bugfix for ConcatLayer with propagate_down set #2972
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
propagate_down[0] == false correctness)
jeffdonahue
force-pushed
the
concat-backward-fix
branch
from
edb0011
to
4bed0ac
Compare
if propagate_down[i] was set, offset_concat_axis was not correctly updated for subsequent bottoms i+1, i+2, ...
|
The test added in the first commit should cause the build to fail, and even did so in the build on my public fork, but somehow passes here... Not sure what's up with that. Regardless, we should probably merge this ASAP. |
|
Thanks for the fix @jeffdonahue and reporting the issue @bharath272. To echo Jeff, nets with a concat layer that does not backprop to all inputs should be re-examined! |
shelhamer
added a commit
that referenced
this pull request
Aug 25, 2015
[fix] properly backprop through ConcatLayer with propagate_down set
myfavouritekk
added a commit
to myfavouritekk/caffe
that referenced
this pull request
Aug 15, 2016
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Currently, if
propagate_down[i]isfalse,offset_concat_axisis not correctlyupdated for any subsequent bottoms (i+1, i+2, ...).
This bug is quite serious for anyone whom it affects -- if you use concat layer and some of its inputs don't need backpropagation (i.e. because they have no learned parameters below them), you should pull and try again with this fix. Many thanks to @bharath272 for reporting.
This was my bug -- it's been broken since merging #1970 in which I (tried to) generalize blobs and many layers to N dimensions. Sorry for the trouble.
In the near future I plan to go through all multi-bottom layers and at least glance to make sure they don't have similarEdit: After going through the list, I don't think any other (official) Caffe layers are subject to this flavor of bug, but please let me know if I missed anything.propagate_downrelated bugs, since we don't ever really test thepropagate_downfunctionality.