We are looking into using mlflow to handle our keras models, and we would also like to use mlflow to deploy the models in e.g azure. But the models require some simple preprocessing of the data, in our case the use of a minmax scaler. For the deployed models to answer correctly they must apply the scaler on the input (and inverse on the output). I have not found any way to include the scaling in the persisted/deployed models. Have I overlooked something, or is it not possible?
Asked
Active
Viewed 627 times
1 Answers
-1
Mlflow's mlflow.keras and mlflow logging APIs persists your model and reload it as desired, which you presumably know and are doing it right.
If you reload the model saved in h5 format, and supply your new input processed the minmax scalar, it should work.
On another note, we have a mlflow-user group on google. Please post this there too. You can find the URL here:https://mlflow.org/#community
Cheers Jules
Jules Damji
- 179
- 2
-
The question is about saving model and parameters as one 'object' as they belong together. The answer assumes this is the case. I have the same question on how to keep the connection between model and parameters – donnut Oct 09 '20 at 08:55