@longjon let me know if there's as succinct a way to do this as your earlier code accomplished and if you have general thoughts on what to do with the python wrapper. @sergeyk too.

The code looks fine, although I don't know why we don't use iterators, and in CaffeNet::params I would push the layer blobs directly instead of calling net_->params, since the latter couples the ordering in the wrapper to the ordering in caffe (although of course those should always be the same).

More generally though I'm not sure where the wrapper should land on the spectrum between mirroring exactly the C++ interface and choosing the most Pythonic/convenient interface.

We could:

1. Mirror the C++ interface exactly
2. Mirror the C++ interface, but use properties instead of getters
3. Provide blobs and layers as dict or OrderedDict instead of pairs of lists or lists of pairs
4. Mirror the C++ interface as in 2 but provide a nicer interface as well, implemented in Python

I guess I feel like 3 right now. In any case it seems odd to add the names directly to the blobs, as this changes the semantics of a blob between C++ and Python (plus, as you noted, layer blobs don't have unique names)...

I'm for the pythonic side of the spectrum. If you're looking for the C++ interface, code in C++. I agree with your 3–I'll figure out how to push an OrderedDict through python::boost.

Regarding names in blobs, I did that because it felt awkward to have a list of blob names and a list of blobs and the same for params. Replacing it all with an OrderedDict feels right.

Thanks for the feedback.

I think this is desirable as it is, and refactoring to return an OrderedDict can follow. Let's merge? @longjon

Sure, I have no problem merging this, as I think it is an improvement, although I hope it's clear to users that the interface is not yet stable...

With that in mind, I've merged it to dev @ 37f340b for the adventurous. Thanks for the wrapper discussion in this thread.