search

IntelLabs / IntelNeuromorphicDNSChallenge

Intel Neuromorphic DNS Challenge
MIT License
120 stars 27 forks source link

Baseline training performance #23

Open tao-sun opened 9 months ago

tao-sun commented 9 months ago

I ran the baseline model training but performance was not so good as reported in the paper. The SI-SNR.txt is like

Train Valid

9.808843    10.826404

10.640213 10.979951 10.750848 11.082418 10.770719 11.054980 10.746675 11.096102 10.768912 11.010266 10.769256 10.998541 10.792813 11.049504 10.786952 10.977900 10.750953 10.791486 10.745253 10.933626 10.740764 10.952813 10.736836 11.025456

Both the dataset and code used were downloaded from the repo and unmodified and data checksum passed. I also compare my args.txt with the one under original one under /baseline_solution/sdnn_delays/Trained and all the params were same. I ran training for two times and got similar results.

Anybody know the cause of this issue or encountered the same situation?

Thanks!

BujSet commented 6 months ago

+1: Not sure what the issue is, but I have similar results from running the baseline, achieves 10.68 SI-SNR for training and 11.00 SI-SNR for validation. I also use the unmodified dataset and have matched the args.

tao-sun commented 6 months ago

Setting lr=0.001 can solve the problem.

BujSet commented 6 months ago

Thank you @tao-sun! That seems to have allowed the baseline to converge to 12.44 SI-SNR on the validation set, which is close to the 12.50 that is reported on the metricsboard.