Currently, our implementation of validation/test set split is not ideal. During our training we use only 10 batches (validation set) for evaluation of our model performance. At the end, we should ideally take only the rest of them to test the actual performance on not seen data. Additionally, we should select these sets kind of shuffled based on our different datasets (when I would have multiple different model subsets).
Currently, our implementation of validation/test set split is not ideal. During our training we use only 10 batches (validation set) for evaluation of our model performance. At the end, we should ideally take only the rest of them to test the actual performance on not seen data. Additionally, we should select these sets kind of shuffled based on our different datasets (when I would have multiple different model subsets).