i see, thanks for the explanation, so basically it is initializing the weights to the weights which are saved instead of to the random ones, if i understand correctly? So in this case to prevent this behavior and start from random weights I should just change the name of output file with weights, right?

hm, but why it happens if my new model has different hyperparameters and is separate experiment with separate name? how to avoid this behavior?

@<1566596960691949568:profile|UpsetWalrus59> the hyperparameters will be the ones associated with the experiment. The input model you specified has a path to the actual weights file - this message basically means the same weights file is already registered in a model object and thus a new model object will not be created, and the existing one will be used as an input model

yes, i can see the hyperparams are reflecting correctly, but I mean it shouldn't start from previous model checkpoint if it is different run with different hyperparams

Hi @<1566596960691949568:profile|UpsetWalrus59> , I think this basically means you have an existing model and it's using it as the starting point.

Hyperparameters are connected to the experiment so your config will be right 🙂

hyperparams are only for the task, not the model