seominjoon / memnn-tensorflow

End-to-end memory networks in TensorFlow

End-to-end memory networks in TensorFlow

TensorFlow implementation of Facebook's end-to-end memory networks. The original source code is implemented in MatLab (bAbI QA) and Torch (LM).

Paper: http://arxiv.org/abs/1503.08895

Original source code: https://github.com/facebook/MemNN

The paper presents results on two tasks; bAbI question answering and language modeling. Currently, this implementation runs bAbI QA only.

Requirements

pip install these packages:

tensorflow, progressbar

Quick start

To train and test on task 1, run:

python main.py --task 1

For more info and other options, run:

python main.py --help

Results

On a single run, I got:

Task 1 test error: 0.40%

Task 2 test error: 10.70%

Task 3 test error: 61.90%

Task 4 test error: 18.00%

Task 5 test error: 19.00%

Task 6 test error: 48.30%

Task 7 test error: 25.80%

Task 8 test error: 17.60%

Task 9 test error: 31.00%

Task 10 test error: 37.50%

Task 11 test error: 1.30%

Task 12 test error: 0.50%

Task 13 test error: 0.70%

Task 14 test error: 10.10%

Task 15 test error: 0.00%

Task 16 test error: 52.00%

Task 17 test error: 49.80%

Task 18 test error: 12.00%

Task 19 test error: 91.20%

Task 20 test error: 0.00%

Note that yes/no question task is performing poorly. Please pull request if you find solution to this!

Implementation details

The implementation does not use random noise (RN); reading the paper only, it wasn't clear to me. Also, I haven't tested joint learning, but you can simply run it by modifying the main.py file only.

memnn-tensorflow's Issues

AttributeError: 'NoneType' object has no attribute 'model_checkpoint_path'

envy@ub1404:/media/envy/data1t/os_prj/github/memnn-tensorflow$ python main.py --task 1
Loaded 1000 examples from: qa1_single-supporting-fact_train.txt
Loaded 1000 examples from: qa1_single-supporting-fact_test.txt
{'anneal_period': 25,
'anneal_ratio': 0.5,
'batch_size': 32,
'data_dir': 'data/tasks_1-20_v1-2/en/',
'draft': False,
'eval_num_batches': 3,
'gpu': False,
'hidden_size': 20,
'init_lr': 0.01,
'init_mean': 0,
'init_std': 0.1,
'linear_start': False,
'load': False,
'log_dir': 'log',
'ls_init_lr': 0.005,
'ls_num_epochs': 20,
'max_grad_norm': 40,
'max_ques_size': 3,
'max_sent_size': 6,
'memory_size': 50,
'num_epochs': 100,
'num_layers': 3,
'position_encoding': True,
'progress': True,
'save_dir': 'save',
'save_period': 10,
'task': 1,
'test_num_batches': 31,
'train': False,
'train_num_batches': 28,
'tying': 'adj',
'val_period': 5,
'val_ratio': 0.1,
'vocab_size': 20}
building N2NModel tower ...
/home/envy/os_pri/github/tensorflow/_python_build/tensorflow/python/ops/gradients.py:88: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
"Converting sparse IndexedSlices to a dense Tensor of unknown shape. "
loading model ...
Traceback (most recent call last):
File "main.py", line 97, in
tf.app.run()
File "/home/envy/os_pri/github/tensorflow/_python_build/tensorflow/python/platform/default/_app.py", line 30, in run
sys.exit(main(sys.argv))
File "main.py", line 93, in main
model.load(sess)
File "/media/envy/data1t/os_prj/github/memnn-tensorflow/models/base_model.py", line 112, in load
self.saver.restore(sess, checkpoint.model_checkpoint_path)
AttributeError: 'NoneType' object has no attribute 'model_checkpoint_path'
envy@ub1404:/media/envy/data1t/os_prj/github/memnn-tensorflow$