Request - integrate code for all-in-one nets

[lukeyeager]

BVLC#1245 is nearing completion.

• Add Input Layer to Replace inputs BVLC/caffe#3211 adds support for Input layers
• Expose net stages to pycaffe BVLC/caffe#3736 adds better support for stages

Once all-in-one nets are fully supported, the DIGITS team would like to make use of these new features in order to specify networks all in a single file instead of having separate train_val and deploy networks.

[lukeyeager]

Updates:

• Add Input Layer to Replace inputs BVLC/caffe#3211 is included in the caffe-0.15 branch
• Expose all netstate options (for all-in-one nets) BVLC/caffe#3863 will be included in the caffe-0.16 branch

[drnikolaev]

@lukeyeager may I close it?

[lukeyeager]

Actually, it doesn't look like we've pulled in BVLC#3863 after all. You can close this as will-not-fix if you like, but without those changes we still can't really do all-in-one nets.

Of course, based on the lack of progress in BVLC#3864, it doesn't seem like there's much of a desire for this feature anyway.

[KeithWM]

I'm new to caffe, but I don't understand why there isn't more enthusiasm about this feature. To me it seems unnecessarily convoluted and error prone to require maintaining two nigh identical prototxt files. Or am I missing some alternative way of keeping my train and deploy (or test) networks properly aligned?

[drnikolaev]

Keeping this in TODO list..

[drnikolaev]

Done in v0.17.0