Added param skip_propagate_down to LayerParameter

[mtamburrano]

Added skip_propagate_down parameter as discussed with @jeffdonahue in #2054.

The parameter can be specified either 0 or N times, where N is the number of bottom blobs.
if skip_propagate_down[i] == True then nothing will be backpropagated to bottom[i].
If force_backward is specified, then skip_propagate_down has no effect.

Tests are included in this PR in test.net.cpp

[bhack]

@jeffdonahue I see that somebody is tagging PR as "ready for review". Can you tag also this one?

[bhack]

@jeffdonahue Can you pass here?

[jeffdonahue]

There's no reason to make skip_propagate a new input -- just get it from layer_param in AppendBottom. (I don't think there should be any changes to Net::Init if you do it this way, but I might have overlooked something.)

[mtamburrano]

I can change it if you want, but in this way this check needs to be done in the AppendBottom function, that it is called Bottom.size() times.
I know that this is not a computational relevant part, but don't you think that will be a little bit dirty?

[jeffdonahue]

Leaving the check in Net::Init is fine (though I'd also be fine with moving it to AppendBottom). The rest of the logic is more natural to move to AppendBottom (and its signature shouldn't change) since it's associated specifically with a new bottom blob, and Net::Init is already longer than a single method should be.

[mtamburrano]

Ok, I'll move the param retrieval into AppendBottom and I'll leave the check into Net::Init

[mtamburrano]

@jeffdonahue I apported the changes you suggested:

• AppendBottom doesn't take anymore additional params
• At the end of Net::Init, when the graph is visited top-down, layer_need_backward_ is updated to be consistent with bottom_need_backward_. So if all the top blobs of a Layer don't need backprop, the Layer itself doesn't need back prop and his parents will notice it, to skip backprop themselves if each of their children will not propagate anything.
  In this way if a Layer has two children, with one of them able to backprop and the other not, the Layer still will be correctly needing backprop

[mtamburrano]

hi @jeffdonahue, any news?

[jeffdonahue]

All the CHECK_EQ(.., true)s in this test should be EXPECT_TRUE (and CHECK_EQ(..., false) -> EXPECT_FALSE) -- the tests should fail but not die if the behavior being tested is incorrect.

[bhack]

@jeffdonahue We want to program our activity to reserve a little bit of time for this. Do you need other changes or is it mergeable?

[jeffdonahue]

LGTM -- please squash to a single commit and I'll merge. Thanks @mtamburrano!

[bhack]

Done

[jeffdonahue]

Thanks @bhack and @mtamburrano!

[bhack]

Another brick in the (merge) wall :)