-
Notifications
You must be signed in to change notification settings - Fork 0
/
scheduler_hypers.txt
25 lines (22 loc) · 1.21 KB
/
scheduler_hypers.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
Resnet101:
(Run 5) ASWT Model 1: gamma=0.6, n = 17, slackProp=0.35, init_lr = 0.1 Repeat Run 18-26 (X)
(Run 6) ASWT Model 2: gamma=0.6, n = 17, slackProp=0.05, init_lr = 0.1 Repeat Run 27-35 (X)
GoogleNet:
(Run 5) ASWT Model 1: gamma=0.76, n = 19, slackProp=.05, init_lr = 0.1 Repeat Run 18-26 (X)
(Run 6) ASWT Model 2: gamma=0.76, n = 19, slackProp=.35, init_lr = 0.1 Repeat Run 27-35 (X)
StepLR:
(Run 7) gamma = 0.1, stepsize=50, init_lr = 0.1
(Run 11) gamma = 0.5, stepsize=50, init_lr = 0.1
(Run 14) gamma = 0.5, stepsize=50, init_lr = 0.25 Repeat Run 36-42, 150, 250 (ResNet: 40, 41, 250, GoogLeNet: 150, 250)
ExponentialLR:
(Run 8) gamma = 0.9, init_lr = 0.1
(Run 12) gamma = 0.95, init_lr = 0.1
ReduceLR:
(Run 9) gamma = 0.1, patience = 10, init_lr = 0.1
(Run 13) gamma = 0.1, patience = 20, init_lr = 0.1
(Run 15) gamma = 0.1, patience = 20, init_lr = 0.25 Repeat Run 43-51 (ResNet: X, GoogLeNet: X)
CyclicLR:
"cyclic1" (Run 16) mode = triangular2, base_lr = 0.01, max_lr = 0.026, stepsizeup = 30000 Repeat Run 52-60 (ResNet:58, 59, 60, GoogLeNet: X)
"cyclic2" (Run 17) mode = triangular, base_lr = 0.1, max_lr = 0.3, stepsizeup = default
ADAM:
(Run 10) LR = 0.1