In case of typos in configurations, the `analyze latent` reloads a model and finds the wrong dimensions. Should it be possible to always retrain? Definitely log that a pre-trained model was found.