ssui-liu / learnable-embed-sizes-for-recsys Goto Github PK

Hi, this paper works great for embedding compressed. However, if there are many rows of all 0 in the pruned matrix V, does it mean that all difficult samples are discarded and there is a possibility of fitting?

Hello, thank you for sharing the codes. I have one question.
In init function of class engine, you use setup_factorizer(opt) function where there will be a initialization of self.model.
Then in train_an_episode function, you call self._factorizer.init_episode() function to init self.model.
Why you init self.model twice? Is that a duplication?

Hi Siyi,

I just want to say that your idea of using soft threshold for embedding pruning is a simple yet effective strategy. I'm currently developing an alternative soft threshold pruning technique for my honors thesis and I would also like to use some of the baseline models for testing. But I don't seem to be able to find the implementation of MGQE and DartsEmb online so just wonder if it is possible to get your baseline testing scripts as well? Thanks.