Saved model backward compatibility

[zeeshan]

Hello! I trained a model and saved it using the latest TF 2.15. I am trying to figure out if there is any way for me to use this saved model to run inference on a RaspberryPi (32 bit ARM architecture)? I tried loading the same model on the RPi using tf.keras.models.load_model but I get errors of sort TypeError keyword not understood: keepdims.

I wonder if there is any format for saving TF models which are backward compatible for all (or most) tensorflow versions? Or is there any simpler way to do this- running inference of a model saved using the latest tensorflow, on an old (v2.5) setup?

[Kiran_Sai_Ramineni]

@zeeshan

here is any format for saving TF models which are backward compatible

Hi @zeeshan, The backward compatibility is used to update model parameters during training based upon the difference between model predictions and actual ground truth values. Once the model is saved, the saved file contains the model architecture and weights that are updated during training, if you want to retrain the saved model then the backward compatibility will be available during the retraining phase. Thank You.