I believe the default behavior of the ImageClassifierData.from_csv method is to use a random 20% of the dataset for validation. Now since most whales in this dataset have only one image, you almost certainly delegate one of those whales entirely into the validation set. This means there will be classes in the validation set which were not present in the training set. That is certainly unpleasant, and according to some people on this thread:
Kaggle's Whale Competition,
it is the likely cause of the problem.
2 Likes