Initial commit

This commit is contained in:
Laurent El Shafey 2024-12-10 08:56:11 -08:00
commit 9fdd561586
246 changed files with 58283 additions and 0 deletions

157
layers/layer-params-100.cfg Normal file
View file

@ -0,0 +1,157 @@
[conv1a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv1b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv2a]
epsW=0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
[conv2b]
epsW=0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
[conv3a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.001
pow=0.25
[rnorm1b]
scale=0.001
pow=0.25
[rnorm2a]
scale=0.001
pow=0.25
[rnorm2b]
scale=0.001
pow=0.25
# on guppy7
# this is like #97 (on gpu) but with different rnorm coeffs
# /nobackup/kriz/tmp/ConvNet__2012-06-27_14.03.18
# epoch 15: set epsw to 0.001 from 0.01
# epoch 43: killed, seems slightly worse than using my old rnorm coeffs

184
layers/layer-params-106.cfg Normal file
View file

@ -0,0 +1,184 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
# this is like #101 but with contrast normalization layers over rnorm2
# on guppy7
# logs/layers-106.log
# /nobackup/kriz/tmp/ConvNet__2012-07-07_21.11.34
# epoch 22: set epsw to 0.001 from 0.01
# epoch 31: killed, turns out weight contrast normalization is better
# restart after fixing cnorm
# on guppy9
# logs/layers-106a.log
# /nobackup/kriz/tmp/ConvNet__2012-07-17_19.06.09
# epoch 21: set epsw to 0.001 from 0.01
# restart with proper learning rate
# logs/layers-106b.log
# /nobackup/kriz/tmp/ConvNet__2012-07-19_04.15.40
# epoch 23: set epsw to 0.001 from 0.01
# epoch 46: set epsw to 0.0001 from 0.001
# epoch 61: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 72: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 79: set epsw to 0.00001 from 0.0001
# epoch 93: killed
# [1.5942473039940013, 0.3705782743769917, 0.16672222296297284]

167
layers/layer-params-107.cfg Normal file
View file

@ -0,0 +1,167 @@
[conv1a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
wcnorm=0.00
wcNormMin=0.001
wcNormMax=0.002
[conv1b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
wcnorm=0.00
wcNormMin=0.001
wcNormMax=0.002
[conv2a]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2b]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv3a]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# this is like #101 but uses wcnorm in conv1/conv2. so it uses def file #101.
# it's also like #104, but #104 only does wcnorm on conv2
# on guppy7
# logs/layers-107.log
# /nobackup/kriz/tmp/ConvNet__2012-07-09_19.20.14

187
layers/layer-params-109.cfg Normal file
View file

@ -0,0 +1,187 @@
[conv1a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
wcnorm=0.00
[conv1b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
wcnorm=0.00
[conv2a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv3a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1536a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1536b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1536ba]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1536bb]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1536ca]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1536cb]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs3a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[hs3b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# this is like #101 but uses wcnorm in conv2 and also has 3 fc layers.
# on guppy9
# logs/layers-109.log
# /nobackup/kriz/tmp/ConvNet__2012-07-10_00.46.52
# epoch 17: set epsw to 0.001 from 0.01
# epoch 26: enabled dropout on hs3a,hs3b
# epoch 27: killed -- overfitting as feared

146
layers/layer-params-110.cfg Normal file
View file

@ -0,0 +1,146 @@
[conv1a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
# this is like #101 but without rnorm. it's about time i found out how helpful it is to modern nets
# on guppy7
# logs/layers-110.log
# /nobackup/kriz/tmp/ConvNet__2012-07-11_00.26.55
# epoch 19: set epsw to 0.001 from 0.01
# epoch 46: set epsw to 0.0001 from 0.001
# epoch 67: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 66: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 75: killed, it looks to be about 1% worse than #101

187
layers/layer-params-111.cfg Normal file
View file

@ -0,0 +1,187 @@
[conv1a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
wcnorm=0.00
[conv1b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
wcnorm=0.00
[conv2a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv3a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ca]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048cb]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs3a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[hs3b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# this is like #101 but uses wcnorm in conv2 and also has 3 fc layers.
# its also like #109 but uses wider fc layers with dropout in all cos 109 overfit
# on guppy9
# logs/layers-111.log
# /nobackup/kriz/tmp/ConvNet__2012-07-12_23.59.48
# epoch 19: set epsw to 0.001 from 0.01
# epoch 42: this is quite a bit worse, and in an underfitting way, so i'm starting #104, which will be like this but the fc layers will be 3072 each instead of 2048

163
layers/layer-params-112.cfg Normal file
View file

@ -0,0 +1,163 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# on guppy7
# this is like #101 but with rnorm region of size 5 instead of 9
# logs/layers-112.log
# epoch 22: set epsw to 0.001 from 0.01
# epoch 46: set epsw to 0.0001 from 0.001
# epoch 66: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 71: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 79: set epsw to 0.00001 from 0.0001
# epoch 90: killed
# [1.6064990917001289, 0.37237829837731168, 0.16815557540767209]

154
layers/layer-params-113.cfg Normal file
View file

@ -0,0 +1,154 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[conv3b]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# this is like #101 but with conv3 taking both conv2 and conv1 as input, and conv2 just taking the low res img as input
# on guppy9

187
layers/layer-params-114.cfg Normal file
View file

@ -0,0 +1,187 @@
[conv1a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
wcnorm=0.00
[conv1b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
wcnorm=0.00
[conv2a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv3a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc3072a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc3072b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc3072ba]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc3072bb]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc3072ca]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc3072cb]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs3a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[hs3b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# this is like #101 but uses wcnorm in conv2 and also has 3 fc layers, with width 6144.
# on guppy9
# logs/layers-114.log
# 140523240 params (incl biases)
# /nobackup/kriz/tmp/ConvNet__2012-07-15_14.56.24
# epoch 20: set epsw to 0.001 from 0.01
# epoch 40: killed, doing worse than 115 which is the same but has only 2 fc layers

View file

@ -0,0 +1,181 @@
[conv1a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
wcnorm=0.00
[conv1b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
wcnorm=0.00
[conv2a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv3a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc3072a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc3072b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc3072ba]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc3072bb]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc3072ca]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc3072cb]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# this is like 115 (on gpu) but trained on ilya's new imgnet-2010 jpeg
# on guppy7
# logs/layers-115-jpeg.log
# /nobackup/kriz/tmp/ConvNet__2012-07-18_20.56.13
# epoch 22: set epsw to 0.001 from 0.01
# epoch 48: set epsw to 0.0001 from 0.001
# epoch 58: killed, since this was a duplicate (jpeg) of a suboptimal net anyway

303
layers/layer-params-116.cfg Normal file
View file

@ -0,0 +1,303 @@
[conv1a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1c]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1d]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.01,0.01
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2b]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2c]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2d]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv3a]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3c]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3d]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4c]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4d]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5c]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5d]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1024a]
epsW=0.01,0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024b]
epsW=0.01,0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024c]
epsW=0.01,0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024d]
epsW=0.01,0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2a]
epsW=0.01,0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2b]
epsW=0.01,0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2c]
epsW=0.01,0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2d]
epsW=0.01,0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1000]
epsW=0.01,0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[hs1c]
enable=true
[hs2c]
enable=true
[hs1d]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm1c]
scale=0.0001
pow=0.75
[rnorm1d]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[rnorm2c]
scale=0.0001
pow=0.75
[rnorm2d]
scale=0.0001
pow=0.75
# on guppy8
# this is like 112, but has wcnorm in conv2, and also its on 4 gpus

279
layers/layer-params-117.cfg Normal file
View file

@ -0,0 +1,279 @@
[conv1a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1c]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1d]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.01,0.01
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2b]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2c]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2d]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv3a]
epsW=0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
[conv3b]
epsW=0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
[conv3c]
epsW=0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
[conv4a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4c]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4d]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5c]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5d]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1408a]
epsW=0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
[fc1408b]
epsW=0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
[fc1408c]
epsW=0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
[fc1408-2a]
epsW=0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
[fc1408-2b]
epsW=0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
[fc1408-2c]
epsW=0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
[fc1000]
epsW=0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[hs1c]
enable=true
[hs2c]
enable=true
[hs1d]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm1c]
scale=0.0001
pow=0.75
[rnorm1d]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[rnorm2c]
scale=0.0001
pow=0.75
[rnorm2d]
scale=0.0001
pow=0.75
# on guppy8
# this is like 112, but has wcnorm in conv2, and also its on 4 gpus

168
layers/layer-params-118.cfg Normal file
View file

@ -0,0 +1,168 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# on guppy7
# this is like #112 but with wcnorm on conv2, and also trained on jpeg
# logs/layers-118.log
# /nobackup/kriz/tmp/ConvNet__2012-07-19_18.35.31
# epoch 23: set epsw to 0.001 from 0.01
# epoch 46: set epsw to 0.0001 from 0.001
# epoch 65: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 75: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 84: set epsw to 0.00001 from 0.0001
# epcoh 98: killed
# [1.640873252105713, 0.37831333333333333, 0.17355999999999999]

View file

@ -0,0 +1,174 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
# this is like #120 (so uses def file #120) but trained on lsvrc-2012 (full)
# on gpu
# /storage/tmp/ConvNet__2012-07-26_04.06.44
# logs/layers-120-2012-full.log
# epoch 23: set epsw to 0.001 from 0.01
# epoch 38: moved to guppy9
# /nobackup/kriz/tmp/ConvNet__2012-07-26_04.06.44
# epoch 49: set epsw to 0.0001 from 0.001
# epoch 66: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 73: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 87: set epsw to 0.00001 from 0.0001
# epoch 94: killed
#

View file

@ -0,0 +1,173 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
# this is like #120 (so uses def file #120) but trained on lsvrc-2012 (non-full)
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-07-24_23.16.15
# epoch 22: set epsw to 0.001 from 0.01
# epoch 49: set epsw to 0.0001 from 0.001
# epoch 66: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 73: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 81: set epsw to 0.00001 from 0.0001
# epoch 95: killed
# validation multiview error:
# logprob: 1.765247, 0.410440, 0.187140

View file

@ -0,0 +1,313 @@
[conv1a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1c]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1d]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2c]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2d]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3c]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3d]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4c]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4d]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5c]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5d]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1024-1a]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1b]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1c]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1d]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2a]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2b]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2c]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2d]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1000]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm1c]
scale=0.0001
pow=0.75
[rnorm1d]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[rnorm2c]
scale=0.0001
pow=0.75
[rnorm2d]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like #120 but on 4 gpus. trained on 2012 (non-full)
# on guppy
# logs/layers-120-4gpu.log
# /ais/gobi3/u/kriz/tmp/ConvNet__2012-08-03_14.28.23

View file

@ -0,0 +1,313 @@
[conv1a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1c]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1d]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2c]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2d]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3c]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3d]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4c]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4d]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5c]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5d]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1024-1a]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1b]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1c]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1d]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2a]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2b]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2c]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2d]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1000]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm1c]
scale=0.0001
pow=0.75
[rnorm1d]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[rnorm2c]
scale=0.0001
pow=0.75
[rnorm2d]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like #120 but on 4 gpus. trained on 2012 (non-full)
# on guppy
# logs/layers-120-4gpu.log
# /ais/gobi3/u/kriz/tmp/ConvNet__2012-08-03_14.28.23

View file

@ -0,0 +1,313 @@
[conv1a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1c]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1d]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2c]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2d]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3c]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3d]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4c]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4d]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5c]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5d]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1024-1a]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1b]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1c]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1d]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2a]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2b]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2c]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2d]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1000]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm1c]
scale=0.0001
pow=0.75
[rnorm1d]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[rnorm2c]
scale=0.0001
pow=0.75
[rnorm2d]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like #120 but on 4 gpus. trained on 2012 (non-full)
# on guppy
# logs/layers-120-4gpu.log
# /ais/gobi3/u/kriz/tmp/ConvNet__2012-08-03_14.28.23

View file

@ -0,0 +1,313 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1c]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1d]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2c]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2d]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3c]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3d]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4c]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4d]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5c]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5d]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1024-1a]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1b]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1c]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1d]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2a]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2b]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2c]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2d]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1000]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm1c]
scale=0.0001
pow=0.75
[rnorm1d]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[rnorm2c]
scale=0.0001
pow=0.75
[rnorm2d]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like #120 but on 4 gpus. trained on 2012 (non-full)
# on guppy
# logs/layers-120-4gpu.log
# /ais/gobi3/u/kriz/tmp/ConvNet__2012-08-03_14.28.23

View file

@ -0,0 +1,313 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1c]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1d]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2c]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2d]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3c]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3d]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4c]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4d]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5c]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5d]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1024-1a]
epsW=0.00001,0.00001,0.00001,0.00001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1b]
epsW=0.00001,0.00001,0.00001,0.00001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1c]
epsW=0.00001,0.00001,0.00001,0.00001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1d]
epsW=0.00001,0.00001,0.00001,0.00001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2a]
epsW=0.00001,0.00001,0.00001,0.00001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2b]
epsW=0.00001,0.00001,0.00001,0.00001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2c]
epsW=0.00001,0.00001,0.00001,0.00001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2d]
epsW=0.00001,0.00001,0.00001,0.00001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1000]
epsW=0.00001,0.00001,0.00001,0.00001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm1c]
scale=0.0001
pow=0.75
[rnorm1d]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[rnorm2c]
scale=0.0001
pow=0.75
[rnorm2d]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like #120 but on 4 gpus. trained on 2012 (non-full)
# on guppy
# logs/layers-120-4gpu.log
# /ais/gobi3/u/kriz/tmp/ConvNet__2012-08-03_14.28.23

View file

@ -0,0 +1,314 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1c]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1d]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2c]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2d]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[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
[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
[conv3c]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3d]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4c]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4d]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5c]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5d]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1024-1a]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1b]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1c]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1d]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2a]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2b]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2c]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2d]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1000]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm1c]
scale=0.0001
pow=0.75
[rnorm1d]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[rnorm2c]
scale=0.0001
pow=0.75
[rnorm2d]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like #120 but on 4 gpus. trained on 2012 (non-full)
# on guppy
# logs/layers-120-4gpu.log
# /ais/gobi3/u/kriz/tmp/ConvNet__2012-08-03_14.28.23
# epoch 56: killed, this is overfitting. will try reducing the # of params.

