Skip to content

Instantly share code, notes, and snippets.

@aneesh-joshi

aneesh-joshi/QA Transfer.md

Created July 19, 2018 14:59
Show Gist options
  • Star 0 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save aneesh-joshi/f2fe3459714ee3b57ef87820dad2b083 to your computer and use it in GitHub Desktop.
Save aneesh-joshi/f2fe3459714ee3b57ef87820dad2b083 to your computer and use it in GitHub Desktop.
Download ZIP
Raw
QA Transfer.md

Resources on QA-Transfer Model

QA-Transfer Model uses:

  • SQUAD-T dataset
  • BiDAF model (with end layers changed)

BiDAF moedel has 3 open source implementations:

I am currently working on getting a working BiDAF, hopefully in keras.

The original code

  • in tf 0.12.1
  • rains very slowly (6 seconds per iteration) without GPU
  • version mismatch of tf 0.11 and recent CUDA drivers
  • tf code is difficult to read and maintain

AlenAI code

  • is part of the DeepQA toolkit(now archived/deprecated/closed)
  • uses a lot of internal code (custom layer, models, etc)
  • very well documented

PyTorch code

  • small and contained within itself
  • written in pytorch, so, hard to read and maintain
  • I have less faimiliarity with PyTorch

Currently, I tried/am trying to port the original code from tf 0.12.1 to tf 1.3.0 There are some functions which have been removed in 1.3.0 and are making porting difficult. If this feels like it'll fail, I will move to the AllenAI code. Then the PyTorch code.

Links to papers

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment