search

mravanelli / SincNet

SincNet is a neural architecture for efficiently processing raw audio samples.
MIT License
1.11k stars 260 forks source link

Cannot replicate the cumulative frequency response of the SincNet filters #88

Closed tuananh0305 closed 4 years ago

tuananh0305 commented 4 years ago

Hi Marco,

Thanks for sharing your great work!

I am trying to replicate the cumulative response of the SincNet filter like the figure 3 in the paper (https://arxiv.org/pdf/1808.00158.pdf).

I load your pre-trained model and computed the Fourier transforms of all the learned filters like figure below. ask_Figure_2

Then I averaged the sum of all frequency response of the learned filters to get the cumulative response of the SincNet filter like figure below ask_Figure_1

What is your opinion about this problem?

mravanelli commented 4 years ago

Hi, I'm aware of it. In practice, after the submission we did a code refactoring and for some reason that I have to investigate better (probably due to filter amplitude normalization) the gradient is smaller than before and the filters are more static than those of the original version. I will work on that to find exactly the modification that caused that. For now, I think you can add a separate optimizer for the sincnet layer and use higher learning rate for it. This will encourage the filters to change more and eventually reach some configuration like the one shown in the original paper.

On Tue, 5 May 2020 at 17:09, Dao Anh Tuan notifications@github.com wrote:

Hi Marco,

Thanks for sharing your great work!

I am trying to replicate the cumulative response of the SincNet filter like the figure 3 in the paper (https://arxiv.org/pdf/1808.00158.pdf).

I load your pre-trained model and computed the Fourier transforms of all the learned filters like figure below. [image: ask_Figure_2] https://user-images.githubusercontent.com/35039289/81115154-8e2a8500-8f23-11ea-95f7-fcafd0132250.png

Then I averaged the sum of all frequency response of the learned filters to get the cumulative response of the SincNet filter like figure below [image: ask_Figure_1] https://user-images.githubusercontent.com/35039289/81115740-7bfd1680-8f24-11ea-9609-a14d72807d78.png

What is your opinion about this problem?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/mravanelli/SincNet/issues/88, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEA2ZVQ5MABWBELVN474KPLRQB6CPANCNFSM4MZ5GL3Q .

tuananh0305 commented 4 years ago

Hi Marco,

Thanks for your quick response!

I will try your suggestion. Thank you!

JLC7 commented 2 years ago

I am getting small changes to the input layer. Any luck on using larger learning rates?