fix bugs

tensor2tensor/models/transformer.py

@@ -858,6 +858,7 @@ def transformer_before_shared25():
     hparams.learning_rate_warmup_steps = 8000
     hparams.optimizer = "MultistepAdam"
     hparams.optimizer_multistep_accumulate_steps = 4
+    hparams.encoder_layers = 25
     # it's likely to oom when you train deep transformer-pre-norm within 4096 batch_size
     hparams.batch_size = 2048
     return hparams