In my 3fold cross validation, the training loss decreases but the validation loss increases during the first cycle. But during the 3rd cycle, (last one in 3fold) the training and validation loss and accuracy graphs match (same stable line). So does it mean the final validation set has already been seen by the first 2 trainings? If so, is the model insufficient to face real world data?
与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…