source code for dual memory neural computer
arXiv version: https://arxiv.org/abs/1802.00662
KDD version: http://www.kdd.org/kdd2018/accepted-papers/view/dual-memory-neural-computer-for-asynchronous-two-view-sequential-learning
reference: https://github.com/Mostafa-Samir/DNC-tensorflow
cd to dual_task folder
run command examples:
train concatenate LSTM>>python3 sum2_run.py --name=single --mode=train --seq_len=10 --use_mem=False
train concatenate DNC>>python3 sum2_run.py --name=single --mode=train --seq_len=10 --use_mem=True
train Dual LSTM>>python3 sum2_run.py --mode=train --seq_len=10 --use_mem=False --attend=0
train WLAS>>python3 sum2_run.py --mode=train --seq_len=10 --use_mem=False --attend=128
train DMNC_l>>python3 sum2_run.py --mode=train --seq_len=10 --use_mem=True --share_mem=False
train DMNC_e>>python3 sum2_run.py --mode=train --seq_len=10 --use_mem=True --share_mem=True
test: use --mode=test
Learning curves:
Please prepare the EMR data as described in the paper, which includes:
- list of 3 dictionaries mapping from token to view code
- list of 3 dictionaries mapping from view code to token
- list of patients, each consists of list of admissions, each consists of 3 sequences corresponding to 3 views
Please modify the code in emr_run.py to point to your data location
run command examples:
train Dual LSTM>>python3 emr_run.py --mode=train --seq_len=10 --use_mem=False --attend=0
train WLAS>>python3 emr_run.py --mode=train --seq_len=10 --use_mem=False --attend=128
train DMNC_l>>python3 emr_run.py --mode=train --use_mem=True --share_mem=False
train DMNC_e>>python3 emr_run.py --mode=train --use_mem=True --share_mem=True
test: use --mode=test --from_checkpoint=default
Feel free to modify the hyper-parameters