-
Notifications
You must be signed in to change notification settings - Fork 4
Submission History
limiao edited this page Dec 11, 2017
·
17 revisions
- 使用随机预测多次的方式进行 voting,每个模型使用的 voting 票数如下
NO | model | 手动调权 | 自动调权 |
---|---|---|---|
1 | Visual Attention RL 1080k | 20 | 14 |
2 | Visual Attention RL 840k | 20 | 6 |
3 | Visual-Semantic Attention | 16 | 14 |
4 | Stacked Visual Attention MC-RL | 20 | 20 |
5 | Stacked Visual Attention RL | 20 | 0 |
6 | Stacked Visual Attention | 16 | 4 |
7 | Localized Visual Attention | 8 | 4 |
8 | Semantic Attention | 8 | 2 |
第一次提交: 手工调权 20 20 16 20 20 16 8 8
第二次提交: 自动调权 14 6 14 20 0 4 4 2
- 开发集结果
NO | model | BLEU_4 | CIDEr | METEOR | ROUGE_L |
---|---|---|---|---|---|
1 | 手工调权 voting | 0.60829 | 1.95772 | 0.41610 | 0.70414 |
2 | 自动调权 voting | 0.60836 | 1.96232 | 0.41654 | 0.70383 |
- B 集结果
NO | model | score | BLEU_4 | CIDEr | METEOR | ROUGE_L |
---|---|---|---|---|---|---|
1 | 最终结果 | 5.68515 | 0.72666 | 2.00006 | 0.41685 | 0.70614 |
NO | model | step | score | BLEU_4 | CIDEr | METEOR | ROUGE_L |
---|---|---|---|---|---|---|---|
1 | show_and_tell | 230000 | 6.54236 | 0.73105 | 1.23151 | 0.35082 | 0.60045 |
2 | show_and_tell | 529346 | 6.57050 | 0.73298 | 1.24599 | 0.35132 | 0.60499 |
3 | show_and_tell finetune | 615359 | 6.49749 | 0.73934 | 1.28029 | 0.35515 | 0.61231 |
4 | show_and_tell finetune* | 840000 | 6.57731 | 0.59671 | 1.37921 | 0.36678 | 0.62941 |
5 | show_and_tell finetune** | 840000 | 6.52121 | 0.61224 | 1.44411 | 0.37362 | 0.63947 |
6 | show_and_tell advanced*** (Visual Attention, finetune with decay) | 580699 | 6.14249 | 0.68483 | 1.79038 | 0.40753 | 0.6829 |
7 | show_and_tell advanced_da*** (Visual Attention, finetune with decay, data augmentation) | 640713 | 6.67553 | 0.69240 | 1.83234 | 0.41117 | 0.68795 |
8 | show_and_tell advanced_da_rl**** (Visual Attention, finetune with decay, data augmentation, RL training) | 840423 | 6.53832 | 0.70151 | 1.88851 | 0.4071 | 0.69014 |
* 这是 ShowAndTellInGraphModel 的结果,finetune 时学习率调大到 0.001
** 这是 ShowAndTellInGraphModel 的结果,finetune 时学习率接续训练时的学习率开始 decay
*** 从 105000 开始 finetune, 学习率接续训练时学习率继续 decay, initlr=1.0, decay=0.6
**** 基于上次提交最好的结果,使用RL训练,使用Adam优化,开始学习率为5e-5,每3个epoch decay一次,decay=0.8