174
layers/layer-params-120.cfg Normal file
View file

@ -0,0 +1,174 @@
[conv1a]
epsW=0.0000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.0
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.0
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
# this is like #106 but with rnorm of size 5, also train on jpegs
# on gpu
# logs/layers-120.log
# /storage/tmp/ConvNet__2012-07-22_04.40.34
# moving to guppy7
# /nobackup/kriz/tmp/ConvNet__2012-07-22_04.40.34/
# epoch 26: set epsw to 0.001 from 0.01
# epoch 47: set epsw to 0.0001 from 0.001
# epoch 66: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 72: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 82: set epsw to 0.00001 from 0.0001
# epoch 106: killed
# logprob: 1.634692, 0.378533, 0.172360

179
layers/layer-params-121.cfg Normal file
View file

@ -0,0 +1,179 @@
[conv1a]
epsW=0.0000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.0
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.0
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm1a]
scale=0.001
pow=0.75
[cnorm1b]
scale=0.001
pow=0.75
# this is like #120 but with cnorm over conv1 as well
# on guppy8
# logs/layers-121.log
# /nobackup/kriz/tmp/ConvNet__2012-07-22_15.59.00
# epoch 25: set epsw to 0.001 from 0.01
# epoch 51: set epsw to 0.0001 from 0.001
# epoch 63: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 76: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 90: set epsw to 0.00001 from 0.0001
# worse than 120

View file

@ -0,0 +1,165 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[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
[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
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
# this is like #120-2012-full but also has horiz reflection for gpu2
# on guppy8
# logs/layers-126.log
# /nobackup/kriz/tmp/ConvNet__2012-07-31_22.55.59
# killed after 19 epochs..seems no good, and also full sucks we now know

174
layers/layer-params-127.cfg Normal file
View file

@ -0,0 +1,174 @@
[conv1a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv3b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
# trained on lsvrc-2012 (full), like #120 but
# this examines whether communication is necessary at conv3
# .. meaning it has no communication at conv3
# on gpu
# /storage/tmp/ConvNet__2012-08-01_02.35.01
# logs/layers-127.log
# killed, since we know now that full sucks.
# trained on lsvrc-2012 (non-full). like #120 but now also make conv3,conv4 wider to compensate for lost connections
# on guppy8
# logs/layers-127a.log
# /ais/gobi3/u/kriz/tmp/ConvNet__2012-08-02_00.18.36
# epoch 21: set epsw to 0.001 from 0.01
# epoch 36: killed, significantly worse than 120

167
layers/layer-params-128.cfg Normal file
View file

@ -0,0 +1,167 @@
[conv1a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv5a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
# this is like #120
# and has narrower columns which communicate more. i'm running this because #127 suggests
# that communication is good
# on guppy9
# logs/layers-128.log
# epoch 25: set epsw to 0.001 from 0.01
# on hold

316
layers/layer-params-129.cfg Normal file
View file

@ -0,0 +1,316 @@
[conv1a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1c]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1d]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2c]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2d]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3c]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3d]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4c]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4d]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5c]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5d]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1024-1a]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1b]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1c]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1d]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2a]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2b]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2c]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2d]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1000]
epsW=0.001,0.001,0.001,0.001
epsB=0.002
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm1c]
scale=0.0001
pow=0.75
[rnorm1d]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[rnorm2c]
scale=0.0001
pow=0.75
[rnorm2d]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# 4 gpus, based on 120
# on guppy7
# logs/layers-129.log
# /nobackup/kriz/tmp/ConvNet__2012-08-06_22.23.16
# epoch 22: set epsw to 0.001 from 0.01
# uhh.. relu wiped this. nice.

320
layers/layer-params-130.cfg Normal file
View file

@ -0,0 +1,320 @@
[conv1a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1c]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1d]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2c]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2d]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3c]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3d]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4c]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4d]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5c]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5d]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1024-1a]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1b]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1c]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-1d]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2a]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2b]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2c]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1024-2d]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[fc1000]
epsW=0.0001,0.0001,0.0001,0.0001
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm1c]
scale=0.0001
pow=0.75
[rnorm1d]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[rnorm2c]
scale=0.0001
pow=0.75
[rnorm2d]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like #129, but with 2x as many filters in conv2
# on guppy8
# /nobackup/kriz/tmp/ConvNet__2012-08-07_13.31.34
# logs/layers-130.log
# uhh.. relu wiped this. nice.
# on guppy9
# logs/layers-130a.log
# /nobackup/kriz/tmp/ConvNet__2012-08-09_14.09.20
# epoch 22: set epsw to 0.001 from 0.01
# epoch 46: set epsw to 0.0001 from 0.001
# epoch 62: killed. surprisingly, this is hardly (if at all) better than 2-gpu net

View file

@ -0,0 +1,172 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[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
[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
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
# on guppy9
# logs/layers-131-2009.log
# /nobackup/kriz/tmp/ConvNet__2012-08-18_15.41.20
# epoch 7: set epsw to 0.001 from 0.01
# epoch 14: set epsw to 0.0001 from 0.001
# epoch 20: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 24: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 31: set epsw to 0.00001 from 0.0001
# epoch 36: killed
# logprob: 3.466260, 0.694209, 0.437308
# a bit worse than previous 2009 thing!

175
layers/layer-params-131.cfg Normal file
View file

@ -0,0 +1,175 @@
[conv1a]
epsW=0.0000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.0
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.0
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
# this is like #120, but puts rnorm1 right over conv1 (trained on 2012-nonfull)
# on gpu
# /storage/tmp/ConvNet__2012-08-09_12.33.33
# logs/layers-131.log
# moved to guppy7
# /nobackup/kriz/tmp/ConvNet__2012-08-09_12.33.33/
# epoch 22: set epsw to 0.001 from 0.01
# epoch 46: set epsw to 0.0001 from 0.001
# epoch 66: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 75: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 81: set epsw to 0.00001 from 0.0001
# epoch 100: killed
# validation multiview error:
# logprob: 1.755725, 0.409340, 0.185740

179
layers/layer-params-132.cfg Normal file
View file

@ -0,0 +1,179 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv3b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv5a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
# this is like 120 but with communication in conv4 instead of conv3
# on gpu
# logs/layers-132.log
# /storage/tmp/ConvNet__2012-08-11_02.23.36
# epoch 20: set epsw to 0.001 from 0.01
# epoch 44: set epsw to 0.0001 from 0.001
# moved to guppy9
# @#$%&!, killed, i accidentally trained this on full
# restart:
# /nobackup/kriz/tmp/ConvNet__2012-08-13_16.47.07
# logs/layers-132a.log
# epoch 23: set epsw to 0.001 from 0.01
# epoch 4x: set epsw to 0.0001 from 0.001
# epoch 65: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 71: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 88: killed, worse than 131

167
layers/layer-params-133.cfg Normal file
View file

@ -0,0 +1,167 @@
[conv1a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv3b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv5a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
# this is a hybrid of 131 and 132: so it's like 120, but has communication in conv4 instead of conv3, and it also puts rnorm1 directly over conv1
# on guppy7
# logs/layers-133.log
# /nobackup/kriz/tmp/ConvNet__2012-08-15_16.08.23
# epoch 21: set epsw to 0.001 from 0.01
# epoch 48: set epsw to 0.0001 from 0.001
# epoch 50: killed, worse than 131

169
layers/layer-params-134.cfg Normal file
View file

@ -0,0 +1,169 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[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
[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
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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=0.25
[rnorm1b]
scale=0.0001
pow=0.75
minDiv=0.25
[rnorm2a]
scale=0.0001
pow=0.75
minDiv=0.25
[rnorm2b]
scale=0.0001
pow=0.75
minDiv=0.25
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
# this is like #131, but with minDiv of 0.25 on rnorms
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-08-20_23.26.41
# logs/layers-134.log
# epoch 13: on hold

View file

@ -0,0 +1,199 @@
[conv1a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 trains 135 on 2012, initialized from 2009 1-8800
# on guppy9
# init epsw 0.001
# logs/layers-135-2012-pretrain-2009.log
# /nobackup/kriz/tmp/ConvNet__2012-09-09_15.20.47
# epoch 22: set epsw to 0.0001 from 0.001
# epoch 23: putting on hold to train softmax tree
# this is doing worse than 141-2009 anyway, which has an extra 6th conv layer (1.97 vs 2.00)
# 135 notes:
# this is like #131, but with minDiv of 2 on rnorms
# on guppy8
# /nobackup/kriz/tmp/ConvNet__2012-08-21_01.49.23
# logs/layers-135.log
# epoch 20: set epsw to 0.001 from 0.01
# epoch 47: set epsw to 0.0001 from 0.001
# epoch 66: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 75: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 81: set epsw to 0.00001 from 0.0001
# epoch 96: killed
# validation multiview:
# logprob: 1.757653, 0.410700, 0.184160
# now let's train on 2009 1-8800
# logs/layers-135-2009-bigtrain.log
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-08-26_22.39.45
# epoch 4.7822: set epsw to 0.001 from 0.01
# epoch 8.1299: set epsw to 0.0001 from 0.001
# epoch 10.3697: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 11.4731: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 14.3906: set epsw to 0.00001 from 0.0001
# epoch 17: killed

View file

@ -0,0 +1,189 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 #131, but with minDiv of 2 on rnorms
# on guppy8
# /nobackup/kriz/tmp/ConvNet__2012-08-21_01.49.23
# logs/layers-135.log
# epoch 20: set epsw to 0.001 from 0.01
# epoch 47: set epsw to 0.0001 from 0.001
# epoch 66: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 75: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 81: set epsw to 0.00001 from 0.0001
# epoch 96: killed
# validation multiview:
# logprob: 1.757653, 0.410700, 0.184160
# now let's train on 2009 1-8800
# logs/layers-135-2009-bigtrain.log
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-08-26_22.39.45
# epoch 4.7822: set epsw to 0.001 from 0.01
# epoch 8.1299: set epsw to 0.0001 from 0.001
# epoch 10.3697: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 11.4731: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 14.3906: set epsw to 0.00001 from 0.0001
# epoch 17: killed

177
layers/layer-params-135.cfg Normal file
View file

@ -0,0 +1,177 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 #131, but with minDiv of 2 on rnorms
# on guppy8
# /nobackup/kriz/tmp/ConvNet__2012-08-21_01.49.23
# logs/layers-135.log
# epoch 20: set epsw to 0.001 from 0.01
# epoch 47: set epsw to 0.0001 from 0.001
# epoch 66: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 75: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 81: set epsw to 0.00001 from 0.0001
# epoch 96: killed
# validation multiview:
# logprob: 1.757653, 0.410700, 0.184160

169
layers/layer-params-136.cfg Normal file
View file

@ -0,0 +1,169 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[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
[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
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 #135 (so uses def file 135), but subtracts scalar mean
# on guppy7
# logs/layers-136.log
# /nobackup/kriz/tmp/ConvNet__2012-08-23_04.38.51
# epoch 15: eh, this is no better, and has no reason to be better. screw it.

