search

qiuqiangkong / sampleRNN_acoustic_scene_generation

Other
13 stars 3 forks source link

Training cost remains high #9

Closed guozixunnicolas closed 4 years ago

guozixunnicolas commented 4 years ago

Hi Qiuqiang,

Thank you for your brilliant work! Im currently trying to re-implement the whole thing in tensorflow.

However my training loss remains high( around 2 ). I've noticed that the training/validation/test cost is around 2.7 in your implementation. I wonder can the model produce useful result with such high loss value?

If it's convenient for you, would you mind sharing with me a screenshot of training loss curve?

Many thanks!

Best regards,

ZX

qiuqiangkong commented 4 years ago

Hi ZX,

We did not keep the training curve in our disk. For acoustic scene generation, we observed higher loss (around 2) than generating music and speech (around 1.2). I guess it is because of the randomness in acoustic scenes. Though the loss is high, the system can still generate acoustic scenes.

Best wishes,

Qiuqiang

From: zguo008 notifications@github.com Sent: 10 December 2019 15:16 To: qiuqiangkong/sampleRNN_acoustic_scene_generation sampleRNN_acoustic_scene_generation@noreply.github.com Cc: Subscribed subscribed@noreply.github.com Subject: [qiuqiangkong/sampleRNN_acoustic_scene_generation] Training cost remains high (#9)

Hi Qiuqiang,

Thank you for your brilliant work! Im currently trying to re-implement the whole thing in tensorflow.

However my training loss remains high( around 2 ). I've noticed that the training/validation/test cost is around 2.7 in your implementation. I wonder can the model produce useful result with such high loss value?

If it's convenient for you, would you mind sharing with me a screenshot of training loss curve?

Many thanks!

Best regards,

ZX

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://eur02.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fqiuqiangkong%2FsampleRNN_acoustic_scene_generation%2Fissues%2F9%3Femail_source%3Dnotifications%26email_token%3DADFXTSI6HYPB75XUPUGSJR3QX6XFRA5CNFSM4JZAER4KYY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4H7PTIPA&data=02%7C01%7Cq.kong%40surrey.ac.uk%7C580b241990de48e6e72f08d77d83f594%7C6b902693107440aa9e21d89446a2ebb5%7C0%7C0%7C637115878026542675&sdata=FYHZIBWco8UnYHaQkQb19GikdWiulP3LRQYX5X8XNtU%3D&reserved=0, or unsubscribehttps://eur02.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FADFXTSMQHDICQJBNPOMGSJ3QX6XFRANCNFSM4JZAER4A&data=02%7C01%7Cq.kong%40surrey.ac.uk%7C580b241990de48e6e72f08d77d83f594%7C6b902693107440aa9e21d89446a2ebb5%7C0%7C0%7C637115878026542675&sdata=f%2F%2F7wfyPfGVDj3m2vhFIP%2BkcbLAD9G48fSbVcSDUhy0%3D&reserved=0.

guozixunnicolas commented 4 years ago

Hi Qiuqiang,

Thank you for your answer! That helps:)

Have a nice day!

Best regards,

ZX

guozixunnicolas commented 4 years ago

@qiuqiangkong Hi qiuqiang, just a short follow-up question, may I know how long does it take for the model to generate 1 second of audio? Mine needs around 6hour(with a GPU) to generate 1 seconds of audio which seems to be wrong. Thanks in advance:)