how to predict with the model
szq261299 opened this issue · comments
The code shows how to train and evaluate model ,but there seems is not enough information to predict with this model.
Hi. Sorry for the late response. You are right. In the current implementation the Candidate Sampling generation and the features retrieval for negative samples occurs within the train / eval loop. To be able to provide predictions (e.g. deploying this model in production), it would be necessary to move the candidate sampling outside the train/eval loop, and the model would receive the features for both positive and negative samples as input, and would output the predicted scores for each of the negative samples.
I have been drafted that implementation, and was able to deploy the CHAMELEON on TF Serving. But need to clean that code before pushing to this repo, as it required a major refactory.