View file

@ -0,0 +1,196 @@
[conv1a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[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 137, but with treefc
# on guppy9
# init epsw 0.01 -- this run does not sale epsw by node size
# /nobackup/kriz/tmp/ConvNet__2012-09-10_22.47.57
# logs/layers-137-tree.log
# epoch 14: set epsw to 0.001 from 0.01
# epoch 38: killed..its stuck at 2.17 nats.. should be nearer to 2.06. perhaps resume later
# 137 notes:
# this is like #135, but changes the cnorm layers to rnorm
# on guppy8
# logs/layers-137.log
# /nobackup/kriz/tmp/ConvNet__2012-08-25_05.39.04
# epoch 26: set epsw to 0.001 from 0.01
# epoch 50: set epsw to 0.0001 from 0.001
# epoch 75: set epsw to 0 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 84: set epsw to 0.00001 from 0.0001
# epoch 92: made backup to /nobackup/kriz/tmp/ConvNet__2012-08-25_05.39.04.bak
# set epsw to 0.0001 from 0.00001 (conv1/2 still 0)
# using BRIGHTNESS NOISE of 0.2 (in other words i zeroed out the other components of the color noise)
# epoch 101: set color (brightness) noise to 0 from 0.2
# epoch 105: set epsw to 0.00001 from 0.0001
# experiment a failure. going back to training /nobackup/kriz/tmp/ConvNet__2012-08-25_05.39.04.bak
# epoch 99: killed
# logprob: 1.751138, 0.407820, 0.183440
# batch size 128 x 8:
# /nobackup/kriz/tmp/ConvNet__2012-09-07_17.08.47
# epoch 25: set epsw to 0.001 from 0.01
# made backup to /nobackup/kriz/tmp/ConvNet__2012-09-07_17.08.47.bak
# epoch 34: killed, its not good

207
layers/layer-params-137.cfg Normal file
View file

@ -0,0 +1,207 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 #135, but changes the cnorm layers to rnorm
# on lsvrc-2010:
# logs/layers-137-2010.log
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-09-13_02.47.12
# epoch 25: set epsw to 0.001 from 0.01
# epoch 49: set epsw to 0.0001 from 0.001
# epoch 81: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 85: set epsw to 0 from 0.00001 on conv1,conv2
# set epsw to 0.00001 from 0.0001 elsewhere
# epoch 103: killed
# validation:
# logprob: 1.727592, 0.394153, 0.182784
# validation multiview:
# logprob: 1.632875, 0.377960, 0.171020
# test multiview:
# logprob: 1.623185, 0.376167, 0.171247
# on lsvrc-2012:
# on guppy8
# logs/layers-137.log
# /nobackup/kriz/tmp/ConvNet__2012-08-25_05.39.04
# epoch 26: set epsw to 0.001 from 0.01
# epoch 50: set epsw to 0.0001 from 0.001
# epoch 75: set epsw to 0 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 84: set epsw to 0.00001 from 0.0001
# epoch 92: made backup to /nobackup/kriz/tmp/ConvNet__2012-08-25_05.39.04.bak
# set epsw to 0.0001 from 0.00001 (conv1/2 still 0)
# using BRIGHTNESS NOISE of 0.2 (in other words i zeroed out the other components of the color noise)
# epoch 101: set color (brightness) noise to 0 from 0.2
# epoch 105: set epsw to 0.00001 from 0.0001
# experiment a failure. going back to training /nobackup/kriz/tmp/ConvNet__2012-08-25_05.39.04.bak
# epoch 99: killed
# logprob: 1.751138, 0.407820, 0.183440
# batch size 128 x 8:
# /nobackup/kriz/tmp/ConvNet__2012-09-07_17.08.47
# epoch 25: set epsw to 0.001 from 0.01
# made backup to /nobackup/kriz/tmp/ConvNet__2012-09-07_17.08.47.bak
# epoch 34: killed, its not good

172
layers/layer-params-139.cfg Normal file
View file

@ -0,0 +1,172 @@
[conv1a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv2b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
wcNormMin=0.001,0
wcNormMax=0.002,0
[conv3a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 #137 (hence uses same file) but has wcnorm on conv2[0]
# epoch 19: set epsw to 0.001 from 0.01
# epoch 49: set epsw to 0.0001 from 0.001
# epoch 62: killed, about 0.01 nat worse than 137 (which is pretty significant at this stage)

View file

@ -0,0 +1,203 @@
[conv1a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv6a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv6b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048a]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.0001,0.0001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 141, but trained on half of 2009 imgnet, to be comparable to google's results
# logs/layers-141-2009-half.log
# /nobackup/kriz/tmp/ConvNet__2012-09-09_00.26.31
# on guppy9
# epoch 6.2600: set epsw to 0.001 from 0.01
# epoch 13.3361: set epsw to 0.0001 from 0.001
# epoch 18.2396: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 21.1949: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 25.3718: set epsw to 0.00001 from 0.0001
# epoch 28.3271: killed
# ok test erro rate is a bit worse than 131, restarting with epsw 0.001, color noise 0.1
# epoch 44.183: set epsw to 0.0001 from 0.001
# epoch 56: eek, it started getting worse on validation :/
# 141 notes:
# this is like #137 but with conv6, also communication in conv6
# /nobackup/kriz/tmp/ConvNet__2012-09-03_16.27.48
# logs/layers-141.log
# epoch 23: set epsw to 0.001 from 0.01
# epoch 48: set epsw to 0.0001 from 0.001
# epoch 60: this seems overfitty....killing
# but will use these weights to initialize a net on 2009... why the hell not?

View file

@ -0,0 +1,231 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv6a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv6b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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
# training on lsvrc-2010
# initialized from 141 trained on lsvrc-2012, then 2009
# using def file layers-141-2009-2010.cfg
# /nobackup/kriz/tmp/ConvNet__2012-09-12_01.06.32
# on guppy8
# init epsw 0.001
# logs/layers-141-2010-pretrain-2009-pretrain-2012.log
# epoch 14: set epsw to 0.0001 from 0.001
# epoch 30: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 36: set epsw to 0 on conv1/2
# epoch 47: set epsw to 0.00001 from 0.0001
# epoch 54: killed
# logprob: 1.511725, 0.356707, 0.154893
# training on lsvrc-2012
# initialized from 141 trained on lsvrc-2012, then 2009
# using def file layers-141-2009-2012.cfg
# init epsw 0.001
# logs/layers-141-2012-pretrain-2009-pretrain-2012.log
# /nobackup/kriz/tmp/ConvNet__2012-09-09_03.36.13
# backup: /ais/gobi3/u/kriz/tmp/ConvNet__2012-09-09_03.36.13
# also /ais/gobi3/u/kriz/net-backups/
# on guppy8
# epoch 13: set epsw to 0.0001 from 0.001
# epoch 26: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 32: set epsw to 0 on conv1/2
# epoch 43: set epsw to 0.00001 from 0.0001
# epoch 54: killed
# python convnet.py -f /nobackup/kriz/tmp/ConvNet__2012-09-09_03.36.13 --test-only=1 --test-one=0 --multiview-test=1
# logprob: 1.671316, 0.395620, 0.172060
#python convnet.py -f /nobackup/kriz/tmp/ConvNet__2012-09-09_03.36.13 --test-only=1 --test-one=0 --multiview-test=0
# logprob: 1.779082, 0.415920, 0.186780
# 141-2009 notes, before going back to 2012:
# initialized from 141 trained on lsvrc-2012
# init epsw 0.001
# logs/layers-141-2009-pretrain-2012.log
# /nobackup/kriz/tmp/ConvNet__2012-09-07_05.22.51
# epoch 4.1189: set epsw to 0.0001 from 0.001
# epoch 5.1596: killed, not improving much. lets go back to training on lsvrc-2012 with these weights now.
#
# 141 notes:
# this is like #137 but with conv6, also communication in conv6
# /nobackup/kriz/tmp/ConvNet__2012-09-03_16.27.48
# logs/layers-141.log
# epoch 23: set epsw to 0.001 from 0.01
# epoch 48: set epsw to 0.0001 from 0.001
# epoch 60: this seems overfitty....killing
# but will use these weights to initialize a net on 2009... why the hell not?

187
layers/layer-params-141.cfg Normal file
View file

@ -0,0 +1,187 @@
[conv1a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv6a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv6b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 #137 but with conv6, also communication in conv6
# /nobackup/kriz/tmp/ConvNet__2012-09-03_16.27.48
# logs/layers-141.log
# epoch 23: set epsw to 0.001 from 0.01
# epoch 48: set epsw to 0.0001 from 0.001
# epoch 60: this seems overfitty....killing
# but will use these weights to initialize a net on 2009... why the hell not?

View file

@ -0,0 +1,206 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 #137 but without low-res stuff
# on lsvrc-2010:
# guppy9
# logs/layers-145-2010.log
# /nobackup/kriz/tmp/ConvNet__2012-09-27_12.39.44
# epoch 23: set epsw to 0.001 from 0.01
# epoch 51: set epsw to 0.0001 from 0.001
# epoch 68: set epsw to 0.00001 from 0.0001 on conv1
# set color noise to 0 from 0.1
# epoch 72: set epsw to 0 on conv1
# epoch 78: set epsw to 0.00001 from 0.0001
# epoch 93: killed
# test multliview:
# logprob: 1.614660, 0.374727, 0.169987
# test center patch:
# logprob: 1.706031, 0.390247, 0.182953 (NOTE, NOT MULTIVIEW!!)
# on gpu (now guppy8)
# logs/layers-145.log
# /storage/tmp/ConvNet__2012-09-13_03.43.56
# epoch 25: set epsw to 0.001 from 0.01
# epoch 36: paused for localization experiments
# resuming on guppy9
# logs/layers-145-cont.log
# /nobackup/kriz/tmp/ConvNet__2012-09-13_03.43.56
# epoch 51: set epsw to 0.0001 from 0.001
# epoch 58: paused for imgnet-20k experiments
# moved to guppy8
# epoch 67: set epsw to 0.00001 from 0.0001 on conv1
# set color noise to 0 from 0.1
# epoch 72: set epsw to 0 on conv1
# epoch 79: set epsw to 0.00001 from 0.0001
# epoch 91: killed
# logprob: 1.741473, 0.406640, 0.182100
# on 2012-full:
# on guppy7
# logs/layers-145-full.log
# /nobackup/kriz/tmp/ConvNet__2012-09-23_19.38.45
# epoch 19: set epsw to 0.001 from 0.01
# epoch 47: set epsw to 0.0001 from 0.001
# epoch 61: moved to gpu

View file

@ -0,0 +1,100 @@
[conv1a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc4096a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc4096ba]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1000]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
minDiv=2
[rnorm2a]
scale=0.0001
pow=0.75
minDiv=2
[cnorm2a]
scale=0.001
pow=0.75
# this is #145 but only one column, although conv5 is as wide as 2 columns because otherwise this net would have about half as many parameters as the 2-column net, which wouldnt make for a fair comparison.
# on guppy9
# trained on 2010
# logs/layers-145-half.log
# /nobackup/kriz/tmp/ConvNet__2012-11-03_01.00.35
# epoch 20: set epstw ot 0.001 from 0.01
# epoch 48: set epstw ot 0.0001 from 0.001
# epoch 66: set epsw to 0.00001 from 0.0001 on conv1, set color noise to 0 from 0.1
# epoch 72: set epsw to 0 from 0.00001 on conv1
# epoch 96: killed
# test multiview logprob: 1.702802, 0.391680, 0.182287

204
layers/layer-params-145.cfg Normal file
View file

@ -0,0 +1,204 @@
[conv1a]
epsW=0.000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 #137 but without low-res stuff
# on gpu (now guppy9)
# logs/layers-145.log
# /storage/tmp/ConvNet__2012-09-13_03.43.56
# epoch 25: set epsw to 0.001 from 0.01
# epoch 36: paused for localization experiments
# resuming on guppy9
# logs/layers-145-cont.log
# /nobackup/kriz/tmp/ConvNet__2012-09-13_03.43.56
# epoch 51: set epsw to 0.0001 from 0.001
# epoch 58: paused for imgnet-20k experiments
# moved to guppy8
# epoch 67: set epsw to 0.00001 from 0.0001 on conv1
# set color noise to 0 from 0.1
# epoch 72: set epsw to 0 on conv1
# epoch 79: set epsw to 0.00001 from 0.0001
# epoch 91: killed
# logprob: 1.741473, 0.406640, 0.182100
# on 2012-full:
# on guppy7
# logs/layers-145-full.log
# /nobackup/kriz/tmp/ConvNet__2012-09-23_19.38.45
# epoch 19: set epsw to 0.001 from 0.01
# epoch 47: set epsw to 0.0001 from 0.001
# epoch 61: moved to gpu
# pushing learning rate back up:
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-09-13_03.43.56.2
# start epsw 0.001 (still 0 on conv1)
# color noise put back to 0.1
# logs/layers-145.log
# epoch 107: set epsw to 0.0001 from 0.001
# epoch 124: set epsw to 0.00001 from 0.0001
# set color noise to 0 from 0.1
# epoch 135: killed
# multiview test:
# logprob: 1.725738, 0.402500, 0.179940

View file

@ -0,0 +1,182 @@
[conv1a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv6a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv6b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[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 146-2009 but with tree
# epoch 9: set epsw to 0.001 from 0.01

View file

@ -0,0 +1,188 @@
[conv1a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv6a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv6b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 145, and initialized from 145 (up to conv5) but with also conv6 and trained on 2009-10k
# on guppy8
# /nobackup/kriz/tmp/ConvNet__2012-09-14_22.48.00
# initialized from /nobackup/kriz/tmp/ConvNet__2012-09-13_03.43.56.bak
# init epsw 0.001 on conv1-5, 0.01 on fc, conv6
# logs/layers-146-2009.log
# epoch 2.1600: set epsw to 0.001 from 0.01 on fc, conv6
# epoch 6.6491: killed

View file

@ -0,0 +1,221 @@
[conv1a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv6a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv6b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 146, trained on 2011-20k, initialized from 2009-10k (conv layers only)
# init epw 0.001 on conv layers, 0.01 on fc
# logs/layers-146-2011.log
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-09-19_23.29.04
# epoch 6: set epsw to 0.001 from 0.01
# epoch 8.11295: killed
# 146-2009 notes:
# this is 145, and initialized from 145 (up to conv5) but with also conv6 and trained on 2009-10k
# on guppy8
# /nobackup/kriz/tmp/ConvNet__2012-09-14_22.48.00
# initialized from /nobackup/kriz/tmp/ConvNet__2012-09-13_03.43.56.bak
# init epsw 0.001 on conv1-5, 0.01 on fc, conv6
# logs/layers-146-2009.log
# epoch 2.1600: set epsw to 0.001 from 0.01 on fc, conv6
# epoch 6.6491: killed

View file

@ -0,0 +1,203 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv6a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv6b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 146 but pre-trained on 2009-10k (notes below), so initialized from below net
# init epsw 0.001, 0.0001 on conv1
# on guppy8
# logs/layers-146-2012-2009.log
# /nobackup/kriz/tmp/ConvNet__2012-09-17_17.01.42
# epoch 3: set conv1 epsw to 0.00001 from 0.0001
# epoch 4: set conv1 epsw to 0 from 0.00001
# epoch 22: set epsw to 0.0001 from 0.001
# epoch 38: set color noise to 0 from 0.1
# epoch 42: set epsw to 0.00001 from 0.0001
# epoch 52: killed
# multiview validation:
# logprob: 1.646452, 0.391000, 0.168760
# 146-2009 notes:
# this is 145, and initialized from 145 (up to conv5) but with also conv6 and trained on 2009-10k
# /nobackup/kriz/tmp/ConvNet__2012-09-14_22.48.00
# on guppy8
# initialized from /nobackup/kriz/tmp/ConvNet__2012-09-13_03.43.56.bak
# init epsw 0.001 on conv1-5, 0.01 on fc, conv6
# logs/layers-146-2009.log
# epoch 2.1600: set epsw to 0.001 from 0.01 on fc, conv6
# epoch 6.6491: killed

View file

@ -0,0 +1,205 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv6a]
epsW=0.00001,0.00001
epsB=0.022
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv6b]
epsW=0.00001,0.00001
epsB=0.022
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 146 but pretrained on 2011, now training on 2012
# init epsw 0.0001 on conv1, 0.001 elsewhere
# guppy9
# logs/layers-146-2012-2011.log
# /nobackup/kriz/tmp/ConvNet__2012-09-23_14.57.12
# 300 batches: set epsw to 0.01 from 0.001 on fc1000
# epoch 5: set conv1 epsw to 0.00001 from 0.0001
# epoch 7: set conv1 epsw to 0 from 0.00001
# epoch 9: set fc1000 epsw to 0.001 from 0.01
# epoch 22: set epsw to 0.0001 from 0.001
# epoch 38: set color noise to 0 from 0.1
# epoch 42: set epsw to 0.00001 from 0.0001
# epoch 54: killed
# multiview validation:
# logprob: 1.633191, 0.389900, 0.166220
# 146-2011 notes:
# this is 146, trained on 2011-20k, initialized from 2009-10k (conv layers only)
# init epw 0.001 on conv layers, 0.01 on fc
# logs/layers-146-2011.log
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-09-19_23.29.04
# epoch 6: set epsw to 0.001 from 0.01
# epoch 8.11295: killed

174
layers/layer-params-147.cfg Normal file
View file

@ -0,0 +1,174 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 #145 but with non-overlapping pooling
# on guppy9
# logs/layers-147.log
# /nobackup/kriz/tmp/ConvNet__2012-10-07_23.42.30
# epoch 23: set epsw to 0.001 from 0.01
# epoch 51: set epsw to 0.0001 from 0.001
# epoch 79: set epsw to 0 on conv1, 0.00001 elsewhere
# epoch 90: killed
# validation multiview:
# logprob: 1.757644, 0.410580, 0.185100

View file

@ -0,0 +1,169 @@
ano[conv1a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.001,0.001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 #145 but with non-overlapping pooling
# on guppy9
# logs/layers-147.log
# /nobackup/kriz/tmp/ConvNet__2012-10-07_23.42.30
# epoch 23: set epsw to 0.001 from 0.01

148
layers/layer-params-148.cfg Normal file
View file

@ -0,0 +1,148 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
# this is like #145 but without normalization layers
# on guppy9
# logs/layers-148.log
# /nobackup/kriz/tmp/ConvNet__2012-11-02_23.33.30
# epoch 22: set epsw to 0.001 from 0.01
# epoch 44: set epsw to 0.0001 from 0.001
# epoch 69: set epsw to 0.00001 from 0.0001 on conv1, set color noise to 0 from 0.1
# epoch 73: set epsw to 0 from 0.00001 on conv1
# epoch 86: set epsw to 0.00001 from 0.0001
# epoch 97: killed
# validation multiview:
# logprob: 1.822358, 0.420340, 0.193620 (1.4% top-1 worse than 145)

182
layers/layer-params-149.cfg Normal file
View file

@ -0,0 +1,182 @@
[conv1a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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
[rnorm5a]
scale=0.0001
pow=0.75
minDiv=2
[rnorm5b]
scale=0.0001
pow=0.75
minDiv=2
# this is like #145 but with rnorm over conv5
# on guppy
# /nobackup_a/kriz/tmp/ConvNet__2012-11-13_23.21.47
# logs/layers-149.log
# epoch 21: set epsw to 0.001 from 0.01
# epoch 54: set epsw to 0.0001 from 0.001
# epoch 66: set epsw to 0.00001 from 0.0001 on conv1
# set color noise to 0 from 0.1

180
layers/layer-params-150.cfg Normal file
View file

@ -0,0 +1,180 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv3b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv4a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc4096a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc4096b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048ba]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048bb]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc1000]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[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 #145 but with no column communication, trained on 2012
# guppy9
# /nobackup/kriz/tmp/ConvNet__2012-11-20_01.25.08
# logs/layers-150.log
# moved to gpu
# epoch 25: set epsw to 0.001 from 0.01
# epoch 49: set epsw to 0.0001 from 0.001
# epoch 50: move back to guppy9
# epoch 66: set epsw to 0.00001 from 0.0001 on conv1
# set color noise to 0 from 0.1
# epoch 72: set epsw to 0 on conv1
# epoch 79: set epsw to 0.00001 from 0.0001
# epoch 92: killed
# validation multiview:
# logprob: 1.811173, 0.418280, 0.193300

184
layers/layer-params-153.cfg Normal file
View file

@ -0,0 +1,184 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=linear[1000]
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=linear[1000]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=linear[1000]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=linear[1000]
[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
sched=linear[1000]
[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
sched=linear[1000]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=linear[1000]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=linear[1000]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=linear[1000]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=linear[1000]
[fc2048a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=linear[1000]
[fc2048b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=linear[1000]
[fc2048ba]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=linear[1000]
[fc2048bb]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=linear[1000]
[fc1000]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=linear[1000]
[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 #145 but with linear learning rate schedule
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-12-18_22.39.10
# logs/layers-153.log
# epoch 18: killed. i realized linear learning rate schedule is completely mental.

372
layers/layer-params-154.cfg Normal file
View file

@ -0,0 +1,372 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[1000]
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[1000]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[1000]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[1000]
[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
sched=exp[1000]
[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
sched=exp[1000]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[1000]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[1000]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[1000]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[1000]
[fc2048a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[1000]
[fc2048b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[1000]
[fc2048ba]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[1000]
[fc2048bb]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[1000]
[fc1000]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[1000]
[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 exp learning rate schedule
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-12-18_23.40.29
# logs/layers-154.log
# moved to gpu
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[1000]
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[1000]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[1000]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[1000]
[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
sched=exp[1000]
[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
sched=exp[1000]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[1000]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[1000]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[1000]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[1000]
[fc2048a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[1000]
[fc2048b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[1000]
[fc2048ba]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[1000]
[fc2048bb]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[1000]
[fc1000]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[1000]
[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 exp learning rate schedule
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-12-18_23.40.29
# logs/layers-154.log
# moved to gpu
# /storage/tmp/ConvNet__2012-12-18_23.40.29
# i think something got corrupted
# resuming from epoch 10 on guppy7
# /nobackup/kriz/tmp/ConvNet__2012-12-18_23.40.29

190
layers/layer-params-155.cfg Normal file
View file

@ -0,0 +1,190 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[4000]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[4000]
[conv2a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[4000]
[conv2b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[4000]
[conv3a]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[4000]
[conv3b]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[4000]
[conv4a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[4000]
[conv4b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[4000]
[conv5a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[4000]
[conv5b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[4000]
[fc2048a]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[4000]
[fc2048b]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[4000]
[fc2048ba]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[4000]
[fc2048bb]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[4000]
[fc1000]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[4000]
[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 exp learning rate schedule
# its also like #154, but with learning rates in the range 0.02 to 0.000005
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-12-19_22.13.12
# logs/layers-155.log
# epoch 61: set color noise to 0 from 0.1
# epoch 74: set epsw to 0 from 0.02 on conv1
# validation:
# logprob: 1.861853, 0.426530, 0.199652
# validation multiview:
# logprob: 1.750063, 0.407440, 0.185240

190
layers/layer-params-156.cfg Normal file
View file

@ -0,0 +1,190 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[2000]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[2000]
[conv2a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[2000]
[conv2b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[2000]
[conv3a]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[2000]
[conv3b]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[2000]
[conv4a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[2000]
[conv4b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[2000]
[conv5a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[2000]
[conv5b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[2000]
[fc2048a]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[2000]
[fc2048b]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[2000]
[fc2048ba]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[2000]
[fc2048bb]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[2000]
[fc1000]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[2000]
[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 exp learning rate schedule
# its also like #154 and #155, but with learning rates in the range 0.02 to 0.00001
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2012-12-20_01.29.32
# logs/layers-156.log
# epoch 61: set color noise to 0 from 0.1
# epoch 72: set epsw to 0 from 0.02 on conv1
# validation:
# logprob: 1.870253, 0.428933, 0.198336
# validation multiview:
# logprob: 1.751178, 0.407640, 0.183500

188
layers/layer-params-157.cfg Normal file
View file

@ -0,0 +1,188 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[8000]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[8000]
[conv2a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[8000]
[conv2b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[8000]
[conv3a]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[8000]
[conv3b]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[8000]
[conv4a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[8000]
[conv4b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[8000]
[conv5a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[8000]
[conv5b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[8000]
[fc2048a]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[8000]
[fc2048b]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[8000]
[fc2048ba]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[8000]
[fc2048bb]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[8000]
[fc1000]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[8000]
[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 exp learning rate schedule
# its also like #155, but with learning rates in the range 0.04 to 0.000005
# on guppy8
# logs/layers-157.log
# /nobackup/kriz/tmp/ConvNet__2012-12-23_02.12.31
# epoch 62: set color noise to 0 from 0.1
# epoch 73: set conv1 epsw to 0 from 0.04
# valid: logprob: 1.880485, 0.431177, 0.203271
# multiview valid: logprob: 1.767696, 0.411140, 0.187040

189
layers/layer-params-158.cfg Normal file
View file

@ -0,0 +1,189 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[20000]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[20000]
[conv2a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[20000]
[conv2b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=exp[20000]
[conv3a]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[20000]
[conv3b]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[20000]
[conv4a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[20000]
[conv4b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[20000]
[conv5a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[20000]
[conv5b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=exp[20000]
[fc2048a]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[20000]
[fc2048b]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[20000]
[fc2048ba]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[20000]
[fc2048bb]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[20000]
[fc1000]
epsW=0.04,0.04
epsB=0.08
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=exp[20000]
[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 exp learning rate schedule
# its also like #155, but with learning rates in the range 0.04 to 0.000002
# on guppy8
# logs/layers-158.log
# /nobackup/kriz/tmp/ConvNet__2012-12-23_17.34.48
# epoch 63: set color noise to 0 from 0.1
# epoch 77: set epsw to 0 from 0.04 on conv1
# validation: logprob: 1.862656, 0.428884, 0.199910
# validation multiview: logprob: 1.757155, 0.410260, 0.185380

