DEV-28879 Lifted Structured Embedding Code from rksltnl#11
DEV-28879 Lifted Structured Embedding Code from rksltnl#11lou-k merged 1 commit intocuralate:masterfrom lou-k:DEV-28879
Conversation
jes-bz
left a comment
jes-bz
left a comment
There was a problem hiding this comment.
Nothing blocking if the code is just c/p from another working branch, Just two things I noticed.
| } | ||
|
|
||
| //#ifdef CPU_ONLY | ||
| //STUB_GPU(LiftedStructSimilaritySoftmaxLossLayer); |
There was a problem hiding this comment.
I see there is no .cu version of the layer to implement the forward and backwards pass on the GPU. I checked and the papers repo doesn't have it either. This layer should only be used in training correct? So it probably will not be an issue if we end up using it in Prod/looking for max speed.
| blob_pos_diff_.Reshape(bottom[0]->channels(), 1, 1, 1); | ||
| blob_neg_diff_.Reshape(bottom[0]->channels(), 1, 1, 1); | ||
| } | ||
|
|
There was a problem hiding this comment.
This layer is not implementing the reshape fiction? I was under the impression this was required for a layer https://github.com/BVLC/caffe/wiki/Development. I guess this will not cause an issue as long as the net is never resized. Context on the addition of the reshape functionality into Caffe BVLC#594
2 participants
Changed files from https://github.com/rksltnl/Caffe-Deep-Metric-Learning-CVPR16
Project files are here https://github.com/rksltnl/Deep-Metric-Learning-CVPR16