Skip to content

Conversation

@tnarihi
Copy link
Contributor

@tnarihi tnarihi commented Feb 16, 2015

I implemented parameterized ReLU unit, which is described in http://arxiv.org/abs/1502.01852 . In my caffe fork, I tried RReLU for CIFAR10 quick examples and the performance was improved about 3%. I am happy If it is worth merging. I don't care even if it is not because I implemented it just for my purpose.

described in Kaiming He et al, "Delving Deep into Rectifiers: Surpassing
Human-Level Performance on ImageNet Classification", arxiv 2015.

Conflicts:
	src/caffe/proto/caffe.proto
@pannous
Copy link

pannous commented Feb 16, 2015

Doesn't relu_layer already have negative_slope, which could be the same as the 'a' parameter in the paper?
** https://github.com/BVLC/caffe/blob/dev/src/caffe/layers/relu_layer.cpp

@tnarihi
Copy link
Contributor Author

tnarihi commented Feb 16, 2015

@pannous The differences between them are that

  1. slopes can be learned.
  2. slopes can vary across channels.

@pannous
Copy link

pannous commented Feb 16, 2015

I see, thanks. Nice PR

This was referenced Feb 16, 2015
@tnarihi
Copy link
Contributor Author

tnarihi commented Feb 19, 2015

Modified to allow in-place computation.

@tnarihi tnarihi force-pushed the prelu branch 2 times, most recently from 6bc51c0 to 366c332 Compare February 22, 2015 17:50
@tnarihi tnarihi mentioned this pull request Feb 22, 2015
@tnarihi
Copy link
Contributor Author

tnarihi commented Feb 22, 2015

Replaced by #1940

@tnarihi tnarihi closed this Feb 22, 2015
@tnarihi tnarihi deleted the prelu branch March 12, 2015 22:21
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants