[mlpack-git] [mlpack] Create a function to visualize the features learned by sparse autoencoder (#465)

Ryan Curtin notifications at github.com
Sun Nov 8 21:36:20 EST 2015


Great, thanks!  Everything looks good; thanks for taking the time to write the test and make these changes.  I only have a couple questions before I merge it.

 - Do you think `MaximizeHiddenUnit()` should be exposed to the user instead of in a hidden namespace?  I don't see a reason to keep it in a hidden namespace, since it could potentially be useful to the user.

 - Why does `MaximalInputs()` return the suggested size?  I think the heuristic for guessing a size is kind of brittle here and again restricted to the case where users are training networks on close-to-square images (in which case, wouldn't they know the size of their images anyway, so the suggested size would not be useful?).  I hate to be so picky on this one point, but I really think it's important to ensure both that each function mlpack provides is general and flexible, but also that each function mlpack provides is concise and does not provide too many options that are specific to very few use cases (this balance is very hard to achieve).

 - In `ColumnsToBlocks()` I'd suggest that the margin (`buf`) and value of the margin (-1) be configurable by the user.  Suppose I have an image that has values between 250 and 255, and I set `scale = true`... then `outputs.max() - outputs.min()` will be 256, not 5, and the scaled output will be nothing like the user expects.  Maybe the width of the margin is a less useful parameter, but definitely being able to set the value will be useful for preventing unexpected behavior.

---
Reply to this email directly or view it on GitHub:
https://github.com/mlpack/mlpack/pull/465#issuecomment-154904925
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack-git/attachments/20151108/a005c25a/attachment.html>


More information about the mlpack-git mailing list