why is my final validation accuracy much lower than the validation accuracy during training?

my FINAL validation is: 21%, while most of my validation accuracy measurements is ~70. what could be the reason for that?

Answer 1

[Raunak Gupta]

Hi,

This can happen due to presence of batchNormalizationlayer in the Layer graph. Since in batch normalization layers the mean and variance of data is calculated for whole training data at the end of the training it can produce different result than that seen in training phase (because there these statistics are calculated for mini batches only). Also, I see there is no increase in the loss in the last iteration which suggest that the model hasn’t changed much (So, this decrease looks only due to the batchNormalizationlayer). This shows that the training data is not shuffled while training and, I think you can train the model for more epochs.

You may set the Shuffle option to ‘every-epoch’ for better distribution of data while training.