search

GitHub-HongweiZhang / prediction-flow

Deep-Learning based CTR models implemented by PyTorch
MIT License
250 stars 52 forks source link

Incremental training #29

Closed YahyaL closed 4 years ago

YahyaL commented 4 years ago

I am interested to use the DIEN model. What are the strategies to do incremental training to fit new user interactions and new items without having to retrain the whole model from scratch ?

I read the paper of the model and they only discuss how to reduce latency for model serving and not how to do incremental training.

GitHub-HongweiZhang commented 4 years ago

@YahyaL

incremental training.

I think the easiest way is to save and load the model many times. when new data comes, load trained model, re-train using new data and save it.