yulewang97
commented
1 month ago Thank you for your thoughtful reminder! The code was released when the paper was first submitted to arXiv, and at that time, there was no established benchmark for the distribution alignment task, and we had not encountered the NaN value error. We will conduct a detailed review of the code, repeat the experiments, and re-upload the updated code afterward. We welcome any further feedback you may have and appreciate your recognition of our work.
Thank you for this exciting work. I am wondering if I could get your help in reproducing your experimental results. I ensured that the code, data, and hyperparameters for the experimental setup were exactly the same as those in your paper, but I faced an error during the alignment process. The gradients turned into NaN values after the first backpropagation step. This issue did not occur when I used your pretrained VAE model for training the MLA model. Additionally, I noticed that when I changed the random seed for VAE training (the original code set it to 21, and I changed it to 666), the NaN problem was resolved, but the alignment results were not ideal. Specifically, the R² value dropped to -25.58%, and the RMSE increased to 10.00. Given this, could you please provide the specific settings used for training the VAE? I am also curious if the model might be sensitive to certain parameters and how to adjust them for more stable training results. I look forward to your response and thank you once again for your contributions to this research.