Conversation
|
It gives 78.5% final accuracy on valid set (w/o data augmentation) which is 4% more than similar Keras setup. |
|
Can you rebase this against kaldi_52? Github is showing some changes that I think I already merged. |
|
Rebased it |
|
I merged this, but I was surprised that you didn't combine dropout with batchnorm, e.g. conv-relu-batchnorm-dropout-layer. |
|
I will do that. First I wanted to try the similar setup as Keras (which does not use batchnorm). |
|
I'm trying myself adding batchnorm to all 4 of your 1c layers.
…On Fri, Apr 28, 2017 at 12:16 AM, Hossein Hadian ***@***.***> wrote:
I will do that. First I wanted to try the similar setup as Keras (which
does not use batchnorm).
—
You are receiving this because you modified the open/close state.
Reply to this email directly, view it on GitHub
<#1589 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/ADJVuxoWiches2AwbERWkJQ9pv800XCRks5r0Wg2gaJpZM4NK-pr>
.
|
|
OK. |
|
BTW, I also tried 1c setup w/o dropout, and the final accuracy was 68% and there was significantly more overfitting. |
|
Adding batch-norm to 1c seems to help, I am nearly finished training (as
1d), and test accuracy is 0.82 (vs 0.78 in 1c)
…On Fri, Apr 28, 2017 at 12:53 AM, Hossein Hadian ***@***.***> wrote:
BTW, I also tried 1c setup w/o dropout, and the final accuracy was 68% and
there was significantly more overfitting.
—
You are receiving this because you modified the open/close state.
Reply to this email directly, view it on GitHub
<#1589 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/ADJVu3eVkywFCTBM_C2ul-UjaPFiOE9Yks5r0XC3gaJpZM4NK-pr>
.
|
|
That's interesting. It is also interesting that 1c was very fast to train too (even though it has similar setup to Keras with a rather big fully connected layer 8192x512). I wonder why Keras takes so long to train. |
2 participants
No description provided.