Unable to Recreate Performance Achieved in Training Mode in Evaluation Mode

[DimitriosChristopoulos]

I'm currently training mobileone on a dataset, but unfortunately the model is performing significantly worse in .eval() mode than it does on .train() mode. The model's performance fluctuates greatly from epoch to epoch on the validation set (when it is in .eval() mode) and I've determined that it is because the model is only predicting one class. Is this or the cause of this issue known?