https://kedro.org/ logo
#questions
Title
# questions
f

Francis Duval

01/12/2024, 2:50 PM
Hello all! Is there any documentation about tuning neural networks (or other machine learning algorithms) hyperparameters using Kedro? I don't really know where to start. Sure, I want to create a new pipeline which will be given a training and validation set. I was wondering about the best way(s) to proceed. This is what I want: • I would like the pipeline to output a text file or table containing the validation loss for all hyperaparameters combinations. • Ideally, I would like to track the experiments (so if I change my hyperparameters grid, I want to still have access to former results). • Could MLFlow help me with this? I know that Kedro has its own experiment tracking system. • Could the Kedro parameters help me achieve my goal? • It would be nice to visualize the results in Kedro viz. Thanks!
n

Nok Lam Chan

01/12/2024, 2:56 PM
Answer to some of these questions. For tracking, you can use Kedro's own experiment tracking, MLflow would also work. Checkout the kedro-mlflow plugin which is maintained by our community.
👍 1
for outputing text file, I guess it's quite simiilar to what you want from experiment tracking (or any similar tool)? To do this, you can use hook
👍 1
What exactly do you want to visaulise? Kedro-viz has some capacity to show plots/table. If you use experiment tracking then you get a few more advanced charts and ability to compare between runs.
For hyperparameter tuning, maybe @marrrcin will have some good idea?
f

Francis Duval

01/12/2024, 3:00 PM
Mostly tables showing the loss for each hyperparameter combination.
Looks like kedro-mlflow is what I need! I could log the tuning results as an artifact. An since MLFlow tracks the runs, I'll be able to see the results of previous run. Thanks!
😁 1
n

Nok Lam Chan

01/12/2024, 3:44 PM
that's great:) there is a #plugins-integrations channel and plenty of kedro-mlflow users and the authors of the plugin in this Slack group.
😮 1
m

marrrcin

01/12/2024, 7:39 PM
For hypermarams search I don't think that there's a golden/community approved standard yet. I'd love to hear some ideas or approaches people use/try to use
m

Marc Gris

02/09/2024, 5:54 PM
Hi everyone, It’s been a while ! Regarding Hyperparameter Search: I’m looking into using Optuna… Has any of you guys used it in a kedro pipeline + kedro-mlflow ? 🙂 Have a nice week-end, Cheers M
2 Views