[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[500,4]
schedB=dexp[10,2]

[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[500,4]
schedB=dexp[10,2]

[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[500,4]
schedB=dexp[10,2]

[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[500,4]
schedB=dexp[10,2]

[conv3a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[500,4]
schedB=dexp[10,2]

[conv3b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[500,4]
schedB=dexp[10,2]

[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[500,4]
schedB=dexp[10,2]

[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[500,4]
schedB=dexp[10,2]

[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[500,4]
schedB=dexp[10,2]

[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[500,4]
schedB=dexp[10,2]

[fc2048a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.001,0.001
wball=0,0
schedW=dexp[500,4]
schedB=dexp[10,2]

[fc2048b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.001,0.001
wball=0,0
schedW=dexp[500,4]
schedB=dexp[10,2]

[fc2048ba]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.001,0.001
wball=0,0
schedW=dexp[500,4]
schedB=dexp[10,2]

[fc2048bb]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.001,0.001
wball=0,0
schedW=dexp[500,4]
schedB=dexp[10,2]

[fc1000]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.001,0.001
wball=0,0
schedW=dexp[500,4]
schedB=dexp[10,2]

[logprob]
coeff=1
topk=5

[hs1a]
enable=true

[hs2a]
enable=true

[hs1b]
enable=true

[hs2b]
enable=true

[rnorm1a]
scale=0.0001
pow=0.75
minDiv=2

[rnorm1b]
scale=0.0001
pow=0.75
minDiv=2

[rnorm2a]
scale=0.0001
pow=0.75
minDiv=2

[rnorm2b]
scale=0.0001
pow=0.75
minDiv=2

[cnorm2a]
scale=0.001
pow=0.75

[cnorm2b]
scale=0.001
pow=0.75

# this is like #153 (so uses same file) but with dexp learning rate schedule
# its also like 163 but final learning rate is 0.00002 instead of 0.00001
# on guppy7
# logs/layers-165.log
# /nobackup/kriz/tmp/ConvNet__2012-12-30_18.42.56
# NOTE: performance to be compared with 163
# epoch 63: set color noise to 0 from 0.1
# epoch 78: set conv1 epsw to 0 from 0.01
#           logprob:  1.847919, 0.427840, 0.198452 
# multiview logprob:  1.757196, 0.409820, 0.183920