I tried cutting out two layers of the codec and found that its score on the dataset plummeted dramatically.