[caffe学习笔记][04][生成solver文件]

说明：

caffe通过配置文件prototxt来描设置训练参数，通过Python接口来生成solver配置文件比较简单。

步骤：

1.生成配置文件

touch create_solver_prototxt.py

spyder create_solver_prototxt.py

#!/usr/bin/env python
# coding: utf-8
"""
yuandanfei Editor

This is a temporary script file.
"""
path = '/home/yuandanfei/work/caffe/mnist/out/'  #solver prototxt path
solver_file = path + 'solver.prototxt'

sp = {}
sp['train_net'] = '"' + path + 'train.prototxt"' #train prototxt path
sp['test_net'] = '"' + path + 'test.prototxt"'   #test prototxt path

sp['test_interval'] = '938'                      #test interval = train sample numbers/train batch numbers
sp['test_iter'] = '313'                          #test iter = test sample numbers/test batch numbers
sp['max_iter'] = '93800'                         #train numbers = max iter/test interval

sp['lr_policy'] = '"step"'                       #learning rate change strategy = base_lr*gamma^(floor(iter/stepsize))
sp['base_lr'] = '0.001'                          #base learning rate
sp['gamma'] = '0.1'                              #learning rate change index
sp['stepsize'] = '31267'                         #learning rate change frequece = max iter/stepsize
sp['momentum'] = '0.9'                           #learnig momentum
sp['weight_decay'] = '0.0005'                    #weight decay

sp['display'] = '938'                            #display log interval = display/test interval
sp['snapshot'] = '9380'                          #save model interval = snapshot/test interval
sp['snapshot_prefix'] = '"snapshot"'             #save model prefix

sp['solver_type'] = 'SGD'                        #optimization algorithm
sp['solver_mode'] = 'GPU'                        #use GPU

def write_solver():
    with open(solver_file, 'w') as f:
        for key, value in sorted(sp.items()):
            if not(type(value) is str):
                raise TypeError('All solver parameters must be strings')
            f.write('%s: %s\n' % (key, value))
            
if __name__ == '__main__':
    write_solver()

参考资料：

https://www.cnblogs.com/denny402/p/5679154.html

https://www.cnblogs.com/denny402/p/5074049.html

https://www.cnblogs.com/denny402/p/5074212.html