search

yluo42 / TAC

transform-average-concatenate (TAC) method for end-to-end microphone permutation and number invariant ad-hoc beamforming.
247 stars 54 forks source link

How to train the model? #4

Open framsc opened 4 years ago

framsc commented 4 years ago

Could you please share the training script? I can not reproduce the results after reading FaSNet, https://doi.org/10.1109/ASRU46091.2019.9003849 TAC, https://doi.org/10.1109/ICASSP40776.2020.9054177

Thanks in advance.

yluo42 commented 3 years ago

Sorry for my late reply as I missed this issue. I will probably prepare one script for training by early Nov.

On the other hand, some recent literatures have used the FaSNet+TAC model as baselines (e.g. https://arxiv.org/abs/2010.01703 and https://arxiv.org/abs/2010.06007), so the models should be able to be reproduced.

framsc commented 3 years ago

Thanks for the reply. I have read these papers too. It seems that the researchers are quite interested in FaSNet and TAC. Maybe you can provide a pre-trained model thus we have the same model to talk about. So are the best hyper-parameters, such as batch size, learning rate.

Nice job. Hope to hear from you soon.

ghost commented 1 month ago

How to train the fasnet model after getting the dataset generated.