[<FedScale component: utility calculation in torch_client.py.>]

[Bossbap]

Description

Hi, I noticed a small discrepancy between the implementation and the Oort paper.

In torch_client.py, the utility sent back to the server is based on the loss from only the last mini-batch of local training. However, in the Oort paper, the statistical utility is meant to reflect the overall difficulty of a client's data — it should be based on the average of the squared losses across all samples seen during the round.

Using only the last batch underestimates the utility and can introduce noise into client selection. Just wanted to flag this in case it's unintentional.