How minibatch accuracy can go beyond 100% while training convolutional neural network?

I am training CNN on expression data and I am getting sharp spikes in accuracy that goes beyond 100%, and loss goes less than 0. But when curve smoothed both accuracy and loss are in the limits (i.e. accuracy less than 100% and loss is greater than 0). How it is possible?

Total number of samples: 20500

Epoch: 5

Minibatch size: 25.

What am I doing wrong?

[KSSV]

I don't think it is an issue. You check your model with test data and then decide.

[Dedi Suharman]

this already solved ?

[Ankita Jain]

No...