Too difficult to reproduce training results with tuner

[TonyCongqianWang]

It would be nice, if one could conveniently save the hyperparameters used to train a model. In other words, if there was a function to create custom hyperparameter templates so to speak

So right now I am using either model.describe() or the hyperparameter tuning logs to extract the best hyperparameters and merge them together with the fixed hyperparameters. It would be nice, if there was a function to just extract a dictionary with all of the parameters the same way they are saved for the hyperparameter templates.

[achoum]

You are right to note that the hyper-parameters used to train a model are not saved in the model (YDF separate clearly the learner and the model concepts).

The hyper-parameters are available in the learner's hyperparameters method.

In addition, if the model was trained with a tuner, the tested hyper-parameters (and their respective scores) are available in a UI (model.describe) and programmatically (model.hyperparameter_optimizer_logs).

For example, here is the dictionary of hyper-parameters used to train a model with a tuner:

logs = model.hyperparameter_optimizer_logs()
top_score = max(t.score for t in logs.trials)
selected_trial = [t for t in logs.trials if t.score == top_score][0] 
print(selected_trial.params)  # This is a dictionary

[TonyCongqianWang]

Thanks for the reply and help!

logs = model.hyperparameter_optimizer_logs()
top_score = max(t.score for t in logs.trials)
selected_trial = [t for t in logs.trials if t.score == top_score][0] 
print(selected_trial.params)  # This is a dictionary

this is exactly what I have been looking for. It would be really nice, if it was possible to get this dictionary easier! I also got the parameters from the describe UI but that was slow. I didn't figure out a fast way to get the dict like you posted. One of the reasons is that I had no idea logs was a list of objects! I thought it was simply a string! (I guess I could have checked it, but I didnt). It would be nice if that was a hint in the documentation.