Generally in supervised machine learning algorithms, the model performance is measured splitting train, test, validation set.

But in case of unsupervised method , like autoencoder, is it necessary to have such split ? Can't the training dataset itself be a test set to evaluate its performance ?