richardhahahaha / SpatialMaxFeatureMap

Torch implementation of spatial max feature map

SpatialMaxFeatureMap

Torch implementation of the spatial max feature map layer described in Xiang Wu, Ran He, Zhenan Sun A Lightened CNN for Deep Face Representation http://arxiv.org/abs/1511.02683

Install:

$ luarocks install https://github.com/richardhahahaha/SpatialMaxFeatureMap/raw/master/spatialmaxfeaturemap-scm-1.rockspec

Examples:

require"nn" 
require"spatialmaxfeaturemap"

m=nn.SpatialMaxFeatureMap(2)

X=torch.randn(3,4,3,2)
print(X[3], m(X)[3])

The output is:

(1,.,.) = 
 -0.6997 -1.9338
 -1.4314 -1.4688
 -0.0060  0.1059

(2,.,.) = 
  0.1561  0.7453
  1.5406 -1.4663
  0.1656  0.4614

(3,.,.) = 
  0.8322  1.2153
  1.7452  1.4126
  1.0120  0.1972

(4,.,.) = 
  0.9815  0.2779
  1.8398 -2.0486
 -0.5690  0.9355
[torch.FloatTensor of size 4x3x2]

(1,.,.) = 
  0.8322  1.2153
  1.7452  1.4126
  1.0120  0.1972

(2,.,.) = 
  0.9815  0.7453
  1.8398 -1.4663
  0.1656  0.9355
[torch.FloatTensor of size 2x3x2]