Question about controller detail

[zhangcx93]

I noticed that in your controller, the prev_c and prev_h never changed, are zeros all the time. For the LSTM, only the chosen action's embedding is feed into the next step as inputs. Is it intended or a bug? If it is intended, is it necessary to use a 2 layer lstm rather than a simple dense layer with tanh activation? Because output of the cell is only related to the chosen embedding, not related to the whole process of the decisions.

[hyhieu]

It's a bug. Thank you very much for spotting it. We are super grateful! We went from using a tf.while_loop into using a for loop in the controller, so we missed this change. It has been fixed.

[Wronskia]

The new commit seems to sample only two activation functions because of the num_funcs=2. I think it should be set to 4.

[hyhieu]

You can set num_funcs to 4 when you construct the controller class.