search

mxbastidasr / DAWN_WACV2020

It contains the code for the paper Deep Adaptive Wavelet Network
64 stars 15 forks source link

DAWN (no init.) #8

Open ZPP2000 opened 1 year ago

ZPP2000 commented 1 year ago

Im this paper, I notice DAWN (no init.), DAWN (no init.) mean what?

beltegeuse commented 1 year ago

It was an early experiment of pre-initializing the network with some specific weights to produce regular haar wavelet coefficients. This strategy could have helped to improve the network's accuracy.

The default DAWN arguments are used inside the paper (+/- some hyperparameters explained in the results section).

ZPP2000 commented 1 year ago

but there is no Haar wavelet involved in the dawn network, only the lifting scheme to obtain high and low frequency information, DAWN (no init.) and DAWN (16 init.), 16 is the args.first_conv=16？

ZPP2000 commented 1 year ago

Thank you for your answer: I am from China. Some statements may not be accurate. Please forgive me I also have the following questions: First, The training set is one of the four samples in each category as the training set (a total of 1188 pictures), and the rest as the test set (a total of 3564 pictures). The training accuracy is 100%. Why, is my data set wrong? Second, in the visualization process, does the visualization part need major changes? I just adjust the parameter monkey to True Third, the number of multi-scale layers is not set to one layer, why should it be set according to the size of the picture? Is one layer of information extraction inadequate? Fourth, can this code implement classification on the MNist dataset and achieve good classification accuracy Fifth, is this extracted feature strictly high and low frequency information

I look forward to your reply

曾盼盼  @.***

 

------------------ 原始邮件 ------------------ 发件人: "mxbastidasr/DAWN_WACV2020" @.>; 发送时间: 2023年7月7日(星期五) 晚上10:16 @.>; 抄送: "曾盼盼 @.**@.>; 主题: Re: [mxbastidasr/DAWN_WACV2020] DAWN (no init.) (Issue #8)

It was an early experiment of pre-initializing the network with some specific weights to produce regular haar wavelet coefficients. This strategy could have helped to improve the network's accuracy.

The default DAWN arguments are used inside the paper (+/- some hyperparameters explained in the results section).

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>