Improve pre-trained embeddings loading

[JohnGiorgi]

Previously, pre-trained embeddings were loaded by simply opening and looping over the file with python. This was a problem for a few reasons, mainly because it required the conversion of embeddings in C binary format (.bin) to C plain text format (.txt), which were much much larger.

TODO

• Use gensim's KeyedVectors.load_word2vec_format to load pre-trained embeddings into memory from a file
• Do a sanity check by running the model with this new loading mechanism and comparing performance to previously achieved performance
• Need to add back in the debug functionality, whereby passing the argument --debug would only load some N number of embeddings. Gensim has a mechanism for only loading this top N number of embeddings, use this!
• Unit tests are breaking, likely because my dummy_word_embeddings are not properly formatted. Maybe just use the Google300 ones?

Changed

• Pre-trained embeddings are now loaded using gensim's KeyedVectors.load_word2vec_format method, no need to convert from .bin to .txt. Saber can load any word embeddings in the word2vec format.
• ~43% speed up in loading pre-trained embeddings (one of Sabers big bottlenecks)
• .bin file is (at least in our case) is ~65% smaller than the corresponding .txt file
• No more confusing instructions for user about converting pre-trained embeddings from .bin to .txt format.

Resources

• this blog
• gensim documentation

Closes #15.

[coveralls]

Pull Request Test Coverage Report for Build 122

• 15 of 15 (100.0%) changed or added relevant lines in 2 files are covered.
• 2 unchanged lines in 1 file lost coverage.
• Overall coverage increased (+0.3%) to 75.236%

---

Files with Coverage Reduction	New Missed Lines	%
saber/sequence_processor.py	2	70.65%

Totals
Change from base Build 118:	0.3%
Covered Lines:	1197
Relevant Lines:	1591

---

💛 - Coveralls