187
layers/layer-params-160.cfg Normal file
View file

@ -0,0 +1,187 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=dexp[2000,4]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=dexp[2000,4]
[conv2a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=dexp[2000,4]
[conv2b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=dexp[2000,4]
[conv3a]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=dexp[2000,4]
[conv3b]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=dexp[2000,4]
[conv4a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=dexp[2000,4]
[conv4b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=dexp[2000,4]
[conv5a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=dexp[2000,4]
[conv5b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=dexp[2000,4]
[fc2048a]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=dexp[2000,4]
[fc2048b]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=dexp[2000,4]
[fc2048ba]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=dexp[2000,4]
[fc2048bb]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=dexp[2000,4]
[fc1000]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=dexp[2000,4]
[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 159 but with rates in the range 0.02 to 0.00001
# on guppy7
# logs/layers-160.log
# /nobackup/kriz/tmp/ConvNet__2012-12-24_17.07.46
# epoch 61: set color noise to 0 from 0.1
# validation: 1.884187, 0.433855, 0.205452
# validation multiview: : 1.789202, 0.413740, 0.190360

183
layers/layer-params-161.cfg Normal file
View file

@ -0,0 +1,183 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=jdexp[1000,4]
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=jdexp[1000,4]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=jdexp[1000,4]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=jdexp[1000,4]
[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
sched=jdexp[1000,4]
[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
sched=jdexp[1000,4]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=jdexp[1000,4]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=jdexp[1000,4]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=jdexp[1000,4]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=jdexp[1000,4]
[fc2048a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=jdexp[1000,4]
[fc2048b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=jdexp[1000,4]
[fc2048ba]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=jdexp[1000,4]
[fc2048bb]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=jdexp[1000,4]
[fc1000]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=jdexp[1000,4]
[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 jumpy dexp learning rate schedule
# on guppy7
# logs/layers-161.log
# (guppy7 is dead for now so doing nothing)

187
layers/layer-params-162.cfg Normal file
View file

@ -0,0 +1,187 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=dexp[1000,3]
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=dexp[1000,3]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=dexp[1000,3]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=dexp[1000,3]
[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
sched=dexp[1000,3]
[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
sched=dexp[1000,3]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=dexp[1000,3]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=dexp[1000,3]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=dexp[1000,3]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=dexp[1000,3]
[fc2048a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=dexp[1000,3]
[fc2048b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=dexp[1000,3]
[fc2048ba]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=dexp[1000,3]
[fc2048bb]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=dexp[1000,3]
[fc1000]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
sched=dexp[1000,3]
[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 159 but with 3 levels of learning rates in the range 0.01 to 0.00001
# on guppy9
# logs/layers-162.log
# /nobackup/kriz/tmp/ConvNet__2012-12-25_22.41.00
# epoch 61: set color noise to 0 from 0.1
# validation: logprob: 1.894451, 0.438533, 0.207935
# validation multiview: ah screw it, it'll suck

188
layers/layer-params-163.cfg Normal file
View file

@ -0,0 +1,188 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=dexp[1000,4]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=dexp[1000,4]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=dexp[1000,4]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
sched=dexp[1000,4]
[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
sched=dexp[1000,4]
[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
sched=dexp[1000,4]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=dexp[1000,4]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=dexp[1000,4]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=dexp[1000,4]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
sched=dexp[1000,4]
[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
sched=dexp[1000,4]
[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
sched=dexp[1000,4]
[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
sched=dexp[1000,4]
[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
sched=dexp[1000,4]
[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
sched=dexp[1000,4]
[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 159 but with 2x the wc on fc layers
# on guppy9
# logs/layers-163.log
# /nobackup/kriz/tmp/ConvNet__2012-12-26_01.15.38
# epoch 61: set color noise to 0 from 0.1
# epoch 73: set conv1 epsw to 0 from 0.01
# validation: logprob: 1.849131, 0.429085, 0.199072
# validation multiview: ah screw it, it'll suck

204
layers/layer-params-165.cfg Normal file
View file

@ -0,0 +1,204 @@
[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

204
layers/layer-params-166.cfg Normal file
View file

@ -0,0 +1,204 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[1000,4]
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[1000,4]
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[1000,4]
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[1000,4]
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[1000,4]
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[1000,4]
schedB=dexp[10,2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[1000,4]
schedB=dexp[10,2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[1000,4]
schedB=dexp[10,2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[1000,4]
schedB=dexp[10,2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[1000,4]
schedB=dexp[10,2]
[fc2048a]
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[1000,4]
schedB=dexp[10,2]
[fc2048b]
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[1000,4]
schedB=dexp[10,2]
[fc2048ba]
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[1000,4]
schedB=dexp[10,2]
[fc2048bb]
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[1000,4]
schedB=dexp[10,2]
[fc1000]
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[1000,4]
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 #159 but with separate schedb and also it puts the 2nd layer normalization *below* relu
# on guppy9
# logs/layers-166.log
# /storage/tmp/ConvNet__2012-12-23_19.38.31
#
# epoch 64: set color noise to 0 from 0.1
# epoch 73: set epsw conv1 to 0 from 0.01
# logprob: 1.863109, 0.428586, 0.201039
# multiview logprob: 1.768124, 0.410960, 0.186740

187
layers/layer-params-167.cfg Normal file
View file

@ -0,0 +1,187 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[1000,4]
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[1000,4]
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[1000,4]
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[1000,4]
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[1000,4]
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[1000,4]
schedB=dexp[10,2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[1000,4]
schedB=dexp[10,2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[1000,4]
schedB=dexp[10,2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[1000,4]
schedB=dexp[10,2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[1000,4]
schedB=dexp[10,2]
[fc2048a]
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[1000,4]
schedB=dexp[10,2]
[fc2048b]
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[1000,4]
schedB=dexp[10,2]
[fc2048ba]
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[1000,4]
schedB=dexp[10,2]
[fc2048bb]
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[1000,4]
schedB=dexp[10,2]
[fc1000]
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[1000,4]
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
# this is like #153 but with dexp learning rate schedule
# also deletes rnorm over conv2
# on guppy8
# logs/layers-167.log
# /nobackup/kriz/tmp/ConvNet__2013-01-03_23.39.35
# epoch 63: set color noise to 0 from 0.1
# epoch 73: set conv1 epsw to 0 from 0.01
# logprob: 1.851845, 0.426772, 0.197590
# multiview logprob: 1.738715, 0.404880, 0.181180

212
layers/layer-params-169.cfg Normal file
View file

@ -0,0 +1,212 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[500,0,4]
schedB=dexp[10,0,2]
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[500,0,4]
schedB=dexp[10,0,2]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[500,0,4]
schedB=dexp[10,0,2]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[500,0,4]
schedB=dexp[10,0,2]
[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[500,0,4]
schedB=dexp[10,0,2]
[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[500,0,4]
schedB=dexp[10,0,2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[500,0,4]
schedB=dexp[10,0,2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[500,0,4]
schedB=dexp[10,0,2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[500,0,4]
schedB=dexp[10,0,2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[500,0,4]
schedB=dexp[10,0,2]
[fc2048a]
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[500,0,4]
schedB=dexp[10,0,2]
[fc2048b]
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[500,0,4]
schedB=dexp[10,0,2]
[fc2048ba]
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[500,0,4]
schedB=dexp[10,0,2]
[fc2048bb]
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[500,0,4]
schedB=dexp[10,0,2]
[fc1000]
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[500,0,4]
schedB=dexp[10,0,2]
[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 #159 (so uses same def file: 153) but with learning rate decaying to 0.00002
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2013-01-07_04.03.15
# logs/layers-169.log
# epoch 60: set color noise to 0 from 0.1
# epoch 73: set conv1 epsw to 0 from 0.01
# logprob: 1.841218, 0.425298, 0.195489
# multiview logprob: 1.735123, 0.405960, 0.181120
# now lets run this for 80 epochs instead of 95:
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2013-01-13_15.31.51
# logs/layers-169-80.log
# epoch 46: paused for experiments with dropout in conv layers
# epoch 53: set color noise to 0 from 0.1
# epoch 74: set epsw to 0 from 0.01 on conv1
# epoch 80: killed :(
# logprob: 1.866277, 0.429784, 0.200178

View file

@ -0,0 +1,201 @@
[conv1a]
epsW=0.015
epsB=0.03
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv1b]
epsW=0.015
epsB=0.03
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv2a]
epsW=0.015
epsB=0.03
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv2b]
epsW=0.015
epsB=0.03
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv3a]
epsW=0.015,0.015
epsB=0.03
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv3b]
epsW=0.015,0.015
epsB=0.03
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv4a]
epsW=0.015
epsB=0.03
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv4b]
epsW=0.015
epsB=0.03
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv5a]
epsW=0.015
epsB=0.03
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv5b]
epsW=0.015
epsB=0.03
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048a]
epsW=0.015,0.015
epsB=0.03
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048b]
epsW=0.015,0.015
epsB=0.03
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048ba]
epsW=0.015,0.015
epsB=0.03
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048bb]
epsW=0.015,0.015
epsB=0.03
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc1000]
epsW=0.015,0.015
epsB=0.03
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[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 170 but running with minibatch 256 on krunch, also setting eps to 0.015:
# use def file 153
# not doing fading
# krunch
# /nobackup/kriz/tmp/ConvNet__2013-02-10_10.41.07
# logs/layers-170-256-0.015.log
# epoch 25: killed to do 4gpu experiments

View file

@ -0,0 +1,203 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv2a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv2b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv3a]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv3b]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv4a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv4b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv5a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv5b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048a]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048b]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048ba]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048bb]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc1000]
epsW=0.02,0.02
epsB=0.04
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[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 170 but running with minibatch 256 on krunch, also doubling eps to compensate for big batch size:
# use def file 153
# not doing fading
# /nobackup/kriz/tmp/ConvNet__2013-02-05_13.04.30
# logs/layers-170-256-double-eps.log
# epoch 61: set color noise to 0 from 0.1
# epoch 73: set conv1 epsw to 0 from 0.02
# after 84 epochs: logprob: 1.840643, 0.422726, 0.196598
# this matches #170 exactly, so im killing it to run 4gpu experiments

View file

@ -0,0 +1,199 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[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[250,0,4]
schedB=dexp[10,0,2]
[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[250,0,4]
schedB=dexp[10,0,2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048a]
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[250,0,4]
schedB=dexp[10,0,2]
[fc2048b]
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[250,0,4]
schedB=dexp[10,0,2]
[fc2048ba]
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[250,0,4]
schedB=dexp[10,0,2]
[fc2048bb]
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[250,0,4]
schedB=dexp[10,0,2]
[fc1000]
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[250,0,4]
schedB=dexp[10,0,2]
[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 170 but running with minibatch 256 on krunch:
# use def file 153
# not doing fading
# /nobackup/kriz/tmp/ConvNet__2013-02-05_12.50.10
# logs/layers-170-256.log

View file

@ -0,0 +1,412 @@
[conv1a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv1b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv1c]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv1d]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv2a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv2b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv2c]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv2d]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv3a]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv3b]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv3c]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv3d]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv4a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv4b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv4c]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv4d]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv5a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv5b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv5c]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv5d]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024a]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024b]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024c]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024d]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024ba]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024bb]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024bc]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024bd]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1000a]
epsW=0.04,0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1000b]
epsW=0.04,0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1000c]
epsW=0.04,0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1000d]
epsW=0.04,0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1b]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1d]
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
[rnorm2c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm2d]
scale=0.0001
pow=0.75
minDiv=2
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like 170-4gpu but with exp leraning rate
# uses config file #153-4gpu
# on krunch
# logs/layers-170-4gpu-exp.log
# /nobackup/kriz/tmp/ConvNet__2013-02-24_23.19.05
# epoch 14: moved to guppy9
# /nobackup/kriz/tmp/ConvNet__2013-02-24_23.19.05
# killed, exp is bad

