Is pi_iter necessary?

[hcwang95]

Hi, I am also trying to implement this paper, and your code really helps.

One question is about the pi_iter, in your code you may update \pi for about 20 times in one epoch, but in the paper, it seems when theta is fixed, pi could be updated to the smallest only applying the equation once.

I removed the pi_iter and try the example mnist.py and it seems has better accuracy and is less prone to overfit. I am not sure if you could test with more example to see if the performance will increase.

By the way, I am trying to implement this decision tree to other platforms such as MXnet, but it's quite hard for me since I just kick out my research career.

So have you tried other platform playing with this model? And I will greatly be appreciated it if you could provide the examples with combining deep neural forest with conv net if you've done some of them.

Thanks!

[hcwang95]

Oh, forget to mention the figure of accuracy.

Using pi_iter to mnist example: Peek accuracy: around 95.4% and after 500 epoch is about 92% Only update pi_iter once per epoch: Peek accuracy: about 97.1%