DNC-theano : A theano implementation of DeepMind's DNC model

What's this

This is a naive implementation of DeepMind's Differentiable Neural Computer model.

Requirements

• python3

• numpy

• theano

Usage

Clone this repository first. Then run dnc.py script in interactive mode.

NOTE: If you don't run in interactive mode it will just quit silently.

$ python3 -i dnc.py
Compiling ... done
>>>

Compilation in theano is a bit slow so it may take a while. Once compiled, you can just call python functions defined in dnc.py. The default training task is the conditional copy task, the network must produce input signal in same or reversed order, depending on the first "condition" signal. You may want to modify the code to introduce new tasks.

Most useful functionalities are shown below:

>>> train() #train a new model, model is saved in "params.pkl" every 100 iters
Iter 100/10000 loss: 0.040833
Iter 200/10000 loss: 0.034119
Iter 300/10000 loss: 0.030109
Iter 400/10000 loss: 0.028633
...
Iter 10000/10000 loss: 0.004308
>>> g_optimizer.lr = 5e-4 #change learning rate
>>> train(4,6) #customize training plan
Iter 6/24 loss: 0.004296
...
Iter 24/24 loss: 0.004265
>>> save_params() #save model parameters manually
>>> exit()
$ python3 -i dnc.py #go back
>>> load_params() #load previously saved model parameters
>>> X,Y = gen_episode() #generate a test data point
>>> fn_rst() #reset DNC states
>>> Y_pred = predict(X) #make a prediction with DNC
>>> import matplotlib.pyplot as plt
>>> plt.imshow(Y_pred+X); plt.show() #make a visualization of DNC prediction

Bugs/Issues

• 'nan' crash on rare occasion