Typically the model is pre-trained and you're just training the last layer, or adding a new layer, which is much much less computationally expensive.