周一工作:
- 改善dssm网络,添加matrix attention模块,效果有改善,当前inference结果:acc:0.709 auc:0.715 precision:0.945;
- 对mv-dssm的训练集使用hard sample的方式进行重新采样,由于资源紧张,这一部分工作现在只完成了一小部分;
- 修改mv-dssm模型,添加attention模块。模型已经修改完成,训练集生成后可以开始训练;
- 准备转正答辩的ppt
今天计划:
-
总结实习期间工作,完成转正答辩ppt
-
减少图片数据样本量,希望今天可以提取完图片特征
-
改善dssm网络,添加query attention模块,效果有改善,当前inference结果:
acc:0.708772770853 auc:0.71492940136 precision:0.944751381215 -
运行过程:
-
20190826153707434gdl84wyi2
-
20190826155030305gw3rss69
-
20190826163104894gs34vyvj2
-
20190826163227106g9uywtyi2
-
20190826163439847g4srss69
-
20190826163935315gysj2xyi2
-
20190826164035889gkco4m39
-
20190826164938426gb09vyvj2
-
20190826171326155gu4ctuvj2
insert overwrite table hs_tmp_320 partition (part = 9) select item_id, url, index, search_offline:yuyan_udf_resnet50_fullcate_fc(search_offline:Imgto1d_Yuyan_python(search_offline:single_img_get(url))) as pic_ws from hs_tmp_318 where part = 9;
create table hs_tmp_322 as select * from hs_tmp_318 order by item_id;
网友评论