For example, when transfer to inpainting task, shall I retrain autoencoder with mask images as input?