View file

@ -0,0 +1,416 @@
[conv1a]
epsW=0.0000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=default
schedB=default
[conv1b]
epsW=0.0000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=default
schedB=default
[conv1c]
epsW=0.0000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=default
schedB=default
[conv1d]
epsW=0.0000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=default
schedB=default
[conv2a]
epsW=0.00008
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=default
schedB=default
[conv2b]
epsW=0.00008
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=default
schedB=default
[conv2c]
epsW=0.00008
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=default
schedB=default
[conv2d]
epsW=0.00008
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=default
schedB=default
[conv3a]
epsW=0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=default
schedB=default
[conv3b]
epsW=0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=default
schedB=default
[conv3c]
epsW=0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=default
schedB=default
[conv3d]
epsW=0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=default
schedB=default
[conv4a]
epsW=0.00008
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=default
schedB=default
[conv4b]
epsW=0.00008
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=default
schedB=default
[conv4c]
epsW=0.00008
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=default
schedB=default
[conv4d]
epsW=0.00008
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=default
schedB=default
[conv5a]
epsW=0.00008
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=default
schedB=default
[conv5b]
epsW=0.00008
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=default
schedB=default
[conv5c]
epsW=0.00008
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=default
schedB=default
[conv5d]
epsW=0.00008
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=default
schedB=default
[fc1024a]
epsW=0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=default
schedB=default
[fc1024b]
epsW=0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=default
schedB=default
[fc1024c]
epsW=0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=default
schedB=default
[fc1024d]
epsW=0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=default
schedB=default
[fc1024ba]
epsW=0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=default
schedB=default
[fc1024bb]
epsW=0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=default
schedB=default
[fc1024bc]
epsW=0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=default
schedB=default
[fc1024bd]
epsW=0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=default
schedB=default
[fc1000a]
epsW=0.00008,0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=default
schedB=default
[fc1000b]
epsW=0.00008,0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=default
schedB=default
[fc1000c]
epsW=0.00008,0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=default
schedB=default
[fc1000d]
epsW=0.00008,0.00008,0.00008,0.00008
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=default
schedB=default
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1b]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1d]
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
[rnorm2c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm2d]
scale=0.0001
pow=0.75
minDiv=2
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# uses config file #153-4gpu
# on krunch
# /nobackup/kriz/tmp/ConvNet__2013-02-07_16.20.22
# logs/layers-170-4gpu.log
# epoch 5: set mini to 256 from 128, epsw to 0.02 from 0.01
# epoch 7: set epsw to 0.003 from 0.02
# epoch 35: made backup to .bak
# set epsw to 0.0005 from 0.003
# epoch 55: it seems strangely bad, restarting from 35
# epoch 49: set color noise 0 from 0.1, set epsw to 0.00008 from 0.0005 on conv1
# epoch 52: set epsw to 0.00008 from 0.0005 everywhere, set epsw to 0 from 0.00008 on conv1
# epoch 63: killed.. its no better than 146

View file

@ -0,0 +1,234 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc2048a]
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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc2048b]
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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc2048ba]
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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc2048bb]
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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000]
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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=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
[pool2a]
quantF=half
quantB=half
[pool2b]
quantF=half
quantB=half
[pool3a]
quantF=half
quantB=half
[pool3b]
quantF=half
quantB=half
[hs1a]
quantF=half
quantB=half
[hs1b]
quantF=half
quantB=half
[hs2a]
quantF=half
quantB=half
[hs2b]
quantF=half
quantB=half
# this is like #170 (def file 153) but uses half quantization on all communicated layers
# logs/layers-170-quant.log
# /nobackup/kriz/tmp/ConvNet__2013-02-21_22.33.24
# guppy9
# epoch 68: set color noise to 0 from 0.1
# epoch 74: set epsw conv1 to 0 from 0.01
# logprob: 1.861123, 0.425700, 0.199970
# multiview logprob: 1.749848, 0.405240, 0.184160

242
layers/layer-params-170.cfg Normal file
View file

@ -0,0 +1,242 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc2048a]
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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc2048b]
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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc2048ba]
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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc2048bb]
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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000]
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[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=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 #159 (so uses same def file: 153) but with learning rate decaying to 0.00004
# not starting this because its first ~24 epochs are gonna be completley identical to 169, so might as well wait for that one
# 169 backup at 23 epochs: /nobackup/kriz/tmp/ConvNet__2013-01-07_04.03.15.bak (guppy9)
# on guppy8
# /nobackup/kriz/tmp/ConvNet__2013-01-07_04.03.15.layer-170
# logs/layers-170.log
# epoch 68: set color noise to 0 from 0.1
#NOTE: free gpus 2,3 after done
# epoch 74: set conv1 epsw to 0 from 0.01
# logprob: 1.838933, 0.423407, 0.195827
# multiview logprob: 1.730076, 0.405280, 0.180480
# resuming for 55 more epochs, this time adding in the fade-in of images with prob 5%:
# /nobackup/kriz/tmp/ConvNet__2013-01-07_04.03.15.layer-170.fade
# guppy8
# logs/layers-170-contfade.log
# epoch 107: set epochs to 300 to increase learning rate, also turned on 0.1
# epoch 127: set epsw to 0.0025148669 from 0.01, set schedw to default from dexp[250,0,4]
# epoch 127: set epsw to 0.00063245555 from 0.0025148669
# epoch 127: set epsw to 0.0015874011 -- which is the correct one for thi slevel
# epoch 127: set epsw to 0.00025198421
# epoch 137: set color noise to 0 from 0.1, also turned off fade-in
# restart from scratch with fade:
# logs/layers-170-fade.log
# /nobackup/kriz/tmp/ConvNet__2013-01-31_18.52.37
# guppy8
# epoch 61: set color noise to 0 from 0.1
# epoch 76: set conv1 epsw to 0 from 0.01
# logprob: 1.831731, 0.422944, 0.195334
# multiview: logprob: 1.726111, 0.402500, 0.181160
# now with fade probability 0.15 instead of 0.05:
# krunch
# logs/layers-170-fade-0.15.log
# /nobackup/kriz/tmp/ConvNet__2013-02-06_08.27.55
# moving to guppy9
# epoch 68: set color noise to 0 from 0.1
# epoch 74: set conv1 epsw to 0 from 0.1, set fade prob to 0.05 from 0.15
# logprob: 1.836181, 0.425008, 0.196351
# restart normal 170 but on 2 gpus which can't talk to each other -- just to make sure it'll work fine, and to measure the effect of quantization (#170-quant)
# guppy9
# logs/layers-170-gpu-1-2.log
# /nobackup/kriz/tmp/ConvNet__2013-02-23_12.16.26
# epoch 60: set color noise to 0 from 0.1
# epoch 63: moved to guppy5
# /nobackup_a/kriz/tmp/ConvNet__2013-02-23_12.16.26/

202
layers/layer-params-171.cfg Normal file
View file

@ -0,0 +1,202 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[500,5]
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[500,5]
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[500,5]
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[500,5]
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[500,5]
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[500,5]
schedB=dexp[10,2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[500,5]
schedB=dexp[10,2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[500,5]
schedB=dexp[10,2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[500,5]
schedB=dexp[10,2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[500,5]
schedB=dexp[10,2]
[fc2048a]
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[500,5]
schedB=dexp[10,2]
[fc2048b]
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[500,5]
schedB=dexp[10,2]
[fc2048ba]
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[500,5]
schedB=dexp[10,2]
[fc2048bb]
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[500,5]
schedB=dexp[10,2]
[fc1000]
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[500,5]
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 #169 (so uses same def file: 153) but with 5 epsw levels
# on guppy7
# /nobackup/kriz/tmp/ConvNet__2013-01-07_04.07.16
# logs/layers-171.log
# epoch 60: set color noise to 0 from 0.1
# epoch 72: set conv1 epsw to 0 from 0.01
# logprob: 1.843676, 0.423356, 0.197664
# multiview logprob: 1.735358, 0.404200, 0.181400

202
layers/layer-params-172.cfg Normal file
View file

@ -0,0 +1,202 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=jdexp[1000,4]
schedB=dexp[10,2]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=jdexp[1000,4]
schedB=dexp[10,2]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=jdexp[1000,4]
schedB=dexp[10,2]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=jdexp[1000,4]
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=jdexp[1000,4]
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=jdexp[1000,4]
schedB=dexp[10,2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=jdexp[1000,4]
schedB=dexp[10,2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=jdexp[1000,4]
schedB=dexp[10,2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=jdexp[1000,4]
schedB=dexp[10,2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=jdexp[1000,4]
schedB=dexp[10,2]
[fc2048a]
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=jdexp[1000,4]
schedB=dexp[10,2]
[fc2048b]
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=jdexp[1000,4]
schedB=dexp[10,2]
[fc2048ba]
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=jdexp[1000,4]
schedB=dexp[10,2]
[fc2048bb]
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=jdexp[1000,4]
schedB=dexp[10,2]
[fc1000]
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=jdexp[1000,4]
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 #159 (so uses same def file: 153) but with jdexp
# on guppy9
# /nobackup/kriz/tmp/ConvNet__2013-01-07_06.07.30
# logs/layers-172.log
# epoch 60: set color noise to 0 from 0.1
# epoch 64: made backup to /nobackup/kriz/tmp/ConvNet__2013-01-07_06.07.30.bak
# epoch 76: set conv1 epsw to 0 from 0.01
# epoch 84: killed, its at 1.898

209
layers/layer-params-174.cfg Normal file
View file

@ -0,0 +1,209 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=jdexp[500,0,4]
schedB=dexp[10,0,2]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=jdexp[500,0,4]
schedB=dexp[10,0,2]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=jdexp[500,0,4]
schedB=dexp[10,0,2]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=jdexp[500,0,4]
schedB=dexp[10,0,2]
[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=jdexp[500,0,4]
schedB=dexp[10,0,2]
[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=jdexp[500,0,4]
schedB=dexp[10,0,2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=jdexp[500,0,4]
schedB=dexp[10,0,2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=jdexp[500,0,4]
schedB=dexp[10,0,2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=jdexp[500,0,4]
schedB=dexp[10,0,2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=jdexp[500,0,4]
schedB=dexp[10,0,2]
[fc2048a]
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=jdexp[500,0,4]
schedB=dexp[10,0,2]
[fc2048b]
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=jdexp[500,0,4]
schedB=dexp[10,0,2]
[fc2048ba]
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=jdexp[500,0,4]
schedB=dexp[10,0,2]
[fc2048bb]
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=jdexp[500,0,4]
schedB=dexp[10,0,2]
[fc1000]
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=jdexp[500,0,4]
schedB=dexp[10,0,2]
[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 #159 (so uses same def file: 153) but with jdexp
# this is also like 172, which uses jdexp, but here i'm usign the kind of jdexp that jumps to geometric mean instead of previous level (and also 172 decayed learning rate by factor of 1000 instead of 500)
# its also like 169, in that it decays learning rate to 0.00002, but this one's jumpy. so it should be compared to 169.
# on gpu
# initialized from 24 epochs of #169
# logs/layers-174.log
# /storage/tmp/ConvNet__2013-01-07_04.03.15.layers-174
# epoch 58: made backup to /storage/tmp/ConvNet__2013-01-07_04.03.15.layers-174.bak
# epoch 62: set color noise to 0 from 0.1
# epoch 74: set conv1 epsw to 0 from 0.01
# epoch 86: killed for noisy epsw experiments
# moved to guppy9
# logprob: 1.825121, 0.425222, 0.193946
# multiview logprob: 1.729982, 0.404380, 0.180420

202
layers/layer-params-175.cfg Normal file
View file

@ -0,0 +1,202 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[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[250,0,4]
schedB=dexp[10,0,2]
[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[250,0,4]
schedB=dexp[10,0,2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048a]
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[250,0,4]
schedB=dexp[10,0,2]
[fc2048b]
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[250,0,4]
schedB=dexp[10,0,2]
[fc2048ba]
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[250,0,4]
schedB=dexp[10,0,2]
[fc2048bb]
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[250,0,4]
schedB=dexp[10,0,2]
[fc1000]
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[250,0,4]
schedB=dexp[10,0,2]
[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 #169 (so uses same def file: 153) but with learning rate decaying to 0.00008
# on guppy8
# logs/layers-175.log
# /nobackup/kriz/tmp/ConvNet__2013-01-07_04.03.15.layers-175
# epoch 60: set color noise to 0 from 0.1
# epoch 75: set conv1 epsw to 0 from 0.01
# epoch 91: changed schedw factor to 250 from 125
# logprob: 1.842557, 0.425160, 0.194863

210
layers/layer-params-177.cfg Normal file
View file

@ -0,0 +1,210 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[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[250,0,4]
schedB=dexp[10,0,2]
[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[250,0,4]
schedB=dexp[10,0,2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048a]
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[250,0,4]
schedB=dexp[10,0,2]
[fc2048b]
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[250,0,4]
schedB=dexp[10,0,2]
[fc2048ba]
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[250,0,4]
schedB=dexp[10,0,2]
[fc2048bb]
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[250,0,4]
schedB=dexp[10,0,2]
[fc1000]
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[250,0,4]
schedB=dexp[10,0,2]
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[hsconv4a]
enable=true
[hsconv4b]
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 #170 but with dropout over conv4
# /nobackup/kriz/tmp/ConvNet__2013-01-16_11.40.03
# on guppy9
# logs/layers-177.log
# epoch 60: set color noise to 0 from 0.1
# epoch 73: set epsw to 0 from 0.01 on conv1
# epoch 73: made backup to /nobackup/kriz/tmp/ConvNet__2013-01-16_11.40.03.bak
# consider using this backup to see what happens if i turn off conv dropout
# logprob: 1.836304, 0.424376, 0.197056
# multiview logprob: 1.744414, 0.406900, 0.183040

