Layer exclusion naming convention applied to classification nets [DON'T MERGE YET]

[gheinrich]

In case someone hits a problem like that mentioned in #601 for a classification network.
Once Caffe implements input layers and phase control from Python we should be able to remove those workarounds.

[lukeyeager]

In the generic path, I'm stripping the prefix from the final layer name:
https://github.com/NVIDIA/DIGITS/blob/v3.0.0/digits/model/tasks/caffe_train.py#L556

Can we do the same here?

[lukeyeager]

I tried using this to integrate my Python accuracy layer (here), and got this error:

ERROR: Layer 'prob' references bottom 'accuracy' at the TEST stage however this blob is not included at that stage. Please consider using an include directive to limit the scope of this layer.

That's because this layer gets added to the deploy network:

layer {
  name: "prob"
  type: "Softmax"
  bottom: "accuracy"
  top: "prob"
}

The bottom should be ip2 instead of accuracy.

Looking at the code, I don't see a great way around this. I spent some time trying to find a programmatic solution to all this mess a while back (rough code here) and I gave up on it because there was simply no way to do it elegantly. I'd love to get BVLC/caffe#1245 integrated and update our default networks so that the networks can be fully, correctly specified without needing to do any magic behind the user's back.

[gheinrich]

Updated to strip prefix but as @lukeyeager mentioned there is an issue with the helper code in DIGITS which adds a softmax layer to the deploy network.
Needs more work.

[lukeyeager]

Closed in favor of #623 (which has been closed in favor of #628 !).