wangzhe

个人理解： 1.不做增量训练，预测时通过新的item的side embedding直接得出新item的embedding。 2.根据用户行为生成包含新item的session，对于新增的session送入训练好的模型做增量训练得到新item的embedding。

First you need to define which items belong to the cold items，then get the side embedding through training other items，the result of the cold item should be average sum of...

自己动手，欢迎提mr

> 目前在SDM模型代码60行打印不支持dense feature，raise ValueError("Now SDM don't support dense feature")，能否加上？ 建议先尝试离散化

refer to examples/run_sdm.py

I know it is resnet101,but I think it should be faster than 2fps, because when I run faster rcnn which is writed with pytorch in the same machine(GTX 1080), the...

the repo is https://github.com/jwyang/faster-rcnn.pytorch the backbone is resnet101, proposal number is 300 when test, test size is 600, maxsize is 1200