And here we are at the moment of truth. Let's see how I did. More importantly, let's compare GloVe vectors to learned vectors for this problem.

The orange line in the following screenshot corresponds to the learned embedded layer and the blue line corresponds to the GloVe vectors:

>

Not only does the GloVe pretrained network learn faster, but it also performs better, throughout every epoch. Overall these networks seem to do a good job learning the document classification task. They're both beginning to overfit after about the fifth epoch; however, the GloVe model is more robust against overfitting than the network trained without GloVe.

If you're working though these examples with me, I would recommend that you attempt the same problem with an LSTM. I think you'll find the problem more difficult to solve, and harder to manage overfitting, when using an LSTM.