Deep Learning with Keras

Fine-tune learned embeddings from GloVe

Fine tuning using pre-trained GloVe embeddings is very similar to fine tuning using pre-trained word2vec embeddings. In fact, all of the code, except for the block that builds the weight matrix for the embedding layer, is identical. Since we have already seen this code twice, I will just focus on the block of code that builds the weight matrix from the GloVe embeddings.

GloVe embeddings come in various flavors. We use the model pre-trained on 6 billion tokens from the English Wikipedia and the gigaword corpus. The vocabulary size for the model is about 400,000, and the download provides vectors of dimensions 50, 100, 200, and 300. We will use embeddings from the 300 dimensional model.

The only thing ...