Use only 1 hasher

[guillaume-chevalier]

Instead of using a FeatureUnion over T=80 random hashers of d=14 dimensions (80*14=1120 word features), use only one hasher (1x1120), which results in a dramatic speedup.

You can see the fix here: https://github.com/guillaume-chevalier/NLP-TP3

[guillaume-chevalier]

(note: the linked repo will become public soon and isn't yet public)

[guillaume-chevalier]

I profiled the execution time of the pipeline. The speedup to using only 1 hasher rather than 80 is quite good.

Also, it's weird that using threads on hashers (rather than 1 thread) was performing slower. I tested on a 32 core computer to use many threads for the hashers in parallel (n_jobs=-1 param in feature union), and it was even slower than using no thread.

[tnlin]

Hi, have you replicate the model in their paper?
I found their experiment results are too good to be true. In my experiment, I use BERT with context to achieve 81% accuracy, so I wonder how did they achieve their score (SwDA acc=83% without context?)
Besides, their talk on EMNLP said ATIS has intent "purchase", but I never see any ATIS dataset have "purchase" intent...

[guillaume-chevalier]

@tnlin I did not implement the neural network layer on top of the projection layer. This is only the projection, and it also differs a little bit from the paper.

[glicerico]

Hey @tnlin , did you manage to find the real performance of SGNN? I can only achieve 71% with their architecture.