215
layers/layer-params-178.cfg Normal file
View file

@ -0,0 +1,215 @@
[conv1a]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv1b]
epsW=0.0
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[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[250,0,4]
schedB=dexp[10,0,2]
[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[250,0,4]
schedB=dexp[10,0,2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048a]
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[250,0,4]
schedB=dexp[10,0,2]
[fc2048b]
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[250,0,4]
schedB=dexp[10,0,2]
[fc2048ba]
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[250,0,4]
schedB=dexp[10,0,2]
[fc2048bb]
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[250,0,4]
schedB=dexp[10,0,2]
[fc1000]
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[250,0,4]
schedB=dexp[10,0,2]
[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
[pool2a]
doMax=true
[pool2b]
doMax=true
# this is like #170 but with pool2 now rand instead of max
# on guppy8
# logs/layers-178.log
# /nobackup/kriz/tmp/ConvNet__2013-01-19_01.39.21
# epoch 68: set color noise to 0 from 0.1
# epoch 69: set pool2a,pool2b domax=true from false
# 1.85 :(
# restart from 31, turning off rnadomness
# logs/layers-178-domax31.log
# /nobackup/kriz/tmp/ConvNet__2013-01-19_01.39.21.restart-31/
# epoch 60: set color noise to 0 from 0.1
# epoch 75: set epsw conv1 to 0 from 0.01
# logprob: 1.845021, 0.424038, 0.197017

202
layers/layer-params-180.cfg Normal file
View file

@ -0,0 +1,202 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=jdexp[250,0,4]
schedB=dexp[10,0,2]
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=jdexp[250,0,4]
schedB=dexp[10,0,2]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=jdexp[250,0,4]
schedB=dexp[10,0,2]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=jdexp[250,0,4]
schedB=dexp[10,0,2]
[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=jdexp[250,0,4]
schedB=dexp[10,0,2]
[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=jdexp[250,0,4]
schedB=dexp[10,0,2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=jdexp[250,0,4]
schedB=dexp[10,0,2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=jdexp[250,0,4]
schedB=dexp[10,0,2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=jdexp[250,0,4]
schedB=dexp[10,0,2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=jdexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048a]
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=jdexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048b]
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=jdexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048ba]
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=jdexp[250,0,4]
schedB=dexp[10,0,2]
[fc2048bb]
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=jdexp[250,0,4]
schedB=dexp[10,0,2]
[fc1000]
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=jdexp[250,0,4]
schedB=dexp[10,0,2]
[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 #159 (so uses same def file: 153) but with jdexp
# this is also like 174, which uses jdexp, but 172 decayed learning rate by factor of 500 instead of 250
# so this one should be compared to 174 (which decays learning rate less) and 170 (which decays learning rate the same, but is not jumpy)
# on gpu
# initialized from 24 epochs of #169
# logs/layers-180.log
# /storage/tmp/ConvNet__2013-01-07_04.03.15.layers-180

View file

@ -0,0 +1,451 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1c]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1d]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2c]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2d]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3a]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3b]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3c]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3d]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4c]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4d]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5c]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5d]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6a]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6b]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6c]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6d]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024a]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024b]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024c]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024d]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024ba]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024bb]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024bc]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024bd]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000a]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000b]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000c]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000d]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1b]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1d]
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
[rnorm2c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm2d]
scale=0.0001
pow=0.75
minDiv=2
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like 170-4gpu but with 6th conv layer and also 48 low-level filters
# this one is also meant to be trained for 26 epochs and with mini 128 -- lets see what happens
# on guppy9
# logs/layers-183-dexp-26epc.log
# /nobackup/kriz/tmp/ConvNet__2013-03-03_23.17.25
# # killed in favor of 184
# this is obvioulsy the wrong net to run -- it has way too many params.

View file

@ -0,0 +1,450 @@
[conv1a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv1b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv1c]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv1d]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv2a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv2b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv2c]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv2d]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv3a]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv3b]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv3c]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv3d]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv4a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv4b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv4c]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv4d]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv5a]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv5b]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv5c]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv5d]
epsW=0.04
epsB=0.08
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv6a]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv6b]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv6c]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[conv6d]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024a]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024b]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024c]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024d]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024ba]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024bb]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024bc]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1024bd]
epsW=0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1000a]
epsW=0.04,0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1000b]
epsW=0.04,0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1000c]
epsW=0.04,0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[fc1000d]
epsW=0.04,0.04,0.04,0.04
epsB=0.08
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=exp[tgtFactor=1000,noiseStdev=0]
schedB=exp[tgtFactor=10,noiseStdev=0]
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1b]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1d]
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
[rnorm2c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm2d]
scale=0.0001
pow=0.75
minDiv=2
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like 170-4gpu-exp but with 6th conv layer and also 48 low-level filters
# on krunch
# /nobackup/kriz/tmp/ConvNet__2013-02-26_20.53.34
# logs/layers-183.log
# killed, exp is bad

View file

@ -0,0 +1,450 @@
[conv1a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1c]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1d]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2c]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2d]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3a]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3b]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3c]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3d]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4c]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4d]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5c]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5d]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6a]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6b]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6c]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6d]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024a]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024b]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024c]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024d]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024ba]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024bb]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024bc]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024bd]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000a]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000b]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000c]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000d]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1b]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1d]
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
[rnorm2c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm2d]
scale=0.0001
pow=0.75
minDiv=2
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like 170-4gpu but with 6th conv layer and also 48 low-level filters
# on krunch
# logs/layers-183-dexp.log
# /nobackup/kriz/tmp/ConvNet__2013-03-03_23.14.01
# killed in favor of 184
# this is obvioulsy the wrong net to run -- it has way too many params.

View file

@ -0,0 +1,448 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1c]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1d]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2c]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2d]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3a]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3b]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3c]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3d]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4c]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4d]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5c]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5d]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6a]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6b]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6c]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6d]
epsW=0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024a]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024b]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024c]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024d]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024ba]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024bb]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024bc]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024bd]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000a]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000b]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000c]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000d]
epsW=0.01,0.01,0.01,0.01
epsB=0.02
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1b]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1d]
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
[rnorm2c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm2d]
scale=0.0001
pow=0.75
minDiv=2
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like 184 but for 26 epochs
# guppy9
# logs/layers-184-26epc.log
# /nobackup/kriz/tmp/ConvNet__2013-03-04_04.08.39

View file

@ -0,0 +1,450 @@
[conv1a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1c]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv1d]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2c]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv2d]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3a]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3b]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3c]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv3d]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4c]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv4d]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5a]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5b]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5c]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv5d]
epsW=0.02
epsB=0.04
momW=0.9
momB=0.9
wc=0.0005
wball=0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6a]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6b]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6c]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[conv6d]
epsW=0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024a]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024b]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024c]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024d]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024ba]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024bb]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024bc]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1024bd]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000a]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000b]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000c]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[fc1000d]
epsW=0.02,0.02,0.02,0.02
epsB=0.04
momW=0.9,0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005,0.0005
wball=0,0,0,0
schedW=dexp[tgtFactor=250,noiseStdev=0,numSteps=4]
schedB=dexp[tgtFactor=10,noiseStdev=0,numSteps=2]
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs1b]
enable=true
[hs1c]
enable=true
[hs1d]
enable=true
[hs2a]
enable=true
[hs2b]
enable=true
[hs2c]
enable=true
[hs2d]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1b]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm1d]
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
[rnorm2c]
scale=0.0001
pow=0.75
minDiv=2
[rnorm2d]
scale=0.0001
pow=0.75
minDiv=2
[cnorm2a]
scale=0.001
pow=0.75
[cnorm2b]
scale=0.001
pow=0.75
[cnorm2c]
scale=0.001
pow=0.75
[cnorm2d]
scale=0.001
pow=0.75
# this is like 183 but with half as many top-level conv filters so the # of params is not @#$%&!
# also the fc layers are not connected .. fully (not to 3)
# on krunch
# 50 epochs
# logs/layers-184.log
# /nobackup/kriz/tmp/ConvNet__2013-03-04_04.05.18

View file

@ -0,0 +1,162 @@
[conv1a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc10184]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# on guppy9
# this is like #101 (on gpu) but this is trained on imgnet-2009
# epoch 6: set epsw to 0.001 from 0.01
# epoch 14: set epsw to 0.0001 from 0.001
# epoch 19: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 27: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 30: set epsw to 0.00001 from 0.0001
# epoch 33: killed
# [3.4620055494832287, 0.69382157140195966, 0.43646610858041701]

View file

@ -0,0 +1,156 @@
[conv1a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.001,0.001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# on guppy7
# logs/layers-96-16k.log
# /ais/gobi3/u/kriz/tmp/ConvNet__2012-06-24_02.01.57
# epoch 5: set epsw to 0.001 from 0.01
# epoch 6: enabled dropout

View file

@ -0,0 +1,164 @@
[conv1a]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.00
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.00
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# on guppy9
# logs/layers-98-16kinit.log
# weights initialized from net trained on 16k imgnet for a few epochs: /ais/gobi3/u/kriz/tmp/ConvNet__2012-06-24_02.01.57
# /nobackup/kriz/tmp/ConvNet__2012-06-25_17.55.06
# logs/layers-98-16kinit.log
# epoch 30: set epsw to 0.0001 from 0.001
# epoch 44: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 51: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 64: set epsw to 0.00001 from 0.0001
# epoch 71: killed
# (294, 0.37132068707483007, 0.1679778095238095)

162
layers/layer-params-99.cfg Normal file
View file

@ -0,0 +1,162 @@
[conv1a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.005
[conv1b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0.005
[conv2a]
epsW=0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0.003,0.003,0.003
[conv2b]
epsW=0.01,0.01,0.01
epsB=0.002
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0.003,0.003,0.003
[conv3a]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.01
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.01,0.01
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# on guppy7
# this is like #97 (on gpu) but with different rnorm2
# logs/layers-99.log
# /nobackup/kriz/tmp/ConvNet__2012-06-26_20.35.00
# diff rnorm2 doesnt seem to stop conv2 filters from dying
# now trying wball on conv1, conv2
# logs/layers-99a.log
# /nobackup/kriz/tmp/ConvNet__2012-06-26_23.41.56
# /nobackup/kriz/tmp/ConvNet__2012-06-27_03.57.56
# lot of filters seem to remain random on conv2

View file

@ -0,0 +1,158 @@
[conv1a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0001,0.0001,0.001
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0.00,0.00,0.00
[conv2b]
epsW=0.0001,0.0001,0.001
epsB=0.02
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0.00,0.00,0.00
[conv3a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc10003]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[crossent]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# on guppy7
# this is like #97, but on flickr
# also initialized from #97 on imgnet
# /ais/gobi3/u/kriz/tmp/ConvNet__2012-06-30_23.41.20
# epoch 59: set epsw to 0.0001 from 0.001
# epoch 78: killed because i realized its not really fair

View file

@ -0,0 +1,161 @@
[conv1a]
epsW=0.0000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.0
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000,0.0000
epsB=0.0
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0.00,0.00,0.00
[conv2b]
epsW=0.0000,0.0000,0.0000
epsB=0.0
momW=0.9,0.9,0.9
momB=0.9
wc=0.0005,0.0005,0.0005
wball=0.00,0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc10003]
epsW=0.00001,0.00001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[crossent]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# on guppy7
# this is like #97, but on flickr
# /ais/gobi3/u/kriz/tmp/ConvNet__2012-06-30_17.00.32
# epoch 85: set epsw to 0.0001 from 0.001
# epoch 108: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 120: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 136: set epsw to 0.00001 from 0.0001
# epoch 162: killed

View file

@ -0,0 +1,155 @@
[conv1a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.00005
wball=0.00
[conv1b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.00005
wball=0.00
[conv2a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.00005,0.00005
wball=0.00,0.00
[conv2b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.00005,0.00005
wball=0.00,0.00
[conv3a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.00005,0.00005
wball=0,0
[conv3b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.00005,0.00005
wball=0,0
[conv4a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.00005
wball=0
[conv4b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.00005
wball=0
[conv5a]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.00005
wball=0
[conv5b]
epsW=0.01
epsB=0.02
momW=0.9
momB=0.9
wc=0.00005
wball=0
[fc2048a]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.00005,0.00005
wball=0,0
[fc2048b]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.00005,0.00005
wball=0,0
[fc2048ba]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.00005,0.00005
wball=0,0
[fc2048bb]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.00005,0.00005
wball=0,0
[fc10003]
epsW=0.01,0.01
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.00005,0.00005
wball=0,0
[rcost]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# on guppy9
# this is like #101, but on flickr, and with robust flickr cost
#

