Dynamic-memory-networks-plus-Pytorch
DMN+ implementation in Pytorch for question answering on the bAbI 10k dataset.
Contents
| file | description |
|---|---|
babi_loader.py |
declaration of bAbI Pytorch Dataset class |
babi_main.py |
contains DMN+ model and training code |
fetch_data.sh |
shell script to fetch bAbI tasks (from DMNs in Theano) |
Usage
Install Pytorch v0.1.12 and Python 3.6.x (for Literal String Interpolation)
Run the included shell script to fetch the data
chmod +x fetch_data.sh
./fetch_data.sh
Run the main python code
python babi_main.py
Benchmarks
Low accuracies compared to Xiong et al's are may due to different weight decay setting or the model's instability.
On some tasks, the accuracy was not stable across multiple runs. This was particularly problematic on QA3, QA17, and QA18. To solve this, we repeated training 10 times using random initializations and evaluated the model that achieved the lowest validation set loss.
You can find pretrained models here
| Task ID | This Repo | Xiong et al |
|---|---|---|
| 1 | 100% | 100% |
| 2 | 96.8% | 99.7% |
| 3 | 89.2% | 98.9% |
| 4 | 100% | 100% |
| 5 | 99.5% | 99.5% |
| 6 | 100% | 100% |
| 7 | 97.8% | 97.6% |
| 8 | 100% | 100% |
| 9 | 100% | 100% |
| 10 | 100% | 100% |
| 11 | 100% | 100% |
| 12 | 100% | 100% |
| 13 | 100% | 100% |
| 14 | 99% | 99.8% |
| 15 | 100% | 100% |
| 16 | 51.6% | 54.7% |
| 17 | 86.4% | 95.8% |
| 18 | 97.9% | 97.9% |
| 19 | 99.7% | 100% |
| 20 | 100% | 100% |