Resuming checpoint leads to a higher accuracy than keep it running

Hello experienced ones,

When the training is progressing, I see the test accuracy does not change a lot and is almost steady. However, when I save the checkpoint and resume the iteration from the checkpoint the training show a jump in the test accuracy with a significant margin.

I wonder if you guys know what is going on and possibly provide an explanation for such a random behavior.

It’s hard to understand the situation in detail without any code; are you referring to test accuracy that is run on a separate test dataset?

It might be useful to verify that the test accuracy is consistent when saving a checkpoint vs. when loading the same checkpoint (without any additional training).

Yes, it is accuracy on the test set.

The test accuracy at the event of saving is equal to the one at the event of resuming.

Are the other hyperparameters the same upon resuming? It might be easier to understand if there were some concrete code/outputs shown.

The hyperparameters are the same. Preparing a code snippet due to semi-supervised nature of the training which is performed.