模型輸入特徵問題請教
JohnChen1124 opened this issue · 7 comments
JohnChen1124 commented
您好:
非常感謝你們所完成的程式，希望能抽空回答一個小小的問題。模型輸入特徵取完頻譜大小之後有做絕對值，但好像沒有取平方，但在A Regression Approach to Speech Enhancement Based on Deep Neural Networks這篇paper中特徵為log-power spectral，這是有什麼原因嗎?感謝你們~~~
qiuqiangkong commented
您好！您看到的代码是论文之后复现的，我们发现绝对值比平方PESQ有0.1左右的提升。所以在代码中使用了绝对值loss。
Best wishes,
Qiuqiang
…________________________________
From: JohnChen1124 <notifications@github.com>
Sent: 23 June 2018 11:07:35
To: yongxuUSTC/sednn
Cc: Subscribed
Subject: [yongxuUSTC/sednn] 模型�入特徵��}�教 (#13)
您好:
非常感x你��所完成的程式，希望能抽空回答一��小小的��}。模型�入特徵取完lV大小之後有做^�χ担孟�]有取平方，但在A Regression Approach to Speech Enhancement Based on Deep Neural Networks@篇paper中特徵�log-power spectral，@是有什N原因�?感x你��~~~
―
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub<#13>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AMt5yT4IvzbJm4ReqatBS2CSkLF1jUR1ks5t_hNngaJpZM4U0uwB>.
JohnChen1124 commented
我了解了~所以代碼中的loss function是使用mean absolute error而非mean square error。
在擷取音檔特徵的過程中，將音檔轉換成spectrogram大小值之後有做絕對值，也有取log，但好像沒有取平方，是因為loss function改變而做的調整嗎?感謝您~
qiuqiangkong commented
Loss 是一个值得深入研究的方向。在代码中我们用了log Mel spectrogram，没有取平方。Loss可以选取Lp范数。当p=1时就是abs，当p=2时就是平方。可以尝试调。
…________________________________
From: JohnChen1124 <notifications@github.com>
Sent: 23 June 2018 14:56:12
To: yongxuUSTC/sednn
Cc: Kong Q Mr (PG/R - Elec Electronic Eng); Comment
Subject: Re: [yongxuUSTC/sednn] 模型�入特徵��}�教 (#13)
我了解了~所以代a中的loss function是使用mean absolute error而非mean square error。
在�X取音�n特徵的^程中，�⒁�nD�Q成spectrogram大小值之後有做^�χ担灿腥log，但好像�]有取平方，是因�loss function改�而做的{整�?
―
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#13 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AMt5yWlmzUM4kbl0KYdZLr8AArDhb7fNks5t_kj8gaJpZM4U0uwB>.
Nickkk1124 commented
真的很感謝您的回覆，您說的log Mel spectrogram與Mel Filter Bank有關係嗎？log Mel spectrogram沒有取平方也是因為PESQ的測量結果嗎？
qiuqiangkong commented
不好意思说错了，我们用的是log spectrogram，不需要Mel。上一封邮件应为：
"
Loss 是一个值得深入研究的方向。在代码中我们用了log spectrogram，没有取平方。Loss可以选取Lp范数。当p=1时就是abs，当p=2时就是平方。可以尝试调。
"
…________________________________
From: Nickkk1124 <notifications@github.com>
Sent: 24 June 2018 07:18:16
To: yongxuUSTC/sednn
Cc: Kong Q Mr (PG/R - Elec Electronic Eng); Comment
Subject: Re: [yongxuUSTC/sednn] 模型�入特徵��}�教 (#13)
真的很感x您的回覆，您f的log Mel spectrogramcMel Filter Bank有P�S�幔
―
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#13 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AMt5ydGbRXaNL3LWueV4bCM2WVEhxpqxks5t_y8ogaJpZM4U0uwB>.
Nickkk1124 commented
log Mel spectrogram沒有取平方也是因為PESQ的測量結果嗎？還是因為loss改變而跟著改變呢？感謝您～
qiuqiangkong commented
是的，是根据???果?用的。
Get Outlook for Android<https://aka.ms/ghei36>
…________________________________
From: Nickkk1124 <notifications@github.com>
Sent: Sunday, June 24, 2018 5:35:27 PM
To: yongxuUSTC/sednn
Cc: Kong Q Mr (PG/R - Elec Electronic Eng); Comment
Subject: Re: [yongxuUSTC/sednn] 模型輸入特徵問題請教 (#13)
log Mel spectrogram沒有取平方也是因為PESQ的測量結果嗎？還是因為loss改變而跟著改變呢？感謝您～
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub<#13 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AMt5yRHjKHn0MvXb0MzBJ9a_hrY6z-Izks5t_7_PgaJpZM4U0uwB>.