about class StateSpace

[gzpan]

At class StateSpace in Controller.py, many functions e.g. embedding_encode(self, id, value) use state = self[id], I am confused that if it is legal in python ? how could we index self?
another question: why the variable state here is a Dictionary? shouldn't it be a StateSpace object?

[Somshubra Majumdar]

I've declared the getitem method inside of StateSpace. After this, the entire class can behave as an array, and therefore be indexed just as above. This is just a convenience function, you can of course simply uses the expanded version as below.

def __getitem__(self, id):
        return self.states[id % self.size]

The benefit is that when the id is being received by the caller, it can be from multiple cells in a loop, so I don't need to worry about which cyclic index this is when I use self[id]. It's a somewhat roundabout way to do something safely without having the client worry about checks.

As to the second question, the StateSpace is a wrapper over all of the states it holds. It internally uses a dictionary. Therefore the client only needs one StateSpace, register all the states to it and then use it to get the states ready for use by the RNNs or for viewing.

Can you list the line where you feel it should be replaced?

[gzpan]

Thanks for your fast reply! I make a mistake that I ignore the getitem method. Now the two questions are solved. Thanks again

[gzpan]

I forgot another question.
At StateSpace.embedding_encode(self, id, value) function, line 85 in controller.py,
:one_hot[np.arange(1), value_idx] = value_idx + 1,
I think the expression should be : one_hot[np.arange(1), value_idx] = 1 because one hot vector should be
binary. what do you think so

[Somshubra Majumdar]

Ah ordinarily you would be correct. Originally, the code did one hot the data, and then passed it to the RNN without an embedding step, as in the data was passed as a one hot over time.

The paper suggests using embedding matrices to simplify learning and make it correct.

Now that "one hot" function actually returns an integer to index the embedding matrix instead.

I should rename that function. It's not a one hot anymore after all.