search

Jungwoo4021 / OS-KDFT

10 stars 0 forks source link

Pretrained Model Loading #3

Open gancx opened 2 months ago

gancx commented 2 months ago

Thanks for your work. In the paper, you mentioned that the teacher model is wavLM + ECAPA-TDNN. However, in your implementation, I only found you loaded the weights of wavLM (line 13 in train/experiment/model.py). Could you explain this further?

chan-yeong0519 commented 2 months ago

Thank you for your interest in our study, and I apologize for the delayed response. I would like to clarify a point of possible misunderstanding.

In our study, we used WavLM + ECAPA-TDNN as the baseline model, and WavLM-Base+ as the teacher model. Therefore, we only loaded the weights of WavLM and did not utilize the weights of ECAPA-TDNN separately for the teacher model.

Furthermore, regarding the baseline, previous studies have used ECAPA-TDNN without pre-training, and we have constructed our baseline in the same manner.

gancx commented 2 months ago

Thank you for your interest in our study, and I apologize for the delayed response. I would like to clarify a point of possible misunderstanding.

In our study, we used WavLM + ECAPA-TDNN as the baseline model, and WavLM-Base+ as the teacher model. Therefore, we only loaded the weights of WavLM and did not utilize the weights of ECAPA-TDNN separately for the teacher model.

Furthermore, regarding the baseline, previous studies have used ECAPA-TDNN without pre-training, and we have constructed our baseline in the same manner.

I see. Thanks for your clarification. I wonder if you have investigated the results evaluated on voxceleb1-E and H using Voxceleb2 dev for training with augmentations. Because I have not found any related results in the paper. Thank you.

Jungwoo4021 commented 2 months ago

Hello @gancx, We have not yet conducted experiments on evaluating the results on voxceleb1-E and H using Voxceleb2 dev for training with augmentations. However, we are currently working on an improved version of OS-KDFT, and we plan to carry out these experiments as part of this work. We aim to submit a related paper on our findings as soon as possible.

gancx commented 1 month ago

Hello @gancx, We have not yet conducted experiments on evaluating the results on voxceleb1-E and H using Voxceleb2 dev for training with augmentations. However, we are currently working on an improved version of OS-KDFT, and we plan to carry out these experiments as part of this work. We aim to submit a related paper on our findings as soon as possible.

Sure. Looking forward to your future work. I wonder if you can share your pre-trained model trained on voxceleb2-dev. I would like to evaluate your pretrained model on voxceleb1-E and H for comparison.