caffe training solver configuration

Learning rate

NET: "xxx.prototxt"
test_iter: The total number of test image # / test batches
test_interval: 2000
test_initialization: false

base_lr: 0.01
lr_policy: "multistep" #lr_policy: "fixed"
gamma: 0.1
stepvalue: 100000
stepvalue: 140000
stepvalue: 160000
max_iter: 200000

iter_siz to: 2

weight_decay: 0.0005
display: 1000
momentum: 0.9

snapshot: 5000
snapshot_prefix: "./models/xxx_"
solver_mode: GPU

Guess you like

Origin www.cnblogs.com/codingstorm/p/11275563.html