Entry for machine learning
Jupyter Notebook Python
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.

README.md

Siraj_Chatbot_Challenge

Entry for machine learning tutorial How to Make a Chatbot - Intro to Deep Learning #12

Dependencies

  • python (>=3.5) - Developed on 3.5. Untested on 2.7, 3.4 or 3.6, but ought to work on v. >=3.5
  • pandas (>=0.19) Needed for some data sorting operations
  • keras (2.0.2/2.0.3) - Deep learning FTW
  • tensorflow (1.0.1) - My preferred Keras backend. Tensorboard is now integrated, so it'll need some tweaking to be Theano-compatible again
  • h5py - for model checkpointing and saving weights
  • keras-tqdm - because my Jupyter notebooks freezes on the default Keras progbar. Also, it's awesome.

Running

To run the command line interface, just type: python main.py. I have provided pretrained weights for the first model, challenge 1 (default location ./models/c1/dmn00.hdf5). If this is the first time you are running the program with flags, or you just created a new model, you'll have to train it first, which you can do directly from the menu. If you do not have any trained models yet, you can select f from the menu to fit the model.

Arguments

  • -m {modelname} - Set the name of the model and weight save file
  • -c {N} - Run challenge mode N. 1 is Single context bAbI, 2 is Double context bAbI. You can now select from any of the 20 bAbI Q/A tasks!
  • -a {N}- Run architecture N. 1 is bAbI DMN, 2 a more conventional convolutional LSTM (warning: hard on memory)
  • -v - Verbose flag

Example uses:
python main.py -c 2 to switch to double supporting facts dataset
python main.py -m modelname.hdf5 to specify a custom model name. Note that the software automatically places these in the folders models/c1/ or models/c2/ depending on the dataset.

If you want to suppress some of the TF notifications and the progbars, you can append 2> /dev/null to redirect that junk.

Note on Challenge 2: Two Supporting Facts

There are actually two challenges that came with the Q/A task, the single supporting fact, and the double supporting facts. The former is pretty easy to knock out of the park, while the latter has proven quite stubborn. I was able to get >95% training accuracy but only 35-40% validation accuracy, a surefire sign of overfitting. I tried some clever hacks with the network but I was not able to improve results. The authors claim that they aced the two supporting fact problem, but the Keras code as provided seems to fall short. Meh.

Network Improvements:

Here are some improvements I made to the demo network:

Convo LSTM

I added the option to compare against a convolutional LSTM architecture. So far, kind of middling results. Needs to be configured for minibatch.

Bidirectional LSTM

The single forward-pass LSTM was converted to bidirectional layer with the Bidirectional wrapper. Yuuuuuge improvement on double-context task - 84.7% (training acc) after 260 epochs with single, improved to 90% after only 110 epochs with bidirectional. Nice! Asymptoted to 95% after about 150 epochs. However, I later realized these figures were pretty misleading, as the validation was not keeping pace with the training accuracy.

The Single-context task got to 90% validation accuracy after 60 vs 85 epochs, modest improvement.

Time-Distributed Dense

Adding a TDD layer before the LSTM gave an additional jump in terms of training time and overall accuracy, reaching 95% valacc after 65 epochs on single-context (with default 32 nodes).

Not-so-improvements

Conv1D in the Match layer

Who doesn't love convo layers? Hoping to get better context recognition, I put a convolayer after the Match dot product part of the network. It didn't hurt the performance, but it didn't give the gains in the Challenge 2 I was looking for.

Third LSTM layer

Adding a forward pass after the bidirectional pair did not give improvements, in fact it caused the network to stall out around 55%. I've seen towers of LSTMs used to good effect in other NLP papers. Maybe they have some secret sauce I don't.

Navigating

The main menu looks like this:

------------------------------
Sandra went back to the kitchen.
Sandra journeyed to the garden.
Mary went back to the kitchen.
Sandra went to the kitchen.
------------------------------

..: Back
 1: Load Random Story
 2: Query
 3: Query (loop)
 f: Fit for N epochs
 q: Quit
Enter menu selection: 

The currently loaded story is shown at the top. Enter letters to navigate the menu.

  • .. is currently non-functional
  • 1 loads a new story.
  • 2 lets you type in a query. It goes back to the main menu after
  • 3 like 2, but brings you back to the query prompt after, for convenience
  • f lets you enter in a number to fit the model for that many epochs.