MemN2N

Implementation of End-To-End Memory Networks with sklearn-like interface using Tensorflow. Tasks are from the bAbl dataset.

Get Started

git clone git@github.com:domluna/memn2n.git

mkdir ./memn2n/data/
cd ./memn2n/data/
wget http://www.thespermwhale.com/jaseweston/babi/tasks_1-20_v1-2.tar.gz
tar xzvf ./tasks_1-20_v1-2.tar.gz

cd ../
python single.py

Examples

Running a single bAbI task

Running a joint model on all bAbI tasks

These files are also a good example of usage.

Requirements

tensorflow 0.8
scikit-learn 0.17.1
six 1.10.0

Results

For a task to pass it has to meet 95%+ testing accuracy. Measured on single tasks on the 1k data.

pass: 1,4,12,15,20

Several other tasks have 80%+ testing accuracy.

Unless specified, the Adam optimizer was used.

The following params were used:

epochs: 200
learning_rate: 0.01
epsilon: 1e-8
embedding_size: 20

A joint model was also run with the following params:

epochs: 100
learning_rate: 0.01
epsilon: 1.0
embedding_size: 40

Full results.

Task	Training Accuracy	Validation Accuracy	Testing Accuracy
1	0.97	0.95	0.96
2	0.84	0.68	0.61
3	0.5	0.4	0.3
4	0.93	0.94	0.93
5	0.91	0.76	0.81
6	0.78	0.8	0.72
7	0.87	0.86	0.8
8	0.82	0.8	0.77
9	0.76	0.73	0.72
10	0.71	0.66	0.63
11	0.9	0.87	0.89
12	0.93	0.92	0.92
13	0.93	0.88	0.93
14	0.88	0.87	0.76
15	1.0	1.0	1.0
16	0.57	0.42	0.46
17	0.7	0.61	0.57
18	0.93	0.96	0.9
19	0.12	0.07	0.09
20	1.0	1.0	1.0

Notes

I didn't play around with the epsilon param in Adam until after my initial results but values of 1.0 and 0.1 seem to help convergence and overfitting.