kaituoxu
Why the result I train can only reach the loss:( -14 valid and -16 train), the evaluate result is not 15.45 too
作者您好，请问一下我训练了100个epoch, 可是trainloss一直在-14，validloss一直在-16， 评估的sdr什么的也一直在12或13左右。数据集我都是用生成的min文件夹里的，20000张训练集那个。模型超参我也是用的一样的。 我尝试用你训练的最好模型再测试了一遍，结果是Average SDR improvement: 12.93， Average SISNR improvement: 12.50 我现在去很想问一下这种情况该怎么办，我不知道怎么才能提升sisnr，谢谢作者。
Hello, author. May I ask you that I have trained 100 epoch, but train loss is at -14, valid loss is at -16 for so long and never went down, and the assessed SDR has been around 12 . I used the generated min folder as your instruction said , 20,000 training wsj0 sets. I use the same model super-parameters. I tried to test again with the best model you trained and uploaded, and the result was Average SDR improvement: 12.93, Average SISNR improvement: 12.50 Now I would like to ask what to do about this situation. I don't know how to improve sisnr. The result I trained is not as what you said 15. Thank you very much.
I also have the similar results; SDRi is 13.49, SI-SNRi is 12.94. And separated sources seem to be so distorted compared to audio samples by the author.
I also got similar results. Couldn't improve above ~13.4 SDRi. Regarding the distorted samples you should set norm =True in the librosa write function.
I also got similar results. Couldn't improve above ~13.4 SDRi. Regarding the distorted samples you should set norm =True in the librosa write function.
thanks a lot.Without norm=True the separated files are a useless.
em...I'm not sure. Did you use the default config in egs/wjs0/run.sh?
yes,every config is set as your file egs/wjs0/run.sh
Training the model with the same configuration as in egs/wjs0/run.sh. The WSJ0 dataset obtain with the same manner in tools folder. Training & evaluation loss looks exactly the as the loss.png
but the SDRi/SI-SDRi is poor: SDRi is 12.49, SI-SNRi is 12.54.
Training the model with the same configuration as in egs/wjs0/run.sh. The WSJ0 dataset obtain with the same manner in tools folder. Training & evaluation loss looks exactly the as the loss.png
but the SDRi/SI-SDRi is poor: SDRi is 12.49, SI-SNRi is 12.54.
Hi @enk100, is your SI-SNRi tested on the data/tt? mine can only get SISNRi around 10, even with author's trained model. but the training loss and cv loss are the same.
yes my test set is on data/tt .SI-SNRi is around 12.00 using author's trained model ''final.pth.tar'' which is posted on his github. If I use the 100 epoch result model trained by myself, it only got SI-SNRi 12.51.
------------------ 原始邮件 ------------------ 发件人: "deciding"<[email protected]>; 发送时间: 2019年12月29日(星期天) 下午4:53 收件人: "kaituoxu/Conv-TasNet"<[email protected]>; 抄送: "490656285"<[email protected]>;"Author"<[email protected]>; 主题: Re: [kaituoxu/Conv-TasNet] Why the result I train can only reach the loss:( -14 valid and -16 train), the evaluate result is not 15.45 too (#8)
Training the model with the same configuration as in egs/wjs0/run.sh. The WSJ0 dataset obtain with the same manner in tools folder. Training & evaluation loss looks exactly the as the loss.png
but the SDRi/SI-SDRi is poor: SDRi is 12.49, SI-SNRi is 12.54.
Hi, is your SI-SNRi tested on the data/tt? mine can only get SISNRi around 10, even with author's trained model. but the training loss and cv loss are the same.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.
yes my test set is on data/tt .SI-SNRi is around 12.00 using author's trained model ''final.pth.tar'' which is posted on his github. If I use the 100 epoch result model trained by myself, it only got SI-SNRi 12.51. …
Thanks for the reply! Can I have a look at your conv-tasnet pip freeze, I think maybe the package versions cause the problem.
python 3.5 librosa 0.7.1 torch 0.4.1
------------------ 原始邮件 ------------------ 发件人: "deciding"<[email protected]>; 发送时间: 2019年12月30日(星期一) 上午10:38 收件人: "kaituoxu/Conv-TasNet"<[email protected]>; 抄送: "490656285"<[email protected]>; "Author"<[email protected]>; 主题: Re: [kaituoxu/Conv-TasNet] Why the result I train can only reach the loss:( -14 valid and -16 train), the evaluate result is not 15.45 too (#8)
yes my test set is on data/tt .SI-SNRi is around 12.00 using author's trained model ''final.pth.tar'' which is posted on his github. If I use the 100 epoch result model trained by myself, it only got SI-SNRi 12.51. … Thanks for the reply! Can I have a look at your conv-tasnet pip freeze, I think maybe the package versions cause the problem.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.