Question About Initial Learning Rate Adjustment

[Bigsealion]

Description:

During training, I noticed that the learning rate is not consistent with the initial learning rate I set. After inspecting the code, I found the following adjustment being made to the initial learning rate:

self.lr = args.lr * args.batch_size / 256

Question:

What is the purpose of this adjustment by args.batch_size / 256? Is this done to accommodate a specific training strategy? I would appreciate an explanation of the reasoning behind this adjustment.

Thank you!