Chao Xing 2016-02-01

来自cslt Wiki
2016年2月1日 (一) 02:21Xingchao讨论 | 贡献的版本

(差异) ←上一版本 | 最后版本 (差异) | 下一版本→ (差异)
跳转至: 导航搜索

LAST WEEK


Done:

  • Multi-task transfer learning during transfer step pretrain.
  • Find some strategies among each hyper-parameters.
    • Steps determine the train parameters information content, which means larger steps need larger iterations and larger learning rate.
    • Learning rate should not decay in this program.
    • Larger learning may obtain better performance.

THIS WEEK


Want to do:

  • In my test:
    • Steps must consist among source images pretrain, target images pretrain, transfer pretrain and fine tune.
    • In transfer pretrain learning rate could have a larger number lead iteration count could be small, for last two days experiment, I choose 1e-2 as transfer learning rate.
    • In fine tune step, the learning rate should be as small as possible but not too small, in this experiment I choose 1e-3.
  • So this week I will focus on this experiment.