A question about warm-up period #5
Description
Hi @Cheonglok，thanks for your interesting work !
For the trace regularization, I find the paper said, to encourage CMs to be diagonally dominant, "we initialize them with identity matrices by training the annotation network to maximise the trace for sufficient iterations as a warm-up period". While, I do not find the warm-up period in the code. Could you describe the learning way during warm-up period in more detail?
It will help me a lot. Thanks !