2019-08-26工作进展

上周五工作：

1. 使用品类关键词来进行dssm网络负样本的采样，在新的训练集上进行dssm网络的训练，inference效果：acc:0.699 auc:0.703 precision:0.941；
2. 对文章Gated Attentive-Autoencoder for Content-Aware Recommendation之中用到的使用attention矩阵代替attention向量的方式进行了测试，并在其基础上增加了mask，使用该结构代替原网络中的self-attention模块，将da参数设置为20的情况下，inference效果：acc:0.635 auc:0.611 precision:0.911。说明这一方式有着不错的效果，可以将其添加到mv-dssm中以尝试改善mv-dssm网络的效果。

今天计划：

1. 在当前mv-dssm网络的基础上添加attention层；

2. 重新采样mv-dssm训练数据。

3. mv-dssm & attention

pai -name tensorflow140 -Dscript="file:///home/hengsong/origin_deep_cluster_odps_8.tar.gz" -DentryFile="train_inference_v10.py" -Dcluster='{"worker":{"count":50, "cpu":200, "memory":4000}, "ps":{"count":10, "cpu":200, "memory":5000}}' -DuseSparseClusterSchema=True -DenableDynamicCluster=True -Dtables="odps://graph_embedding/tables/hs_train_data_dssm_v2_7,odps://graph_embedding/tables/hs_test_data_dssm_v2_7,odps://graph_embedding/tables/hs_tmp_267" -Doutputs="odps://graph_embedding/tables/hs_dssm_result_5" -DcheckpointDir="oss://bucket-automl/hengsong/?role_arn=acs:ram::1293303983251548:role/graph2018&host=cn-hangzhou.oss-internal.aliyun-inc.com" -DuserDefinedParameters="--learning_rate=3e-4 --batch_size=1024 --is_save_model=True --attention_type=1 --num_epochs=1 --ckpt=hs_ugc_video_2e_1.ckpt" -DuseSparseClusterSchema=True;

pai -name tensorflow140 -Dscript="file:///home/hengsong/origin_deep_cluster_odps_8.tar.gz" -DentryFile="inference_v10.py" -Dcluster='{"worker":{"count":1, "cpu":200, "memory":4000}, "ps":{"count":1, "cpu":200, "memory":5000}}' -DuseSparseClusterSchema=True -DenableDynamicCluster=True -Dtables="odps://graph_embedding/tables/hs_train_data_dssm_v2_7,odps://graph_embedding/tables/hs_test_data_dssm_v2_7,odps://graph_embedding/tables/hs_tmp_267" -Doutputs="odps://graph_embedding/tables/hs_dssm_result_5" -DcheckpointDir="oss://bucket-automl/hengsong/?role_arn=acs:ram::1293303983251548:role/graph2018&host=cn-hangzhou.oss-internal.aliyun-inc.com" -DuserDefinedParameters="--learning_rate=3e-4 --batch_size=1024 --is_save_model=True --attention_type=1 --num_epochs=1 --ckpt=hs_ugc_video_2e_1.ckpt-3" -DuseSparseClusterSchema=True;

2. mv-dssm数据构造

create table hs_tmp_315 as select item_id, concat('https://img.alicdn.com/imgextra/',image_url) as url, image_position, type from tbcdm.dim_tb_itm_image where ds=max_pt('tbcdm.dim_tb_itm_image');

create table hs_tmp_316 as select distinct a.title_id, b.item_id, a.title from hs_query_title_inference_gt_3 a join hs_dssm_dic_title_11 b on a.title == b.title;

create table hs_tmp_317 as select distinct a.* from hs_tmp_315 a join hs_tmp_303 b on a.item_id == b.item_id;

create table hs_tmp_318 as select a.* from hs_tmp_315 a join hs_tmp_303 b on a.item_id == b.item_id;

3. 测试mv-dssm网络

se_keyword_mainse_ws,title_mainse_ws, pic_mainse_ws, label

create table hs_tmp_319 as select *, se_keyword_mainse_ws as pic_mainse_ws from hs_test_data_dssm_v2_7 limit 1000000;

create table graph_embedding.hs_tmp_320 as select * from graph_embedding.hs_tmp_319 limit 10000;