Open ghost opened 8 years ago
Thanks for the PR!
However, your changes to the examples are incorrect. We updated the name of the outputs for layers like SigmoidCE
to be predictions
instead of probabilities
for consistency (see layer implementations). If we indicate otherwise anywhere, it should be changed.
Also, the criterion
argument is important to pay attention to: it indicates if the quantity of interest is expected to be maximized or minimized. It should not be changed to name
. The name for the hook can be additionally specified, if the default name is not desirable.
Aye, my bad. I was still using the original brainstorm code I had downloaded a few weeks ago before you had made those changes to the examples. Will be more careful next time! Thanks for the clarifications.
I made corrections to the MNIST and Hutter Prize examples, and updated the README with instructions for installing on Windows machines.