Hey, let’s say you have a small dataset with 200 data points. When testing the model, I get 100% accuracy, which seems false. However, considering I test in very small batches with batch size = 8-10, it makes sense to me that occasionally every prediction matches the label. So should you even test in such small batch sizes?
The batch size during testing shouldn’t matter, as you would have to use a separate test dataset and only test the final model once on it (the validation dataset is usually used once per training epoch).
In case you are using the training data to “test” the model, note that you are leaking data and the “test” accuracy is indeed the training accuracy of a single batch.