NaN from trans_encoder in concept branch during MELD-DA reproduction

[ffmmqq]

During reproduction on MELD-DA, I traced the NaN source to trans_encoder
in the concept branch (HIER.py). Bypassing it made the pipeline stable.

Could you clarify:

1. Was trans_encoder numerically stable in your original setup, and was
   there any specific initialization, dtype handling, or stabilization applied
   that is not reflected in the released code?

2. If trans_encoder is bypassed, how much would you expect it to affect
   the final performance?

[guyunjin]

1. In our original setup, trans_encoder was numerically stable. We did not apply any special initialization, dtype handling, or extra stabilization beyond what is already in the released code. The NaN issue you observed is most likely caused by using a newer transformers version, so we recommend transformers==4.46.1.
2. If trans_encoder is bypassed, the impact on final performance should be relatively small. We would expect only a slight degradation, not a major drop.

[ffmmqq]

Thank you for the detailed response. Just to confirm, I am already using transformers==4.46.1 with trans_encoder bypassed, so the setup should be aligned with your recommendation.

After bypassing, my reproduced results on MELD-DA are as follows:

• Accuracy: 58.76%
• F1: 43.46%
• Recall: 41.52%

I may well be missing something on my end. Would you be willing to share your pretrained weights? It would be really helpful for identifying where the discrepancy might be coming from. No pressure at all if that's not feasible, and any additional hints would also be greatly appreciated.