-
Notifications
You must be signed in to change notification settings - Fork 1
Home
whereaswhile edited this page Dec 14, 2014
·
5 revisions
The project extends Cuda-ConvNet for sparse coding based image super-resolution.
Some functions added include:
- new neuron type: ShluNeuron (sign(x)*max(0, |x|-a))
- new layer type: cmgrnorm (cross map global response normalization)
- new layer type: eltprod (element-wise product)
- new layer type: allmax (max of all elements in the response)
- extended dimensions parameter for eltsum/eltprod (A[d]*B[j]=C[j] for all j)
- new data provider types: multinstcls/regret/multiregret/videoreg
For the purpose of LSVRC, the following layer is added for bounding box regression:
- BoundBoxOverlapLayer cost layer