thaihungle / UW-DNC

Uniform Writing & Cached Uniform Writing
MIT License
16 stars 3 forks source link

UW-DNC

source code for paper Learning to Remember More with Less Memorization https://openreview.net/forum?id=r1xlvi0qYm
reference: https://github.com/Mostafa-Samir/DNC-tensorflow

Synthetic tasks

Training:

  1. regular DNC:
    python synthetic_task.py --task=copy --mode=train
  2. UW DNC:
    python synthetic_task.py --task=copy --mode=train --hold_mem_mode=2
  3. CUW DNC: python synthetic_task.py --task=copy --mode=train --hold_mem_mode=2 --cache_attend_dim=16 --cache_size=10 --hidden_dim=100

Testing:
--mode=test

Random writing:
--memo_type=random

Feel free to play with other parameters