ixaxaar

Same problem tried `1.30.1` and `1.29.2`. Specs: ``` $ uname -a Linux localhost 4.4.179-Xiaomi_SDM660-LA.UM.7.2.r1-06900+ #1 SMP PREEMPT Tue May 7 14:19:19 UTC 2019 aarch64 ``` ``` # cat build.prop #...

Same here, specifying custom `path` or `file` prevents saving state. Also state is only saved on closing the app, To fix this issue I took [this code](https://github.com/szwacz/electron-boilerplate/blob/master/src/helpers/window.js) and modified it...

+1 having the same issue accompanied by the head mashing.

Hey there. I think the reason behind erasing the memory after each training is that the point is for the network to use the memory. If you let memory uninitialised,...

@chrispugmire heya, sorry for the delay, i was kinda stuck while on vacation in kashmir (under curfew now). Care to review the changes? Its been a while since I've worked...

Closing this cause rewriting the whole thing

AFAIK I did not really use the cpu training except while some testing. Anyway, I need to update pytorch support, lemme look at it this weekend.

The examples are from the paper. Personally, I did use the controller hidden vectors as well as let the memory remain uncleared for entire training time while doing various experiments...

Yes, one has to use them within timesteps otherwise the whole thing is just a bunch of vanilla RNNs i think. What I was talking about is entire training time...

I guess no one thought about letting the python lib have a pure python backoff instead of forcefully shoving the C lib at everyone. "Performance first" even if it does...