Loss Graph spike when epoch starting point

[cxz1418]

Hello~ I'm learning with my custom dataset with arcface_torch code but,

Even though it is not an LR adjustment interval, the loss drops significantly and then rises at the start of the Epoch. What's the problem?

[marlowinnovations]

My theory is the batch size and the size of your dataset. If your final mini-batch is not the same size as the others, the loss will be different because it's based on a smaller mini-batch size. Then at the start of the epoch, it's back to normal mini-batch sizes and the loss starts to go back to normal again. I could be wrong though, but I've seen similar drops/spikes in my own training.