caffe不同lr_policy参数设置方法，计算公式，超参数设置

# the definition of neural network model
net: "t_v.prototxt"
            # test_iter* batchsize=test
test_iter: 355 
            # train batch*test_interval to test                        
test_interval: 200                     
test_initialization: false
            # train display interval      
display: 200  
            # train100iteration average,defalut 1 batch loss                          
average_loss: 100                      
base_lr: 0.001
lr_policy: "poly"
stepsize: 6000
gamma: 0.96
            # The max number of iterations
max_iter: 100000                      
power: 1.0
momentum: 0.9
            # weight decay item, in case of overfitting
weight_decay: 0.0002 
            # save once every 50 training iterations                  
snapshot: 1000   
            # save path                      
snapshot_prefix: "inception-v1-sa-"    
solver_mode: GPU
            #  batchsize * itersize竧rue gradient decent
#iter_size:2    

net: "two_train.prototxt"
test_initialization: false
test_iter: 5000
test_interval: 400
base_lr: 0.08

#lr_policy: "step"
#gamma: 0.1
#stepsize: 50000

lr_policy: "multistep"
gamma: 0.1
stepvalue: 75000
stepvalue: 130000
stepvalue: 170000

display: 200
average_loss: 100
max_iter: 200000
momentum: 0.9
weight_decay: 0.0005
snapshot: 600
snapshot_prefix: "caffemodel/two"
#type:"Adam"
solver_mode: GPU

caffe不同lr_policy参数设置方法
每种学习率的图形表示
https://blog.csdn.net/zong596568821xp/article/details/80917387