Closed mtewes closed 7 years ago
Will address this, as well as other minor points in a branch "post-xmas-fixes", opening now.
I've fixed a random seed issue in training (pushed to current PR): all processes where using exactly the same minibatches (and maybe even worse, not sure if there is some "random" involved in bfgs).
Okay, I'll update it!
This "minimized" version could be written inside of the current workdir automatically after each training. Maybe even with a timestamp, so that previous training results can be easily reused.
All done, except for the multi-cpu predictions (not that urgent), closing.
Best is probably to really just shutil copy it there.