SVHN normalization issue about vision_networks

[ikhlestov]

Hi!
What do you mean with better results? Do such approach provide faster learning/converge and better results or will be run faster?

[balodhi]

better results mean that it provides faster convergence and higher accuracy. which the previous normalization technique was unable to achieve.

[ikhlestov]

ok
It will be great if you add one more flag to the main script, add this new normalization and provide pull request. Are you ok with this?

[balodhi]

Yes i was asking the permission to have a pull request. Let me just make a graph comparison of first 6 epochs of both methods.

[balodhi]

[ikhlestov]

Hm, so huge difference.
Ok, thank you for your pull request - I will run it during next days and if the results be similar I will merge it.

[ZhenyF]

It seems it image is substracted by 255 then divided by 2*255, why this one works better?
I tried the method that the image is only divided by 255 (as the paper). And I got 3.4% error rate. Have you got the same accuracy if only trained on using the original norm method and only on train dataset(without extra)?

[illarion-rl]

Hi ZhenyF! Actually I still not compare normalization methods - so I have no answer for your question. You may try both approaches and tell about the difference