-
Notifications
You must be signed in to change notification settings - Fork 4
持续优化方向
Heda Wang edited this page Oct 25, 2017
·
10 revisions
- 水平镜像数据扩增 @limiao
- Visual Attention @suying @heda
效果:
Visual Attention 在同等训练条件下提升明显ROUGE_L 0.672 --> 0.685
@heda
复现 show-attend-and-tell 论文,固定 seq_embedding,ROUGE_L 0.672 --> 0.662
,可能是固定 embedding 导致的 @suying
- Semantic Attention @limiao @heda
2.1. 预测 Concept,使用 concept embedding 作为 semantic memory
2.2. Attribute-based Attention
-
Lexical embedding 词义信息矩阵 @heda
-
多尺度的 Attention
-
何时做 Attention, Know when to look @heda
-
其他的 Image Model
-
其他的 RNNCell (如多层的、Res、Fast-forward)
-
Cascade Caption, 用 Caption 再进行一次 Caption,第一次的 Caption 用于 Attention 输入
- Reranking (DSSM, etc) @heda
- Fine-tuning @heda
效果:
固定 image model 训练 20 万步之后基本收敛,但是 finetune 可以进一步提升效果ROUGE_L 0.606 --> 0.621
Fine-tuning 调节学习率0.0005 --> 0.001
可进一步提升效果ROUGE_L 0.619 --> 0.629
.
- Scheduled Sampling @heda
效果: 线性的增加采样率 train 阶段 0.0 --> 0.25, finetune 阶段 0.0 --> 0.5 收益持平。
-
Badcase 分析
-
OOV 所占的 Coverage