Performance gap on raf_furnished dataset with 0.32s RIR – is additional configuration required?

[363054777]

Hi, thanks for your great work!

I’ve been trying to reproduce your results on the raf_furnished dataset, specifically for the 0.32s RIR setting. I followed the provided setup and only changed signal_output_dim to 5120 as required for this configuration.

However, I noticed that the results I obtained are significantly worse than those reported in the paper. My result on the 0.32s RIR did not outperform the 0.1s one, so I would like to investigate the reason.

Whether any additional changes (e.g., configuration file, code) are required to reproduce the reported results for 0.32s RIR on raf_furnished?

[363054777]

I see, in the paper, the T60, C50, and EDT values for the 0.32s RIR on the raf_furnished dataset are reported as 4.95, 0.55, and 19.8 respectively.

However, in my reproduction (only changing signal_output_dim to 5120), I obtained 7.67, 0.84, and 24.1.

[ZitongLan]

Hi thank you for reaching out.

By default I set n_samples==32, n_azi==36 and n_ele==18 in the config file for common gpu configurations. And it is normal in this case that you get worse results. Depending on your GPU memory, I would recommend increasing these parameters and using multi gpu to train this model. For example, make n_samples==64, n_azi==64, n_ele==32