QA-Transfer Model uses:
- SQUAD-T dataset
- BiDAF model (with end layers changed)
BiDAF moedel has 3 open source implementations:
- AllenAI-keras
- Original-BiDAF-tf-0.11 and Original-QA-Transfer-tf-0.11 (QA-Transfer essentially forks the first repo and makes some changes to it.)
- PyTorch
I am currently working on getting a working BiDAF, hopefully in keras.
- in tf 0.12.1
- trains 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 (look at code in the link above to get an idea)
- is part of the DeepQA toolkit(now archived/deprecated/closed)
- uses a lot of internal code (custom layer, models, etc)
- very well documented
- small and contained within itself
- written in pytorch, so, hard to read and maintain(look at code in the link above to get an idea)
- 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.
To get a working QA-Transfer model before the meeting on 23 July, 2018.