CS224n (2017) assignment1 coding problems
Some of this code is adapted from anothers' solutions to the problem set from two or three years ago. (I'll upload my original version when I have time to annotate it.) I had hoped that the instant code would be more tractable to parallelizing, but that achievement remains elusive.
Notes:
The cross-entropy cost function, with the skip-gram model, seems to plateau at around 9.4. Costs with the cbow model are substantially lower, but that may stem from a coding error. Note that to run the cbow model you must replace 'skipgram' with 'cbow' in the invocation of the word2vec_sgd_wrapper contained in q3_run.py.
None of this may be correct.
For the sentiment problem, the best regularization (determined on dev set) gave test accuracies of 29.5% (skipgram) and 28.9 (cbow) compared to 37.1% with pretrained word vectors.