Problems with the semi phase of training with vit-huge

[yyyouy]

Hi, when I use vit-huge to train the semi stage, after 12 epochs, loss_u became 0, and the classification accuracy of ema model suddenly became low at 11 epochs, and it is expected that the training will definitely fail. I would like to ask the authors whether you have encountered this problem

[zhaoweicai]

have you successfully reproduced the base/large model results?

[yyyouy]

have you successfully reproduced the base/large model results?
I haven't tried it yet, and I'm wondering if it's possible for the huge model to have a sudden deterioration of the ema model midway through the process. Thank you for your reply

[yyyouy]

I found out that this problem occurs if you use full precision, but not if you change to mixed precision

[yyyouy]

have you successfully reproduced the base/large model results?
I haven't tried it yet, and I'm wondering if it's possible for the huge model to have a sudden deterioration of the ema model midway through the process. Thank you for your reply