks good, at least seeing basic functionality of freeze-unfreeze under certain condition will be already good for people using tutorial.

[bentrevett]

OK, I've implemented a notebook appendix C which covers how to freeze/unfreeze embeddings (and a few other things).

In the implementation it freezes the embeddings for the first 5 epoches and then unfreezes them for the next/last 5. I also added a code snippet on how to unfreeze when the validation loss stops decreasing.

Let me know what you think.