Update README.md

README.md

@@ -8,9 +8,17 @@ In the paper, DMN does bAbi task well with strongly supervised settings. But in
 I've chosen weakly supervised setting - training DMN without supporting facts information in bAbI.
 
 ## Prerequisites
-- Tensorflow
+- Python 3.x
+- Tensorflow 0.8+
 - Numpy
-- [tqdm](https://pypi.python.org/pypi/tqdm)
+- [tqdm](https://pypi.python.org/pypi/tqdm) - Progress bar module
+
+## Usage
+First, Install it.
+```
+sudo pip install tqdm
+git clone https://github.com/therne/dmn-tensorflow
+```
 
 ### Training the model
 ```
@@ -23,7 +31,46 @@ I've chosen weakly supervised setting - training DMN without supporting facts in
 ```
 
 ### Results
-TODO: Add test result
+
+*Model trained with 40 hidden units, semantic memory, 5 memory step, 300 epoches and learning rate 2e-2.*
+
+Task    | Result
+--------|-------
+1. Single supporting facts | 100%
+2. Two supporting facts    | 92.44%
+3. Three supporting facts  | 92.44%
+4. Two arguments relations | 98.89%
+5. Three arguments relations | 98.79%
+6. Yes-No Questions        | 99.09%
+7. Counting                | 98.39%
+8. List/Sets               | 99.19%
+9. Simple negotiation      | 98.49%
+10. Indefinite knowledge   | 94.86%
+11. Basic coreference	     | 99.50%
+12. Conjuction             | 98.69%
+13. Compound coreference   | 99.09%
+14. Time reasoning         | 98.19%
+15. Basic deduction        | 93.75%
+16. Basic induction        | 94.05%
+17. Positional reasoning   | 96.17%
+18. Size reasoning         | 100%
+19. Path finding           | 77.72%
+20. Agent’s motivations    | 100%
+Average                    | 96.48%
+
+Result comes close to the fully-supervised settings. 
 
 ### References
-- [Implementing Dynamic memory networks by YerevaNN](https://yerevann.github.io/2016/02/05/implementing-dynamic-memory-networks/) - Original, and great article
+- [Implementing Dynamic memory networks by YerevaNN][impl-dmn-yerevann] - Original work, great article that helped me a lot
+- [Dynamic-memory-networks-in-Theano][dmn-in-theano] - Original work.
+- [memnn-tensorflow][memnn-tensorflow] - Base code (trainer, loader)
+
+### TO-DO
+- Use [`tf.nn.dynamic_rnn`][dynamic-rnn-docs] for memory optimization
+- Use GRU in answer module to generate answer sequence
+- Interactive mode?
+
+[impl-dmn-yerevann]: https://yerevann.github.io/2016/02/05/implementing-dynamic-memory-networks/
+[dmn-in-theano]: https://github.com/YerevaNN/Dynamic-memory-networks-in-Theano
+[memnn-tensorflow]: https://github.com/seominjoon/memnn-tensorflow
+[dynamic-rnn-docs]: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/g3doc/api_docs/python/functions_and_classes/shard8/tf.nn.dynamic_rnn.md