Can the "input data normalization" of "trainNetwork" be done separatedly?

Question

Jaime Almonacid-Caballer 2020-9-11

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/592177-can-the-input-data-normalization-of-trainnetwork-be-done-separatedly

评论： Jaime Almonacid-Caballer 2020-9-16

Hi,

I am begining with Convolutional Nural Networks in Matlab following the available examples.

I have prepared the input data and parameters. When I have run the training of the networks (trainNetworks) it has began with the 'input data normalization' (during more than 2 hours). Once it has been done, when it should have began the iterations, Matlab has failed (Gpu out of memory).

Would any way to have the normalization done before the training so that I could get the errors (logical errors while I am learning) without expending the previous 2 hours each time?

Thanks a lot,

Jaime

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Madhav Thakker 2020-9-14

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/592177-can-the-input-data-normalization-of-trainnetwork-be-done-separatedly#answer_494242

在 MATLAB Online 中打开

Hi Jaime,

I understand that you want to stop the inbuilt data normalization. You can do so by creating your own input data layer and setting normalization to none. I was able to disable normalization in https://www.mathworks.com/help/deeplearning/ug/create-simple-deep-learning-network-for-classification.html by calling

imageInputLayer([28 28 1], 'Normalization', 'none')

instead of

imageInputLayer([28 28 1])

when defining the network.

Hope this helps.

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Jaime Almonacid-Caballer 2020-9-16

在 MATLAB Online 中打开

Thanks a lot!!,

I have found the solution. I didn't want to preprocess all my images and save them in disk because of their volume. I expected to have a function pre-trainNetwork so that the image would be normalized. I have found the solution in an option of imageInputLayer:

imageInputLayer([128 128 4],'Normalization',@funcN)

Where funcN: