Use a pretrained model for training another model

[haertelkar]

Hi everyone,

I want to build a two part architecture. To put it simple the environment is a 2d grid with some special positions. At every position lies some information about the coordinates around it. The goal is to do as few steps as possible while hitting all positions. I don't want to use reinforcement learning for this one for other reasons.

I already trained on network to get information from 9 (for now random) positions and predict where all special positions are. Now I want to train a second network that iteratively chooses the best positions for the already trained network starting from the center. The already trained network should not be trained any further.

Do I just load the trained model from a checkpoint and use it with model.eval() in the forward pass of the other model?

Thanks!

[haertelkar]

I just realized what I want to do doesnt make much sense from a practical point of view, sorry for bothering you :)