David Warde-Farley

Can you try with the R glmnet package? It might just be the Fortran code...

The context manager approach is quite similar to my [alternative interface for managing batch normalization in graphs](https://github.com/mila-udem/blocks/blob/master/blocks/graph/bn.py#L31-L100). It's a bit differently focused and solves slightly different class of problems. (The...

I've also had the problems you list with graph replacements. The other problem I've had is that early in my usage of it, I wasn't quite filtering for the right...

It's the same thing as the mnist_lenet example. Elapsed time, iteration and epoch number, nothing else. I think it's because it's looking for a `num_examples` attribute. The iteration schemes have...

Relevant: http://deeplearning.net/software/pylearn2/library/devtools.html#pylearn2.devtools.nan_guard.NanGuardMode

You will probably want to be very careful about that to make sure you avoid partially serializing something before you throw that exception.

Reading this thread up until now I tend to think that `push_initialization_config` should grow a 'force' boolean keyword that is `False` by default. initialize() should then do a "soft push"...

Either way, I think it is clearly an "explicit is better than implicit". If you are leaving something alone that's there, that's fine. If you are clobbering something that the...

@bartvm How would you feel about simply refusing to overwrite anything that isn't NoneInitialization? That's already an improvement that does away with a big "gotcha".

I have a possible solution that would keep the complexity out of each Brick's `_push_allocation_config`. What if we defined these methods on `Brick`: ``` python def _pull_config(self, _sentinel, **kwargs): pulled...