View file

@ -0,0 +1,156 @@
[conv1a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.0001
epsB=0.02
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc10003]
epsW=0.0001,0.0001
epsB=0.02
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[sqdiff]
coeff=1
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# on guppy9
# this is like #101, but on flickr, and with sqdiff objective
# /ais/gobi3/u/kriz/tmp/ConvNet__2012-07-04_23.30.19
# epoch 15: set wc to 0.0005 from 0.00005, set epsw to 0.001 from 0.01
# epoch 93: set epsw to 0.0001 from 0.001

View file

@ -0,0 +1,164 @@
[conv1a]
epsW=0.0000
epsB=0.000
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv1b]
epsW=0.0000
epsB=0.000
momW=0.9
momB=0.9
wc=0.0005
wball=0.00
[conv2a]
epsW=0.0000,0.0000
epsB=0.000
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv2b]
epsW=0.0000,0.0000
epsB=0.000
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0.00,0.00
[conv3a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv3b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[conv4a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv4b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5a]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[conv5b]
epsW=0.00001
epsB=0.002
momW=0.9
momB=0.9
wc=0.0005
wball=0
[fc2048a]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048b]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048ba]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc2048bb]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[fc1000]
epsW=0.00001,0.00001
epsB=0.002
momW=0.9,0.9
momB=0.9
wc=0.0005,0.0005
wball=0,0
[logprob]
coeff=1
topk=5
[hs1a]
enable=true
[hs2a]
enable=true
[hs1b]
enable=true
[hs2b]
enable=true
[rnorm1a]
scale=0.0001
pow=0.75
[rnorm1b]
scale=0.0001
pow=0.75
[rnorm2a]
scale=0.0001
pow=0.75
[rnorm2b]
scale=0.0001
pow=0.75
# on guppy9
# logs/layers-inet-5layer-conv94-2gpu.log
# /nobackup/kriz/tmp/ConvNet__2012-06-18_18.34.17
# logs/layers-inet-5layer-conv94-2gpu.log
# epoch 13: set epsw to 0.001 from 0.01
# epoch 46: set epsw to 0.0001 from 0.001
# epoch 55: set epsw to 0.00001 from 0.0001 on conv1,conv2
# set color noise to 0 from 0.1
# epoch 62: set epsw to 0 from 0.00001 on conv1,conv2
# epoch 84: set epsw to 0.00001 from 0.0001
# epoch 90: killed
# 0.38107167346938753, 0.17608947619047613

314
layers/layers-100.cfg Normal file
View file

@ -0,0 +1,314 @@
[data]
type=data
dataIdx=0
[labels]
type=data
dataIdx=1
[blur0]
type=blur
inputs=data
stdev=4
filterSize=9
channels=3
gpu=0
[nails0]
type=nailbed
inputs=blur0
stride=4
channels=3
[conv1a]
type=conv
inputs=data
channels=3
filters=32
padding=0
stride=4
filterSize=11
initW=0.01
partialSum=5
sharedBiases=1
gpu=0
[conv1b]
type=conv
inputs=data
channels=3
filters=32
padding=0
stride=4
filterSize=11
initW=0.01
partialSum=5
sharedBiases=1
gpu=1
[pool1a]
type=pool
pool=max
inputs=conv1a
sizeX=3
stride=2
channels=32
neuron=relu
[pool1b]
type=pool
pool=max
inputs=conv1b
sizeX=3
stride=2
channels=32
neuron=relu
[rnorm1a]
type=cmrnorm
inputs=pool1a
channels=32
size=9
[rnorm1b]
type=cmrnorm
inputs=pool1b
channels=32
size=9
[conv2a]
type=conv
inputs=nails0,rnorm1a,rnorm1b
filters=128,64,64
padding=0,2,2
stride=2,1,1
filterSize=5,5,5
channels=3,32,32
initW=0.01,0.01,0.01
initB=1
partialSum=3
sharedBiases=1
neuron=relu
randSparse=0,1,1
groups=1,2,2
gpu=0
[conv2b]
type=conv
inputs=nails0,rnorm1a,rnorm1b
filters=128,64,64
padding=0,2,2
stride=2,1,1
filterSize=5,5,5
channels=3,32,32
initW=0.01,0.01,0.01
initB=1
partialSum=3
sharedBiases=1
neuron=relu
randSparse=0,1,1
groups=1,2,2
gpu=1
[rnorm2a]
type=cmrnorm
inputs=conv2a
channels=128
size=9
[rnorm2b]
type=cmrnorm
inputs=conv2b
channels=128
size=9
[pool2a]
type=pool
pool=max
inputs=rnorm2a
sizeX=3
stride=2
channels=128
[pool2b]
type=pool
pool=max
inputs=rnorm2b
sizeX=3
stride=2
channels=128
[conv3a]
type=conv
inputs=pool2a,pool2b
filters=192,192
padding=1,1
stride=1,1
filterSize=3,3
channels=128,128
initW=0.03,0.03
partialSum=13
sharedBiases=1
neuron=relu
gpu=0
[conv3b]
type=conv
inputs=pool2a,pool2b
filters=192,192
padding=1,1
stride=1,1
filterSize=3,3
channels=128,128
initW=0.03,0.03
partialSum=13
sharedBiases=1
neuron=relu
gpu=1
[conv4a]
type=conv
inputs=conv3a
filters=192
padding=1
stride=1
filterSize=3
channels=192
neuron=relu
initW=0.03
initB=1
partialSum=13
sharedBiases=1
[conv4b]
type=conv
inputs=conv3b
filters=192
padding=1
stride=1
filterSize=3
channels=192
neuron=relu
initW=0.03
initB=1
partialSum=13
sharedBiases=1
[conv5a]
type=conv
inputs=conv4a
filters=128
padding=1
stride=1
filterSize=3
channels=192
initW=0.03
initB=1
partialSum=13
groups=1
randSparse=0
[conv5b]
type=conv
inputs=conv4b
filters=128
padding=1
stride=1
filterSize=3
channels=192
initW=0.03
initB=1
partialSum=13
groups=1
randSparse=0
[pool3a]
type=pool
pool=max
inputs=conv5a
sizeX=3
stride=2
channels=128
neuron=relu
[pool3b]
type=pool
pool=max
inputs=conv5b
sizeX=3
stride=2
channels=128
neuron=relu
[fc2048a]
type=fc
inputs=pool3a,pool3b
outputs=2048
initW=0.01,0.01
initB=1
neuron=relu
gpu=0
[fc2048b]
type=fc
inputs=pool3a,pool3b
outputs=2048
initW=0.01,0.01
initB=1
neuron=relu
gpu=1
[hs1a]
type=hs
keep=0.5
inputs=fc2048a
[hs1b]
type=hs
keep=0.5
inputs=fc2048b
[fc2048ba]
type=fc
inputs=hs1a,hs1b
outputs=2048
initW=0.01,0.01
initB=1
neuron=relu
gpu=0
[fc2048bb]
type=fc
inputs=hs1b,hs1a
outputs=2048
initW=0.01,0.01
initB=1
neuron=relu
gpu=1
[hs2a]
type=hs
keep=0.5
inputs=fc2048ba
[hs2b]
type=hs
keep=0.5
inputs=fc2048bb
[fc1000]
type=fc
outputs=1000
inputs=hs2a,hs2b
initW=0.01,0.01
gpu=1
[probs]
type=softmax
inputs=fc1000
[logprob]
type=cost.logreg
inputs=labels,probs
gpu=1

322
layers/layers-106.cfg Normal file
View file

@ -0,0 +1,322 @@
[data]
type=data
dataIdx=0
[labels]
type=data
dataIdx=1
[blur0]
type=blur
inputs=data
stdev=4
filterSize=9
channels=3
gpu=0
[nails0]
type=nailbed
inputs=blur0
stride=4
channels=3
[conv1a]
type=conv
inputs=data
channels=3
filters=32
padding=0
stride=4
filterSize=11
initW=0.01
partialSum=5
sharedBiases=1
gpu=0
[conv1b]
type=conv
inputs=data
channels=3
filters=32
padding=0
stride=4
filterSize=11
initW=0.01
partialSum=5
sharedBiases=1
gpu=1
[pool1a]
type=pool
pool=max
inputs=conv1a
sizeX=3
stride=2
channels=32
neuron=relu
[pool1b]
type=pool
pool=max
inputs=conv1b
sizeX=3
stride=2
channels=32
neuron=relu
[rnorm1a]
type=cmrnorm
inputs=pool1a
channels=32
size=9
[rnorm1b]
type=cmrnorm
inputs=pool1b
channels=32
size=9
[conv2a]
type=conv
inputs=nails0,rnorm1a
filters=128,128
padding=0,2
stride=2,1
filterSize=5,5
channels=3,32
initW=0.01,0.01
initB=1
partialSum=3
sharedBiases=1
neuron=relu
gpu=0
[conv2b]
type=conv
inputs=nails0,rnorm1b
filters=128,128
padding=0,2
stride=2,1
filterSize=5,5
channels=3,32
initW=0.01,0.01
initB=1
partialSum=3
sharedBiases=1
neuron=relu
gpu=1
[rnorm2a]
type=cmrnorm
inputs=conv2a
channels=128
size=9
[rnorm2b]
type=cmrnorm
inputs=conv2b
channels=128
size=9
[cnorm2a]
type=cnorm
inputs=rnorm2a
channels=128
size=5
[cnorm2b]
type=cnorm
inputs=rnorm2b
channels=128
size=5
[pool2a]
type=pool
pool=max
inputs=cnorm2a
sizeX=3
stride=2
channels=128
[pool2b]
type=pool
pool=max
inputs=cnorm2b
sizeX=3
stride=2
channels=128
[conv3a]
type=conv
inputs=pool2a,pool2b
filters=192,192
padding=1,1
stride=1,1
filterSize=3,3
channels=128,128
initW=0.03,0.03
partialSum=13
sharedBiases=1
neuron=relu
gpu=0
[conv3b]
type=conv
inputs=pool2a,pool2b
filters=192,192
padding=1,1
stride=1,1
filterSize=3,3
channels=128,128
initW=0.03,0.03
partialSum=13
sharedBiases=1
neuron=relu
gpu=1
[conv4a]
type=conv
inputs=conv3a
filters=192
padding=1
stride=1
filterSize=3
channels=192
neuron=relu
initW=0.03
initB=1
partialSum=13
sharedBiases=1
[conv4b]
type=conv
inputs=conv3b
filters=192
padding=1
stride=1
filterSize=3
channels=192
neuron=relu
initW=0.03
initB=1
partialSum=13
sharedBiases=1
[conv5a]
type=conv
inputs=conv4a
filters=128
padding=1
stride=1
filterSize=3
channels=192
initW=0.03
initB=1
partialSum=13
groups=1
randSparse=0
[conv5b]
type=conv
inputs=conv4b
filters=128
padding=1
stride=1
filterSize=3
channels=192
initW=0.03
initB=1
partialSum=13
groups=1
randSparse=0
[pool3a]
type=pool
pool=max
inputs=conv5a
sizeX=3
stride=2
channels=128
neuron=relu
[pool3b]
type=pool
pool=max
inputs=conv5b
sizeX=3
stride=2
channels=128
neuron=relu
[fc2048a]
type=fc
inputs=pool3a,pool3b
outputs=2048
initW=0.01,0.01
initB=1
neuron=relu
gpu=0
[fc2048b]
type=fc
inputs=pool3a,pool3b
outputs=2048
initW=0.01,0.01
initB=1
neuron=relu
gpu=1
[hs1a]
type=hs
keep=0.5
inputs=fc2048a
[hs1b]
type=hs
keep=0.5
inputs=fc2048b
[fc2048ba]
type=fc
inputs=hs1a,hs1b
outputs=2048
initW=0.01,0.01
initB=1
neuron=relu
gpu=0
[fc2048bb]
type=fc
inputs=hs1b,hs1a
outputs=2048
initW=0.01,0.01
initB=1
neuron=relu
gpu=1
[hs2a]
type=hs
keep=0.5
inputs=fc2048ba
[hs2b]
type=hs
keep=0.5
inputs=fc2048bb
[fc1000]
type=fc
outputs=1000
inputs=hs2a,hs2b
initW=0.01,0.01
gpu=1
[probs]
type=softmax
inputs=fc1000
[logprob]
type=cost.logreg
inputs=labels,probs
gpu=1

Some files were not shown because too many files have changed in this diff Show more