how to export the model?

[jfkirk]

how to export the model?

[jfkirk]

Hey @Composmentis ! We are still building convenience methods for saving and loading models. We'll be tracking that work on #8

I expect that to be merged within the week.

Closing as duplicate

[borisrev]

Hi James,

I'm still unclear on @Composmentis 's initial question, i.e. how does saving/loading models enable feeding a previously unseen user to an already trained model with tensorrec?

Thanks!

[jfkirk]

Hey @borisrev -- A user is simply a collection of feature values, so a TensorRec does not know or care if it has seen the user in training.

I interpreted the initial question to mean "I want to train a TensorRec model and save it. In the future, I will want to load it and get predictions for new users." The missing link, at the time of the initial question, is that TensorRec did not have methods for saving and loading the models.

[borisrev]

Let's say you're using an identity matrix for the user matrix when training as in the MovieLens example. After training, if you want to pass a new user to predict() along with same item_features used in training, what dimensions would then be appropriate for the new user_features matrix? A row of all 0s with the same number of columns as the user matrix used during training?

Thanks for the quick reply!

[jfkirk]

Ahh sorry I was unclear -- A TensorRec model does not know or care if it has seen the user in training unless the system is built with "Indicator Features". In the scenario you've outlined, the identity matrix for users is a set of indicator features, so all users will need to be present at fit() time.

The number of user/item features must be the same both at fit() and predict() time.

The counter-example would be the Book Crossing dataset with user_indicator and item_indicator set to False. In this case, the features are only metadata and the TensorRec model would not care if the user/item was present during training.