If the network contains batch normalization layers, the final validation metrics are often different from the validation metrics evaluated during training. This is because the network undergoes a 'finalization' step after the last iteration to compute the batch normalization layer statistics on the entire training data, while during training the batch normalization statistics are computed from the mini-batches.
If in addition to batch normalization layers the network contains dropout layers, the interaction between these two layers can aggravate this issue, as described here: https://arxiv.org/abs/1801.05134
If one removes the batch normalization (and dropout) layers from the network, the 'final' accuracy should be the same as the last iteration accuracy.
Increasing the size of the mini-batches can also alleviate this issue, since the statistics from a larger mini-batch may be better estimates of the entire training data statistics.
